Optimized incremental mining of customer buying behavior using temporal association rules

Abstract

In retail business, customers’ behavior analytics is a study of customers’ buying behavior for a better understanding of customer needs to be able to provide service accordingly. The buying behavior is majorly influenced by the preferences of a customer. However, preferences of a customer change over a period of time due to various factors like change in income, taste, culture or newer products, etc. Understanding these changes in customer behavior is a very challenging task especially in a dynamic, ever-changing environment. There are various customer behavior mining models and techniques available in the data mining domain that are designed to work on static and dynamic databases. The traditional incremental mining techniques consider all the previous datasets in order to update the patterns. However, in a dynamic database, the size of the database grows with every update. To mine customers’ behavior in a time-variant database, the re-mining of the updated database is required that further increases processing cost in terms of execution time and memory space with every update. The purpose of this paper is to propose a method that can analyze the changes in customers’ behavior in time-variant databases without mining all the transactions. In this paper, an optimized incremental technique is proposed that utilizes temporal association rule mining in a time-variant database for mining customer behavioral patterns in an updated database. The proposed algorithm named ‘Autoregressive Moving Average model-based Incremental Temporal Association Rules Mining (ARMA-ITARM)’ utilizes the ARMA model to substantially reduce the database and maintains temporal frequent patterns in the updated database. Inspired by sliding window and pre-large concepts, the algorithm utilizes past frequent itemsets and probable frequent itemsets from customers’ purchased history along with frequent itemsets and probable frequent itemsets that reduce search space. Consequently, the entire database is scanned only once to count the frequency of occurrence of a few candidate itemsets. In effect, execution time memory need of the algorithm is very small. Experimental results demonstrate that our proposed technique performs better over recent techniques like ITARM, SWF, etc.