Discovering exceptional information from customer inquiry by association rule miner

Abstract

This paper reports the results of our experimental study on a new method of applying an association nile miner to discover useful information from a text database. It has been claimed that association rule mining is not suited for text mining. To overcome this problem, we propose (1) to generate a sequential data set of words with dependency structure from a Japanese text database, and (2) to employ a new method for extracting meaningful association rules by applying a new rule selection criterion. Each inquiry was converted to a list of word pairs, having dependency relationship in the original sentence. The association rules were acquired regarding each pair of words as an item. The rule selection criterion derived from our principle of giving heavier weights to co-occurrence of multiple items than to single item occurrence. We regarded a rule as important if the existence of the items in the rule body significantly affected the occurrence of the item in the rule head. Based on this method, we conducted experiments on a customer inquiry database in a call center of a company and successfully acquired practical meaningful rules, which were not too general nor appeared only rarely. Also, they were not acquired by only simple keyword retrieval. Additionally, inquiries with multiple aspects were properly classified into corresponding multiple categories. Furthermore, we compared (i) rules obtained from a sequential data set of words with dependency structure, which we propose in this paper, and those without dependency structure, as well as (ii) rules acquired through the association rule selection criterion and those through the conventional criteria. As a result, discovery of meaningful rules increased 14.3-fold in the first comparison, and we confirmed that our criterion enables to obtain rules according to the objectives more precisely in the second comparison. © Springer-Verlag Berlin Heidelberg 2003.