Provenance Network Analytics
Abstract
Provenance network analytics is a novel data analytics approach that helps infer properties of data, such as quality or importance, from their provenance. Instead of analysing application data, which are typically domaindependent, it analyses the data’s provenance as represented using the World Wide Web Consortium’s domainagnostic PROV data model. Specifically, the approach proposes a number of network metrics for provenance data and applies established machine learning techniques over such metrics to build predictive models for some key properties of data. Applying this method to the provenance of realworld data from three different applications, we show that it can successfully identify the owners of provenance documents, assess the quality of crowdsourced data, and identify instructions from chat messages in an alternatereality game with high levels of accuracy. By so doing, we demonstrate the different ways the proposed provenance network metrics can be used in analysing data, providing the foundation for provenancebased data analytics.
Keywords
Data provenance Data analytics Network metrics Graph classification1 Introduction
Provenance, a description of what influenced the generation of a piece of information or data, has become an important topic in several communities since it exposes how information flows in systems, providing the means to make them accountable and helping users decide whether information is to be trusted (Moreau 2010). Provenance has been recorded in an increasing number of applications, from legal notices,^{1} climate science (Ma et al. 2014), medical applications,^{2} scientific workflows (Alper et al. 2013; Silva et al. 2011; Davidson et al. 2007; Altintas et al. 2006), computational reproducibility (Chirigati et al. 2013), emergency response (Ramchurn et al. 2016), and in the geospatial domain.^{3}
As a provenance description ‘links’ artefacts with their influences, it can be represented in a graph, called a provenance graph, whose nodes represent the artefacts/influences and whose edges their relations with one another. Studying such graphs, e.g. by visualising them, can facilitate understanding of the provenance information they contain. However, in a typical application, provenance graphs can quickly become very large and complex; this makes it difficult to interpret their information manually. For instance, as an indication, the 2014 edition of the United States’ National Climate Assessment report^{4} was published with full provenance information linking its data and recommendations to 242 authors and over 500 distinct technical inputs (Tilmes et al. 2013). The scale is a few magnitudes larger with automated applications. CollabMap (Ramchurn et al. 2013), an online crowdsourcing platform, recorded more than 5000 provenance graphs over 3 months running, many of which contain 30–200 nodes, 50–700 edges. Scientific workflows (e.g. Wolstencroft et al. 2013; Silva et al. 2011; Gil et al. 2011; Bowers et al. 2008) being applied to petascale problems, are also generating vast amount of provenance information. Such large and complex graphs are overwhelming for manual interpretation or verification (of data correctness, for instance). Therefore, an automated and principled way to analyse provenance data of such scales and, more importantly, to understand what they convey with respect to the data they describe, is much needed.
 1.
First, we adapt a number of existing network metrics (Newman 2010) to suit provenance graphs and define provenancespecific ones to summarise the topological structure of provenance graphs. The provenance network metrics can be computed in a generic manner from provenance records and are independent of domainspecific information. Therefore, they provide the basis for analysing and/or comparing provenance graphs quantitatively, even those from different applications.
 2.
Second, we make use of the provenance network metrics to construct predictive models on provenance information based on known ground truths to relate provenance information with properties of data, such as their quality or importance. Once successfully trained for an application, those predictive models operate without relying on domainspecific information. By so doing, we devise a novel analytics method that analyses data using their provenance, not the data themselves. Thanks to the generic nature of the proposed provenance network metrics, our approach can be used to study data, via the means of their provenance graphs, in applications where provenance information is recorded.
 3.
Finally, we report the successful application of the above method on the provenance of realworld data from three different applications: identifying owners of provenance documents, assessing the quality of crowdgenerated data in CollabMap, and identifying instructions from chat messages in an alternatereality game. The applications were selected in part because they allow us to verify the accuracy of the proposed analytics via known ground truths or via an alternative method. In these applications, our analytic method achieved high levels of accuracy classifying data based on the provenance of such data. By so doing, we also demonstrate how the provenance network analytics approach can be concretely applied in specific contexts as a generic tool for data analytics.
2 Provenance network metrics
Since provenance information describes how various elements were related to, or influenced by, one another, it can be viewed as a directed graph in which those elements (i.e. entities, activities, agents) are represented as nodes, and the relations between them (e.g. used, wasGeneratedBy, wasDerivedFrom) as directed edges. Such a graph is called a provenance graph. Given that some provenance graphs can be very large, the challenge is how to extract useful information and knowledge from complex provenance graphs. In that respect, we turn to the established field of graph theory for principled methods to analyse graphs. Specifically, we are interested in network metrics that allow us to summarise the topological characteristics of a provenance graph, such as its shape, its size, or how its nodes tend to connect to one another. Such network metrics are generic and can be calculated on any graphs, including provenance ones. They provide us a way to summarise provenance graphs into a set of generic network features. As a result, they allow for the comparisons of provenance graphs, even those from different domains or applications, without the need for the knowledge required to interpret domainspecific information contained therein.
In the following subsections, we enumerate the network metrics we employ for our analysis of provenance graphs and provide their formal definitions. Section 2.1 describes the generic network metrics which we adapted to work with provenance graphs. We then define provenancespecific network metrics in Sect. 2.2 to take advantage of provenancespecific information readily available in a provenance graph such as the types of nodes and the relations between them.
2.1 Generic network metrics

Number of nodes \(n=\left V\right \), which is also the number of provenance elements in \(G\).

Number of edges \(e=\left E\right \), which is also the number of provenance relations in \(G\).
 Graph diameter \(d_{G}\) is the longest distance in a graph \(G\), where the distance between two vertices \(u\) and \(v\) is defined as the length of the shortest path between them, denoted by \(d\left( u,v\right) \). The graph diameter reflects how “spread out” the provenance graph \(G\) is.Since nodes in provenance graphs are separated by directed edges, thereby preventing some nodes from forming a path to certain others, strictly speaking, the diameter of each graph is, in many cases, infinite. However, by temporarily assuming the edges are undirected, we are able to calculate the diameter of a provenance graph. Hence, let \(G^{u}=\left( V,E^{u}\right) \) be the undirected counterpart of \(G\), i.e. whose edges are the same as those in \(G\) but undirected: \(E^{u} = E \cup \left\{ \left( v, u\right)  \left( u, v\right) \in E \right\} \). The diameter of a provenance graph \(G\) is then defined as \(d_{G^{u}}\). For the sake of brevity, we simply use \(d\) to denote the graph diameter of \(G^{u}\).$$\begin{aligned} d_{G}=\underset{u,v\in V_{G}}{\max }d\left( u,v\right) \end{aligned}$$(4)

Assortativity coefficient \(r\): Assortativity, or assortative mixing, is the tendency for vertices in networks to be connected to other vertices that are like them in some way (Newman 2003). The assortativity coefficient is the Pearson correlation coefficient \(r\) of degree between pairs of linked nodes. Positive values of \(r\) indicate a correlation between nodes of similar degree, while negative values indicate relationships between nodes of different degree. \(r\) is defined as per Eq. 24 in Newman (2003).
 Average clustering coefficient \(\mathsf {ACC}\): The local clustering coefficient \(c_{v}\) of a vertex in a graph quantifies how close its neighbours are to being a clique (complete graph) (Watts and Strogatz 1998) and was introduced to determine whether a graph is a smallworld network:where \(\deg _{v}\) is the number of \(v\)’s neighbours and \(\varGamma _{v}\) is the number of edges between the neighbours. The average \(c_{v}\) of all vertices in \(G\) represents the extent of neighbourhood clustering in \(G\). In order to avoid biased assessment of the metric, following Kaiser (2008), we exclude leaf and isolated nodes (i.e. \(\deg _{v} \leqslant 1\)) in our calculation:$$\begin{aligned} c_{v}=\frac{\varGamma _{v}}{\deg _{v}\left( \deg _{v}1\right) } \end{aligned}$$(5)$$\begin{aligned} \mathsf {ACC}= \overline{C} \text {, where } C=\left\{ c_{v}v\in V \wedge \deg _{v} > 1 \right\} \end{aligned}$$(6)

Degree distribution: For many realworld graphs, the degree distribution follows a ‘power law’ such that the number of vertices \(N_{k}\) with degree \(k\) is given by \(N_{k}\propto k^{\alpha }\), where \(\alpha > 0\) is usually called the powerlaw exponent. We examine the degree distribution of an entire provenance graph to determine whether the distribution fits a power law as per the method of Clauset et al. (2009) and, if so, the degreedistribution powerlaw exponent (DPE). For provenance graphs whose degree distribution does not fit a power law and \(\alpha \) is therefore undefined, we manually set \(\alpha = 1\).
2.2 Provenancespecific network metrics

Numbers of entities \(n_\mathsf {e}\), activities \(n_\mathsf {a}\), and agents \(n_{\mathsf {ag}}\) in a provenance graph.
 Maximum finite distance (MFD): Since provenance relations represent a form of influence (Moreau and Missier 2013), the length of the longest chain of influence in a provenance graph is a useful characteristic of the graph. It can be viewed in the same vein as the graph diameter in the previous section but now on the directed graph \(G\). In more detail, given two vertex sets \(X, Y \subset V\), let \(L_{X \rightarrow Y}\) the set of all finite distances separating a vertex in \(X\) with another in \(Y\): \(L_{X \rightarrow Y} = \left\{ d \left( u, v \right)  u \in X \wedge v \in Y \wedge d\left( u,v\right) \ne \infty \right\} \). The MFD between \(X\) and \(Y\) in \(G\), denoted as \(\mathrm {mfd}_{X \rightarrow Y}\), is defined as follows:As the kind of influence between different types of provenance elements is quite different, it is interesting to know the MFD between one node type and another. Considering only the distances between entities, for example, the MFD would reflect how far a piece of data was derived from, or somehow influenced by, another; while considering only the distances between agents might reveal how far delegation between them went. Since there are three different node types in a provenance graph, we define nine different MFD metrics, one for each pair of node types: \(\mathrm {mfd}_{t_{s} \rightarrow t_{e}}\), \(t_{s},t_{e} \in \left\{ \mathsf {e},\mathsf {a},\mathsf {ag}\right\} \), where \(\mathsf {e}\), \(\mathsf {a}\), and \(\mathsf {ag}\) are our shorthand notation to denote \(V\)’s subsets whose elements are , and , respectively.$$\begin{aligned} \mathrm {mfd}_{X \rightarrow Y} = {\left\{ \begin{array}{ll} 1 &{} \text {if } L_{X \rightarrow Y} = \emptyset \\ \max L_{X \rightarrow Y} &{} \text {otherwise} \end{array}\right. } \end{aligned}$$(7)
 MFD of derivations (\(\mathrm {mfd}_\mathsf {der}\)): Since is the only influence relation in PROV that has a strict time ordering (Cheney et al. 2013), we calculate additionally the MFD over this relation to examine the longest chain of derivations in a provenance graph. For this, we consider \(G_\mathsf {der}= (V, E_\mathsf {der})\), where \(E_\mathsf {der}= \left\{ e  e \in E \wedge {\text {type}}(e) = \mathsf {Derivation}\right\} \), i.e. the subgraph that contains only \(\mathsf {Derivation}\) relations from \(G\).where \(\mathrm {mfd}_\mathsf {der}\) is set to \(1\) if there is no derivation relation in the graph.$$\begin{aligned} \mathrm {mfd}_\mathsf {der}= {\left\{ \begin{array}{ll} 1 &{} \text {if } E_\mathsf {der}= \emptyset \\ \max \left\{ d_{G_\mathsf {der}}(u, v)  u, v \in V \wedge d_{G_\mathsf {der}}(u, v) \ne \infty \right\} &{} \text {otherwise} \end{array}\right. } \end{aligned}$$(8)
 Average clustering coefficients by node type (\(\mathsf {ACC}_{t}\)): This is a variation of the ACC metric (6); it is calculated from the local clustering coefficients of vertices of a given node type \(t \in \left\{ \mathsf {e}, \mathsf {a}, \mathsf {ag}\right\} \):As there are three different provenance node types, there are also three provenancespecific ACC metrics: \(\mathsf {ACC}_{\mathsf {e}}\), \(\mathsf {ACC}_{\mathsf {a}}\), and \(\mathsf {ACC}_{\mathsf {ag}}\).$$\begin{aligned} \mathsf {ACC}_t=\overline{C}_t \text {, where } C_{t}=\left\{ c_{v}v\in V\wedge {\text {type}}\left( v\right) =t\wedge \deg _{v}>1\right\} \end{aligned}$$(9)
Glossary of provenance network metrics
Metric name  Symbol  Variants  Number of metrics 

Number of elements  \(n\) and \(n_t\)  \(t \in \left\{ \mathsf {e}, \mathsf {a},\mathsf {ag}\right\} \)  4 
Number of relations  \(e\)  1  
Graph diameter  \(d\)  1  
Assortativity coefficient  \(r\)  1  
Average clustering coefficients  \(\mathsf {ACC}\) and \(\mathsf {ACC}_{t}\)  \(t \in \left\{ \mathsf {e}, \mathsf {a},\mathsf {ag}\right\} \)  4 
Degreedistribution powerlaw exponent  \(\alpha \)  1  
Maximum finite distance  \(\mathrm {mfd}_{t_{s} \rightarrow t_{e}}\)  \(t_{s}, t_{e} \in \left\{ \mathsf {e}, \mathsf {a}, \mathsf {ag}\right\} \)  9 
Maximum finite distance of derivations  \(\mathrm {mfd}_\mathsf {der}\)  1 
2.3 Summary
3 Provenance network analytics
The provenance network metrics defined in Sect. 2 can help us summarise a provenance graph’s topological characteristics and allow for the quantitative comparison of provenance graphs. The metrics can tell us a graph with \(n = 4\), \(e = 3\), and \(d = 3\) (i.e. a linear graph), for example, has a much different shape compared to a graph with \(n = 4\), \(e = 3\), and \(d = 2\) (i.e. a star graph). However, without the ability to relate those values to domainspecific interpretation, say, the former is the result of a valid run while the latter is not, the network metrics alone would not help us to gain useful information contained in such graphs.
 Design The purpose of this phase is to define the classification problem and to curate the required training data.
 1.
Define the classification labels: This step formalises the classification problem into a discrete set of labels \(\mathcal {L}\). Given a piece of data \(x\) from the application domain, the classification problem becomes that of predicting the label of \(x\): \(l_{x}\in \mathcal {L}\). For example, if we want to determine whether an application run is valid or not, we could have \(\mathcal {L} = \left\{ \textit{valid}, \textit{invalid} \right\} \); if we want to assess the quality of a data entity, we could have \(\mathcal {L} = \left\{ \textit{good}, \textit{bad}, \textit{uncertain} \right\} \).
 2.
Define the input provenance graph: Since we aim to use the provenance network metrics as inputs for a predictive model, we need to have the provenance graph of \(x\) to produce the metrics. As a provenance graph can record provenance of multiple entities, spanning from a few relations to the full history of an application run, choosing an appropriate extent of the input provenance graph \(G_x\) of \(x\) such that it sufficiently covers \(x\)’s related history to be considered but not too broad, is a key decision. If the chosen provenance graph is too small, it may not include relevant relations that could determine the label of \(x\); on the contrary, a too broad provenance graph, would have redundant information (i.e. noise) that could confuse a learning algorithm. Some knowledge of the application domain is useful for this step. In the above example of applicationrun validity, for instance, one might choose the whole provenance graph recorded from one run as the input, while a much smaller graph covering the generation and usages of a data entity might be more appropriate for the assessment of its quality. Concrete examples of this step are later provided in Sects. 5, 6, and 7, with the last showing how this step can be automated.
 3.
Curate training data: As this method relies on supervised learning techniques, a curated set of labelled training data \(S = \{ (x, l_x)  l_x \in \mathcal {L} \} \) is required (i.e. \(l_{x}\) is defined for all \(x\) in \(S\)).
 1.
 Training Having defined the label set \(\mathcal {L}\), defined the input provenance graph \(G_x\) for all \(x\) to be classified, and curated the training data set \(S\), we build the predictive model which is, in essence, a function that maps \(\mathcal {P}\left( G_x\right) \) (Eq. 10) to \(\mathcal {L}\):
 1.
Choose a supervised learning algorithm^{5} that suits \(\mathcal {L}\) and the given data set.
 2.
Calculate the network metrics for the provenance graphs of the labelled data and transform them into feature vectors with classification labels suitable as inputs to the chosen learning algorithm: \(I=\left\{ \left( \mathcal {P} \left( G_x \right) ,l_{x}\right)  (x, l_x) \in S \right\} \).
 3.
Assess the accuracy of the learning algorithm on the input labelled data \(I\).
 4.
If the accuracy in obtained in Step 3 is sufficiently high,^{6} build the classifier for \(\mathcal {L}\) from \(I\) with the chosen learning algorithm and proceed to the Prediction phase.
 1.

Prediction Use the classifier from the Training phase to predict the labels of unseen data from their provenance.
4 Empirical evaluation
As a tool for data analytics, the provenance network analytics method aims to discover correlations between provenance information and properties of the data it describes. In order to demonstrate the approach, we apply the method to the provenance of realworld data from three different applications and report its performance in the following sections. Before doing that, however, we first describe the common methodology for evaluating the method.
Learning algorithm We use the CART (Breiman et al. 1984) algorithm to train decision tree classifiers (specifically the Scikitlearn implementation by Pedregosa et al. 2011). Empirically, we find decision tree classifiers perform sufficiently well and were fast, although not always producing the highest accuracy. For the three selected applications, other learning algorithms we tested could only marginally improve classification accuracy while incurring significant increases in computing cost, in many cases several magnitudes higher, compared to that of the decision tree classifier (see the Extra 1 experiment in the online Supplementary Materials for more details). In addition, a decision tree classifier is able to explain its classification with decision rules, which may provide useful clues to understand the correlation between an application’s provenance data and the interested data properties.
Balancing learning data To avoid producing biased classifiers, for datasets whose samples are unbalanced, we balance the input dataset \(I\) using the SMOTE method (Chawla et al. 2011), which oversamples the minority samples such that each label has roughly the same number of samples in \(I\).
Assessing accuracy In order to benefit from all the available labelled data, which are small in some cases, we use 10fold crossvalidation (Kohavi 1995). In particular, with \(I\) randomly split into 10 equal subsets, we perform 10 rounds of learning; on each round a \(\frac{1}{10}\) subset is held out as the test set and the remaining are used as training data. To further minimise the potential chance impact of random data splitting, we repeat the above crossvalidation procedure 100 times, hence collecting 1000 accuracy scores in each experiment. We report the mean accuracy score alongside its 95% confidence interval in parentheses, for example, 98.13% (\(\pm \,0.01\% \)).
In addition to the accuracy of classifiers, we also evaluate the following.
Relevance of metrics From each round of learning in the crossvalidation procedure, the trained classifier automatically calculates the relevance of each input feature (i.e. each of the 22 network metrics in Eq. 10) given the training data. In practice, this information will help us selectively reduce the number of metrics to be considered within a specific application (if required). We report the three most relevant network metrics for each application, i.e. those with the highest average relevance values.
Generic versus provenancespecific metrics We repeat the above process (i.e. the Training phase and evaluation) in two further experiments—one using only the generic network metrics (Sect. 2.1) and the other only the provenancespecific network metrics (Sect. 2.2). Comparing the mean accuracy scores from the two experiments will help understand whether the network metrics based on provenance types bring added benefits to the classification application being discussed.
In the following sections, we report the exercise of provenance network analytics to build provenancebased classifiers for three different applications: identifying the owner of provenance documents on ProvStore (Huynh and Moreau 2015) (Sect. 5), assessing the quality of crowdsourced data in CollabMap (Sect. 6), and identifying instructions from chat messages in the Radiation Response Game (Sect. 7). In each application, since the Prediction phase, in which a classifier is run on unseen data, is straightforward, we will not discuss it but will focus on the Design and Training phases of the method, which are to be followed by an evaluation as outlined above. The datasets and code to produce the results and figures in the following sections are provided with this article in the Supplementary Materials (see “Appendix A” for more details).
5 Application 1: Identifying owner of provenance documents
As the PROV data model provides a vocabulary for provenance information, it is likely that each provenance producer has its own “style” of writing provenance using the vocabulary. For example, a user or an application may produce provenance graphs with chains of derivations that can be long or short, with or without attributions to agents, etc. Such individual styles will manifest in different topological characteristics of the resulting graphs and, hence, differences in the graphs’ provenance network metrics. Our hypothesis is that the metrics could be used to identify the user or the application that produced a provenance graph. In order to verify this, we analyse the provenance network metrics of provenance documents deposited by the public at ProvStore, which is a public repository for provenance documents where a user can sign up for an account and store their provenance online for sharing or visualisation purposes (Huynh and Moreau 2015). We apply the provenance network analytics method of Sect. 3 on those provenance documents to check how well it is able identify the documents’ owners, here used as a proxy for the application that generated the provenance. Note that those provenance graphs typically do not contain any information about the users who uploaded the graphs to ProvStore, so it is not possible to identify such users simply from querying the graphs.
5.1 Design phase
Graph labels We define the label set \(\mathcal {L} = \left\{ u_1, u_2, \ldots , u_n \right\} \), where \(l_{x} = u_i\) if the provenance document \(x\) belongs to user \(u_i\) and \(n\) is the total number of users.
Input graphs Since we want to identify the owner of a provenance graph based on its characteristics, we use the whole graph as the input graph, i.e. \(X = x\).
Training data In order to upload a provenance document to ProvStore, the document’s owner needs to register for a user account there. As a result, the owner of each document on ProvStore is known and, hence, a curated labelled data set containing all those documents is readily available. Since each user owns a different number of documents, in order to ensure that there are sufficient samples to represent a user’s provenance documents the Training phase, we limit our experiment to users who have at least 20 documents. There are fourteen such users (the authors were excluded to avoid bias), who we named \(u_{1}, u_{2}, \ldots , u_{14}\); hence, there are 14 labels in \(\mathcal {L}\). Their numbers of documents range between 21 and 6,745, with the total number of documents in the data set is 13,870.
5.2 Training phase
As described in Sect. 4, we train a single decision tree classifier to identify the owner of a given provenance document from the dataset. The tenfold cross validation shows that the classifier can identify owners of provenance documents on ProvStore with a mean accuracy of \(98.13\% \) (\(\pm \,0.01\% \)), compared to the baseline of \(7.14\% \) from selecting a random label from 14. This result strongly supports our hypothesis that the provenance network metrics represents the “signature” of provenance graphs, reflecting how the user or application that produces them models and records provenance information.
5.3 Discussion
After the Training phase, a decision tree classifier is able to explain its classification rules in the form of a decision tree. As an example, the decision tree for identifying document owners above is shown in Fig. 2, whose depth, however, was limited to three to fit the paper. From the decision tree, it is apparent that the most influential metrics selected by the algorithm are provenancespecific ones. The most important metrics, in this case, is \(\mathrm {mfd}_\mathsf {der}\); the tree splits the documents on ProvStore into two subsets: ones without derivation relation (i.e. \(\mathrm {mfd}_\mathsf {der}= 1\), see Eq. 8) and ones with at least one derivation (the right branch). The tree shows the next most important metrics to distinguish provenance documents in this dataset are \(\mathrm {mfd}_{e \rightarrow e}\) and \(\mathsf {ACC}_\mathsf {ag}\). From such information, we can see that half of the selected ProvStore users did not record derivations in their provenance at all. We can also learn that, for example, provenance documents uploaded by user \(u_2\) contains no derivation, has 4 or fewer PROV entities, and the distances between those entities, if any, are less than 2. In addition, knowing which metrics are most relevant within an application or a dataset, one can make informed decision on which features that can safely be ignored (to save computation cost) in an automated manner, e.g. as per the method by Kohavi and John (1997).
In order to confirm that provenance type information indeed contributes into the above classification performance as suggested by the decision tree in Fig. 2, we repeat the exercise, training a decision tree with the same dataset, but this time with only the six generic network metrics (i.e. \(n\), \(e\), \(d\), \(r\), \(\mathsf {ACC}\), \(\alpha \)) and later only the provenancespecific metrics. Compared to the first experiment, which makes use of all the available network metrics, using only the generic metrics achieves a lower accuracy at \(92.32\% \) (\(\pm \, 0.02\% \)), while using only the provenancespecific metrics produces a similar accuracy at \(98.11\% \) (\(\pm \,0.01\% \)). These results suggest that provenance type information, as captured by the provenancespecific network metrics, indeed helps with identifying the originator of a provenance graph, and ignoring such information will result in a lower performance. Nevertheless, even with only six generic network metrics, the trained classifier still achieves a very high level of accuracy (\(92.32\% \)). Therefore, we believe that characterising provenance information by their network metrics is a very promising approach, which can be effective even with a small set of metrics. In the next section, we develop this approach further to assess the usage of crowdgenerated data to infer about their quality, using only their provenance information.
6 Application 2: Assessing the usage of crowdsourced data
The provenance of a piece of data tells us the history that led to its creation. Analysing its provenance may help ascertain the data’s origin and that its production process was appropriate. It is, however, more challenging to infer about the data’s quality or significance from its history without knowing the quality, reliability, or trustworthiness of the data’s originator(s). Thus, instead of examining the data’s historical provenance, we propose an alternative approach that examines the data’s “forward provenance”—the records of how the data is used following its creation (also captured in an application’s provenance traces). In this section, we apply the provenance network analytics on such “forward provenance” of crowdsourced data from CollabMap (Ramchurn et al. 2013) to analyse their usage and, ultimately, their quality.
In order to support auditing the quality of its data, the provenance of crowd activities in CollabMap was fully recorded: the data entities that were shown to users in each microtask, the new entities generated therein, and their interrelationships (see Fig. 3 for a small example with two microtasks). In 2012, CollabMap was deployed to help map the area around the Fawley Oil refinery in the United Kingdom. It generated descriptions for 5,175 buildings, 4,997 routes, and 4,710 route sets. In this application, we apply the provenance network analytics method to construct three classifiers in order to assess the quality of CollabMap data from their provenance, one for each type of data.
6.1 Design phase
Graph labels The main aim of assessing the quality of CollabMap data is to determine which of them are sufficiently trustworthy to be included in the final evacuation map. For a data entity \(x\), we wanted to know whether \(x\) can be trusted to be correct (\(l_{x} = \textit{trusted}\)) or we are unsure about its quality (\(l_{x} = \textit{uncertain}\)). Thus, we define the label set \(\mathcal {L} = \left\{ \textit{trusted}, \textit{uncertain} \right\} \).
Input graphs Whereas Application 1 had provenance graphs deposited separately and, hence, discretely split, CollabMap ran continuously and generated data whose provenance is interwoven with one another’s. As a result, a CollabMap provenance graph contains the provenance of many data entities. As mentioned above, we propose to analyse the usage of a piece of data to infer its significance; for a given data entity \(x\), we extract the provenance graph that contains all the activities and entities that were influenced by \(x\). The intuition of the approach is similar to that of evaluating a publication from its citations. A highly cited academic paper, for example, is generally considered of high value thanks to its citations, or in other words, the relations it has with other papers. Such relations show how many times the paper was used in the generation of, or had an influence on, later papers.
CollabMap data quality classification results
Data type  Trusted  Uncertain  Accuracy (%)  95% confidence interval 

Building  4,491  684  90.03  ± 0.06% 
Route  3,908  1,089  96.98  ± 0.04% 
Route Set  3,019  1,691  95.70  ± 0.05% 
6.2 Training phase
We followed the methodology in Sect. 4, using decision tree classifiers and the same training and evaluation process. In this application, we trained three classifiers to classify the quality of buildings, routes, and route sets, one for each CollabMap data type. The accuracy of the classifiers is presented in Table 2 with the 95% confidence intervals and the number of samples available for each data type. The results show that the classifiers trained on the provenance network metrics of dependency graphs predict the trust labels for buildings, routes and route sets in the test sets with a high level of accuracy: 90% for buildings, 97% for routes, and 96% and route sets.
6.3 Discussion
With such high accuracy levels achieved by the classifiers, it is important to note that our method did not rely on any domainspecific information from CollabMap but only on generic, domainindependent provenance network metrics. The strong correlation between the provenance network metrics and data quality in CollabMap discovered by the classifiers suggests that analysing network metrics of provenance graphs is a promising approach to making sense of the (realworld) activities and data they describe, such as classifying crowdgenerated data into trust categories as in this case. The use of provenance network analytics in applications like CollabMap could potentially reduce significantly the number of required verification tasks (which incur a cost in resources and/or time). In such cases, only a much smaller set of verification tasks would need to be carried out to generate enough training data for building the quality classifiers as shown above. While the provenance of a piece of data is traditionally examined to study its history, the successful application of provenance network analytics over “forward provenance” to analyse data’s usage and significance in CollabMap shows that this can be an alternative useful approach for provenance analytics.
Generic versus provenancespecific metrics We retrained the three classifiers first using only the generic network metrics and later using only the provenancespecific metrics. The results (provided in Fig. 4) show that the classifiers trained only on the generic network metrics performed better than those trained only on the provenancespecific metrics in classifying buildings and routes but not route sets. However, the highest accuracy in this application were achieved by making use of the full set of provenance network metrics across the three data types.
7 Application 3: Identifying instruction messages
In the previous application, we introduce a method to extract the dependency graph of an entity of interest from a bigger provenance graphs to analyse its usage after creation. In this section, we show how the method can be further optimised to achieve the highest classification performance, in this case, inferring the significance of a chat message in the Radiation Response Game (RRG) (Fischer et al. 2014) from its “forward provenance”.

Agents: the field responders participating in the game and the headquarters.

Entities: the targets and the messages communicated in the game.

Activities: sending a message, picking up, transporting, and dropping off a target.
Since RRG was designed to study team coordination, the communications among participants are of particular interest as they can reveal when teams are formed and what led to their formation. Therefore, in addition to automatically tracked game logs, each participant’s voice communication is also recorded by individual recorders and their actions captured by video cameras. In a typical RRG game, there are eight to ten audio streams (one per responder), and four video cameras capturing the actions of the headquarters and the field responders over 30 min. Hence, posthoc analysis of these audio and video recordings to learn about when and how team coordination happened requires significant human efforts. In practice, Fischer et al. (2014) relied on the chat messages as a source to identify when teaming decisions were made and where to focus their investigation in the audio and/or video recordings. In order to do so, they first manually classified the chat messages into a number of different categories, the most interesting being that of directives as their study aims to determine whether, when, and why an instruction is followed or rejected (Fischer et al. 2014). Our intuition is that such instruction messages, either followed or rejected, would lead to various activities by the game’s participants following the moment the messages were received. For example, the participants could do as they were instructed, or they could send back more messages either to reject the instruction or to request further information. Since a RRG provenance graph also captures those activities, we believe that analysing the “forward provenance” of a chat message can help identify its role in the game. Hence, in what follows, we seek to apply the provenance network analytics method to identify instruction messages from their dependency (provenance) graphs.
7.1 Design phase
Graph labels For each message in a RRG, Fischer et al. (2014) classified them into one of the six categories: directives, assertives, expressives, declarations, commissives, and requests. Directives in RRG are typically instructions from the headquarters allocating tasks to the responder teams on the ground and are the targets for the classifier. Therefore, we label a message with directive, if it is one, or other, otherwise: \(\mathcal {L} = \{ \textit{directive}, \textit{other} \} \).
Training data We recorded a single provenance graph for the RRG game reported by Fischer et al. (2014), where there were 69 messages sent, 32 of which were categorised as directives. The dataset for this application is, hence, relatively balanced between the two labels (46 vs 54%); hence, no data rebalancing was carried out. For the training, each directive message is labelled as directive, while the rest as other.
7.2 Training phase
The results show that the classifier correctly identified directive messages on an average three out of four times for \(k \ge 3\). It performed slightly worse at \(k=2\) and did no better than the base line at \(k=1\), suggesting that little or no useful network information was contained in such shallow dependency graphs. The exploration procedure discovers that \(k = 11\) yields the top performance at 85.13% (± 0.79%).
7.3 Discussion
In this application, we show how dependency graphs can be parameterised by their maximum depth (\(k\)) to help extract relevant input graphs for network analytics (in applications whose provenance graphs are too large and encompassing). The optimal value of \(k\) can be discovered in an automated manner by trialling different values and checking the classifier’s performance with each value as shown earlier. Although the message classifier’s accuracy is not as high as in the previous two applications, this level of accuracy (i.e. 85%) is sufficiently high for it to be useful as an automated analytic tool for the RRG study. It could assist posthoc analysis of future RRG studies by labelling instructions from chat messages based on their provenance, significantly reducing manual efforts in classifying them as it has hitherto been done. Since our analytics method does not rely on the domain data but only on the provenance of activities in an RRG game, it can certainly be applied in other studies where similar provenance is captured. In those cases, the method can help identify points of interest in a study for further investigations, saving time and efforts of researchers going through voice and video recordings. In scenarios where instructions are already identified (e.g. tasks allocation, military orders), the result from this application suggests that analysing the “forward provenance” of such instructions could help determine their compliance with a predictive model.
Relevance of metrics As with the previous applications, we examine the relevance of the network metrics; however, given the numerous configurations of \(k\), we only give a summary of the results here. The detailed results are available in the Supplementary Material. Across various values of \(k\), we found that the most influential metrics was the number of edges \(e\), followed by the number of entities \(n_e\) and \(\mathrm {mfd}_{e \rightarrow a}\). This is compatible with our earlier intuition that a directive message generally would generate more game activities, manifesting in more entities and provenance relations (i.e. edges) in the message’s dependency graphs.
Generic versus provenancespecific metrics Comparing the accuracy of classifiers trained only on the generic network metrics and that of those trained only on the provenancespecific metrics across the 18 values of \(k\), the result is mixed. As shown in Fig. 6, both perform similarly in 7 cases, using provenancespecific metrics outperforms using only generic metrics in 7 cases, and in the remaining 4 cases, the reverse is true. It is difficult to draw a clearcut conclusion from this. However, the result indicates that the provenancespecific metrics still plays a significant role in this application. Finally, both types of network metrics perform equally well in with \(k = 11\), delivering the top accuracy for this application.
8 Related work
Our work is conducted within the context of the descriptive analysis of network graph characteristics (Kolaczyk 2009). It has been shown that when studying a complex system such as a longterm crowdsourcing application or any program giving rise to a large amount of data (provenance or otherwise), various questions of interest can be rephrased usefully as questions regarding some aspect of the structure or characteristics of the corresponding network graph (Brandes and Erlebach 2005). For example, particular notions of the importance of individual system elements may be captured by measurements related to the corresponding vertices in the network. Indeed, Vaz de Melo et al. (2012) provide a compelling example of why the inputs to a predictive algorithm should sometimes be based on network topology (such as those related to changes in the relationships among sports players and coaches) rather than node attributes (namely a player’s performance statistics).
The field is continually evolving, and graphs can be viewed in a growing number of ways; provenance data itself can be interpreted as collaboration networks (Altintas et al. 2010) or otherwise. Recently, Margo and Smogor (2010) examined a provenance graph based on components of a filestore, to show that provenance and other metadata can successfully predict semantic attributes: in particular, they predicted file extensions in a filehistory graph. (As in earlier sections, ‘predict’ refers not to the temporal sense of the word, but to the reinferring of removed data.) Although their particular choice of attribute to predict “has few applications”, the study functioned as a useful proof of concept. The authors employed the C4.5 decision tree algorithm on their provenance graph, with the network structure and artefact attributes as input; the levels of accuracy achieved were comparable to our own, even though in the present work we examine provenance graphs of a different topology and size. The authors recognised that further exploration of the feature space over provenance graphs was called for; among other things, our methodology extends the types of features used in such analyses.
Related to our work in categorising provenance graphs, Cheah and Plale (2012) proposed a method to check for “structural flaws” in provenance graphs from workflow execution in order to detect anomalies. A component of the method relies on counting the number of nodes and edges from a set of provenance graphs to identify graphs that have too few or too many nodes/edges compared to their quantiles computed from the whole population. This approach, however, is only effective in the cases where provenance graphs are recorded from a workflow that is consistent in its outputs. Our approach, instead, employs machine learning techniques to reveal subtler and more complex correlations between such metrics and data properties. Moreover, it makes use of many more network metrics and also takes into account provenance type information. In order to cope with the potential complexity of provenance graphs (containing both structural information and node/edge provenance attributes), instead of directly analysing their network topology (like we did in this work), Chen et al. (2014) proposed partitioning provenance graphs into subsets of vertices according to their temporal ordering. Scalar features (e.g. vertex type, the number of nodes in the subset, the average number of characters in node names) can then be collected for each subset of a provenance graph to represent the graph in tasks such as graph clustering, graph classifying, and rule mining. The key difference here is that the number of features can vary greatly depending how many subsets a provenance graph is partitioned into according to the Logical P algorithm by the same authors (Chen et al. 2014). Therefore, the aforementioned data mining tasks can only be performed with graphs having the same number of subsets (i.e. the same number of features). Our provenance network metrics, on the contrary, are calculated on whole provenance graphs, and, hence, provide the same number of features regardless of graph size. In addition, the Logical P algorithm was designed to work with the Open Provenance Model (Moreau et al. 2011) while our method was based on the later PROV Data Model (Moreau and Missier 2013) standardised by the World Wide Web Consortium. Given that not all PROV relations have temporal constraints associated with them, the Logical P algorithm may not work with certain PROV graphs.
Similar to our data quality assessment in Application 2 (Sect. 6), Ceolin et al. (2014) sought to assess trustworthiness of crowdsourced data using provenance information. They relied, however, on node attributes (such as timestamps and typing speeds) rather than the network topology, in a different application area to ours (annotation of museum collections), achieving accuracies of approximately 80%. CrowdTruth (Inel et al. 2014) is another crowdsourcing annotation application that sought to derive the quality crowdsourced data from a set of metrics on disagreements within the collected data. This work derived the metrics from the actual content of the data, not from analysing the relationships between them as per our method.
Our method provides a broader type of analysis than certain previous work on hyperlink network analysis (Park 2003) in which the links between web pages were studied to estimate the value of websites (e.g. their credibility) or to identify social networks. In the former case, the previous work only counted the number of links and did not investigate the network connections further than one link away (in contrast with the size of dependency graphs in our analyses). In the latter, the focus was on clustering similar nodes or detecting outliers, e.g. isolated nodes or those with few links, not on predicting node attributes as in this work. Also relatedly, Varlamis and Louta (2009) count network links in order to express trustworthiness, as an example application of the principle described by Yu and Singh (2000) that propagation can be considered one of the properties of trust (along with symmetry, transitivity, selfreinforcement, etc.) However, Varlamis and Louta (2009) did not have voting data available in order to assess the accuracy of their model in the way we could as in the CollabMap application.
More generally speaking, graphs, as a generic and flexible data representation, are ubiquitous in describing computation. Analysing graph data is, hence, an active research topic of multiple communities in a variety of fields such as graphbased semisupervised learning (Subramanya and Talukdar 2014), graph mining (Aggarwal and Wang 2010), and more. The latter includes frequent pattern mining (Cheng et al. 2014), graph clustering (Gaertler 2005), graph classification (Tsuda and Saigo 2010), etc. Our work largely falls into the last area by providing a method for predicting the label of a whole provenance graph, as opposed to predicting the label of a node in the graph, also known as “label propagation” (Bengio et al. 2006). However, compared to other graph classification techniques, our method makes use of network metrics instead of graph kernels (Vishwanathan et al. 2010) or boosting (Saigo et al. 2009); and the metrics were specifically constructed to work with PROV provenance graphs. Such provenance network metrics have not been studied before and our work is the first to propose employing them for characterising realworld properties of data in an automated manner.
9 Conclusions
Characterising properties of data, such as their quality or importance, can be challenging, especially with those generated by human contributors (like crowdsourced data or chat messages). It is usually a manual process that requires retrospection by experts who understand well the concerned application domain; in some other cases, it instead relies on the opinions of the participants (e.g. via a votinglike mechanism). In this work, we propose applying machine learning techniques on the network metrics of provenance graphs to explore and automate data characterisation. In particular, we have presented a generic and principled data analytics method for analysing data and applications based on their provenance graphs. Using this method, via the means of offtheshelf machine learning algorithms, it is now possible to explore and learn about some properties of the data from their provenance in an automated manner. Since the method employs common network analyses and machine learning techniques on generic provenance graphs, it can be used in a wide range of applications where provenance are captured (or can be generated from application data/logs). Indeed, we have demonstrated the applicability of this method within three different applications: (1) identifying the owners of provenance documents on ProvStore, (2) classifying the trust labels for buildings, routes, and route sets drawn by crowd contributors in CollabMap, and (3) identify instructions from chat messages in the RRG; all resulted with high levels of accuracy. At the same time, we show how the method can be customised and optimised to suit a particular application context. Particularly, the results of Application 3 (Sect. 7) also led us to believe that the provenance network analytics method can be a useful analytic tool for studying human activities or determining their compliance.
While avoiding using domainspecific information allows the provenance network metrics to be generically applied, we also appreciate the potential value of applicationspecific data in improving classification performance. In addition, certain applications may produce provenance graphs having the same topological characteristics, resulting the same set of network metrics values, confusing predictive models based solely on those. Therefore, in another future direction, we plan to extend our proposed metrics to utilise domainspecific information recorded in provenance information. We expect such customised provenance network metrics, albeit no longer generic, will help improve accuracy in analysing an application’s data and will work with provenance graphs of highly similar topology.
In the three applications reported in this paper, we were able to collect the full provenance information recorded by them. This may, however, not always be the case. There can be applications where the provenance records available for analytics are incomplete or corrupted, or parts of them might be intentionally hidden or transformed to protect sensitive information (Cheney and Perera 2015; Danger et al. 2015; Missier et al. 2015; Hussein et al. 2016). It is an open question how resilient a predictive model based on provenance network metrics performs against such variances in the input provenance graphs. An extension of this work, thus, could be on studying models of imperfect provenance information and their effects on provenance network analytics.
With an increasing number of applications continuously generating provenance,^{9} we can quickly get overwhelmed with torrents of provenance data requiring our attention. The provenance network analytics method presented here can potentially be applied on provenance graph summaries, such as those produced by graph summarisation techniques (e.g. Moreau 2015; Riondato et al. 2016), as significantly smaller proxies of the original provenance graphs, in order to lower computation costs. The analyses can also be extended to study the provenance network metrics that characterise the evolution of provenance graphs (like those introduced by Ebden et al. 2012), which reflect the development of the tasks they represent. Such an extension could help us to understand developing dynamic behaviours, and to allow for appropriate onthefly interventions (in order to stop an undesirable behaviour from progressing, for instance).
In a wider context, provenance graphs do not only describe the origin of data, but they also reveal the interactions of agents in connected activities and how the activities themselves unfolded at the same time. The provenance network metrics presented in this work, therefore, could find useful applications in other areas in addition to those presented here. Analysing the influence of agents in the provenance graph of a collaborative task could identify the most valuable team member. Studying the distances between the agents in the graph could reveal close collaboration or team breakdown; or finding frequent patterns (Kuramochi and Karypis 2005; Yan et al. 2008) in provenance graphs may show how they usually work together. In addition, focusing on the activities in a graph could help detect bottlenecks, important data, and activities that were crucial to the outcome of a task. Given the generic nature of network analysis techniques, the possibilities are highly promising and vast.
Footnotes
 1.
 2.
 3.
 4.
The online version of the report, provided with its provenance, is available at http://nca2014.globalchange.gov/.
 5.
 6.
The required level of accuracy depends on the intended application of the predictive model.
 7.
The RRG provenance graph used later in this section has 1,682 nodes and 4,184 edges.
 8.
The accuracy level declines with \(k > 15\) and, thus, we stop the exploration at \(k = 18\).
 9.
See http://provenanceweek.org/2016/p3yl/programme.html for examples.
References
 Aggarwal CC, Wang H (2010) Graph data management and mining: a survey of algorithms and applications. In: Aggarwal CC, Wang H (eds) Managing and mining graph data, advances in database systems, chap 2, vol 40. Springer, Boston, pp 13–68. https://doi.org/10.1007/9781441960450_2 CrossRefGoogle Scholar
 Akoglu L, Tong H, Koutra D (2015) Graph based anomaly detection and description: a survey. Data Min Knowl Discov 29(3):626–688. https://doi.org/10.1007/s106180140365y MathSciNetCrossRefGoogle Scholar
 Alper P, Belhajjame K, Goble CA, Karagoz P (2013) Enhancing and abstracting scientific workflow provenance for data publishing. In: Proceedings of the joint EDBT/ICDT 2013 workshops, ACM, New York, NY, USA, EDBT’13, pp 313–318. https://doi.org/10.1145/2457317.2457370
 Altintas I, Barney O, JaegerFrank E (2006) Provenance collection support in the Kepler scientific workflow system. In: Proceedings of the 2006 international conference on provenance and annotation of data, Springer, IPAW’06, pp 118–132. https://doi.org/10.1007/11890850_14
 Altintas I, Anand MK, Crawl D, Bowers S, Belloum A, Missier P, Ludäscher B, Goble CA, Sloot PMA (2010) Understanding collaborative studies through interoperable workflow provenance. In: McGuinness DL, Michaelis JR, Moreau L (eds) Provenance and annotation of data and processes. Springer, Berlin, Heidelberg, pp 42–58CrossRefGoogle Scholar
 Bengio Y, Delalleau O, Roux NL (2006) Label propagation and quadratic criterion. In: Olivier C, Schölkopf B, Zien A (eds) Semisupervised learning. MIT Press, Cambridge, pp 193–216. https://doi.org/10.7551/mitpress/9780262033589.003.0011 Google Scholar
 Bowers S, McPhillips T, Riddle S, Anand MK, Ludäscher B (2008) Kepler/pPOD: Scientific workflow and provenance support for assembling the tree of life. In: Freire J, Koop D, Moreau L (eds) Provenance and annotation of data and processes, Lecture Notes in Computer Science, chap 9, vol 5272. Springer, Berlin, pp 70–77. https://doi.org/10.1007/9783540899655_9 CrossRefGoogle Scholar
 Brandes U, Erlebach T (2005) Network analysis: methodological foundations. Springer, BerlinCrossRefzbMATHGoogle Scholar
 Breiman L, Friedman J, Olshen R, Stone C (1984) Classification and regression trees. Wadsworth, BelmontzbMATHGoogle Scholar
 Ceolin D, Nottamkandath A, Fokkink W (2014) Efficient semiautomated assessment of annotations trustworthiness. J Trust Manag 1(3):1–31. https://doi.org/10.1186/2196064X13 Google Scholar
 Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2011) SMOTE: Synthetic minority oversampling technique. J Artif Intell Res 16:321–357. https://doi.org/10.1613/jair.953 zbMATHGoogle Scholar
 Cheah YW, Plale B (2012) Provenance analysis: towards quality provenance. In: 2012 IEEE 8th international conference on escience, IEEE, pp 1–8. https://doi.org/10.1109/eScience.2012.6404480
 Chen P, Plale B, Aktas MS (2014) Temporal representation for mining scientific data provenance. Future Gener Comput Syst 36:363–378. https://doi.org/10.1016/j.future.2013.09.032 CrossRefGoogle Scholar
 Cheney J, Perera R (2015) An analytical survey of provenance sanitization. In: Ludäscher B, Plale B (eds) Provenance and annotation of data and processes. IPAW 2014, Lecture Notes in Computer Science, vol 8628. Springer, Cham, pp 113–126. https://doi.org/10.1007/9783319164625_9 Google Scholar
 Cheney J, Missier P, Moreau L, Nies TD (2013) Constraints of the PROV data model. W3C recommendation RECprovconstraints20130430, World Wide Web Consortium, http://www.w3.org/TR/2013/RECprovconstraints20130430/
 Cheng H, Yan X, Han J (2014) Mining graph patterns. In: Aggarwal CC, Han J (eds) Frequent pattern mining, chap 13. Springer, Berlin, pp 307–338. https://doi.org/10.1007/9783319078212_13 Google Scholar
 Chirigati F, Shasha D, Freire J (2013) Reprozip: using provenance to support computational reproducibility. In: Proceedings of the 5th USENIX conference on theory and practice of provenance, USENIX Association, Berkeley, CA, USAGoogle Scholar
 Clauset A, Shalizi C, Newman M (2009) Powerlaw distributions in empirical data. SIAM Rev 51:661–703MathSciNetCrossRefzbMATHGoogle Scholar
 Danger R, Curcin V, Missier P, Bryans J (2015) Access control and view generation for provenance graphs. Future Gener Comput Syst 49:8–27. https://doi.org/10.1016/j.future.2015.01.014 CrossRefGoogle Scholar
 Davidson SB, Boulakia SC, Eyal A, Ludäscher B, McPhillips TM, Bowers S, Anand MK, Freire J (2007) Provenance in scientific workflow systems. Data Eng Bull 30(4):44–50Google Scholar
 Ebden M, Huynh TD, Moreau L, Ramchurn S, Roberts S (2012) Network analysis on provenance graphs from a crowdsourcing application. In: Groth P, Frew J (eds) Provenance and annotation of data and processes, Lecture Notes in Computer Science, vol 7525. Springer, Berlin, pp 168–182. https://doi.org/10.1007/9783642342226_13 CrossRefGoogle Scholar
 Fischer JE, Jiang W, Kerne A, Greenhalgh C, Ramchurn SD, Reece S, Pantidi N, Rodden T (2014) Supporting team coordination on the ground: requirements from a mixed reality game. In: Rossitto C, Ciolfi L, Martin D, Conein B (eds) COOP 2014—Proceedings of the 11th international conference on the design of cooperative systems. Springer, Nice, pp 49–67. https://doi.org/10.1007/9783319064987_4 Google Scholar
 Gaertler M (2005) Clustering. In: Brandes U, Erlebach T (eds) Network analysis, Lecture Notes in Computer Science, chap 8, vol 3418. Springer, Berlin, pp 178–215. https://doi.org/10.1007/9783540319559_8 Google Scholar
 Gil Y, Ratnakar V, Kim J, GonzalezCalero P, Groth P, Moody J, Deelman E (2011) Wings: intelligent workflowbased design of computational experiments. IEEE Intell Syst 26(1):62–72. https://doi.org/10.1109/MIS.2010.9 CrossRefGoogle Scholar
 Hussein J, Sassone V, Moreau L (2016) A templatebased graph transformation system for the prov data model. In: Seventh international workshop on graph computation models GCM 2016Google Scholar
 Huynh TD, Moreau L (2015) ProvStore: a public provenance repository. In: Ludäscher B, Plale B (eds) 5th international provenance and annotation workshop, IPAW 2014, Lecture Notes in Computer Science, vol 8628. Springer, Cologne, pp 275–277. https://doi.org/10.1007/9783319164625_32 Google Scholar
 Inel O, Khamkham K, Cristea T, Dumitrache A (2014) CrowdTruth: machinehuman computation framework for harnessing disagreement in gathering annotated data. In: Mika P, Tudorache T, Bernstein A, Welty C, Knoblock C, Vrandečić D, Groth P, Noy N, Janowicz K, Goble C (eds) The semantic web—ISWC 2014, Lecture Notes in Computer Science, vol 8797. Springer, Berlin, pp 486–504. https://doi.org/10.1007/9783319119151 Google Scholar
 Kaiser M (2008) Mean clustering coefficients: the role of isolated nodes and leafs on clustering measures for smallworld networks. New J Phys 10(8):083,042. https://doi.org/10.1088/13672630/10/8/083042 CrossRefGoogle Scholar
 Kohavi R (1995) A study of crossvalidation and bootstrap for accuracy estimation and model selection. In: Proceedings of the 14th international joint conference on artificial intelligence, vol 2. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, pp 1137–1143Google Scholar
 Kohavi R, John GH (1997) Wrappers for feature subset selection. Artif Intell 97(1–2):273–324. https://doi.org/10.1016/S00043702(97)00043X CrossRefzbMATHGoogle Scholar
 Kolaczyk E (2009) Statistical analysis of network data. Springer, BerlinCrossRefzbMATHGoogle Scholar
 Kuramochi M, Karypis G (2005) Finding frequent patterns in a large sparse graph. Data Min Knowl Discov 11(3):243–271. https://doi.org/10.1007/s1061800500039 MathSciNetCrossRefGoogle Scholar
 Lebo T, Sahoo S, McGuinness D (2013) PROVO: the PROV ontology. Tech. Rep. RECprovo20130430, World Wide Web Consortium. https://www.w3.org/TR/2013/RECprovo20130430/, W3C Recommendation
 Ma X, Fox P, Tilmes C, Jacobs K, Waple A (2014) Capturing provenance of global change information. Nat Clim Change 4(6):409–413. https://doi.org/10.1038/nclimate2141 CrossRefGoogle Scholar
 Margo D, Smogor R (2010) Using provenance to extract semantic file attributes. In: Proceedings of the 2nd conference on theory and practice of provenance, Berkeley, USA, USENIX AssociationGoogle Scholar
 Marsland S (2014) Machine learning: an algorithmic perspective. Chapman and Hall/CRC, LondonCrossRefGoogle Scholar
 Missier P, Bryans J, Gamble C, Curcin V, Danger R (2015) ProvAbs: Model, policy, and tooling for abstracting prov graphs. Provenance and annotation of data and processes. IPAW 2014, Lecture Notes in Computer Science, vol 8628. Springer, Cham, pp 3–15. https://doi.org/10.1007/9783319164625_1 Google Scholar
 Moreau L (2010) The foundations for provenance on the web. Found Trends Web Sci 2(2—3):99–241. https://doi.org/10.1561/1800000010 CrossRefGoogle Scholar
 Moreau L (2015) Aggregation by provenance types: a technique for summarising provenance graphs. In: Graphs as models 2015, London, UK, pp 129–144. https://doi.org/10.4204/EPTCS.181.9
 Moreau L, Missier P (2013) PROVDM: The PROV data model. Tech. Rep. RECprovdm20130430, World Wide Web Consortium. http://www.w3.org/TR/2013/RECprovdm20130430/, W3C Recommendation
 Moreau L, Clifford B, Freire J, Futrelle J, Gil Y, Groth P, Kwasnikowska N, Miles S, Missier P, Myers J, Plale B, Simmhan Y, Stephan E, Van den Bussche J (2011) The open provenance model core specification (v1.1). Future Gener Comput Syst 27(6):743–756. https://doi.org/10.1016/j.future.2010.07.005 CrossRefGoogle Scholar
 Newman M (2010) Networks: an introduction. Oxford University Press, OxfordCrossRefzbMATHGoogle Scholar
 Newman MEJ (2003) Mixing patterns in networks. Phys Rev E 67(2):026,126. https://doi.org/10.1103/PhysRevE.67.026126 MathSciNetCrossRefGoogle Scholar
 Park H (2003) Hyperlink network analysis: a new method for the study of social structure on the web. Connections 25(1):49–61Google Scholar
 Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay E (2011) Scikitlearn: machine learning in Python. J Mach Learn Res 12:2825–2830MathSciNetzbMATHGoogle Scholar
 Ramchurn SD, Huynh TD, Venanzi M, Shi B (2013) CollabMap: Crowdsourcing maps for emergency planning. In: 5th ACM web science conference (WebSci ’13)Google Scholar
 Ramchurn SD, Huynh TD, Wu F, Ikuno Y, Flann J, Moreau L, Fischer JE, Jiang W, Rodden T, Simpson E, Reece S, Roberts S, Jennings NR (2016) A disaster response system based on humanagent collectives. J Artif Intell Res 57:661–708. https://doi.org/10.1613/jair.5098, http://www.jair.org/papers/paper5098.html
 Riondato M, GarcíaSoriano D, Bonchi F (2016) Graph summarization with quality guarantees. Data Min Knowl Discov 31(2):314–349. https://doi.org/10.1007/s1061801604688 MathSciNetCrossRefGoogle Scholar
 Russell S, Norvig P (2010) Artificial intelligence: a modern approach, 3rd edn. Pearson, LondonzbMATHGoogle Scholar
 Saigo H, Nowozin S, Kadowaki T, Kudo T, Tsuda K (2009) gBoost: A mathematical programming approach to graph classification and regression. Mach Learn 75(1):69–89. https://doi.org/10.1007/s109940085089z CrossRefGoogle Scholar
 Silva CT, Anderson E, Santos E, Freire J (2011) Using VisTrails and provenance for teaching scientific visualization. Comput Graph Forum 30(1):75–84. https://doi.org/10.1111/j.14678659.2010.01830.x CrossRefGoogle Scholar
 Subramanya A, Talukdar PP (2014) Graphbased semisupervised learning, synthesis lectures on artificial intelligence and machine learning, vol 8. Morgan & Claypool, Los Altos. https://doi.org/10.2200/S00590ED1V01Y201408AIM029 zbMATHGoogle Scholar
 Teacy WTL, Patel J, Jennings NR, Luck M (2006) TRAVOS: Trust and reputation in the context of inaccurate information sources. Auton Agents MultiAgent Syst 12(2):183–198CrossRefGoogle Scholar
 Tilmes C, Fox P, Ma X, McGuinness DL, Privette AP, Smith A, Waple A, Zednik S, Zheng JG (2013) Provenance representation for the national climate assessment in the global change information system. IEEE Trans Geosci Remote Sens 51(11):5160–5168. https://doi.org/10.1109/TGRS.2013.2262179 CrossRefGoogle Scholar
 Tsuda K, Saigo H (2010) Graph classification. In: Aggarwal CC, Wang H (eds) Managing and mining graph data, advances in database systems, chap 11, vol 40. Springer, Berlin, pp 337–363. https://doi.org/10.1007/9781441960450_11 CrossRefGoogle Scholar
 Varlamis I, Louta M (2009) Towards a personalized blog site recommendation system: a collaborative rating approach. In: Fourth international workshop on semantic media adaptation and personalization, IEEE, San Sebastian, Spain, pp 38–43. https://doi.org/10.1109/SMAP.2009.17, http://ieeexplore.ieee.org/document/5381709/
 Vaz de Melo POS, Almeida VAF, Loureiro AAF, Faloutsos C (2012) Forecasting in the NBA and other team sports. ACM Trans Knowl Discov from Data 6(3):1–27. https://doi.org/10.1145/2362383.2362387 CrossRefGoogle Scholar
 Vishwanathan SVN, Schraudolph NN, Kondor R, Borgwardt KM (2010) Graph kenrels. J Mach Learn Res 11:1201–1242MathSciNetzbMATHGoogle Scholar
 Watts DJ, Strogatz SH (1998) Collective dynamics of ‘smallworld’ networks. Nature 393(6684):440–2. https://doi.org/10.1038/30918 CrossRefzbMATHGoogle Scholar
 Wolstencroft K, Haines R, Fellows D, Williams A, Withers D, Owen S, SoilandReyes S, Dunlop I, Nenadic A, Fisher P, Bhagat J, Belhajjame K, Bacall F, Hardisty A, Nieva de la Hidalga A, Balcazar Vargas MP, Sufi S, Goble C (2013) The Taverna workflow suite: designing and executing workflows of web services on the desktop, web or in the cloud. Nucleic Acids Res 41(W1):W557–W561. https://doi.org/10.1093/nar/gkt328 CrossRefGoogle Scholar
 Yan X, Cheng H, Han J, Yu PS (2008) Mining significant graph patterns by leap search. In: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, ACM, pp 433–444. https://doi.org/10.1145/1376616.1376662
 Yu B, Singh MP (2000) A social mechanism of reputation management in electronic communities. Cooperative information agents. Springer, Berlin, pp 154–165Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.