Date of Award

2-15-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

association rule mining;Cross-domain recommendations;data mining;e-commerce;recommender systems;social media

Supervisor

Christie Ezeife

Abstract

Ecommerce recommendation accuracy can be improved by mining patterns from other domains such as social media (Facebook), to predict purchase behaviours. The "cross-site cold start problem" arises when traditional recommender systems, relying on e-commerce purchase history, face platforms with no user history. Existing systems such as the GaoLinRec23 (2023) system that employs (CMF) Collective Matrix Factorization to jointly factorize user-item interaction matrices from both domains, WangZhaoRec21 (2021), GaoRec20 (2021), WangHeNeiChuaRec17 (2017), which also incorporate user attribute and social connections from the social media domain have attempted to bridge the gap, however the assumption that specific item embeddings which are the specific product details are shared between these domains does not align with the real-world scenario. Major e-commerce and social media platforms, including Amazon and Facebook, typically do not exchange granular product information. This disconnect poses a critical obstacle for existing recommendation systems in providing accurate suggestions for users starting with no observable e-commerce activity. This thesis proposes Facebook Data Cross Recommendation ‘2023 (FD-CDR ’23) system, which uses the proposed MLTU (Mine Likes and Transactions per User) algorithm to extract likes and purchase history of users from both domains, transforming then into itemsets. A modified association rule mining is applied uncover patterns of frequent co-occurrence between user Facebook post likes and e-commerce transactions as rules. It then uses the proposed HARR (Hybrid Association Rule Recommendation) algorithm to match new user facebook likes to, generated rules such as “Users who typically like cooking posts, buy cooking recipes” without needing to share item embeddings across platforms and still solve the cross-site cold start problem. Experimental results with precision and recall show that the proposed FD-CDR’23 system provides more accurate recommendations than the mentioned existing systems.

Download

Share

COinS