Data Mining: The Textbook

Age[85, 95] ⇒ Checkers

Other demographic attributes, such as the gender or the ZIP code, can be used to determine more refined rules. Such rules are referred to as profile association rules. Profile association

5.4. PUTTING ASSOCIATIONS TO WORK: APPLICATIONS

149

rules are very useful for target marketing decisions because they can be used to identify relevant population segments for specific products. Profile association rules can be viewed in a similar way to classification rules, except that the antecedent of the rule typically identifies a profile segment, and the consequent identifies a population segment for target marketing.

5.4.4 Recommendations and Collaborative Filtering

Both the aforementioned applications are closely related to the generic problem of recom-mendation analysis and collaborative filtering. In collaborative filtering, the idea is to make recommendations to users on the basis of the buying behavior of other similar users. In this context, localized pattern mining is particularly useful. In localized pattern mining, the idea is to cluster the data into segments, and then determine the patterns in these segments. The patterns from each segment are typically more resistant to noise from the global data distribution and provide a clearer idea of the patterns within like-minded customers. For example, in a movie recommendation system, a particular pattern for movie titles, such as {Gladiator, Nero, Julius Caesar}, may not have sufficient support on a global basis. How-ever, within like-minded customers, who are interested in historical movies, such a pattern may have sufficient support. This approach is used in applications such as collaborative filtering. The problem of localized pattern mining is much more challenging because of the need to simultaneously determine the clustered segments and the association rules. The bib-liographic section contains pointers to such localized pattern mining methods. Collaborative filtering is discussed in detail in Sect. 18.5 of Chap. 18.

5.4.5 Web Log Analysis

Web log analysis is a common scenario for pattern mining methods. For example, the set of pages accessed during a session is no different than a market-basket data set containing transactions. When a set of Web pages is accessed together frequently, this provides useful insights about correlations in user behavior with respect to Web pages. Such insights can be leveraged by site-administrators to improve the structure of the Web site. For example, if a pair of Web pages are frequently accessed together in a session but are not currently linked together, it may be helpful to add a link between them. The most sophisticated forms of Web log analysis typically work with the temporal aspects of logs, beyond the set-wise framework of frequent itemset mining. These methods will be discussed in detail in Chaps. 15 and 18.

5.4.6 Bioinformatics

Many new technologies in bioinformatics, such as microarray and mass spectrometry tech-nologies, allow the collection of different kinds of very high-dimensional data sets. A classical example of this kind of data is gene-expression data, which can be expressed as an n × d matrix, where the number of columns d is very large compared with typical market basket applications. It is not uncommon for a microarray application to contain a hundred thou-sand columns. The discovery of frequent patterns in such data has numerous applications in the discovery of key biological properties that are encoded by these data sets. For such cases, long pattern mining methods, such as maximal and closed pattern mining are very useful. In fact, a number of methods, discussed in the bibliographic notes, have specifically been designed for such data sets.

150.