Mining Multiple 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 and user-requirements data on one data source to make recommendations about products or users and can be extended to multiple data sources (MDSs) with ratings computed from sequential patterns of user/item behaviors for comparative querying and higher recommendation accuracy. Fast growing online E-commerce retail industry is essential to Canada's economic growth. More people are choosing online purchases from E-stores like Walmart and Canadian Tire. Shoppers want better deals while retailers want to remain competitive, profitable and increase sales. Shoppers and retailers can benefit from comparative querying and recommendation of products using attributes such as prices, user opinions. Sequential patterns of product features or user behavior can be used to summarize classes of products and their ratings by users for the purposes of recommendations. Existing sequential pattern mining techniques find frequent patterns on single tables and are not able to answer complex pattern queries on multiple tables that may be related through foreign key attributes to enable analyzing data from MDSs. There is need to develop methods for mining groups of patterns (eg. profitable), integrating patterns from MDSs, cleaning and warehousing historical web data, recommendations and handling evolving data from sources. These methods can help users make good purchase choices and help retailers retain customers.******Thus, the goal of the proposed research program is to do comparative querying, mining and recommendations on MDSs by defining new algorithms to (i) mine relevant patterns for answering complex user queries, (ii) recommend and predict most relevant groups of items or users on MDSs. The research methods to be used include for sequential pattern mining transformation of MDSs into a uniform format with foreign key link, and defining methods similar to the GSP-join for mining higher level n-itemset patterns (eg., first transforming the database to have a record of all subsequences where each 1-item occurs in each transaction and having each frequent pattern carry its foreign key attribute link to a related table, and defining methods for mining higher level n-itemset patterns). Extensions of non-Apriori tree-like approaches can also be defined. For recommendation systems, we plan to model user and item data as sequential patterns of data (users, items, ratings) from MDSs using Collaborative filtering (CF) with summarized pattern user/item matrix, making system more extendable, understandable and answering more complex queries with historical, derived and comparative data. The proposed methods for improving the quality of MDSs recommendations in E-Commerce are novel, can benefit Canada's industries and people for comparing prices, opinions and historical data, the constructed data grouping methods can be used in other application domains.