Mining Web Data Sources for Integrated Informative Querying and Recommendation

点击翻译按钮获取中文摘要

Recommendation systems  make use of user preferences, user and item data from web sources, to make recommendations about products or users. Collaborative filtering is a method of recommendation systems that uses user ratings for products in a user-item rating matrix to recommend additional products. Fast growing online E-commerce retail industry is essential to Canada's economic growth as more people choose online purchases  for convenience and better deals. Retailers want to remain competitive, profitable and increase sales. Recommendation systems help users cope with too many product choices by recommending needed, new and diverse items.  The utility (rating) of an item i to a user u, expressed as R(u, i)  can be predicted from transaction data. Sequential pattern mining (SPM) can learn a model of customer purchase behaviour as sequential patterns which can be converted to  utility function R(u, i) for a recommendation system collaborative filtering algorithm for more effective, diverse and accurate results. Existing recommendation systems do not learn sequential patterns of customer purchase behaviour from historical  or click stream data. Thus, the goal of the proposed research program is to use sequential pattern mining or in conjunction with other mining methods on recommendation system input data sources (eg. historical purchase and click-stream data) to discover richer customer interests, such as sequential patterns of purchases so as to improve  recommendation system accuracy and diversity. Algorithms will be developed for discovering, integrating and using mined sequential patterns of purchases to improve i) the quality of user ratings of products, ii) the quantity of ratings previously largely sparse, iii) to extend  sequential pattern mining techniques for single table to handle integrative querying and mining of multiple data sources  related through foreign key attributes. Methods to be used include pre-processing relevant E-Commerce data sources with  SPM algorithms (eg. GSP) to discover sequential (eg. frequent, rare) of purchases. Sequential patterns are then used as the user purchase item vectors to derive a single real rating value from the groups of items. To integrate mining results, we will define both Apriori-like (eg. GSP join techniques) and non-Apriori type techniques (eg. tree based) for finding sequential patterns in related multiple data sources by first transforming the original tables into tuple patterns that contain sub-sequences with foreign key attributes where they occurred, so as to link record occurrences of sequential patterns. The research is novel as no existing work explored this approach. It can improve the quality and quantity of item utility rating function values for more accurate and diverse results, lead to higher sales, user convenience and loyalty. The foreign key linked patterns is novel, contributes to fundamental SPM solution, improved system performance, integrating mined patterns.