A data mining framework for targeted category promotions

This research presents a new approach to derive recommendations for segment-specific, targeted marketing campaigns on the product category level. The proposed methodological framework serves as a decision support tool for customer relationship managers or direct marketers to select attractive product categories for their target marketing efforts, such as segment-specific rewards in loyalty programs, cross-merchandising activities, targeted direct mailings, customized supplements in catalogues, or customized promotions. The proposed methodology requires customers’ multi-category purchase histories as input data and proceeds in a stepwise manner. It combines various data compression techniques and integrates an optimization approach which suggests candidate product categories for segment-specific targeted marketing such that cross-category spillover effects for non-promoted categories are maximized. To demonstrate the empirical performance of our proposed procedure, we examine the transactions from a real-world loyalty program of a major grocery retailer. A simple scenario-based analysis using promotion responsiveness reported in previous empirical studies and prior experience by domain experts suggests that targeted promotions might boost profitability between 15 % and 128 % relative to an undifferentiated standard campaign.


Introduction
Even though their popularity might already have exceeded peak levels, with an average of more than twelve memberships per U.S. household and a reported half of U.S. adults being enrolled in at least one, loyalty programs continue to be a mainstay in customer relationship management (Kivetz and Simonson 2003;Ferguson and Hlavinka 2007;Berry 2013). Many companies invest tremendous amounts of money in both, their online and offline loyalty program environments to strive for building and preserving loyalty of their primary clientele. In the marketing literature, however, there is mixed evidence on the effectiveness of loyalty programs and many recent efforts to improve them concentrate on various program design components, tier or reward structures; excellent reviews are provided by Liu (2007), Liu and Yang (2009) and Zhang and Breugelmans (2012).
In a nutshell, loyalty programs have been mainly criticized for their vanishing ability to gain a competitive advantage in an environment where almost all major competitors in a particular industry rival in ''loyalty wars'' for the most profitable clients (Kumar and Shah 2004;Shugan 2005;Singh et al. 2008). However, some smart companies have learned how to squeeze out valuable customer insights from the vast amount of data permanently accruing from monitoring customer interactions with their touch points and to benefit from personally identifiable purchase history data by customizing targeted marketing activities (Ailawadi et al. 2010;Liu 2007;Bodapati 2008). For example, the U.K.'s biggest and most profitable grocery chain Tesco pioneered data-driven loyalty programs by deriving a lifestyle segmentation of their customer base using behavioral data. At Tesco, these ''lifestyle'' segments are constructed by looking into the composition of their customers' shopping baskets and used for deriving segment-specific targeted mailings, coupons and promotions (Humby et al. 2004). A similar approach is adopted by the French grocery retailer Carrefour.
In this paper, we take the perspective of multi-category retailers like Tesco or Carrefour, who need to manage their category level merchandising and target marketing decisions to increase sales generated by their existing customer base (Chen et al. 1999;Rowley 2005). Such retailers frequently make use of targeted promotions to draw consumers' attention to specific categories. One of the key challenges in such a setting is to decide which categories to promote from among the hundreds or even thousands they offer and to whom, i.e., which customer segment(s) to target. We propose a procedure which addresses both the customer segmentation issue and the task to support managers with selecting attractive categories for deriving segment-specific, targeted category level promotion campaigns 1 . In addition, we will explore the potential benefits of adapting the presented promotional decision support system and compare its empirical performance relative to that of a simple standardized promotion heuristic.
In the next section, we position the proposed framework against prior related research. Section 3 then presents the building blocks of the methodology to determine product categories to be featured in segment-specific promotional campaigns. Section 4 illustrates the empirical application of the framework by analyzing a real-world transaction dataset collected from a retailer's loyalty program. We also provide rough estimates of the expected outcome of our procedure to support target marketing campaigns using a simple scenario-based setting and compare the profitability implications with those anticipated from a standardized promotion heuristic. Finally, Sect. 5 discusses the results and provides an outlook on future enhancements of the proposed approach.

Literature review and research contribution
In our research framework we consider targeted category level promotions in the same manner as Venkatesan and Farris (2012). These authors focus on retailerinitiated coupon campaigns, which are customized to customers' specific preferences (as reflected in their purchase histories) and are targeted to only a subset of the retailer's clientele. Such targeted promotions are typically offered by major retailers as part of their loyalty programs or are distributed by specialized target marketing services like Catalina Marketing for cooperating retailers and manufacturers (Zhang and Wedel 2009;Pancras and Sudhir 2007).
In recent years, the effectiveness of targeted promotions has increasingly been studied by marketing researchers (e.g., Rossi et al. 1996;Shaffer and Zhang 1995;Zhang and Krishnamurthi 2004). There is empirical evidence that compared to conventional (i.e., undifferentiated) ones targeted promotions are capable to increase profits (Khan et al. 2009;Musalem et al. 2008). In an early contribution, Bawa and Shoemaker (1989) show for direct mailing coupons that consumers are more responsive to (coupon) promotions if their prior preference for the promoted brand is higher.
Using survey data, Shoemaker and Tibrewala (1985) also report that consumers indicate higher redemption intentions for brands they are loyal to. Zhang and Wedel (2009) show that profit differences are mainly the result of variations in redemption rates and in offline stores the incremental benefit of individual level targeting is relatively small compared to segment-level customization. In a retailer-customized setting, Venkatesan and Farris (2012) also provide support for higher redemption rates of targeted promotions. Furthermore, beyond a lift in coupon redemption the authors also find a mere exposure effect of customized coupon campaigns. This is consistent with a recent study by Sahni et al. (2014); using data from a set of randomized field experiments the authors find significant carryover effects after the promotions expired and evidence for cross-category spillover to non-promoted items. Summing up, these findings suggest that promotions customized to the prior preferences of customer segments can boost company's profits.
Most prior research on designing targeted promotions has focused on how to detect interesting customer segments to target. For example, the direct approach by Bodapati and Gupta (2004) predicts whether a prospective customer exceeds a predetermined threshold on a defined outcome (e.g., grocery expenditures). Rossi et al. (1996) assess the information content of various information using a target couponing problem that customizes coupons to specific households. Shaffer and Zhang (1995) analytical framework notes the effect of targeting coupons to selected households on firm profits, prices, and coupon face values. Zhang and Krishnamurthi (2004) provide recommendations about when to promote how much to whom, according to the time-varying pattern of purchase behavior and impact of current promotions on future purchases.
Notwithstanding its importance, most of this prior research neglects the selection of which category to promote for the derived customer segment. The approach presented in the next section aims to support decision makers in this respect. Our analytical framework shares some common notions with the approaches introduced by Reutterer et al. (2006) and Boztug and Reutterer (2008). We also segment customers based on the multi-category choices observed in their past purchase history data with the focal company and we also derive the targeted categories based on their aptitude to stimulate cross-category spillover. However, beyond technical aspects, the major differences of the present contribution against these previous studies are as follows: For each customer segment our approach provides the decision maker with a list of candidate categories for segment level targeted promotion campaigns; the list is derived such that the included categories maximize the cross-category spillover effects for non-promoted categories. This task is accomplished by combining various data mining tools with optimization techniques in one integrated analytical framework.
Furthermore, in developing our approach we explicitly distinguish between two types of product categories: The first type contains categories purchased by a significant fraction of a specific company's customer base. Such ''bestsellers'' or top-selling categories show high purchase incidence rates and are very frequently bought compared to the rest of the assortment; we therefore denote these as highfrequency categories (HFCs). In a grocery retailing context, such HFCs typically include every day food categories like fresh milk, vegetables, bread or other fast moving consumer good categories. Using Drèze and Hoch (1998) terminology such ''type 1 categories'' are purchased by customers with the focal company on a regular basis whenever they visit the store.
Thus, such HFCs are perfect for traffic building and useful candidates for undifferentiated (or non-targeted) promotions to draw customers into the store. However, they are less useful for the targeted promotions we aim to derive, because their category expansion effects tend to be modest (Bell et al. 1999). For the purpose of deriving customer segments and selecting categories for segmentspecific targeted promotions, we instead focus on a second type of categories we denote as low-frequency categories (LFCs), i.e., categories which show relatively low purchase incidences on the aggregate level but might be characterized by substantial variation across customers. The underlying rationale of considering such LFCs for target marketing purposes is related to the so-called ''long tail effect'' (Anderson 2006;Elberse 2008), which suggests that multi-category retailers can stimulate previously untapped demand by detecting and promoting specific category combinations that reflect distinctive tastes and preference structures at the individual customer or segment level but are ''averaged out'' (i.e., vanishing in relative small purchase incidences) on the aggregate level. More precisely, differentiating customer preferences and buying habits are more likely to be reflected in their specific multi-category choices in the ''long tail'' (i.e., LFCs) than in categories purchased by the vast majority of a company's clientele. For example, a baby household and a young single household will probably both buy milk, bread and vegetables in combination. However, the latter household is not very likely to purchase any baby hygiene products. Instead, the shopping baskets of young singles might be significantly characterized by convenience food categories, frozen food, etc. Thus, we posit that using characteristic LFC combinations found in customers' purchase histories might enhance the effectiveness of targeted promotions. The next section presents the technical details of the proposed procedure to derive targeted segment level promotions.
3 Methodological framework Figure 1 illustrates the stepwise procedure of our proposed framework for deriving recommendations of cross-category purchase sensitive items for targeted segment level marketing campaigns. To find a suitable customer segmentation which takes customers' past purchase habits into account, step 1 employs a constrained Kcentroids cluster algorithm (KCCA) as introduced by Leisch and Grün (2006). In step 2, an association rule mining (ARM) analysis identifies the segment-specific frequent itemsets in the pooled transactions for each segment detected in the previous step. In accordance with other association rule mining approaches, an additional filter measure separates the statistically interesting frequent itemsets from the less important ones and helps to reduce the number of considered cross-category associations. Finally, in step 3 the itemsets are used as input for an optimization procedure which recommends a list of categories maximizing profits with respect to their own profitability and a profit lift due to expected cross-category purchase associations. Fig. 1 Stepwise procedure of the proposed framework for deriving categories for segment-specific targeted promotions A data mining framework for targeted category promotions 341

Step 1: identifying household segments
It is common practice among marketing analysts conducting exploratory market basket analysis to assume that each customer transaction (i.e., the shopping basket or market basket) reflects the output of a combined multi-category decision process made during a shopping trip (Manchanda et al. 1999;Russell and Petersen 2000;Hruschka 1991;Kwak et al. 2015). Following previous research this is considered as a ''pick-any/J'' decision task where each transaction can be represented as a Jdimensional binary vector x n 2 f1; 0g J of category purchase incidences, with J denoting the number of categories considered 2 . A database of N transactions then gives the data set X N ¼ fx n ; 1 n Ng.
Our approach employs a constrained K-centroids cluster algorithm introduced by Leisch and Grün (2006). In general, K-centroids methods (such as K-means, McQueen 1967) partition data sets by finding a set of K centroids P K ¼ fp k ; 1 k Kg which optimally represent the data set, in the sense that the total distance between the data points x n and their centroids pðx n Þ 2 P K become minimal. Formally, with d(x, p) the distance between x and p, one aims at solving which implies that pðx n Þ should be taken as the p k closest to x n , and hence naturally provides a partition of the data points according to the closest centroids. i.e., the K segments C K ¼ fc 1 ; . . .; c K g obtained are such that c k contains all x n for which p k is the closest centroid from P K . Finding such optimal representations is typically based on heuristics which iterate between computing optimal centroids for the current partition and optimal partitions for the current centroids (see Bock 1999;Leisch 2006 for more information). We follow Leisch (2006, Sect. 3.2) in taking d as the extended Jaccard distance, such that dðx n ; x m Þ is the relative frequency of categories purchased in only one of the transactions n and m (but not in both).
For personalized basket data, each transaction can be linked to a household it originates from. To identify household segments, one could follow Boztug and Reutterer (2008) to employ a two-step voting procedure, which first segments the transactions without taking the household information into account, and then assigns households to segments according to the majority of their transactions. Here, we follow a constrained clustering approach which already employs the household information when clustering the transactions via a so-called ''must-link'' constraint (Wagstaff et al. 2001;Basu et al. 2008), enforcing all transactions corresponding to one household to the same segment. This immediately yields segments of households with similar basket compositions.
The K-centroids approach very conveniently allows to impose such must-link constraints (Leisch and Grün 2006). Write X N;h for the transactions in X N corresponding to household h, and H for the number of households. Then for all h, all transactions x n in X n;h should have the same centroid pðX n;h Þ 2 P K , and constrained K-centroids clustering is performed via solving The thus obtained segmentation yields centroid vectors p k which correspond to the prototypical ''average'' market basket for their corresponding segment c k (and are typically similar to the vectors of category purchase frequencies of the respective segments, Leisch and Grün 2006).
A common challenge in the application of K-centroids based cluster methods is the determination of an appropriate value of K (Aldenderfer and Blashfield 1984;Milligan and Cooper 1985;Kaufman and Rousseeuw 2005). Although this information is not available before the analysis, in most real-world situations Kcentroids partitioning requires the analyst to predefine a priori the number of expected groups in the dataset or to use heuristics like the 'elbow' criterion (Thorndike 1953;Gordon and Vichi 1998), cluster validation indices (Dimitriadou et al. 2002) or index voting. Our approach to choosing K is based on the idea of increasing K until the corresponding partitions no longer markedly change. More precisely, we employ the ''corrected'' Rand (1971) for measuring the agreement of two different partitions of the same data set. We compute the constrained Kcentroids partitions for a suitable range of K values, and then inspect the Rand indices of the partitions using K and K þ 1 centroids. We then choose K large enough to account for all large changes in the sequence of indices (ensuring that the obtained partition is rather stable with respect to increasing K).

Step 2: mining segment-specific frequent itemsets
Whereas the segment centroids p k reflect the market basket structure of an average transaction of the segment, they do not provide any information on which categories are exactly bought in combination within the segments' transactions. As such the set of centroid vectors merely informs the analyst about the specific ''interests'' of the various household segments in certain (combinations of) categories. For example, observing that a segment features rather frequent purchases of both white wines and red wines does not allow to conclude that these purchases occur together (i.e., in the same baskets). However, identifying interesting category associations clearly is a key ingredient to successful personalized target marketing of the type discussed in the previous section. This can be accomplished by employing transaction data mining techniques for finding frequent so-called itemsets and association rules (e.g., Brijs et al. 2004;Reutterer et al. 2007;Kamakura 2012).
In our application context, itemsets correspond to sets of categories. We say that an itemset A is contained in a transaction x, symbolically A x, if x features purchases of all categories in A. The basic measure of interestingness of an itemset is its support, which is the frequency of transactions containing the itemset. Formally, for transactions from segment c k , where |S| denotes the cardinality of a set S (i.e., the number of its elements). If the support value of an itemset is above a user-defined threshold (so-called minimum support), the itemset is referred to as ''frequent'' (Mannila 1997). Even for very large transaction databases, frequent itemsets can efficiently be mined, for example, by using the APRIORI algorithm (Agrawal et al. 1993;Agrawal and Srikant 1994;Bayardo and Agrawal 1999;Zaki et al. 1997).
An association rule A ! B splits an itemset C ¼ A [ B into two non-empty disjoint itemsets A and B, the antecedent and the consequent of the rule. The strength of the association is typically measured by the confidence of the rule, which is the conditional frequency of transactions containing B within the transactions containing A. Formally, for transactions from segment c k , As in general confðA ! BÞ 6 ¼ confðB ! AÞ, confidence provides an asymmetric measure of the statistical strength of the association between two itemsets A and B.
To separate the statistically attractive frequent itemsets from the ones less so, several measures of interest have been developed (Hettich and Hippner 2001;Hahsler et al. 2006). A commonly employed symmetric measure of the overall strength of association within an itemset is the so-called all-confidence (Omiecinski 2003), which computes the minimal confidence of all association rules that can be generated from the itemset. Formally, (itemsets with at most one element have zero all-confidence). Employing all-confidence to measure statistical association is attractive within our application context because it promises particularly good results in transaction datasets which exhibit itemsets with markedly varying support values (Agrawal et al. 1993;Hui et al. 2006).

Step 3: optimization and filtering segment-specific itemsets
If associations are mined with a low minimum support in a dataset showing skewed purchase frequencies, the analyst has to be aware of finding many weakly related cross-support itemsets (Hui et al. 2006). This results from grouping customers with similar interest in purchasing certain categories of the assortment (e.g., customers of a ''baby'' cluster disproportionately often buy baby related products). As outlined in Sect. 3.2, these problems can be addressed using the all-confidence value suggested by Omiecinski (2003) to reduce the output of the APRIORI algorithm. We thus filter the frequent itemsets obtained from step 2 accordingly, retaining those F frequent itemsets with the highest all-confidence values (and hence length at least two), for a suitable value of F. The remaining frequent itemsets are then transferred to the proposed optimization model, generating a list of single categories which should be promoted within the corresponding customer segment.
To select only the most valuable items for customized marketing, we use the generalized PROFSET model introduced by Brijs et al. (2004), which determines the most profitable categories based on their profit lift into frequent itemsets of interest, by solving an all-binary optimization problem. The resulting categories imply a high monetary value and the ability to initiate cross-selling in the respective customer segment.
Let us write J for the J categories to select from, and F for the frequent itemsets of interest based on these categories (in our approach, F consists of the F frequent itemsets with the highest all-confidence values). Write Q j for the binary variable indicating the selection of category j (i.e., Q j ¼ 1 if j is selected, and zero otherwise). The PROFSET optimization selects the U best categories by solving X

A2F
VðAÞ for the binary decision variables Q j and P A , subject to where the first constraint ensures that an itemset can only be selected if all categories it contains are selected, and the second constraints ensures that exactly a prescribed number U of categories are chosen, COST j gives category-specific handling and inventory costs, and V(A) is the ''value'' (profit margin) of the itemset A which is obtained by suitably aggregating the values of the transactions containing it. The value of a single transaction x n is given by with SP(j) and PP(j) the sales and purchase prices, respectively, and f ðj; x n Þ the number of times j was purchased in transaction x n . When aggregating the transaction values into itemset values, care must be taken to avoid that transactions contribute to several itemsets (e.g., all itemsets they contain). The original PROFSET approach thus takes V(A) as the sum of the vðx n Þ over all transactions exactly matching A (i.e., featuring purchases of exactly the categories in A). When employing only frequent itemsets, this may result in excluding many transactions in the value aggregation, and hence under-estimating the actual values (see Section 3.3 in Brijs et al. (2004)). This effect is particularly relevant in our approach which is based on using relatively small numbers of frequent itemsets with interesting crosscategory associations as measured by all-confidence. We thus generalize the PROFSET model as follows: For each transaction under consideration, we determine all frequent itemsets in F it contains, and distribute the value of the transaction to the values of these itemsets weighted according to the support values of the itemsets, either by direct distribution, or alternatively by randomly selecting the itemsets according to these weights. For example, if a transaction contains exactly the two frequent itemsets fvegetables; waterg with support 0.02 and fbottled beer; waterg with support 0.07, then the weight of the first itemset is 0:02=ð0:02 þ 0:07Þ ¼ 2=9, and the weight of the second is 7/9.
Employing the PROFSET approach requires the specification of a pre-defined number U of categories to be selected. Because both marketing budgets and advertising spaces are scarce resources in both on-and offline environments, loyalty program managers tend to be rather interested in focusing their target marketing efforts on a selected few product categories than dealing with a multitude of interrelated itemsets. After using e.g. a branch-and-bound algorithm to solve the allbinary PROFSET optimization problem, the solution determines U variables, which point to the categories to be selected for maximizing the objective function and subsequently used in target marketing actions.

Empirical application
The application of the proposed framework to derive recommendations for segmentspecific, category level targeted promotions is demonstrated below using a realworld data set. We obtained the data from a major grocery retailer who prefers to stay anonymous. The data set contains transactions realized by members of the loyalty program offered by the focal retailer. Using a simple scenario-based analysis, we illustrate and discuss the benefits and comparative effectiveness of our proposed data-driven target marketing approach vis-a-vis undifferentiated standard promotions.

Data description
The data set at hand contains more than 1.4 million transactions made within one year by 56,000 households which are enrolled in the retailer's loyalty program. The records available for the shopping baskets include prices, quantities and average gross profit margins for 268 categories. As illustrated by Fig. 2,  assortment is dominated by a small range of categories that are bought very frequently. Therefore, and in accordance with our conceptual arguments in Sect. 2, the assortment is separated into two distinct types of categories: the ''bestsellers'', which are the 52 HFC and occur at least in 10 % of all shopping baskets (left-hand side of the vertical dotted line in Fig. 2), and the ''long tail'' range of the assortment which are the remaining 216 LFC (right-hand side of the vertical dotted line in Fig. 2).
To illustrate the practical application of our proposed procedure, we drew two disjoint samples from the transaction data base, each containing the transactions of 3,000 randomly selected households. The first sample is used for selecting an appropriate KCCA segmentation model, i.e., to determine an appropriate value for K and the centroids P K of the corresponding constrained cluster solution. The second sample is used for performance evaluation, using months 1-10 to update the KCCA segmentation obtained from sample 1 by performing KCCA with the chosen K and the centroids initialized with P K , and using months 11 and 12 as the hold-out sample for the profitability scenario-based analysis.
In grocery retailing it is typical that households have varying lengths of buying histories. On average, households in sample 1 made around 26 transactions with a basket size of six categories in the observation period from the LFC. To robustify the selection of an appropriate value for K, we only consider households with buying sequences that are sufficiently long but not extremely long. Specifically, we exclude those households with the smallest 20 % and largest 5 % numbers of transactions, leaving 2,250 households in sample 1 to use for selecting an appropriate KCCA model.

Identifying household segments and extracting itemsets for targeting
According to step 1 of our proposed procedure (see Fig. 1), the first goal is to partition the households of sample 1 into segments with the constrained cluster algorithm. To reduce the risk of getting stuck at a weak local optimum every partitioning task is repeated up to fifteen times (Gordon and Vichi 1998;Aldenderfer and Blashfield 1984;Hornik 2005), retaining the partition minimizing

Number of clusters (k)
Rand−index Fig. 3 Cluster agreement by comparing K to K þ 1 cluster solutions using the Rand index. Note that the higher the value the higher the pairwise similarity between two partitions the target function. The value of the Rand index apparently levels off after partitions with K ¼ 11 and K ¼ 12 (cf. Fig. 3). It can also be seen that the arrangement of the data points would not change radically if a further cluster K ¼ 12 (or K [ 15) is added. Hence, we decide on K ¼ 11, which also allows for sufficiently convenient interpretation by retail managers. We next derive the household segments of the second sample based on the transactions for months 1-10 by running the constrained cluster algorithm with the chosen K ¼ 11 and the centroids initialized with those of the corresponding P K (equivalently, the transactions of each household are initially simultaneously assigned to the best matching centroid from P K ). Figure 4 (dark-grey bars) shows the amounts of baskets from months 1-10 in the segments thus obtained. We can see that there exists one large segment (k ¼ 6) containing about 20 % of all baskets and a smaller segment (k ¼ 10) containing less than 5 %. The remaining baskets are assigned to the other nine segments quite equally. After clustering, the generated household segments are labeled according to the most frequent itemsets within each cluster (cf. Fig. 4).
To give an illustrative example of the cross-category purchase interrelationships resulting from the proposed algorithm, Fig. 5 depicts the results for two of the clusters generated from the second sample obtained for the segments k ¼ 8 and k ¼ 1. The different peaks of light-gray bars on the left-hand side in Fig. 5 Fig. 4 Number of household segments for the second sample of months 1-10 as well as the rank level of each cluster density in brackets (1, highest density; 11, lowest density) compared to the number of assigned households and the profit generated for months 11-12 in percentages itemsets in the corresponding segments are investigated (cf. Fig. 5, right-hand side). The households in segment k ¼ 8 (the ''baby'' cluster) seem to focus on baby food and baby care categories since these products are purchased at a substantially higher than average rate. A typical household being represented by the baby cluster buys baby hygiene products with a probability of 35.24 % and adds baby food in glass with a probability of 22.82 % and baby food mush/powder with a probability of 17.62 %. The households in segment k ¼ 1 (the ''wine'' cluster) combine different kinds of-in particular-wine or other alcoholic beverages. In contrast to its overall purchase probability of 3.86 %, red/rosé wines occur at a rate about ten times higher than average in a basket purchased by a wine cluster household (red/rosé wine's group-specific purchase frequency is 32.21 %).
Other clusters contain itemsets related to health food (such as cereals, organic products, whole meal products, frozen ice cream, etc.), meat (beef, chicken, other kinds of meat, etc.) or beverages (soda, lemonade, water, etc.) and are therefore equally easy to label with a generic term. However, some segments obtained do not contain itemsets with such an obvious interpretation. Therefore, to keep the application in this paper simple, these clusters are labeled as ''mixed'' clusters (cf. Fig. 4), referring to household segments with cross-category associations which are not as straightforwardly interpretable as the generically labeled ones.
The number of mined associations depends on the pre-determined minimum support. Usually, analysts prefer low support thresholds to detect less obvious associations within the transaction datasets (Hui et al. 2006). Based on the  Fig. 5 Graphical illustration of the market baskets of segments k ¼ 8 (baby cluster, above) and k ¼ 1 (wine cluster, below). The black solid line (left-hand side) represents the sample's overall relative purchase frequencies, the light-gray bars correspond to the 216 categories purchase frequencies within the clusters; compared to the ten most frequently purchased itemsets of each segment (right-hand side) APRIORI algorithm all frequent itemsets with a minimum support of 2 % are revealed (cf. Sect. 3.2, step 2 in Fig. 1). The resulting 70 frequent itemsets with the highest all-confidence value and a minimum length of two (ignoring circular associations) are passed on the PROFSET optimization model for further examination (cf. Sect. 3.3, step 3 in Fig. 1). Table 1 includes the itemsets which are obtained for the wine, the health food and the baby segments in Fig. 4 (in the present PROFSET application a value of U ¼ 4 was chosen). The categories included in these itemsets exactly correspond to the categories recommended by our proposed procedure to be featured in marketing actions targeted to the respective three segments. Analogous recommendations (which are not displayed here for space reasons) can be obtained for the remaining segments. In comparison, the last column in Table 1 lists the top-four ''bestselling'' (in terms of generated sales values for the same households and observation period under study) categories from the set of HFCs. While these categories were excluded from our segmentation and subsequent frequent itemset mining procedure, they would represent promising candidates for an undifferentiated, traffic-building promotional campaign. Next we further explore the potential performance of a segment-specific targeting approach against a standardized campaign.

Scenario analysis to evaluate profitability implications
After performing the stepwise procedure illustrated above, managers are provided with (a) a set of household segments, corresponding centroids and household assignments to segments as well as (b) for each segment an itemset including the categories recommended for target marketing actions. Now, suppose that a loyalty program manager considers to launch a targeted promotions campaign for the previously identified household segments. Of course, an evaluation procedure of first choice would be to run a series of (randomized) field experiments and to compare the effectiveness of targeted promotions relative to an undifferentiated approach (or doing nothing). Because we do not have access to such experimental data, we discuss some basic and preliminary considerations from a managerial perspective. In fact, prior to costly experimentation both analysts and-more importantly-managers typically wish to gain some initial notions on the prospective chances of success of such an approach and for which segments targeting is most likely to pay off. In doing a preliminary feasibility study, we next adopt a simple scenario-based evaluation of such a strategy by making some assumptions based on prior empirical findings.

Scenario settings
The scenario analysis is pursued by estimating the profit margin generated with the itemsets recommended for the different campaigns (targeted vs. standard) using the empirical basket data set at hand. For this purpose we reutilize the transaction data included in the second sample used to illustrate the empirical application of our approach. Note that the households' transactions for months 1-10 were used to determine segment memberships and for deriving itemset recommendations. This data is now also used to calculate the expected profit margin for the following two months, which serve as a hold-out period. As profit margins are available on all categories, it is possible to determine the profits realized by the focal retailer with a certain itemset. This profit simulation is done using all 268 categories and all 3,000 households in the second sample.
Since we extend the cluster membership of a household determined for the calibration sample to the hold-out period, we also check whether the cluster size coincides relative to the number of included households (cf. Fig. 4, light gray bars) and the generated percentage profit gains (cf. Fig. 4, white bars) for months 11-12. Despite some smaller deviations, the three values obtained mostly correspond to each other. Thus we conclude that the size of the cluster approximately determines the profits generated by the households of the corresponding segments.
We assume that the retailer at hand considers to conduct a segment-specific promotional campaign for a predefined set of four categories per segment (see the examples in Table 1) in the hold-out period. For example, such a campaign could be effected by distributing targeted coupons among the household members for each segment. With targeted coupons, customers typically can earn a discount of a certain monetary amount (or a percentage value equivalent) if they bought at least one product in the mentioned category within a predefined period of time (Kalwani and Yim 1992;DelVecchio et al. 2007). For evaluating the expected effectiveness of such a targeted marketing campaign we use a standardized promotion campaign as a benchmark (i.e., promoting categories with the highest revenue in months 1-10; see the HFC categories listed in the last column of Table 1). To compare the expected profits resulting from both campaign-types we use the expected gains in profit margins for the baskets along the complete set of categories (i.e., the 52 HFC with the previously identified 216 LFC). Following Brijs et al. 2004 we only consider direct product costs and ignore category handling and inventory costs for sake of simplicity. Furthermore, we also do not consider any potential costs to implement a targeted marketing strategy.
To evaluate the relative performance of the two campaign-types in terms of profitability, their respective expected profit increases are calculated in months 11-12 conditional on the selected promotional campaign. Next we add up all the profit margins potentially accruing from two scenarios which we define according to Table 2 and discuss further below. Due to cross-category associations we also have to consider the purchase interrelationships between the promoted itemsets and the rest of the assortment in step 3 of the proposed framework. Therefore, after mining all association rules with a minimum support of 2 % and a length of two (i.e., the major associations) the profits are multiplied with the confidence value of the corresponding rule, revealing the expected indirectly affected accumulated profits generated by selling the promoted categories. Finally, we estimate the percentage profit gain which is expected to be achieved by the corresponding promotion campaign compared to the real profit achieved in month 11-12. Thus, we implicitly assume stationary marketing activities in both the calibration and hold-out periods and do not account for any seasonal or stock-buying effects.
For evaluation purposes, we define the following two scenario settings: The first setting assumes a lower responsiveness for the segment-specific target marketing campaign and higher values for the standardized promotion campaign, while the second setting does the opposite.
The profit lift values expected for specific combinations of itemsets and campaign types under the two scenarios are included in Table. These values are based on prior empirical findings reported in the relevant marketing literature (e.g., Drèze and Hoch 1998;Zhang and Wedel 2009;Venkatesan and Farris 2012;Sahni et al. 2014) and discussions with domain experts working with the focal retailer. For segment-specific targeted coupons, Drèze and Hoch (1998) report a 25 % increase in the promoted categories after the program has been running for six months and taking costs into account. The overall profitability of the campaign depends on the length of the coupon's validity period; this applies even more, if the promoted categories are from the LFCs and average per-basket sales in the these categories usually tend to increase over time (Drèze and Hoch 1998). Thus, we assume a pessimistic value of 10 % for setting no. 1. In contrast to Drèze and Hoch (1998) the promoted categories of our approach are matched to the purchasing behavior of the targeted households taking purchase interrelationships into account. Therefore, a more optimistic profit lift of as much as 15 % in the promoted itemsets is assumed for setting no. 2. In addition, in a second study by Drèze and Hoch (1998) the authors applied cross-merchandising techniques (of the type ''save a certain amount on category B products if you purchase category A products'', Drèze and Hoch 1998, see Section 3) and found sales increases in the targeted category ranging from 6 to 10 %. Since the targeted items in cross-merchandising campaigns correspond to the right-hand side categories of rules derived from frequent itemsets, we adopt these values as proxies in our scenario-based analysis (cf. Table 2). For the standardized promotion campaign, we refer to the meta-analyses by Tellis (1988) and Bijmolt et al. (2005) which summarizes empirical research related to price elasticities on which our scenario assumptions are based on. Since the categories determined by the standardized promotion all come from the grocery domain the projected profit increase of 5 % employed in setting no. 1 is very optimistic. Nevertheless, to avoid overestimating the results from the segmentspecific framework, a growth of 3 % is still estimated for the accumulated profits of the promoted categories when the standardized promotion method is applied for the more pessimistic scenario setting no. 2 3 . It is also possible that the sales of associated itemsets could rise as much as the sales of the promoted categories (5 % in setting no. 1), but in fact, the gain will likely be much smaller. Therefore, we assume as a pessimistic outcome that the profit in the associated itemsets will increase by only 1 %.

Results
The bars in Fig. 6 represent the expected profit margin gains in months eleven and twelve within each household segment derived for the two scenarios in Table 2. For the first scenario, the retailer would expect an overall profit margin gain of 15 %. Figure 6a shows that the segment-specific targeted category level program outperforms the standardized promotion in only four out of the eleven household segments under investigation (in particular the wine and the baby clusters). In other words, only for these segment-specific targeted promotions as derived by our proposed framework using LFCs are the recommended option because of a higher expected gain in profitability compared to the standardized promotion campaign. For segments k ¼ f2; 4; 6; 7; 9; 10; 11g Fig. 6a shows that the profit increase achieved by the standardized promotion will be twice as high as the projected profit lift for the segment-specific case. For these groups, the expected gain in profit by adopting a targeted marketing program are unlikely to compensate for the profit potential not realized by conventional standard promotion techniques. However, segment-specific targeting would still be profitable for 32.2 % of the targeted households, while about two thirds of all households would still need to be addressed with the standardized promotion.
The situation changes for setting no. 2, which recommends the segment-specific target marketing approach as the preferred one (cf. Fig. 6b). Under the conditions described for this scenario setting, our proposed segment-specific campaign clearly outperforms the standardized promotion and can expected to be much more profitable in every household segment. On aggregate, the additional profit lift generated with segment-specific target marketing exceeds the undifferentiated promotion by up to 128 %.

Discussion and future research
We present and empirically demonstrate the performance of a new approach to support loyalty program managers and direct marketers in customizing their segment-specific target marketing activities on a product category level. Our proposed decision-support framework requires customers' past purchase histories as input data, builds on state-of-art data mining techniques and integrates an optimization procedure which provides the decision maker with a list of candidate product categories for segment-level targeted promotion campaigns. This list is derived such that the included categories maximize the cross-category spillover effects for non-promoted categories.
There are many occasions in which marketing managers can benefit from such itemset recommendations for target marketing purposes. These include but are not limited to designing segment-specific rewards in loyalty programs, cross-merchandising activities, targeted direct mailings, customized supplements in catalogues, and customized promotions. For example, the latter can be delivered both offline directly in the store (e.g., by issuing customized check-out-coupons as provided by Catalina Marketing services) or in online environments by sending targeted emails or during shopping trips in online stores.
We demonstrate the application of the stepwise procedure using transaction data from a real-world loyalty program offered by an anonymous major grocery retailer. In the scope of our empirical application study we also explored the projected profitability implications of utilizing the derived recommendations for designing segment-specific, category level targeted promotions. A scenario-based simulation study suggests that the adoption of targeted promotions might boost profitability between 15 % and 128 % relative to an undifferentiated standard campaign and that at least for some segments targeting can be the preferred option even under very conservative assumptions on the effectiveness of segment-specific target marketing It also would be beneficial to combine our approach, which is primarily concerned with data compression, with a more predictive approach for modeling customers' multi-category choice decisions (such as the work by Manchanda et al. 1999;Dippold and Hruschka 2013;Hruschka 2013). Further promising extensions of our approach could also concentrate on making the segmentation approach dynamic in order to account for changes in customers' purchasing habits. Finally, to accommodate larger numbers of categories or for applications on a sub-category or even item-level the proposed approach needs to be made scalable for very highdimensional transaction data. Such an attempt requires some type of variable selection or variable weighting; the contributions by Carmone Jr. et al. (1999) and Brusco and Cradit (2001) might be promising candidates to deal with this kind of challenges.