Date of Award

4-3-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Collaborative filtering;Data mining;Data sparsity;E-commerce Recommender systems;Multi-behavior Recommender system;Sequential pattern mining

Supervisor

C Ezeife

Abstract

In most real-world recommender systems, users interact with items sequentially and multi-behaviorally. There are various user multi-behaviors in practical scenarios, such as clicks, likes, add-to-cart, and purchases. Analyzing the fine-grained relationship of items behind the users’ multi-behavior interactions is critical in improving the performance of recommender systems. Existing methods, such as HSPRec19, DACBRec21, and MBHT22, use customer multi-behaviour information to improve the accuracy of recommendations. HPCRec18 system used purchase frequency and consequential bond between clicks and purchased items data to improve the user-item frequency matrix. HSPRec19 system enhances the user-item rating matrix input to collaborative filtering with sequential purchase patterns by reducing the matrix sparsity. Still, it does not capture the item-level multi-behavior dependencies to further alleviate the data sparsity problems. DCABRec21 system uses multiple user behaviors and negative feedback in the Collaborative Filtering (CF) method. MBHT22 systems is a multi-behavior recommendation system that uses a hypergraph-transformer. This thesis proposes a system called the Multi-Behaviour Sequential Pattern Recommendation System (MBSPRec System), which is an extension of the HSPRec19 system that includes multi-behavior frequent patterns along with frequent click and purchase patterns to improve the accuracy of recommendations and reduce user-item rating data sparsity problem to a larger extent. The proposed MBSPRec generates a Multi-Behaviour Sequential Database for each user behavior type using the Multi-Behaviour Sequential Database Generator (MBSDBG) and Multi-Behaviour Sequential Pattern Miner (MBSPM), which mines multiple user behavior sequential pattern rules to yield additional sequential patterns and further reduce data sparsity of User-Item Matrix and improve the accuracy of the recommendations. The proposed MBSPRec mines approximate sequential data using the ApproxMAP algorithm to improve the Consequential Bond between multiple behavior and purchase sequences to give multi-behavior frequent sequential rules where no purchase has happened. Experimental results show that the proposed MBSPRec achieves more recommendation accuracy and reduces user-item rating data sparsity than the tested existing systems.

Download

Share

COinS