MSc Thesis Proposal Announcement of Md Burhan Uddin:"Multi-Data Source Recommendations with Derived Sequential Pattern Mining"

 

Date: Thursday, April 21st, 2022 

Time:  10:30 am – 12:30 pm 

Meeting URL:      https://us06web.zoom.us/j/81162369010?from=addon 

Passcode: If interested in attending this event, contact the Graduate Secretary at csgradinfo@uwindsor.ca with sufficient notice before the event to obtain the passcode.

     User-based Collaborative Filtering (CF) methods fail to derive preference value for new items with no ratings or purchase information. Using multi-storage solutions (e.g., multiple related tables in a relational database, distributed and multi-platform databases) in E-commerce and the existence of many iterations of the same item provides the opportunity to improve the recommendation process and accuracy through multi-source data integration. For example, with traditional Recommendation System (RS) integrating purchase sequence similarity helps find users with similar interests resolving sparsity issue. Using the item attributes information (e.g., category) to derive user preference on an item helps generate a recommendation for similar items with no rating or purchase history alleviating cold start problem. 

     The existing multi-source RS algorithms, such as ECCF18, address the cold start problem using a Category Co-occurrence Graph (CCG) to extend the CF model and enhance the preference matrix to derive recommendations. ECCF18 fails to generate personalized recombination for users with few to no ratings. The CD-SPM19 model uses item ontology to derive semantic similarity information to enrich the collaborative filtering (CF) step. The main limitation of the CD-SPM19 model is that it fails to capture changes in user preference over time. Both algorithms fail to learn complex sequential relations in user purchase behavior without utilizing sequential purchase information. The HPCRec18 model derives a consequential bond between click to purchase to predict preferences for users with no purchase history but fails to capture the sequential relations. Finally, the HSPRec19 algorithm derives sequential purchase patterns from its generated historical sequential purchase database (SHOD). It then uses the consequential bond in mined patterns to improve the user-item matrix for CF. As they do not consider the categorical or semantic similarity between items in the CF step, both HPCRec18 and HSPRec19 cannot derive recommendations for new items that were not clicked or purchased before.  

     This thesis proposes an algorithm called Multi-source Category Extended Historical Sequential Pattern Recommendation System (MCE-HSPRec) that derives enriched category-based user profiles by analyzing purchase behavior and items categories that are frequently purchased together. MCE-HSPRec first generates E-commerce Sequential Database (SD) using the SHOD algorithm from historical purchase data. Then it generates a Category Sequence Database (CSD) and a user category preference matrix (UC) from the combination of historical purchases and item category information. The thesis mines SD and CSD database tables related through foreign key relationships for sequential patterns to (i) derive an enriched user-item (UI) matrix using the HSPRec19 model, (ii) derive a Category Co-occurrence Graph (CCG) from the CSD. Finally, the thesis applies an extended category-based CF on the enriched CCG, UC, and UI matrix to derive an enriched user-category preference matrix. 

 

Keywords: Multi-Source, Sequential Pattern Mining, Category Co-occurrence, Recommendation System, Implicit feedback 

 

Internal Reader:      Dr. Yung H. Tsin 

External Reader:     Dr. Scott Mundle       

Advisor:                   Dr. Christie Ezeife 

 

 

 

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