The School of Computer Science is pleased to present…

Mining User Facebook Post Likes for Cross Domain Product Recommendations across E-commerce platforms

MSc Thesis Defense by: Emmanuel Ainoo

 

Date: Friday January 5, 2024

Time:  11:00 am – 12:30 pm

Location: Essex Hall, Room 105

 

Abstract:

E-commerce recommendation accuracy can be improved by transferring user preference insights from social media domains like Facebook to the e-commerce platforms, especially when dealing with users who have limited historical purchase activity. An existing cross domain recommendation system GaoLinRec23 system employs (CMF) Collective Matrix Factorization to jointly factorize user-item interaction matrices from both e-commerce and social media domains by first collecting user-item interaction data from both domains such as clicks and purchases, factorizing these matrices to decompose them into underlying latent patterns and using the user-item interaction matrices, which now reveal the underlying latent patterns, as input data for collaborative filtering technique. However, CMF, as used in the GaoLinRec23, and other existing systems ChenWuRec05, GaoRec19, WangHeNieChuaRec17, assumes that specific item such as product details (e.g., [product1: skirt, product2: cell phone]) is common between selected domains, which may not reflect the practical reality. In practice, most e-commerce and social media domains often do not share product details across domains. It then becomes a problem to apply existing system approaches to real-world platforms like Amazon, Facebook which do not share product/item data.

This thesis proposes Facebook Data Cross Recommendation ‘2023 (FD-CDR ’23) system, which mines users’ Facebook activities such as likes, and e-commerce historical transaction data such as purchases for product recommendations for new users. It first uses the MLTU’23 (Mine Likes and Transactions per User) algorithm to mine user facebook likes and historic user purchases from both domains and transform them into a list of itemsets. These itemsets are fed into a modified association rule mining technique that enforces liked posts as antecedents and purchased products as consequents to cater for new users with no purchase history and then extract patterns of frequent co-occurrence between user Facebook post likes and ecommerce transactions. For instance, a rule can uncover that Users (user) who typically like cooking (post) posts on Facebook, purchase cooking recipes (item) from e-commerce”. With such rules, the system can unearth product recommendations using the proposed HARR’23 (Hybrid Association Rule Recommendation) which matches new user facebook likes to the generated rule database sorted by confidence level, without having to share the item details across the domains, as existing systems do. A pattern such as <User_i, (A), (B)> means, whenever〖 User〗_i likes post A, they follow it up with a purchase of product B. In running experiments, facebook-eBay dataset contains a random sample of over 13,000 anonymous eBay users who connected to Facebook in 2012 was used with focus on purchase and likes. With boxplots and other distributions, we demonstrate that there are significant correlations between user social media activity and online purchases. Results on precision, recall, and F1 score values show that the proposed FD-CDR’23 system is indeed effective in improving recommendation accuracy compared to existing systems.

 

Thesis Committee:

Internal Reader: Dr. Shafaq Khan    

External Reader: Dr. Jeffrey Rau

Advisor:  Dr. Christie Ezeife

Chair: Dr. Majid Afshar Noghondari