Accurate recommender systems can enhance consumers' shopping experiences. In retail and many other business environments, extra contextual factors are usually available for building even more accurate recommender systems. The inuence of some factors is controversial in the industry. For instance, consumers' recent online exposure to products can decrease the chance of in-store purchase as consumers may choose to purchase products online. On the other hand, online exposure can be seen as an evidence of consumers' preference on products, which implies a higher chance of in-store purchase. The understanding of true in uence is important for product recommendation in-store in this case. The question is how to evaluate the relevance and the in uence of potential factors for prediction. Existing literature focuses on applying machine learning techniques to identify relevant contextual factors. While these methods are proven to be e ective in some experiments, an alternative approach that can provide easy-to-interpret analysis on relevance and in uence is preferred in many situations. The paper introduces a computationally inexpensive approach to conduct preliminary relevance and in uence analysis for contextual information in retail business. Statistical techniques from medical research eld are applied to analyze relationship between consumers' online exposure to retailer's e-commerce website, i.e., a contextual factor, and their ofine in-store purchase decisions, i.e., the outcome to be predicted, based on a retail dataset provided by a large UK retail business with both online and o ine presence. Unlike machine learning approaches, this analysis can be done even before a recommender system is built by using the proposed approach. This research further shows that the in uence of this contextual factor depends on extraneous attributes, such as consumers' ages and gender. This papers serves as a preliminary step to analyze relevant contextual factors for building context-aware recommender systems.
In this paper, we consider consumers' recent online exposure to brands at a retailer's website as a potential, but controversial, contextual factor. Consumers' online exposure could imply consumers' tendency to purchase products online, so they are less likely to purchase them in-store. Oppositely, it could be seen as a user context that represents consumers' recent preference on brands, which implies a higher chance of in-store purchase. The understanding of the in uence of this contextual factor is important for a number of recommendation scenarios. For simplicity, we focus on a speci c scenario: A large retailer that has both online e-commerce website and o ine stores presence wants to predict which product brands the customers are going to purchase when they enter o ine stores. For a brand that a targeted consumer has browsed online recently, a recommender system can possibly consider three cases of in uence: 1) If it is strong positively, the consumer will likely to purchase products of this brand in-store anyway, so no recommendation is needed; 2) If it is negative or if there is no in uence, the chance of purchase in-store is not high; 3) If it is medium positively, the system may try to nudge the consumer to purchase products of this brand in-store. The design of this recommender system is part of our future work. From the point of view of utilizing contextual information, this paper studies the relevance and in uence between consumers' online exposure to brands at the retailer's website, i.e., a contextual factor, and the probabilities of their instore purchasing decisions, i.e., the outcome to be predicted, for product brands, strati ed by di erent consumer groups. We propose the use of statistical techniques to analyze contextual factor. Unlike machine learning techniques implemented in the literature, this approach is independent from the prediction model of the system. Besides, unlike traditional correlation analysis techniques, such as sign test and chi-squared test, our approach can estimate the in uence of the factor on the probability of in-store purchase on products of brands. This information can possibly be used to improve the prediction model directly in our future work. A challenge of using statistical techniques, namely the issue that basic probabilistic measurement is sensitive to external noise, is presented by a numerical example in this paper. More robust techniques, such as odds ratio and strati ed analysis, are then proposed. The dataset for experiments used in this research is provided by a large UK retail business. Ten product brands are analyzed. It is a 1-year anonymized records of loyalty card holders who have browsed products of the selected brands online at the retailer's website and of those who have purchased products of the selected brands in any store of the retailer in the UK. All data is collected in a real non-experimental setting. In our experiments, whether the consumers have browsed any product page of a targeted brand at the retailer's website within a month is regarded as a binary contextual factor and the outcome to be predicted is consumers' in-store purchase decisions. Results show that the in uence of this contextual factor on the outcome to be predicted varies with consumers' attributes such as age and gender. The rest of the paper is organized as follows. Section 2 is a review of related literature about techniques that identify relevant contextual information in recommender systems. Section 3 describes the challenges of data analysis in a retail scenario. Solutions are then proposed to analyze contextual factors. Section 4 presents experiments conducted based on real retail dataset and section 5 is the discussion and future work.
Our literature review focuses on techniques that identify and
evaluate relevant contextual information in recommender
systems. The technical goal of a recommender system can
generally be seen as the problem of predicting ratings, or
any other outcome, for the items that have not been seen
by a user [
In this paper, we consider a retailer that operates both an online e-commerce website and physical retail stores. Consumers' recent online exposure at the retailer's website is the contextual factor to be evaluated. In particular, we de ne whether the consumers have browsed at least a page of a targeted brand at the retailer's website within a month as the contextual condition, browse = 1 if the condition exists, 0 otherwise. For illustration purposes, we assume that the outcome to be predicted is the purchase decision of any product of the targeted brand at any physical store (in-store purchase), which is a binary variable: purchase = 1 if purchased, 0 otherwise. In order to evaluate the relevance and in uence between this contextual factor and the outcome, we need to compare the probability of in-store purchase of consumers who have browsed and of those who have not, i.e. p(purchase = 1jbrowse = 1) and p(purchase = 1jbrowse = 0). In a population of N potential consumers, we can construct a table to represent the online browsing and in-store purchasing situation of the dataset: browse=1 browse=0 where a11,a10,a01 and a00 are the number of consumers for the corresponding purchasing and browsing situations. a 1 = a11 + a01 is the number of consumers who have purchased in-store, a 0 = a10 + a00 is the number of consumers who have not purchased in-store, a1 = a11 + a10 is the number of consumers who have browsed online and a0 = a01 + a00 is the number of consumers who have not browsed online. A direct way to express the relationship is to compare the two probabilities with relative correlation (RC), where RC = p(purchase = 1jbrowse = 1) p(purchase = 1jbrowse = 0) = a11=a1 a01=a0 (1) There is no correlation if RC = 1, the in uence is positive if RC < 1 and negative if RC > 1. This approach, however, su ers from two problems when the data is collected from a non-experimental retail environment. First, RC is sensitive to the total number of consumers who have purchased and also to the total number of consumers who have not purchased. These two numbers, unfortunately, can be a ected by external irrelevant factors, such as marketing campaigns or product promotions, which should be isolated from this analysis. This problem can be illustrated with a numerical example. Suppose the data looks like the following table when there is no sales promotion: browse=1 browse=0 purchase=1 5 60 65 purchase=0
500 25,000 25,500 Let us assume that a sales promotion successfully attracts new consumers to purchase and the number of purchase increases 10 times as shown in the following table: browse=1 browse=0 purchase=1 50 600 650 purchase=0
500 25,000 25,500 When all things being equal, a temporary sales promotion should not a ect the relationship between the contextual 5=505 factor and the outcome. In reality, however, RC = 60=25060 = 50=550 4:14 in the rst case while RC = 600=25600 = 3:88 in the second one. In another words, RC is sensitive to the change of number of consumers who purchase (a 1). This problem presence in many real-world environments since businesses can always attract new consumers to stores or website dynamically, which a ects N , and thus a 1 and a 1 can be manipulated. An odds ratio technique to estimate RC that is insensitive to the change of N is proposed later in this paper. The second problem is the existence of extraneous attributes, such as age and gender, that potentially a ect the in uence of the targeted contextual factor on the outcome to be predicted. This problem occurs when an attribute is associated with the contextual factor and at the same time such attribute a ects the outcome dependently or independently. This kind of extraneous attribute is called a confounder in the statistics discipline. Strati ed analysis is proposed to evaluate the impact of possible confounding attributes.
Odds ratio is commonly used as a estimator of RC in medical
and epidemiological research for case-control studies where
disease cases are not easy to be obtained [
Extraneous attributes, such as consumers' age and gender,
potentially a ect the in uence of the contextual factor on
the outcome. Strati ed analysis is a computationally
inexpensive solution to reveal their e ects. This technique is
commonly used in medical research when setting up
control group experiments is not feasible and so the existence
of extraneous factors is common [
ORMH =
Pk
i=1 Pk i=1 a11ia00i
Ni a01ia10i
Ni
(4)
where k is the total number of strata in an analysis and i
represents one of them. For this Mantel-Haenszel method of
estimation to be accurate, the overall sample size must be
large. [
SE(lnORMH ) = p(
Pik=1( a10ia01i )2vi
Ni Pk i=1( a10ia01i )2
Ni ) vi =
Our dataset, which is provided by a large UK retail business, is a 1-year anonymized records of loyalty card holders who have browsed the selected products online on the retailer's website and of those who have purchased the selected products in any store of the retailer in the UK. It contains 10,217,972 unique loyalty card holders and 2,939 unique products under 10 selected brands. There are 21,668,137 instore purchase transaction records and 299,070 online browsing records. We associate consumers' online browsing and in-store purchasing behaviors with unique loyalty card numbers. All data is collected in a real non-experimental setting.
This experiment investigates the relevance and in uence between consumers' recent online browsing behaviors and the probabilities of their in-store purchase decisions for ten product brands carried by a large UK retail business nationally. These ten brands are selected randomly, some of them are luxury brands while the others are mid-range brands. We de ne whether a consumer has browsed at least a page of a targeted brand at the retailer's website within a month as the context of the consumer, browse = 1 if the condition exists, 0 otherwise. Odds ratio is used to compare the in uence of this contextual factor on the probabilities of consumers' binary purchase decision of any product of the targeted brand at any physical store (in-store purchase). We pre-process the dataset to lter out consumers who have not visited any page at the retailer's website at least once in the past year. This process ensures that the remaining N consumers have at least successfully accessed the retailer's website recently. We start with a hypothesis that age and gender are two attributes of consumers that may confound the in uence. We conduct monthly strata-speci c measurement of odds ratio based on these two attributes for each brand. Practically, age and gender information is missing in some records. In each analysis, therefore, we analyze a population size of Nage or Ngender which represent the total number of consumers with age information or with gender information respectively. In these experiments, we calculate the monthly crude (unadjusted) odds ratio for each strata for each brand. If the odds ratio for a strata of a brand is X, it means that, in this strata and in this particular month, the probability to purchase at least one product of this brand in-store by consumers who have browsed at least a webpage of this brand online is X times higher than the probability for those who have not browsed so. We also calculate the monthly common strata-adjusted odds ratio as well as the 95% con dence interval (CI).
Results of only three strati ed odds ratios analysis are presented in this paper due to length constraint. All gures show that odds ratio measurements are well above 1, i.e., the in uence is positive for all brands. It means that the probability to purchase at least one product of a selected brand in-store by consumers who have browsed at least a webpage of that brand online at the retailer's website is higher than the probability for those who have not browsed so. The values and patterns are di erent for each brand though, which means that the impact of this contextual factor of online exposure varies with brands.
Figure 1 represents gender-strati ed analysis of brand A. The in uence on female consumers is much stronger than the one on male consumers. An interesting discovery is that the odds ratio measurements for both genders follow a very similar up and down monthly pattern. Both strata have peaked odds ratio in February. This nding hints that time is a contextual factor that should also be considered in future work. Figure 2 shows that the odds ratio range of di erent age groups for brand B are separated clearly. The in uence for consumers of age 18-25 is the highest while the one for consumers of age 26-35 is the lowest. It means that the probability for consumers of age 18-25 is higher than the one for consumers of age 36-45 and both of them are higher than the one for consumers of age 26-35. This nding implies that, for brand B, the age attribute itself can be correlated to consumers' in-store purchase decisions. Figure 3, on the other hand, draws a di erent conclusion for brand C. In this case, the odds ratio measurements of these age groups mixed together in a close range. There is no clear monthly pattern either. It means that age, gender and month are not confounding factors for this brand.
This paper derives a contextual factor from consumers' recent online browsing behaviors on the retailers' website for the prediction of their o ine in-store purchase. A statistical approach is presented to conduct a preliminary analysis on the relevance and in uence between this factor and the o ine purchase decisions on brands using odds ratio and strati ed analysis techniques. The initial uncertainty that consumers who browse online on retailer's website tend to purchase online and therefore they have lower chance to purchase in-store has been proven untrue for the brands we have analyzed. In addition, as expected, the in uence of online exposure on o ine purchases varies with brands and consumers' ages and gender. In our future work, the analysis for non-binary contextual factor will be illustrated. Besides the factor we have evaluated in this paper, it is interesting to see whether other relevant contextual factors can be derived from consumers' recent online behaviors for their in-store purchase decisions. Future work is to build a context-aware recommender system for in-store product recommendation based on these ndings. We are interested in using the OR value directly to improve prediction. Also, a comparison of predictive performance of recommender systems using contextual factors selected by this approach and by existing machine learning techniques is part of our future work.