Action Rules Mining

Introduction

There are two aspects of interestingness of rules that have been studied in data mining literature, objective and subjective measures (Liu et al., 1997), (Adomavicius & Tuzhilin, 1997), (Silberschatz & Tuzhilin, 1995, 1996). Objective measures are data-driven and domain-independent. Generally, they evaluate the rules based on their quality and similarity between them. Subjective measures, including unexpectedness, novelty and actionability, are user-driven and domain-dependent.

A rule is actionable if user can do an action to his/her advantage based on this rule (Liu et al., 1997). This definition, in spite of its importance, is too vague and it leaves open door to a number of different interpretations of actionability. In order to narrow it down, a new class of rules (called action rules) constructed from certain pairs of association rules, has been proposed in (Ras & Wieczorkowska, 2000). Interventions introduced in (Greco et al., 2006) and the concept of information changes proposed in (Skowron & Synak, 2006) are conceptually very similar to action rules. Action rules have been investigated further in (Wang at al., 2002), (Tsay & Ras, 2005, 2006), (Tzacheva & Ras, 2005), (He at al., 2005), (Ras & Dardzinska, 2006), (Dardzinska & Ras, 2006), (Ras & Wyrzykowska, 2007). To give an example justifying the need of action rules, let us assume that a number of customers have closed their accounts at one of the banks. We construct, possibly the simplest, description of that group of people and next search for a new description, similar to the one we have, with a goal to identify a new group of customers from which no-one left that bank. If these descriptions have a form of rules, then they can be seen as actionable rules. Now, by comparing these two descriptions, we may find the cause why these accounts have been closed and formulate an action which if undertaken by the bank, may prevent other customers from closing their accounts. Such actions are stimulated by action rules and they are seen as precise hints for actionability of rules. For example, an action rule may say that by inviting people from a certain group of customers for a glass of wine by a bank, it is guaranteed that these customers will not close their accounts and they do not move to another bank. Sending invitations by regular mail to all these customers or inviting them personally by giving them a call are examples of an action associated with that action rule.

In (Tzacheva & Ras, 2005) the notion of a cost and feasibility of an action rule was introduced. The cost is a subjective measure and feasibility is an objective measure. Usually, a number of action rules or chains of action rules can be applied to re-classify a certain set of objects. The cost associated with changes of values within one attribute is usually different than the cost associated with changes of values within another attribute. The strategy for replacing the initially extracted action rule by a composition of new action rules, dynamically built and leading to the same reclassification goal, was proposed in (Tzacheva & Ras, 2005). This composition of rules uniquely defines a new action rule. Objects supporting the new action rule also support the initial action rule but the cost of reclassifying them is lower or even much lower for the new rule. In (Ras & Dardzinska, 2006) authors present a new algebraic-type top-down strategy for constructing action rules from single classification rules. Algorithm ARAS, proposed in (Ras & Wyrzykowska, 2007), is a bottom-up strategy generating action rules. In (He at al., 2005) authors give a strategy for discovering action rules directly from a database.