SCHOOL OF COMPUTER SCIENCE
The School of Computer Science at the University of Windsor is pleased to present …
Ecommerce Recommendation by an Ensemble of Purchase Matrices with Sequential Patterns
MSc Thesis Proposal by: Mehdi Naseri
Date: Thursday, April 25th, 2019
Time: 11 am – 1 pm
Location: 3105, Lambton Tower
Abstract:
In E-commerce recommendation systems, combining collaborative filtering (CF) and sequential pattern mining of purchase history data will improve accuracy of recommendations and mitigate limitations of using only explicit user ratings for recommendations. Existing E-commerce recommendation systems which have combined CF with some form of sequences from purchase history are LiuRec09, ChioRec12 and HPCRec18. LiuRec09 receives historical purchase data, segments customers based on frequency, recency and monetary values. It predicts transaction clusters of customers in the next period, before using CF in each cluster to select items with highest rating. ChoiRec012 system derives implicit rating value from online transaction data which it uses to improve user-item rating matrix input to CF. HPCRec18 uses similarity between click streams and purchase patterns to fill up more ratings in the user-item matrix before running the CF algorithm. A limitation of these systems is that in collaborative filtering, when they use purchase frequency to fill up user-item matrix they assign the same weight (value) to items which are already purchased by users and none of these systems is using sequential patterns mined to discover already purchased items that the users might purchase again later.
This thesis proposes an E-commerce recommendation system, SEERs (Stacking Ensemble Ecommerce Recommendation system) which receives purchase history dataset as input and recommends items to users with higher accuracy compared to existing systems by integrating mined sequential patterns with previously recommended items. It first runs collaborative filtering to find unknown rating values in user-item-purchase frequency matrix which serves as input to second phase. SEERs also generate sequential patterns from sequential purchase dataset to extract all frequent sequential rules of users. It then creates a user-item rating matrix for rules where user has purchased a descendant item and another user-item rating matrix for rules where user has not purchased descendant items yet. To find the best weight for each one of these four user-item matrices, SEERs runs a test over training dataset with various weights (0 to 10 with step of 0.5) and select the combination which has the highest F1-score. Finally, it combines these four matrices by giving each one, the discovered weight to create final user-item rating matrix. For each user, items with highest ratings will be selected as the recommendation results.
Thesis Committee:
Internal Reader: Dr. Saeed Samet
External Reader: Dr. Zhiguo Hu
Advisor: Dr. Christie Ezeife
Thesis Proposal Announcement
5113 Lambton Tower, 401 Sunset Ave., Windsor ON, N9B 3P4 (519) 253-3000 Ext. 3716, csgradinfo@uwindsor.ca