Ontology-driven Rule Generalization and Categorization for Market Data  by Dongwoo Won

Radio Frequency Identification (RFID) is a wireless technology that nowadays getting large attention from many organizations. RFID can identify objects using radio frequency by storing customized information into RFID tags which consist of antenna that detects radio waves and responds with signals, and a chip that stores and manipulates data. Then there exists a RFID reader that recognizes the stored information. Different from bar code that requires a contact with the reader, RFID, on the other hand, do not need line of sight identification. Therefore, RFID is fast in reading, saves labor cost, and enables multiple reading. It is also possible to modify the stored information and track the location of the tags. Using this information, people can track the movement of the customer, length and the place of staying in particular section, and the type of product that the customer buys. As a result, we can analyze this useful information either to place the goods to prevent customer from assembling in a particular section or to place the goods that are bought together in nearby section to arouse customers’ interest. However, these kinds of tracking processes produce tons of data, so we need an efficient technique to mine the data.
In the past, association rule mining or clustering technique has been used with limitations to provide adequate solution. We present a process for mining large problem space of market data into a hierarchically structured search space that is efficient for analysis. We use association rule mining for three types of supermarket data analysis that we have defined as Section-To-Section, In-Section, and In-Section-To-In-Section analysis. Section-To-Section Analysis is to see the relationship among the section. We do not need the full item list that belongs to that section, but only the generalized concept of that item from our domain ontologies. In-Section Analysis is to see the relationship among the items within one particular section. All the lowest level of children nodes within one general concept are used for this analysis. In-Section-To-In-Section Analysis is to see the relationship among the items in different sections. We need to use the current item list as it is. The second type In-Section Analysis is a kind of In-Section-To-In-Section Analysis. The number of rule reduction is expected for Section-To-Section Analysis.
Based on the three types of analysis, rule generalization uses domain ontologies to merge and simplify items into more general concepts. The usage of ontologies allows us to have pre-knowledge about the data. Ontologies also provide a way to represent information or knowledge that includes the key concepts and the inter-relationships between them. As a result, it produces fewer, but more closely associated rules. Our result shows that this step reduces the total number of rules being generated.

Figure 1. Rule generalization for supermarket data.

Rule Categorization hierarchically groups the rules by relevance into new clusters, called sub-categories, in which reduces the number of rules to be looked at or searched for analysis. With the generalized association rules, we find sub-categories consisting of rules that are more relevant to the generalized association rule by hierarchically clustering association rules by their relevance. The good thing about rule categorization is that once we cluster the original rules under the generalized rule like R1 or R2, we do not need to scan all the association rules every time we try to analyze the data set. Instead, we select a generalized rule and work with the sub-categories and rules that belong to that generalized rule. Also, since the rules are hierarchically clustered by relevance, we can choose the level of detail for working with the rules. To search only within a category instead of the whole association rule list is an enormous plus for efficient analysis.

Figure 2. Examples for three supermarket analysis types using rule categorization