Date of Award

5-17-2022

Publication Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

Keywords

Clickstream Data;Collaborative Filtering;E-commerce;Historical Purchases;Recommendation Systems;Sequential Pattern Mining

Supervisor

Christie Ezeife

Abstract

E-commerce Recommendation Systems (ERS) facilitate customers’ purchase decision by recommending products or services of interest. Designing a recommender system targeted towards an individual customer’s need is crucial for retailers to increase revenue and retain customers’ loyalty. Collaborative Filtering (CF), a common recommendation technique, takes user-item interaction matrix as input which represents interactions either explicitly (users ratings) or implicitly (users’ browsing or buying behavior) and outputs top item recommendations for each target user, by finding similarities among users or items. The input matrix suffers from (i) sparsity (has low user item interactions), (ii) cold start (an item cannot be recommended if no ratings exist). Content Based method, on the other hand, generates recommendations based on the content (features) of the item and suffers from content overspecialization (lack of diversity in recommended products) due to the use of specific features only. Furthermore users’ interests and preferences change with time. The time stamp of a user interaction (click or purchase event) is an important characteristic. Learning the sequential patterns of user interactions based on the timestamps are useful to understand their long and short term preferences and predict the next items for recommendation. Sequential Pattern Mining mines frequent or high utility sequential patterns from a sequential database comprising of historical purchase or click sequences. Conventional recommendation systems (ChoiRec12, SuChen15, HPCRec18, HSPRec19) utilize mining techniques such as clustering, frequent and sequential pattern mining along with click and purchase similarity measures for next item recommendation. However, the performance of these systems is still limited when the matrix is sparse, as the number of items keep increasing rapidly. Additionally, models utilizing sequential pattern mining suffer from (i) lack of personalization:patterns are not targeted for a specific customer, as they infer decisions based on a global view of sequences and (ii) lack of contextual similarities among recommended items: they can only recommend items that exist as a result of a matching rule generated from frequent sequential purchase pattern(s). To better understand users’ preferences and to infer the inherent meaning of the items (e.g., context in which they are used), this thesis, explores the effectiveness of utilizing semantic knowledge (meaningful relationships between items) extracted from items’ meta data (title, description and brand) and customers’ purchase histories to compute semantic similarities between items according to their (a) usage (e.g., products co-purchased or coreviewed) and (b) textual features by finding similarity between products based on their characteristics. The extracted semantic knowledge is then integrated into different phases of recommendation process such as (i) pre-processing, to learn relationships between items, (ii) candidate item generation and (iii) generating semantic rich and sequential next item recommendations. During the candidate item generation phase, techniques developed include (a) mining semantic rich sequential patterns, (b) enriching the item matrix in Collaborative Filtering to select Top-N candidates that show semantic and sequential relationships between items and (c) enhancing the Transition Probability Matrix in the Markov Model method. The third phase of generating semantic rich sequential recommendations is accomplished by using semantic rich (a) Sequential Pattern Rules, (b) item based Collaborative Filtering or (c) Markov Model depending upon the method used to generate candidate items. Thus, the inclusion of semantic knowledge into all phases of recommendation process can address the issues of sparsity, coldstart, content overspecialization and provide recommendations which are diverse, similar in context and better reflect user’s long and short term interests. Experimental results on publicly available e-commerce data sets such as Amazon and Online Retail has shown that the proposed model has improved performance over existing systems.

Download

Share

COinS