Date:  Wednesday, September 23rd, 2020 

 

Time:  2:00 pm – 4:00 pm 

 

Zoom:    https://zoom.us/j/98375081462? 

 

                Recommendation systems play an extremely important role in e-commerce systems. By recommending products that suit the taste of consumers, e-commerce companies (such as Amazon, Netflix, etc.) can generate large profits as well as increase customer loyalty. The most used recommendation systems typically produce a list of recommendations for customers through collaborative or content-based filtering techniques. The main input data for collaborative filtering recommendation system is user-item rating matrix showing the rating for the products/items by each user/customer. This user-item rating matrix is mostly sparse as in real-life scenarios, every customer would have bought only a very small percentage of hundreds of thousands of products in the e-commerce shelf. Thus, the information gathered from customer purchase history, clickstream behavior or patterns mined from them can be used to improve the accuracy of the recommendation system by first enriching the input data of user-item rating matrix. 

 

Existing recommendation systems such as HPCRec18 (Historical Purchase with Clickstream based Recommendation System) and HSPRec19 (Historical Sequential Pattern Recommendation Systems) improved recommendation accuracy by first enriching the user-item matrix both quantitatively (reducing the user-item matrix sparsity problem) and qualitatively (specifying the level of rating for an item as a real number between 0 and 1 rather than binary). These existing systems use customer purchase behavior such as sequential patterns mined from historical customer purchase and click stream data to improve the user-item rating matrix input to the collaborative filtering process.  However, these systems mine frequent sequential patterns and so they do not capture whether the recommended items are also of high utility to the user seller (e.g., are more profitable to the seller). 

 

                This thesis proposes a system called High Utility Sequential Pattern Recommendation System (HUSRec System), which generates a high utility sequential E-commerce database from ACM RecSys Challenge dataset using the HUSDBG (High Utility Sequential Database Generator) and HUSPM (High Utility sequential Pattern Miner) mines the high utility sequential pattern rules which can yield high sales profits for the seller based on quantity and price of items on the daily basis having minimum confidence and minimum sequence utility, this improves the quality and accuracy of the recommendations. HUSRec also mines clickstream sequential data using PrefixSpan algorithm to give frequent sequential rules to recommend items for the clicks where no purchase has happened, decreasing the sparsity of user-item matrix input to the Collaborative Filtering (CF). Therefore, HUSRec improves the user-item matrix for input to the collaborative filtering. 

 

Keywords: Sequential pattern mining, collaborative filtering, historical recommendation system. 

 

 

 

Internal Reader: Dr. Ahmad Biniaz 

 

External Reader: Dr. Eugene Kim 

 

Advisor: Dr. Christie Ezeife