نگاه اکتشافی در مسیر خرید سوپر مارکت

We present analyses of an extraordinary new dataset that reveals the path taken by individual shoppers in an actual grocery store, as provided by RFID (radio frequency identification) tags located on their shopping carts. The analysis is performed using a multivariate clustering algorithm not yet seen in the marketing literature that is able to handle data sets with unique (and numerous) spatial constraints. This allows us to take into account physical impediments (such as the location of aisles and other inaccessible areas of the store) to ensure that we only report feasible centroid paths. We also recognize that time spent in the store plays an important role, leading to different cluster configurations for short, medium, and long trips. The resulting three sets of clusters identify a total of 14 “canonical path types” that are typical of grocery store travel, and we carefully describe (and cross-validate) each set of clusters. These results dispel certain myths about shopper travel behavior that common intuition perpetuates, including behavior related to aisles, end-cap displays, and the “racetrack.” We briefly relate these results to previous research (using much more limited datasets) covering travel behavior in retail stores and other related settings.

Most marketers have a well-established schema for shopper travel behavior within a supermarket—the typical customer is assumed to travel up and down the aisles of the store, stopping at various category locations, deliberating about her consideration set, choosing the best (utility maximizing) option, and then continuing in a similar manner until the path is complete. Despite the common presumption of this scenario, little research has been undertaken to understand actual travel patterns within a supermarket. How do shoppers really travel through the store? Do they go through every aisle, or do they skip from one area to another in a more direct manner? Do they spend much of their time moving around the outer ring of the store (a.k.a. the “racetrack”), or do they spend most of their time in certain store sections? Do most shoppers follow a single, dominant pattern, or are they rather heterogeneous? A rich new data source, as illustrated in Fig. 1, now allows us to examine these and other important behavioral questions.No, Fig. 1 does not represent the random scribblings of a kindergartener. It is a subset of the PathTracker® data collected by Sorensen Associates, an in-store research firm, for the purpose of understanding shopper behavior in the supermarket. Specifically, Sorensen Associates affixed RFID (radio frequency identification) tags to the bottom of every grocery cart in an actual supermarket in the western U.S. These tags emit a signal every 5 seconds that is received by receptors installed at various locations throughout the store. The arrival latencies of the signals at the receptor locations are used to triangulate the position of the grocery cart. Thus, for every shopping path, data are recorded regarding the cart's two-dimensional location coordinates, (xit, yit) for shopper i at his tth observation (hereafter referred to as “blinks”), at 5-second intervals, which can be used to determine each cart's route through the entire store1. While ideally, one might hope to obtain positioning data directly from the shoppers themselves, this is not currently available in an actual commercial setting. Therefore, we use customers' grocery carts as a proxy for their shopping path, since we know the exact shopper location when the grocery cart is moving and a good guess of the general vicinity of the shopper when the grocery cart is stationary. Regardless, the methodology developed in this paper will continue to be applicable as newer and better datasets become available. Finally, the time and location of the cart at the end of each path offers information about the checkout process; point-of-sale data can then be matched with the cart movement records to provide a complete picture of each shopping path. See Sorensen (2003) for more details about the PathTracker® system. The goal of this research is to undertake exploratory analyses, useful for data summarization, inference, and intuition about shopper travel path data. Specifically, we want to identify typical in-store supermarket travel behaviors that will help us understand how shoppers move through a supermarket. Similar research ideas, summarizing large sets of “behavioral” curves as in Fig. 1 have been explored using principal components analysis methods (Bradlow, 2002 and Jones and Rice, 1992); however, our goal here is not to explain the maximal variation across customers with principal curves, but instead to cluster respondents into “types” of shoppers and describe the prototypical path of a general cluster. Unfortunately, there are numerous challenges we face, since the application of standard clustering routines is not feasible due to the extremely large number of spatial constraints imposed by the physical supermarket layout (e.g. people can't walk through store shelves). For this reason, the contribution of this research is not limited to the empirical findings of the in-store path data, but also introduces to the marketing literature a multivariate clustering algorithm that can be applied to other settings with a large number of spatial constraints. Although this new method represents a useful step forward in our ability to analyze multivariate data, we wish to emphasize our exploratory objectives: we want to use this procedure to help us identify predominant patterns that will catalyze future research. Given the newness of this area, we are not yet at the stage of being able to create (or test) formal theories of shopping behavior. In other words, in this paper we will raise more questions than provide answers, and we hope to motivate readers to pursue these research issues with complementary (and more conclusive) research methods. The remainder of the paper is laid out as follows. First, we describe the data in more detail and explain various obstacles in undertaking exploratory analyses on this data (such as the numerous spatial constraints). Next, we detail the new-to-marketing clustering algorithm used to overcome these obstacles. We then present the results of the algorithm and the canonical shopper path profiles that emerge. The results are then displayed in relation to a set of variables that describe the travel areas of each path. We demonstrate that our methods enable us to cluster shopper paths along important dimensions that would be missed using simpler methods, lending support to the value of our techniques. We next perform a cross-validation of our results to assess the reliability of the findings. Finally, we conclude with a discussion section which summarizes the potential impact of the current findings and relates the current work to past and future research.

Our main purpose in presenting these exploratory analyses was to familiarize other researchers with the existence of such data and stimulate ideas for future use. As such, we leave to future researchers the work of detailing the many potential managerial implications of work using this type of data. However, even the exploratory work we have presented here carries useful and actionable information for store managers. A simple examination of the canonical paths resulting from the k-medoids clustering helps dispel a number of myths that our personal schemas about supermarket travel perpetuate. Of particular note is the extremely low occurrence of the pattern commonly thought to dominate store travel—weaving up and down all aisles. We note that most shoppers tend only to travel select aisles, and rarely in the systematic up and down pattern most tend to consider the dominant travel pattern. Those trips that do display extensive aisle travel tend to travel by short excursions into and out of the aisle rather than traversing the entire length of it. This simple observation has important implications for the placement of key products, the use of end-cap displays, etc. Products placed at the center of aisles will receive much less “face time” than those placed toward the ends. Of related interest is a practitioner study that found that placing familiar brands at the end of the aisles served as a “welcome mat” to those aisles, effectively increasing its traffic (Sorensen, 2005). Granted, the previous observations are specific to this particular store, and cannot without caution be applied to all grocery stores, but this template for identifying true store utilization patterns can be equally useful for any store layout. Informed decisions about store layout can only be made through direct observation of the current utilization of the store. Once we observe that the aisles are utilized much less than common folklore leads us to believe, we turn our attention to the areas of the store that pick up this slack. Whereas previous folklore perpetuated the myth that the perimeter of the store was visited incidental to successive aisle traverses, we now know that it often serves as the main thoroughfare, effectively a “home base” from which shoppers take quick trips into the aisles. The relationship between the perimeter travel and aisle travel has sparked substantial practitioner interest. The data and techniques described in this paper form effective first steps at understanding this complex relationship. Shorter trips tend to stick predominantly to the perimeter and convenience store areas. This simple observation provides an important starting point for the targeting of particular shopper segments. While the dataset featured in this paper is quite novel, we acknowledge that other researchers have addressed the general topic of in-store shopping patterns in the past. Every ten years or so, researchers seem to “rediscover” this topic, and have applied very different methods to capture it. One of the earliest such studies of shoppers was a paper by Farley and Ring (1966) who built a stochastic model to study zone-to-zone transitions within a store. Unfortunately, few researchers, to our knowledge, extended or applied their model. Coming from a psychological perspective, Mackay and Olshavsky (1975) examined consumer perceptions of store space, and Park, Iyer, and Smith (1989) sought to understand the impact that store knowledge and time constraints have on unplanned buying, failure to make planned purchases, and other purchase behaviors. Perhaps the most famous study, or series of studies, on in-store shopping behavior is Why We Buy(1999) by Paco Underhill. He uses anthropological methods to uncover a variety of behavioral patterns observed while tracking shoppers in different types of retail stores, but limits the depth of his research findings to basic suggestions about ways to enhance consumer convenience. Of all the facets of shopper behavior explored in previous research, none has focused on the complete shopper path as we have, thus making our research a useful step forward. A natural avenue of investigation would be an effort to tie the results and methods discussed in these earlier psychological and anthropological studies to the broader behaviors illustrated in the present study. Another stream of related research deals not with the grocery store but with spatial movements in general. Some of the most directly related research in this line has examined individual pedestrian movements in museums and shopping malls (Batty, 2003). The chapter presents a number of useful models to describe individual movements that will prove useful to further research on shopper movements. The emphasis in that work is on pedestrian flow, rather than profit from a store. Earlier work in environmental psychology also studied pedestrian traffic flow inside an architectural space, with the hopes of improving architectural design (Winkel & Sasanoff, 1966). Though this work has obvious connections to the present work, its pure focus on traffic flows makes its application to grocery stores not entirely straightforward. Other work in environmental psychology has an entirely different potential application to the current field of study. Anthropological studies about people's impressions of their surrounding neighborhood has postulated that people look for order on their environment out of an inherent need to organize it in their minds (Lynch & Rivkin, 1959). The way in which shoppers organize a store in their minds may have important implications for their subsequent movements. The current work provides a springboard from which this can be studied. The exploratory analyses we have presented on this new realm of shopper behavior research are only a first step in understanding shopping behavior within the store. The present research focuses only on travel patterns without regard to purchase behavior or merchandising tactics. A study of the linkage between travel and purchase behavior seems a logical next step. Linking specific travel patterns to individual purchase decisions may lead to an improved understanding of consumer motivations for purchasing certain items, and can shed light on the complementarity and substitutability of goods in ways that a more traditional “market basket” analysis cannot capture. Further exploration of travel behavior, independent of purchase, also seems another promising route for future research. In this paper, we have presented some exploratory techniques useful for knowledge building and intuition. A more formal model of travel behavior would lead to an increased understanding of shopper heterogeneity of travel and the underlying sources of said heterogeneity. Specifically, one could model travel as a series of “blink-to-blink” choices (with a careful focus on state dependence, since choices made earlier in the path probably have a great deal of influence on later choices). This would allow a more precise study of the key areas of the store—and perhaps merchandising activities—that may influence travel in a particular direction. Before plunging deeply into such a complex model, we felt it was important to first understand this rich new dataset and the behavioral/computational issues it points to. We hope that this exploratory analysis serves as a useful catalyst for future research that will help us better understand the actual shopping patterns – as opposed to the widely accepted folklore – that take place in different types of retail environments.