Advancing Hybrid Recommender Systems for E-Commerce with Knowledge-Enhanced Architecture and Adaptive User Engagement Modeling

Abstract

In the era of information overload, recommender systems have emerged as pivotal tools for helping users navigate vast product catalogs in e-commerce environments. This thesis presents a novel approach to hybrid recommender systems for e-commerce platforms by combining collaborative filtering with content-based methods enhanced by knowledge graph integration. The proposed system addresses key challenges in recommendation quality, diversity, and cold-start scenarios through adaptive user engagement modeling.

The first part of this work introduces a hybrid recommender system that integrates Alternating Least Squares (ALS) collaborative filtering with a two-tower content-based filtering model for e-commerce applications. This approach leverages user-item interactions while incorporating product description features to overcome data sparsity issues commonly found in traditional recommendation approaches. The system adaptively combines predictions from both models using a weighted approach, offering improved recommendation quality and diversity.

Building upon this foundation, the second part presents KAHRec: a Knowledge-Augmented Hybrid Recommender system that enhances the initial model through three key theoretical innovations. First, the system integrates BERT-driven semantic understanding to capture contextual relationships in product descriptions and user reviews, enabling more nuanced feature representations than traditional word embeddings. Second, it constructs structured knowledge graphs from over 148,000 product relationships (also_bought, also_viewed, similar_item) and applies PageRank and degree centrality measures to quantify product importance within the recommendation ecosystem. Third, it implements stratified user engagement modeling that categorizes users into distinct interaction tiers (One/Few/Medium/Many) and applies tier-specific adaptive fusion strategies, with weights ranging from collaborative-dominant (0.75 ALS, 0.25 Two-Tower) for highly engaged users to content-dominant (0.25 ALS, 0.75 Two-Tower) for cold-start scenarios. This theoretical framework enables the system to dynamically balance matrix factorization efficiency with neural personalization based on data availability and user characteristics.

The knowledge-augmented approach incorporates synthetic item generation for cold-start mitigation that combines knowledge graph centrality with BERT feature similarity, providing comprehensive coverage across diverse user scenarios. The adaptive fusion mechanism employs consensus boosting that amplifies recommendations when both models agree, while applying strategic penalties for model disagreements, ensuring robust performance across varying data conditions.

Evaluated on a large-scale dataset containing over 400,000 user-item interactions across 30,445 products, the proposed approaches demonstrate significant improvements over baseline methods and state-of-the-art models in precision, recall, and user satisfaction metrics. The knowledge-augmented KAHRec approach achieves 6% higher precision and 17% higher recall compared to individual baselines, with a 25% relative improvement in cold-start precision scenarios. KAHRec demonstrates consistent superiority over state-of-the-art models, outperforming BERT4Rec and KGAT with 8.5% precision and 8.2% recall improvements respectively. Additional benefits include 3.6% greater recommendation diversity while maintaining robust error performance (MAE: 0.42).

The system shows particular efficacy for medium-engagement users, outperforming alternatives by more than 8% in precision metrics. By strategically balancing matrix factorization efficiency with neural personalization, this work advances adaptable recommendation frameworks for dynamic e-commerce environments requiring both accuracy and responsiveness to catalog evolution.