The University of Windsor has moved to an “essential service only” model. Learn More.

MSc Thesis Defense Announcement by Mehdi Naseri: "E-commerce Recommendation by an Ensemble of Purchase Matrices with Sequential Patterns"

Tuesday, August 27, 2019 - 11:00 to 13:00



The School of Computer Science at the University of Windsor is pleased to present...

MSc Thesis Defense by:
Mehdi Naseri
Date:  Tuesday, Aug. 27, 2019
Time:  11:00 AM – 1:00 PM
Location: 3105, Lambton Tower


In E-commerce recommendation systems, integrating collaborative filtering (CF) and sequential pattern mining (SPM) 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 those referred to as LiuRec09, ChioRec12 and HPCRec18. ChoiRec12 system, HOPE first derives implicit ratings from purchase frequency of users in transaction data which it uses to create user item rating matrix input to CF. Then, it computes the CFPP, the CF-based predicted preference of each target user (u) on an item (i) as its output from the CF process. Similarly, it derives sequential patterns from the historical purchase database from which it obtains the second output matrix of SPAPP, sequential pattern analysis predicted preference of each user for each item. The final predicted preference of each user for each item FPP is obtained by integrating these two matrices by giving 90% to SPAPP and 10% to CFPP so it can recommend items with highest ratings to users. A limitation of HOPE  system is that in user item matrix of CF, it does not distinguish between
purchase frequency and ratings used for CF. Also in SPM, it recommends items, regardless of whether user has purchased that item before or not. 
This thesis proposes an E-commerce recommendation system, SEERs (Stacking Ensemble E-commerce Recommendation system) which improves on HOPE system to make better recommendations in the following two ways:
i) Learning the best minimum support for SPA, best k similar users for CF and the best weights for integrating the four used matrices.
ii)  Separating their two intermediate matrices of CFPP and SPAPP into four intermediate matrices of CF_notpurchased, SPM_purchased, SPM_notpurchased and purchase history matrix which are obtained and merged with the better-learned parameters from (i) above. Experimental results show that by using best weights discovered in training phase, and also separating purchased and not purchased items in CF and sequential pattern mining methods, SEERS provides better precision, recall, F1 score, and accuracy compared to tested systems.


Thesis Committee:

Internal Reader:  Dr. Samet Saeed
External Reader:  Dr. Zhiguo Hu
Advisor:  Dr. Christie Ezeife
Chair: Dr. Arunita Jaekel



Thesis Defense Announcement


5113 Lambton Tower, 401 Sunset Ave., Windsor ON, N9B 3P4, (519) 253-3000 Ext. 3716,