Association Mining in Gradually Changing Domains

Association mining explores algorithms capable of detecting frequently co-occurring items in transactions. A transaction can be identified with a market basket—a list of items a customer pays for at the checkout desk. In this paper, we explore a framework for the detection of changes in the buying patterns, as affected by fashion, season, or the introduction of a new product. We present several versions of our algorithm and experimentally examine their behaviors in domains with gradually changing domains. Introduction Let a database consist of transactions T1, T2, . . . , TN such that ∀i, Ti ⊆ I , where I is a set of items. Let an itemset, X , be defined as a group of items such that ∃i,X ⊆ Ti. A support of itemset X is the number of transactions, Ti, containing this itemset. Suppose a user of an associationmining algorithm submits a minimum support, θ. A popular research issue is how to detect all itemsets whose support is at least equal to θ. We will call them high-support itemsets. In a classical application, the transactions are identified with lists of items a customer pays for at the checkout desk (market baskets). A supermarket then places associated items on neighboring shelves, advertises them in the same catalogues, and avoids discounts on more than one member of the same itemset. However, the paradigm extends well beyond the realm of supermarket data. In the Internet environment, a transaction may consist of hyperlinks pointing to a web page, and high-support itemsets then signal associations among web sites (Noel, Raghavan, & Chu 2001). In a medical application, each transaction can summarize a patient’s history, and association mining may seek co-occurring symptoms or ailments. Here, we are interested in domains where the list of highsupport itemsets is subject to changes in time, for instance as a result of fashion or seasonal impacts (Cheung & Han 1996; Pitkow 1997; Raghavan & Hafez 2000). We will assume the framework of block evolution (Ganti, Gehrke, & Ramakrishnan 2000) where a block of market baskets is periodically added to an existing database as individual stores report their daily business. The task is to adapt to this change Copyright c © 2003, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. and to measure the accuracy of this adaptation. To this end, we recently developed a novel algorithm (Rozsypal & Kubat 2002) and investigated its behavior in domains where the change is abrupt. However, this may be too much of a simplification. More often than not, the “Winter” buying patterns only gradually replace the “Fall” patterns. This is the scenario we address. Somewhat surprisingly, it turns out that, under this circumstance, different operators for change detection and for knowledge update are useful. We report a series of experiments illustrating the point.