Domain object hierarchies inducing multi-level models

Conceptual modeling of domain object hierarchies, such as product hierarchies or organization hierarchies, is difficult due to the intricate nature of nonphysical domain objects organized in such hierarchies. Modeling domain object hierarchies as part-whole hierarchies covers their hierarchical structure, yet to capture their meaning, part-whole hierarchies have to be combined with specialization and multi-level instantiation. To this end we introduce the deep domain object (DDO) multi-level modeling pattern and approach. With the DDO approach, subclasses and metaclasses are induced by and integrated with the part-whole hierarchy. The approach is aligned with the multi-level theory (MLT) and formalized by a metamodel and a set of deductive rules implemented in F-Logic. The proof-of-concept prototype is used for automated application of the pattern and for querying induced multi-level models.


Introduction
Conceptual modeling is the activity of formalizing some aspects of the physical and social world around us for purposes of understanding and communication [26,33], typically in the context of information systems development. With regard to databases, conceptual modeling aims at 'capturing the meaning of a database' [26] independent of its logical and physical implementation.
Before we discuss the aims and contributions of this paper, let us first explain and illustrate our notion of domain object hierarchy. While the approach to modeling domain object hierarchies investigated in this paper is applicable to structural conceptual modeling in general, it is particularly geared towards the conceptual modeling of enterprise databases underlying enterprise information systems, such as ERP and CRM systems.
Enterprise information systems deal, to a large extent, with social worlds where many of the domain objects, i.e., the relevant entities from the world represented in the enterprise Communicated by Adrian Rutle and Manuel Wimmer.
Domain objects may be organized in hierarchies where domain objects at higher levels not only but also act as abstractions of their subordinate objects. For example, sales employees, sales outlets, and sales divisions are organized in a sales organization hierarchy with a sales division also acting as abstraction of its subordinate sales employees and sales outlets. Likewise, product individuals, product models and product categories are arranged in a product hierarchy, with a product category also acting as abstraction of its subordinate product models and product individuals.
To make these abstraction roles of domain objects more tangible, we say a higher-level domain object induces a class for each subordinate level. Figure 1b shows the classes together with their extensions induced in this way by the product hierarchy of Fig. 1a. For example, a higher level domain object such as Car at level Category induces classes such as CarModel and CarIndividual, which have the sets of subordinate domain objects at levels Model and Individual, respectively, as extensions.
With these induced classes and their extensions, the instance-of (set membership) and subclass-of (subset) relationships, which are implicit in the domain object hierarchy, become tangible. For example, VolvoV50 is instance of induced class CarModel, i.e., it is member of the extension of CarModel. The latter, in turn, is a subclass of class ProductModel, that is, its extension is a subset of the extension of ProductModel.
Let us now turn to the metaclasses implicit in a domain object hierarchy. As illustrated in Fig. 1d, induced classes can be collected into metaclasses with regard to the level of their instances and with regard to the class of the domain object by which they are induced. For example, classes CarModel and PhoneModel have instances at level Model and are each induced by a member of class ProductCategory. Hence, they can be collected into an induced metaclass ProductModelClassByCategory.
What we aim for in this paper is a conceptual modeling approach for representing such hierarchies. The approach should make explicit the induced classes and metaclasses so that the modeler can use them for specialization and information aggregation. We are especially interested in supporting the following modeling challenges: (1) specialization along the hierarchy, e.g., product models have a list price, car models additionally have a maximum speed. (2) Regulating specialization along the hierarchy, e.g., every product model is associated with a sales employee acting as product manager. By linking a product category to a sales division, the range of property product manager is refined to employees of that sales division. (3) Information aggregation along the hierarchy, e.g., currently, the average list price of car models is e 21,034.
Based on an initial representation of domain object hierarchies as part-whole hierarchies in UML with levels represented as classes and domain objects as instances of these classes, see Fig. 1c, the paper makes the following contributions: (1) the deep domain object (DDO) pattern is a multi-level modeling pattern for extending domain object hierarchies which are modeled as part-whole hierarchies with induced subclasses and induced metaclasses. The DDO pattern is related to 'promotion with base classes' [24], a solution to the type-object pattern, and is aligned with the multi-level theory (MLT) [8]. (2) The DDO approach is a multi-level modeling approach based on the automatic and recursive application of the DDO pattern together with mechanisms for specialization based on induced subclasses, regulating specialization based on induced metaclasses, and aggregated information provided with classes and metaclasses. (3) A proof-of-concept prototype in F-Logic which can be used to query the induced multi-level model and which is available open-source.
In this paper we introduce the DDO approach on top of a simplified multi-level modeling language. The main simplification is that all properties are single-valued, mandatory and uni-directional. While the DDO pattern is independent of these simplifications, they facilitate an in-depth discussion of specialization, regulating specialization, as well as information aggregation along the hierarchy, without getting drowned by the complexities of optional, multi-valued, or bi-directional properties.
The remainder of this paper is structured as follows. Section 2 introduces a running example consisting of several domain object hierarchies modeled as part-whole hierarchies. Section 3 extends the running example with modeling challenges together with a baseline solution in UML. Section 4 discusses relevant aspects of MLT and Dual Deep Modeling (DDM) [30], which provided the starting point for the development of the DDO approach. Section 5 introduces a set of deductive rules for deriving a multi-level model from a part-whole hierarchy by, in principle, recursive application of the DDO pattern. Section 6 demonstrates the use of deep domain objects for solving the modeling challenges. Section 7 discusses the realization of the DDO approach in F-Logic and demonstrates its use for querying the induced multi-level model, with the complete code provided in Appendix A together with the fully-worked running example in Appendix B. Sect. 8 gives an overview of related work. Section 9 concludes the paper.

Starting Point: Domain Object Hierarchies modeled as Part-Whole Hierarchies
In this section we describe and exemplify the two-level model that serves as core of the running example throughout the paper.
Global two-level model. The starting point for conceptual modeling with domain object hierarchies is a two-level model (depicted in Fig. 2) comprising a schema level (as UML class diagram) and an instance level (as UML object diagram). The modeler has already identified three domain object hierarchies, the Product hierarchy, the Sales organisation hierarchy, and the Engine hierarchy, marked with blue, green, and red background, respectively. Note that there may be classes in the global model that are not part of a domain object hierarchy, for example the class Customer. Domain object hierarchy schema. At the schema level, a domain object hierarchy gets a name and comprises a set of classes that represent the levels of the hierarchy. Each class in a domain object hierarchy gets a level name which is unique within the hierarchy. The set of classes (and hence levels) in a domain object hierarchy is arranged in a path (e.g., the Product hierarchy), a tree (e.g., the Sales hierarchy), or a forest (no example given) by aggregation relationships. When we say a class or level is under another class or level, respectively, we refer to the transitive closure of the aggregation relationships.
For example, the Sales domain object hierarchy comprises classes SalesDivision, SalesEmployee, and SalesOutlet which represent hierarchy levels Division, Employee, and Outlet, respectively. These classes are connected by aggregation relationships expressing that each sales outlet and each sales employee belongs to exactly one sales division. SalesEmployee and SalesOutlet are under SalesDivision.
The Product domain object hierarchy comprises classes ProductCategory, ProductModel, and ProductIndividual which represent hierarchy levels Category, Model, and Individual. These classes are connected by aggregation relationships expressing that each product individual belongs to some product model which in turn belongs to some product category. Product Individual is under ProductModel and under ProductCategory, ProductModel is under Product Category. Properties. A class specifies the schema of its member objects by a set of properties. Each property has as range a class or a simple datatype. For simplicity's sake we only consider properties that are uni-directional, mandatory and functional. For

Modeling challenges: specialization and information aggregation along the hierarchy
In this section we break down the three modeling challenges associated with domain object hierarchies into specific modeling tasks. Preliminary modeling solutions in UML (with extensions) are provided as baseline for the qualitative evaluation of the DDO approach.

Desiderata
Before we introduce the modeling challenges let us discuss desiderata for modeling solutions.
-Non-disruptive: modeling domain object hierarchies as part-whole hierarchies of physical and mainly nonphysical objects is a pragmatic modeling choice reflecting how such hierarchies are often represented in enterprise databases. Modeling solutions should be non-disruptive in that they should not require the replacement of the existing conceptual model but act as an add-on to the part-whole hierarchy. -Conceptual clarity: a domain object in a domain object hierarchy not only represents itself but also acts as an abstraction of its descendants at various levels. Conceptual clarity can be reached by clearly indicating with a property whether it describes (i) the domain object as such, (ii) commonalities of descendants at a particular level, or (iii) aggregated information about the set of descendants at a particular level. -Compactness (modularity, cf. [29] and scoping): All model elements that are introduced with a domain object to describe the domain object itself or to characterize its descendants at various levels should be tied together and it should be possible to use and update them as a selfcontained part of the model without interfering with the rest of the model. -Factoring out commonalities: a modeling solution should factor out commonalities as much as possible to avoid the manual duplication of the same modeling information with various model elements. The modeler should not be forced to model by hand additional classes and additional dependencies that redundantly represent the hierarchical relationships in the domain object hierarchy. The modeler should, for example, not be forced to introduce the same aggregated property multiple times. -Avoid custom constraints: modeling challenges should not be solved by custom constraints in natural language, OCL, or similar, but by applying adequate modeling constructs which serve much better the purposes of conceptual modeling, namely understanding and com-munication, and are less error-prone in development and maintenance.

Specialization along the hierarchy
When adding a new domain object, e.g., a new product category, it should be possible to specialize the schema for descendant domain objects at various hierarchy levels. We will, first, describe the modeling tasks and exemplify them for product category Car in the running example. We will then present a preliminary solution in UML and then discuss the quality of the solution with regard to the desiderata (see Sect. 3.1). The to-be supported modeling tasks are: C1. Derived classes VehiclesEmployee and Vehicles Outlet represent the descendants of sales division Vehicles at levels Employee and Outlet, respectively. Class CarModel refines the range of property productMgr to VehiclesEmployee. Class CarIndividual refines the range of property soldAt to VehiclesOutlet. Assessment of preliminary solution We will now discuss the quality of the preliminary modeling solution with regard to the desiderata (see Sect. 3.1). The preliminary solution is nondisruptive (+) since the original conceptual model remains intact. Conceptual clarity (+) is reached by modeling properties that characterize descendants at a particular level with a derived class. Compactness (−) is weak since the derived classes introduced with domain object Car are not tied together with the domain object and only upon inspection of the classes' derivation rules it becomes clear that they belong to domain object Car. With regard to factoring out commonalities (∼), the modeling information that Car is under Product is manually duplicated as specialization between CarModel and ProductModel and CarIndividual and ProductIndividual. The solution does not avoid custom constraints (−) since the derivation rules for derived classes are encoded manually.

Regulating specialization along the hierarchy
The second challenge is to let the modeler govern specialization at a more generic level. Schema refinements can be modeled separately for each sub-hierarchy as above, but with domain object hierarchies in enterprise information systems it comes natural that these refinements often follow common constraints and rules for every sub-hierarchy. We will, first, describe the modeling tasks and exemplify them along the running example. For the preliminary and only partial solution we will make heavy use of custom constraints on top of the UML class diagram (see Fig. 3) and we will assess the preliminary solution with regard to the desiderata (see Sect. 3.1).
The to-be supported modeling tasks are: C2.1 The modeler should be able to specify that certain range refinements for property p of domain objects at level k must be made with every domain object at some level k (where k is under k ), e.g.: -A product individual cannot be sold at an arbitrary sales outlet but only at sales outlets that belong to a particular sales division. Such a refinement (i.e., restricting the range of property soldAt) must be made with every product category.
C2.2 The modeler should be able to 'automate' the range refinement of property p for objects at level k based on the value of a property p of domain objects at level k (where k is under k ).
-The product manager responsible for a product model must be from the sales division which distributes the product category the product model belongs to.
C2.3 The modeler should be able to specify that the value of property p of domain objects at level k is propagated as shared value of a property p to descendant objects at level k , e.g., -The VAT rate of a product model is the same as the VAT rate of the product category the product model belongs to. -The VAT rate of a product individual, likewise, is the same as the VAT rate of the product category the product individual indirectly belongs to.
C2.4 The modeler should be able to specify such regulations also only for sub-hierarchies, e.g., -The type of an individual car's engine must correspond to the engine type associated with that car's car model.
Partial preliminary solution. In our preliminary solution (see Fig. 3) we found no way to model the obligation to refine a property to a certain level (see C2.1). All other modeling tasks could only be solved by custom constraints (expressed in natural language) attached to the properties. Assessment of preliminary solution. The major shortcoming is obviously that the solution does not at all avoid custom constraints (−). The solution is non-disruptive (+). It comes, arguably, with conceptual clarity (+) because custom constraints are attached to the affected model elements. The solution does not care about compactness (−). There are no shortcomings with regard to factoring out commonalities (+) as there is no need to duplicate model information.

Information aggregation along the hierarchy
The third challenge is to model the aggregated information that should be provided with domain objects about its descendant objects at various hierarchy levels. We will first describe which aggregated information should be provided with the domain objects in the Product hierarchy. As a preliminary solution, we will model the aggregated properties in UML as derived properties. We will then assess this preliminary solution based on the desiderata (see Sect. 3.1). The to-be supported modeling tasks are: C3.1 The modeler should be able to specify an aggregated property p that provides aggregated information about property p of domain objects at some level k and should provide values for aggregated property p with domain objects above level k.
-The sold price of product individuals should be aggregated to average sold price. This aggregated property should be available with every domain object in the Product hierarchy that is above the individual level. -The list price of product models should be aggregated to average list price. This aggregated property should be available with every domain object in the Product hierarchy that is above the model level.
C3.2 It should be possible to define multi-step aggregations, i.e., aggregated properties that are calculated from aggregated properties, e.g., -The average sold price per product model should be further aggregated to a minimum average sold price per model and this aggregated property should be available with every domain object in the Product hierarchy that is above the model level.
C3.3 Aggregated information should be provided also for the domain object hierarchy as a whole.
-Values for the three aggregated properties should be available for the Product hierarchy as a whole.
Preliminary solution. In our preliminary solution (see Fig. 4), we introduce a singleton class ProductHierarchy the instance of which represents the overall product hierarchy. We model 'average sold price' multiple times as derived property avgSoldPrice with classes ProductModel,

Towards multi-level modeling of domain object hierarchies
In this section we discuss how to model aspects of the sample domain object hierarchy using existing multi-level modeling techniques, namely Carvalho et al.'s Multi-Level Theory (MLT) [8] and our previous work on Dual Deep Modeling (DDM) [30]. These two MLM approaches serve as starting point for our further considerations.

MLT-based modeling of domain object hierarchies
Domain object hierarchies modeled as part-whole hierarchies come with a uniform treatment of physical objects like MyCar and more class-like nonphysical objects such as product model VolvoV50. Instead of regarding the latter as classes we treated them as nonphysical objects that have a one-to-one correspondence with a class, e.g., VolvoV50Individual. With MLT one typically takes another approach and represents by VolvoV50 both the product model VolvoV50 and the class VolvoV50 Individual of physical cars of product model Volvo V50. In this subsection let us model the Product hierarchy accordingly (see Fig. 5 for an UML-like representation). Domain objects at the lowest level (e.g., product individual MyCar) are simple objects. Domain objects at higher levels, such as product model VolvoV50 and product category Car, are modeled as first-order classes. Non-bottom levels of the product hierarchy are modeled as 2nd-order classes ProductModel and ProductCategory. The bottom level of the product hierarchy (represented in Fig. 2 by class ProductIndividual) is represented as first-order class Product-this adjustment in naming is to make the models conformant to the naming conventions applied in the examples found in the literature on MLT.
The hierarchy of domain objects that are not at the bottom level is represented as a subclass hierarchy. For example, VolvoV50 is a subclass of Car, which is a subclass of Product, which makes VolvoV50 also a (indirect) subclass of Product. As an instance of class VolvoV50, MyCar is also a member of class Car and of class Product.
Higher-order classes are also organized in subclass hierarchies. Second-order classes ProductCategory and ProductModel are sub-classes of ProductClass with Car being an instance of ProductCategory and, hence, a member of ProductClass. Now let us use MLT to specify in more detail the kinds of second-order and higher-order classes in our multi-level model.
The first kind of metaclass we are interested in is the power type as introduced by Cardelli [7] and incorporated into MLT. The power type of a class (which is the power type's base class) has all subclasses of that class, including that class, as members. For example, in Fig. 5

, metaclass
ProductClass is the power type of class Product. All sub-classes of Product, including Product are members of ProductClass.
The second kind of metaclass relevant to the multi-level modeling of domain object hierarchies is the kind of power type introduced by Odell [31]. We limit the use of Odell's power types to what in MLT [8] is referred to as disjoint and complete categorization, or partitioning. Note, the restriction to disjoint and complete categorization is because we assume every domain object at a lower level in a domain object hierarchy to belong to exactly one domain object at the next higher level. We refer to this kind of metaclass not as power type but to as partitioning metaclass, or simply as metaclass, since in our approach all metaclasses, apart from Cardelli's power types are partitioning metaclasses.
A partitioning metaclass always comes together with a partitioned class, also referred to as base class. Every member of the partitioning metaclass is a (direct or indirect) subclass of its base class and every member of the partitioned class is member of exactly one member of the partitioning metaclass. For example, the metaclass ProductModel partitions class Product. Its members, such as VolvoV50, are sub-classes of Product. Every member of partitioned class Product is member of exactly one member of ProductModel, for example, MyCar is member of VolvoV50.
Also relevant to the multi-level modeling of domain object hierarchies is the notion of subordination between metaclasses, as introduced with MLT [8]. If a metaclass is subordinate to another metaclass, then every member of the former is subclass of some member of the latter. For example, the metaclass ProductModel is subordinate to metaclass ProductCategory. Every member of ProductModel is sub-class of one member of ProductCategory, for example, VolvoV50 is subclass of Car. Discussion. The modeling of higher-level domain objects, such as VolvoV50 and Car, as classes seems very natural. The same is true for the modeling of classes, such as ProductModel and ProductCategory, as metaclasses. One may argue, that the change in semantics from the original model better reflects the ontological nature of things and avoids accidental complexity [4]. Yet, other authors [11] argue quite the opposite, namely that concepts like VolvoV50 are, first and foremost, objects.
What we aim for with the DDO pattern and approach is broader applicability of multi-level modeling without sacrificing conceptual clarity. What motivates our research in this direction is the observation that many domain object hierarchies, like the Sales organisation hierarchy in our running example, are of a very different nature, cannot be adequately modeled using the multi-level modeling style exemplified in this section, yet also induce subclasses and metaclasses. The following example also illustrates this observation: A domain object hierarchy of geographical entities, e.g., Innsbruck is a city in Tyrol, and Tyrol is a state in Austria, induce multilevel models, e.g., Innsbruck is a member of class 'Tyrolean city' which in turn is a subclass of 'Austrian city' as well as a member of metaclass 'Austrian city class by state'.
Our intuition here is that all these different domain object hierarchies are different kinds of part-whole hierarchies and that the modeling of part-whole comes prior to the modeling of subclasses and partitioning metaclasses. This is related to Partridge et al's work [32] which is driven by Kit Fine's theory of part [12]. Fine "takes a very liberal notion of part" [32] with subset and set-membership as kinds of part-whole. With sets considered as wholes, Fine argues for "taking both its members and its subsets to be parts". Also, our distinction between modeled domain objects and induced classes and metaclasses relates well to the works of Partridge et al. and Fine. They distinguish a basis domain, a constructor domain, and a constructed domain (cf. Table 1 in [32]). We consider domain objects and their arrangement in a domain object hierarchy as given, as 'basis domain', and the induced classes and metaclasses as 'constructed domain'.
Currently, we only consider domain object hierarchies where a domain object always belongs to exactly one domain object at the next higher level. That is why we are only interested in partitioning metaclasses and can avoid multiple inheritance. A partial lifting of these constraints, which will make relevant MLT's remaining cross-level relations (categorizes, disjointly categorizes, completely categorizes), is left to future work.

Higher-order partitioning metaclasses
What seems particular relevant to the multi-level modeling of domain object hierarchies are higher-order partitioning metaclasses. Higher levels of a domain object hierarchy not only partition the domain object hierarchy's bottom level but also intermediate levels. With intermediate levels already being represented as second-order classes, this gives rise to 3rd-order classes. This can be applied recursively: a domain object hierarchy's n-th level (from the bottom) can be represented as 2nd . . . nth-order classes, each partitioning a 1st . . . (n − 1)th-order class.
For example, level Category of the domain object hierarchy not only partitions the set of product individuals but also the set of product models. With the latter represented by 2ndorder class ProductModel, we introduce a 3rd-order class ProductModelClassByCategory which partitions it (see Fig. 6) into sub-classes CarModel and PhoneModel. Discussion. So far we have discussed the representation of domain object hierarchies with classes and meta-classes applying some of the constructs of MLT, namely power type, partitioning, and subordination. As already evident with these small examples the modeling of domain object hierarchies using these constructs is elaborate and results in well understandable but verbose models. A major drawback is that a single domain object (e.g., product category Car) and a single hierarchical relationship (e.g., Car belongs to Product) have to be represented by multiple classes and relationships, respectively, in the multi-level model.
This observation led us to the conviction that large parts of multi-level models with precise semantics need not be modeled manually but can be derived from abstraction hierarchies modeled as part-whole hierarchies that do not, as such, come with precise semantics.
One possible limitation of our view on domain object hierarchies is that we treat all hierarchical relationships in a domain object hierarchy the same and distinguish between subset and set membership only with respect to induced classes. This can make the application of the DDO approach unnecessarily heavyweight, since potentially irrelevant classes and metaclasses, as well as their subclass-of and instance-of relationships, are also made explicit.

Multi-level modeling with named levels in DDM
Now let us revisit our previous work on Dual Deep Modeling (DDM) [30]. DDM facilitates more compact multi-level models, yet lacks the conceptual clarity of MLT.
The core construct of the DDM approach is the DDM clabject. DDM clabjects are connected by properties with dual potencies, a source potency and a target potency. The source and target potencies of a property indicate the number of instantiation steps at the source and, respectively, target end of the property to reach an ultimate instance of the property. Properties in DDM can have inverse properties, can be arranged in specialization hierarchies, may be multi-valued and can be further characterized by multi-level cardinality constraints. Although DDM's powerful and flexible property mechanism will guide our future work to develop a fullyfledged property mechanism in DDO, it is clearly beyond the scope of this paper.
What we are interested in this paper is DDM's use of named levels which go back to our work on M-Objects [28].
In Fig. 7 (left part), a product hierarchy is represented by DDM clabject Product with levels Category, Model, and Individual, where Individual is under Model is under Category. Third-order clabject Product is instantiated by a second-order clabject Car at level Category which in turn is instantiated by a first-order clabject VolvoV50 at level Model which in turn is instantiated by a clabject MyCar at level Individual.
What led us to the DDO approach was the realization that a DDM clabject actually represents a nested structure of local classes of ascending order (illustrated by Fig. 7), induced as follows.
Going back upwards the instantiation hierarchy we say, now more precisely, MyCar is an instance of VolvoV50 Individual, the Individual class local to car model VolvoV50. VolvoV50 is an instance of CarModel, the Model class local to product category Car; and Car is an instance of ProductCategory, the Category class local to root object Product. Now, turning to the classification of induced classes along the hierarchy, we say VolvoV50Individual is an instance of CarIndividualClassByModel, the Individual metaclass local to CarModel. Class CarModel  [30]. Note, in DDM there is no distinction between domain object hierarchy levels and classification levels. Right: The product hierarchy modeled with local meta * -classes, the basic idea underlying Dual Deep Modeling [30] (not elaborated as such in [30]). Local meta * -classes get a compound name reflecting their nesting and are shown with their order or potency. Different shades of grey represent the different classification levels of the multi-level model. Nesting of rounded rectangles represents local to relationships, e.g., first-order class CarModel is local to object Car, and 2nd-order class CarIndividualClassByModel is local to class CarModel and is indirectly local to object Car is an instance of ProductModelClassByCategory, which is the Model metaclass local to ProductCategory.
Finally, turning to the classification of induced metaclasses, we can say that metaclass CarIndividualClass ByModel is an instance of ProductIndividual MetaclassByModelByCategory, which is the Individual meta-meta class local to metaclass ProductModel ClassByCategory.
Summarizing, the instantiation of a DDM clabject of order n by a clabject of order n − 1 can be regarded as n closely coupled instantiations at different levels of classification. Discussion. The clear picture of the meaning of DDM clabjects we have painted here, fits perfectly with our intuition behind DDM. However, DDM's highly flexible and powerful property mechanism including support for skipping levels when instantiating properties may obscure this clear picture.
In order to more clearly convey the meaning of induced multi-level models, we will, in this paper, employ a highly simplified and more strict property mechanism, only considering single-valued, uni-directional properties without support for skipping levels.
Another shortcoming of DDM is the lack of support to address superclasses induced along the hierarchy, for example, generalizing VolvoV50Individual to superclass CarIndividual, which in turn generalizes to superclass ProductIndividual . This shortcoming is mitigated by DDM's powerful and flexible property mechanism, not discussed here in detail, which allows to use higher-order classes as superclasses, for example metaclass CarIndividualClassByModel can be used in many regards like it was a superclass CarIndividual, what further diminishes conceptual clarity.
As a major addition in comparison to DDM, the DDO approach will make all the induced superclasses (at various classification levels) directly addressable by the modeler. The explicit representation of induced superclasses adds to the modeling power and improves conceptual clarity.
Furthermore, DDM's reliance on numeric potencies limits its use to the modeling of domain object hierarchies the levels of which are in a total order. By getting rid of numeric potencies and instead fully relying on level names, the DDO approach can also support the multi-level modeling of domain object hierarchies the levels of which are only in a partial order, such as the Sales organisation hierarchy. The use of level names would also facilitate the introduction of intermediate levels in sub-hierarchies (not discussed in this paper) as featured by the M-Objects [28] approach.

From part-whole to multi-level hierarchy
In this section we first introduce the deep domain object (DDO) pattern for representing classes and metaclasses induced by a domain object hierarchy. The DDO approach to multi-level modeling has at its core the automatic and recursive application of the DDO pattern, which we will formalize in this section by a set of deductive rules. When working with the DDO approach, the modeler does not have to care about Fig. 8 The DDO pattern: a domain object hierarchy with two levels modeled as part-whole hierarchy (left) induces a multilevel model (right) comprising local classes, a metaclass, and their subclass-of, direct-instance-of (iof), and partitions dependencies. The rounded rectangles with grey background show the nesting structure of local classes, e.g., class ProductCategory is local to domain object Product, and of metaclasses, e.g, metaclass ProductModelClassByCategory is local to class ProductCategory. The shade of grey indicates the nesting depth which corresponds to the order of the classes, also referred to as classification level The approach is illustrated along our running example. Figure 8 exemplifies the DDO pattern with a small fragment of the product hierarchy. Figure 9 shows the metamodel of domain object hierarchies modeled as part-whole hierarchies and of domain object hierarchies with induced classes and metaclasses. Figures 10 and 11 exemplify the application of the pattern for a fragment of the Product hierarchy (a domain object hierarchy with three levels in a total order) and of the Sales organisation hierarchy (a domain object hierarchy where the levels form a tree and not a path).
For the deductive rules we apply minimal Herbrand model semantics and the closed-world assumption for negation as in Datalog ¬ (cf. [19]) plus existentially quantified rule heads for creating new model elements. Rules are being invoked alternating and recursively until a fixpoint is reached and a minimal Herbrand model found (in our case representing the induced multi-level model).
In this section we focus on deriving the multi-level structures and multi-level dependencies induced by part-whole hierarchies. With the DDO approach we further assume that every (induced) class has a power type which has the class and all its direct and indirect subclasses as members. Since this is not specific to the DDO approach we refrain from providing formalization of power types and do not show them in the multi-level models except when used for modeling aggregated properties. A complete realization of the approach (including power types) and of the running exam- Fig. 10 The DDO pattern applied on the Product hierarchy with its three levels which are arranged in a total order. In comparison to Fig. 8, the additional Individual level gives rise to additional classes, metaclasses, a meta-metaclass, and a subordinate-to dependency between metaclasses. Note, the partitions and subordinateTo dependencies are not an essential part of the induced multi-level model yet clarify its semantics in terms of MLT [8] Fig. 11 The DDO pattern applied on the Sales organisation hierarchy with its tree-like schema (levels Outlet and Employee are directly under level Division) ple is provided in F-Logic (see Appendix) which also serves as proof-of-concept implementation (see also Sect. 7).

The DDO pattern
Let us first introduce in simple terms the Deep Domain Object pattern (DDO pattern) for representing domain object hierarchies and their induced classes and metaclasses. The DDO pattern is a multi-level modeling pattern that builds on 'promotion with base classes' [24], a solution to the type-object pattern, and makes use of metaclasses and cross-level dependencies in accordance with the multi-level theory (MLT) [8].
A key principle of the DDO pattern is that induced classes are modeled local to the domain object by which it is induced. In this way, the DDO pattern fulfills the compactness desideratum of Sect. 3.1.
We will introduce the DDO pattern for a domain object hierarchy fragment with only two levels, each with only one domain object. At the schema level the hierarchy, e.g., Product, is modeled by a component class, e.g., ProductModel, associated via aggregation relationship to a composite class, e.g., ProductCategory, and at the instance level by a component object, e.g., VolvoV50, linked via aggregation to a composite object, e.g., Car. The result of applying the pattern is shown in the right part of Fig. 8 Applying the DDO pattern by hand is cumbersome and not feasible for all but very small domain object hierarchies. In the following we will formalize the pattern's automatic and recursive application.

Preparing the domain object hierarchy
The first step is to prepare the part-whole hierarchies for uniform treatment of (the roots of) hierarchies and subhierarchies as domain objects. Every hierarchy is treated as a domain object that represents the whole hierarchy and acts as root of the hierarchy [rule (A1)]. Then all domain objects at a top-most level of the hierarchy are assigned as parts of that root object (A2). Domain objects are annotated with the hierarchy level represented by their class (A3) and are assigned to classification level 0, i.e., their order is 0 (A4). For example (see Fig. 10), domain object hierarchy Product becomes domain object Product at level Root representing the whole product hierarchy with product category Car as part. Domain objects like Car are annotated with their level, e.g., Category, and are associated with classification level 0. Note, we use predicate instanceOf for the asserted instance-of relation, e.g., VolvoV50: ProductModel is represented as instanceOf(VolvoV50,ProductModel), and use predicate iof for the direct instance of relation which takes into account induced sub-classes, e.g., iof(VolvoV50,CarModel).

Deriving multi-level instantiation hierarchies
In the following we define the rules that derive multi-level instantiation hierarchies of local classes. Given is a domain object o which is part of domain object o and modeled as instance of class c which represents level k. Then domain object o is direct instance of (iof) the class c local to o which also represents level k. For example, product category Car is part of domain object Product and modeled as instance of ProductCategory which represents level Category. Now, since ProductCategory is local to Product, it follows that Car is a direct instance of ProductCateogry.
Given is an object o which is a direct instance of (iof) class c which has class c representing level k as direct or indirect part (partOfSchema + represents the transitive closure of partOfSchema). Then a new class c local to o and representing level k comes into existence. For example, Car is a direct instance of ProductCateogry which has classes ProductModel, representing level Model, and ProductIndividual, representing level Individual, as direct or indirect parts. This induces local class CarModel, representing level Model for the subhierarchy rooted in Car, and class CarIndividual, representing level Individual for the subhierarchy rooted in Car. For the mechanism to generate names of local classes and metaclasses see the F-Logic implementation in the Appendix.
By alternating recursive invocation, rules (A6) and (A5) produce first-order classes local to domain objects further down the hierarchy, such as VolvoV50Individual local to domain object VolvoV50.
Next we define the rules to derive local metaclasses. Consider a class c that has class c representing level k as direct or indirect part and the order of c is lower than a constant maxOrder (which can be set by the modeler). Then a new metaclass c , local to class c and representing level k, comes into existence. For example, the class CarModel has CarIndividual as part which represents level Individual. This induces metaclass CarIndividualClassByModel, which represents level Individual local to class CarModel. Further metaclasses induced in this way are ProductModelClassBy Category, ProductIndividualClassByCategory, ProductIndividualClassByModel, and finally 3rd order class ProductIndividualMetaclassByModel-ByCategory.
With local metaclasses derived we also need to derive the instance of (iof) relationships between classes and metaclasses. Consider a metaclass m representing level k local to class c, and a class c also representing level k local to o which is an instance of c. Now we can derive that c is an instance of m. Consider, for example, the metaclass ProductIndividualClassByCategory which is local to class ProductCategory which has Car as an instance, and class CarIndividual which is local to Car. We can derive that CarIndividual is a direct instance of ProductIndividualClassByCategory.
The order of a local class is the increment of the order of the object the class is local to.
The part-whole hierarchy modeled under a class is propagated to metaclasses local to that class. Consider class c which is above classes c and c in the part-whole hierarchy. When class c is directly under class c then this direct part-whole relationship is propagated to metaclasses local to c . According to (A10), when class c which represents level k is a direct part of c which represents level k and both are direct or indirect parts of c and metaclasses m, representing level k, and m , representing level k , are local to c , then m is part of m . For example, Product-Individual is directly under ProductModel and both are under ProductCategory, thus metaclass Product-IndividualClassByCategory has a direct part-whole relationship with ProductModelClassByCategory at classification level 2.
The part-whole hierarchy modeled with metaclasses is propagated to their instances. According to (A11), when classes c and c are local to the same object o, and c is instance of metaclass m, and c is instance of metaclass m , and m is part of m , then c is derived to be part of c .

Deriving generalization hierarchies
With the multi-level instantiation hierarchies of local classes in place we can now derive subclass-of relationships at all classification levels. We only show how to derive direct sub-classOf relationships. From there one can easily derive the reflexive-transitive closure which is often used instead of only considering direct subclass-of relationships.
A first-order class local to a domain object is subclass of the class that is local to the domain object's parent and represents the same level. According to (A12) when domain object o is part of domain object o and o has a local class c representing level k and o has a local class c representing level k, then c is a subclass of c . For example, VolvoV50 is part of product category Car. Its local class VolvoV50Individual which represents level Individual in the subhierarchy rooted in VolvoV50 is a subclass of CarIndividual which represents level Individual for the subhierarchy rooted in Car.
According to (A13) when c is a subclass of c and c has a local metaclass m representing level k and c has a local metaclass m representing level k, then m is a subclass of m . This applies also to classes of higher orders. For example, CarModel is a subclass of ProductModel, its local metaclasses for level Individual, namely CarIndividual-ClassByModel and ProductIndividualClassBy-Model are, hence, in a subclassOf relationship.
Finally, the members of a class c (of any order) are defined as the instances of that class together with instances of direct or indirect subclasses of c.

Aligning DDO with MLT
To illustrate the semantics of multi-level models induced by domain object hierarchies in terms of the multi-level theory MLT [8] we now define the derivation of partition and subordinate to relationships. This is for informative purposes and does not add to the semantics of the approach. The partitions and subordinateTo relations for the running example are shown in Fig. 10.
An induced fact partitions(m, c) says that metaclass m partitions class c, i.e., every instance of m is a subclass of c, and, further, every member of c is member of exactly one instance of m. An induced fact subordinateTo(m, m ) says that metaclass m is subordinate to m which means that every instance of m is subclass of some instance of m .
According to (A14), when a class c representing level k is local to a domain object o and another class c , also local to o, has a local metaclass m also representing level k, then m partitions c. In this case, c is also referred to as the base class of m. The subordinateTo relationship, according to MLT [8], is a relationship between partitioning metaclasses with the same base class. When a partitioning metaclass is subordinate to another partitioning metaclass then every instance of the former is a subclass of some instance of the latter. Now let us define how subordinateTo relationships are induced from domain object hierarchies. According to (A16), when a domain object o has two local classes c and c with c being directly under c in the part-whole hierarchy schema, and c and c have both a local metaclass, m and m , respectively, which both represent level k, then m is subordinate to m .

Discussion
Let us briefly assess the induced multi-level model with regard to the desiderata of Sect. 3.1. The approach is nondisruptive (+) as the part-whole hierarchy remains intact and the multi-level structures act as add-on. Conceptual clarity (+) is facilitated by decomposing domain objects into their different facets. Compactness (+) is reached through the nesting of induced classes and metaclasses with domain objects acting as roots of localTo-hierarchies. The central part-whole hierarchy factors out commonalities (+) among the multiple generalization and multi-level instantiation hierarchies. Avoiding custom constraints (+) is facilitated by having fine-grained classes and metaclasses readily available which spares the modeler from writing derivation rules by hand.

Modeling with deep domain objects
In this section we introduce and exemplify modeling with deep domain objects along the modeling challenges set out in Sect. 3.
A deep domain object is a domain object together with an aggregated and schematic description of its descendant domain objects together with regulations that govern how the schema is to be specialized along the hierarchy. Every domain object (apart from objects at the bottom level) has an induced multi-level schema characterizing the sub-hierarchy rooted in that domain object.
In the simple case of a domain object hierarchy with a totally ordered set of levels, a deep domain object is similar to a DDM clabject but with the various class/superclass/metaclass facets of the DDM clabject now represented explicitly by classes and metaclasses 'nested' inside the domain object and with support for providing aggregated information with these local meta* classes. These local classes and metaclasses are induced automatically, together with the multi-level dependencies connecting classes which are local to the same object and the dependencies connecting classes local to objects that are in a direct hierarchical relationship in the domain object hierarchy. Furthermore, by encapsulating local classes and metaclasses in deep domain objects, we retain one of the promises of dual deep modeling, namely to avoid the cluttering of models with a myriad of classes and metaclasses.

Solving challenge 1: specialization along the hierarchy
In this subsection we explain how the modeling tasks presented in Sect. 3.2 are solved with deep domain objects (see Fig. 12). C1.1 Range refinements. Induced local classes facilitate the refinement of the range of a property p for the descendants of a domain object o at level k to descendants of a domain object o at level k . For example, with local class CarModel the range of property productMgr is refined to local class Vehicles Employee, that is, product models of product category Car may have as product manager only employees of sales division Vehicles.
With local class CarIndividual the range of property soldAt is refined to local class VehiclesOutlet, that is, individual cars may only be sold at sales outlets of the Vehicles sales division. C1.2 Extending the schema. With local classes the modeler is able to extend the schema of descendants of a domain object at a particular level by introducing additional properties.
For example, with local class CarModel the modeler extends the schema with an additional property engineType with range EngineType (from the Engine hierarchy). With local class CarIndividual an additional property engine with range EngineIndividual is introduced.

C1.3 Shared values.
With local classes the modeler may also specify a shared value for a property of descendant objects at a certain level.
For example, with local class CarModel the modeler specifies a shared value of 20% for property vatRatespecifying that every car model has exactly that value for property vatRate. With local class CarIndividual the vatRate value of 20% is fixed for all individual cars. Assessment of DDO solution to Challenge 1 Let us discuss the quality of the DDO solution with regard to the desiderata (see Sect. 3.1). The solution is non-disruptive (+) since the original conceptual model remains intact. Conceptual clarity (+) is reached by modeling properties that charac- terize descendants at a particular levels with local classes. Compactness (+) is reached since the specializations regarding the sub-hierarchy rooted in domain object Car are all modeled with model elements local to Car. With regard to factoring out commonalities (+), the modeler is not forced to represent the same modeling information twice. The solution does not come with custom constraints (+).

Solving challenge 2: regulating specialization with metaclasses
Domain object hierarchies typically have a rather homogeneous hierarchical structure across their sub-hierarchies. This degree of homogeneity will typically also be reflected in the schema refinements. Metaschema assertions can enforce a certain degree of homogeneity and can also automate schema refinements. Before solving Challenge 2, let us have a closer look at the meaning of local metaclasses from the perspective of domain object hierarchies.
A metaclass, e.g., ProductIndividualClassBy-Category represents a level, e.g., Individual, which we refer to as its member-member level. The metaclass is local to a class, e.g., ProductCategory, which, in turn, represents another level, e.g., Category, which we refer to as the metaclass' member level, and is local to a domain object, e.g., Product. For a particular state of the database, a metaclass represents a metaset, namely the set of descendants of the domain object, e.g., Product, at the metaclass' member-member level, e.g., Individual, grouped by the metaclass' member-level, e.g., Category. Its member classes are induced subclasses local to domain objects at the metaclass' member level.
The modeling tasks concerning regulating specialization are solved with deep domain objects (see Fig. 12) as follows. C2.1 Metarange. A metarange assertion specifies an obligation for members (which are classes) of the metaclass to refine the schema of a property to a certain hierarchy level. The modeler is able to specify with a (local metaclass of a) higher-level object a constraint that a certain refinement should be done by all descendant domain objects at a certain level (namely by local classes of descendant domain objects at that level).
For example, the metaclass ProductIndividual-ClassByCategory further defines SalesOutletClass-ByDivision as metarange of its member-member property soldAt, denoted as soldAt : : SalesOutlet ClassByDivision, meaning that its member classes at the Category level, e.g., CarIndividual must refine the range of soldAt to outlets that belong to a particular sales division. C2.2 Propagate property value as range. A propagate-asrange assertion tells members of the metaclass to use a property value of the domain object they are local to, to refine the range of another property.
Using propagate-as-range assertions the modeler can generically specify the refinement of property ranges for domain objects at lower levels (i.e., the metaclass' membermember level) by property values of domain objects at intermediate levels (i.e., the metaclass' member level).
For example, the metaclass ProductModelClass-ByCategory has a propagate-as-range assertion for its member-member property productMgr, denoted as productMgr : {($salesDivision,Employee)}, defining that its member classes at the category level, such as CarModel, take the value, e.g., Vehicles, of property salesDivision of the product category, e.g., Car, to refine the range of the property to employees of that sales division, e.g. by local class VehiclesEmployee.

C2.3 Propagate property value as shared value.
A propagateas-shared-value assertion tells members of the metaclass to use a property value of the domain object they are local to, to define a shared value for the metaclass' member-member property.
Using propagate-as-shared-value assertion the modeler can generically define the propagation of property values from domain objects at an intermediate level (i.e., the metaclass' member level) to domain objects at lower levels (i.e., the metaclass' member-member level) as shared values.
For example, the metaclass ProductModelClassBy-Category has a propagate-as-shared-value assertion for its member-member property vatRate, denoted as vatRate = { $vatRate }, defining that its member classes, such as CarModel, use the vatRate given with the product category they are local to, e.g., Car, as shared value for their member property vatRate. Metaclass Product-IndividualClassByCategory has a similar assertion, specifying the propagation of the vatRate to the individual level. C2.4 Regulating specialization in sub-hierarchies. With local metaclasses the modeler is able to specify such regulations also only for sub-hierarchies.
For example, a propagate-as-range assertion given with metaclass CarIndividualClassByModel specifies that the engine type, e.g., VD4162T, given with a particular car model, such as VolvoV50, is used to refine the range for the individual class of that car model, e.g., in class VolvoV50Individual the range for property engine gets automatically refined to class VD4162TIndividual (the class of individual engines of engine type VD4162T). Assessment No customs constraints (+) are necessary and the solution is non-disruptive (+). Conceptual clarity (+) is high because each metaschema assertion is made with a model element that specifically represents the two affected levels in a domain object hierarchy or sub-hierarchy. Compactness (+) is reached since the metaschema assertions are made with metaclasses which are tied together with a domain object. The solution avoids custom constraints (+). Regarding factoring out commonalities (+), there is no need to model some information twice.

Solving challenge 3: aggregated information
In the DDO approach, every class has associated its power type. A class' power type is used to define properties that are to be instantiated by the class and also by the class' direct and indirect subclasses, which are all, by definition, members of the power type. Note, in the DDO approach we currently only have use for one layer of power types, that is why there are no power types of power types.
By using the power type of a class representing a certain level (e.g., Model) for the definition of aggregated properties, the modeler makes sure that the aggregated information is available with domain objects above that level without the need to define the aggregated property multiple times.
Aggregated information is also available for the domain object hierarchy as a whole.
The class with which the aggregated information is provided represents the set of objects from which the aggregated information is calculated. The metaclass with which a twostep aggregated information is provided represents the set of objects and its partitioning into a set of set of objects from which the aggregated information is calculated. C3.1 One-step aggregation. Simple aggregated information, i.e., aggregated properties like avgSoldPrice and avgListPrice are defined with a powertype of a firstorder local class and the aggregated vlaues are provided with first-order local classes. C3.2 Multi-step aggregation. Two-step aggregated information (e.g., minAvgSoldPricePerModel) is provided with second-order classes, and so forth (no example given for three step aggregation). C3.3 Aggregated information for the whole hierarchy. By defining aggregated properties with a powertype of a class that is local to a root object, the aggregated information is calculated also for the hierarchy as a whole. Assessment. No custom constraints (+) or complex queries are necessary. Compactness (+) and conceptual clarity (+) is reached since the aggregated information is provided with the class or metaclass it describes which is also tied together with the domain object. Concerning factoring out commonalities (+), there is no need to define an aggregated property multiple times.

Proof-of-concept prototype
The deductive rules introduced in Sect. 5 are realized in F-Logic (see Appendix A) together with the formalization/implementation of the complete DDO approach. The prototype can be used to query induced multi-level models and property values, including aggregated information and shared values, and property ranges, including those propagated from higher-level objects. This is demonstrated by the query results shown in Figs. 14, 15, 16, and 17. One further query result is provided in the appendix. Formating query results in tabular form is implemented in a small Java program which interacts with Flora-2 via command line.
In addition to the derivation of local classes and metaclasses and their multi-level dependencies, the prototype covers: -Specialization along the hierarchy with range refinement, schema extension and shared values. -Regulating specialization with metaclasses by metarange assertions, propagate-as-range assertions, and propagateas-shared-value assertions for domain object hierarchies as a whole and also for sub-hierarchies. -Modeling aggregated properties with powertypes and providing respective aggregated information with local classes and metaclasses. -Structural conformance checks.
Conformance checks are also realized as deductive rules which fill predicate error/2, which can be queried for violations of the structural conformance rules.
The functionality of the prototype is demonstrated with the running example started in Sect. 2 and further developed in Sect. 6. The fully worked example is given in Appendix B.
Note, in the F-Logic implementation we use custom properties subclassOf and iof instead of the built-in notation used by F-Logic ('::' and ':'). This is to avoid an unintended strong integration into F-Logic, since we do not conceive the implementation as an add-on to F-Logic but rather only as a validation of the approach. Nevertheless, to facilitate preliminary integration with F-Logic's built-in instantiation and subclassing we have to consider them in the realization of property range constraints. We further consider in the F-Logic implementation that a class c in the formalization may be a data type or a 'normal' class from the global schema or a local class from some domain object hierarchy.
The proof-of-concept prototype comes with four aggregation operators, namely SUM, MIN, MAX, and AVG (see Appendix A). In the definition of the aggregated property in the metaclass (typically a powertype), these operators take as parameter a member property of their base class. The aggregation operation is then evaluated with regard to an instance Fig. 13 Solving Challenge 3 (aggregated information) using power types of classes and of metaclasses to define aggregated properties of the powertype, that is, with regard to a subclass of the base class, and aggregates all members of that subclass.
Concerning performance we have conducted very preliminary experiments on a HP EliteBook 850 G2 with Intel(R) Core(TM) i7-5600U CPU with 2.60GH with two kernels with 16 GB working memory and running Windows 10 Pro. The time measurements are based on 10 consecutive runs of a custom-made Java program that starts the Flora engine, interacts with the Flora engine via command line, and which transforms the query results to tabular format.
We have experimented with two variants. The first variant (see Appendix A.2) uses structured class IDs for local classes and powertypes, e.g., k(k (Product,Model), Individual). The second variant (see Appendix A.3) generates class names as used throughout the paper, e.g., ProductIndividualClassByModel. With the first variant, starting the Flora2 engine and loading the code takes on average 0.66 sec, the evaluation of the five queries (as shown in Figures 14-18) takes together on average 1.14 sec.
With the second variant, startup took on average 0.81 sec and query evaluation on average 3.52 s.

Related work
In multi-level modeling, the clabject [1] allows metamodeling with arbitrarily many levels of instantiation. Deep characterization allows a clabject to specify the schema of members at arbitrarily deeper levels of instantiation, it can specify attributes to be instantiated by its members or by the members of its members, and so forth.
Telos [27] was one of the first approaches that allowed meta-modeling with an unbounded number of instantiation levels. In Telos, classes and properties are also objects, and objects are instances of other objects. ConceptBase [17] is an implementation of Telos based on the O-Telos [18] dialect.
VODAK [21] was one of the first systems to support deep characterization, although limited to three levels of instantiation. VODAK distinguishes three kinds of objects,  Result of querying all partitioning metaclasses in the multilevel model, together with the base class they are partitioning and, if existing, the metaclass they are subordinate to namely individual objects, classes, and metaclasses, as well as types, which describe the schema of objects in terms of structure and behaviour. An object has an own-type which describes the object's own schema. A class or metaclass further has a member-type which describes the schema of its members. A metaclass further has a member-member-type which describes the schema of the members of its members, realizing a limited form of deep characterization. An object instantiates its own-type, which in turn is a specialization of the member-type of the object's class. The member-type of a class is a specialization of the member-member-type of the metaclass of the class. In this way, the relationship between an object and its class encompasses instantiation and specialization.
Potency-based deep instantiation [2] goes beyond VODAK by allowing to specify the schema of member objects at arbitrarily deeper levels of instantiation by assigning a potency to properties. Fig. 16 Result of querying property values (including aggregated property values, aggregated from descendants at lower levels, and values that were propagated as shared values from higher-level objects). By having both the object to which the property value belongs directly as well as the domain object to which that object belongs, the DDO approach combines compactness and conceptual clarity. Note: different background colors were added manually Fig. 17 Result of querying property ranges (including property ranges propagated according to range-propagation assertions). By having both the class to which the property range belongs directly as well as the domain object to which that class is local to, the DDO approach combines compactness and conceptual clarity. Note: different background colors and horizontal lines were added manually The orthogonal classication architecture distinguishes linguistic and ontological metamodeling [2,3]. Linguistic metamodeling is about the modeling of elements of the modeling language, e.g., Domain Class and Domain Object, and their instantiation by model elements, e.g, ProductModel and VolvoV50, respectively. Ontological metamodeling is about modeling of individuals, classes and metaclasses from the domain of interest and their ontological instance-of relationships, e.g., VolvoV50 is ontological instance of CarModel. In our approach, ontological instantiation is divided into two parts, domain object VolvoV50 is, rst, an instance of domain class ProductModel representing level Model, and, second, under domain object Car, so that it is nally an ontological instance of CarModel.
With regard to Kühne's level cohesion and level segregation principles [22], classification levels in DDO hierarchies are order-synchronized, yet the organization in classification levels is second to the organization along domain object hierarchies.
The DDO pattern is related to materialization [33] which also combines instantiation and specialization. With mate-rialization, a class of more concrete objects, e.g., Product-Model, may be related to a class of more abstract objects, e.g., ProductCategory. Each member, e.g., VolvoV50, of the former materializes a member, e.g, Car, of the latter. Materialization supports level-crossing property propagation.
The power type was introduced to conceptual modeling by Odell [31] as a class whose members are subclasses of another class.Power types were applied and evolved by Gonzalez-Perez and Henderson-Sellers [15] and analyzed in full detail by Carvalho et al. [8,10]. The partitions relationship which underlies our approach is a specific form of a power type relationship in the sense of Odell. Our notion of power types was introduced by Cardelli [7] and revisited by Carvalho et al. [8].
Some of the ideas underlying the current approach were already implicitly realized in the M-Object approach [28] but lacking the foundation in traditional object-oriented modeling and lacking the explicit induction of meta * -classes. An m-object, e.g., Car, has a set of totallyordered named levels, e.g., Category, Model, and Individual. Each level of an m-object implicitly defines a class with the top-level of an m-object implicitly defining a singleton class with the mob-ject itself as member. When a more concrete mobject, e.g., VolvoV50, concretizes a more abstract m-object, e.g., Car, then the classes implicitly defined by the levels of the more concrete m-object are specializations of the classes implicitly defined by the levels of the more abstract m-object, and the more concrete m-object is a member of the class implicitly defined by the second-top level of the more abstract m-object. Other than in the DDO approach, m-objects are not related by properties but by multi-level relationships instantiates its toplevel class, i.e., its own-type. An m-object, e.g., Car, may concretize another m-object, e.g., Product. The concretization relationship between m-objects has an instantiation and specialization facet. The concretizing m-object instantiates the concretized m-object's second-level class and specializes the concretized m-object's other classes. M-relationships relate m-objects at various levels.
The practical relevance of multi-level modeling was demonstrated by examining the application of multilevel modeling patterns in existing metamodels from different domains [23]. We believe that our approach fundamentally extends the practical relevance of multi-level modeling to modeling scenarios where domain objects are organized in part-of hierarchies (or similar) and where the goal is to refine and extend the schema along that hierarchy.
There is an ongoing discussion about the advantages and drawbacks of multi-level modeling with clabjects [5,11]. Various authors [9,34] argue that there is trade-off between compactness (reduction of model size) and understandability (semantic clarity). We claim that with our current approach both goals, namely semantic clarity and reduction of model size (of the model presented to the modeler), are met, first, by the approach's foundation in MLT and, second, by inducing multi-level structures and showing them only on demand.
Multilevel Coupled Model Transformations (MCMTs) [25] are an expressive approach for specifying level-crossing constraints. It remains to be analyzed how MCMTs can be employed on top of our approach or as an extension to our simple expression mechanism. Balaban et al. [6] introduce an executable multi-level modeling approach facilitating interlevel constraints, rules, and queries.
The multi-level theory (MLT) [8][9][10] improves semantic clarity through ontologically well-founded multi-level modeling constructs aligned with Guizzardi's Unified Foundational Ontology UFO [16]. The DDO approach is aligned in parts with MLT but does not support all cross-level structural relations identified in [8], in addition to power type of and partioning, they also propose a more relaxed relation called categorization which may be disjoint and/or complete. The partitions relation (which is fundamental to our approach) is a disjoint and complete categorization relation. It remains open to further analysis how the classes in the global model (which are classes without super-classes) relate to the notion of substance sortals as discussed by Guizzardi [16]).
Local object classes were proposed by Kappel and Schrefl [20] as a means to represent local referential integrity. There, a local object class 'is a set of objects belonging exclusively to some composite object", and the "database is considered an object itself" and "every object class is a local object class".

Conclusion
In this paper we have introduced the DDO approach, which may serve as the nucleus of multi-level modeling approaches that build on traditionally modeled domain object hierarchies. We have defined and implemented the approach in F-Logic/Flora-2. We have exemplified all constructs by a fully worked example. We have assessed the approach with regard to a set of desiderata together with a set of to-be supported modeling tasks. The proof-of-concept implementation of the approach and the fully worked example in F-Logic are provided in the Appendix and are made available as free software.
With the prototype, we have shown the feasibility of the approach and demonstrated its use for querying induced multi-level models. Qualitative evaluation of the approach in terms of usability and comprehensibility remains for future work, as well as systematic performance studies. In future work we will also explore the remaining design space for a full-fledged DDO-based multi-level modeling approach, especially regarding the multi-level modeling of properties and associations.
Funding Open access funding provided by Johannes Kepler University Linz.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.

A Implementation of DDO in F-logic
The implementation of the DDO approach in F-Logic was developed and tested with Flora-2/ErgoLite Reasoner 2.0 (Pyrus nivalis) of 2018-10-14. The prototype implementation is available as free software under https://github. com/bneumayr/DDO. The Flora-2/ErgoLite system is necessary to run the prototype and is available at http://flora. sourceforge.net/. The following code listing contains the complete implementation of the formalization of DDO.     ?c. order < config . maxOrder .

A.3 Alternative: rules for creating named classes class_creation_names.flr
The following rules create local classes and powertypes with names, e.g., CarModel and CarIndividual, as used throughout the paper. These names are generated by string concatenation which comes with performance overhead.

B Fully worked example in F-logic
In the example we use dot-notation to refer to local classes and powertypes. In this way the example facts can be executed with class names as well as with structured class IDs. ].