Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation (1901.08907v1)

Abstract: Collaborative filtering often suffers from sparsity and cold start problems in real recommendation scenarios, therefore, researchers and engineers usually use side information to address the issues and improve the performance of recommender systems. In this paper, we consider knowledge graphs as the source of side information. We propose MKR, a Multi-task feature learning approach for Knowledge graph enhanced Recommendation. MKR is a deep end-to-end framework that utilizes knowledge graph embedding task to assist recommendation task. The two tasks are associated by cross&compress units, which automatically share latent features and learn high-order interactions between items in recommender systems and entities in the knowledge graph. We prove that cross&compress units have sufficient capability of polynomial approximation, and show that MKR is a generalized framework over several representative methods of recommender systems and multi-task learning. Through extensive experiments on real-world datasets, we demonstrate that MKR achieves substantial gains in movie, book, music, and news recommendation, over state-of-the-art baselines. MKR is also shown to be able to maintain a decent performance even if user-item interactions are sparse.

• The paper introduces cross-compress units that enable high-order feature interactions between recommender systems and knowledge graphs.
• It leverages a multi-task learning framework to mitigate data sparsity and cold-start issues with notable AUC improvements.
• Empirical evaluations across movies, books, music, and news demonstrate MKR’s robust real-world recommendation performance.

Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation

The paper "Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation" presents a novel approach aimed at improving the efficiency of recommender systems by leveraging the rich informational context of knowledge graphs (KGs). The proposed model, MKR (Multi-task feature learning for Knowledge graph enhanced Recommendation), employs a multi-task learning framework that integrates a recommender system (RS) task with a knowledge graph embedding (KGE) task, thereby addressing the issues of data sparsity and cold-start problems typically encountered in collaborative filtering.

Key Contributions

The standout contribution of this work is the introduction of cross-compress units, which facilitate high-order interaction learning between items in RS and related entities in KGs. By modeling the latent feature interactions more explicitly, these units allow MKR to effectively transfer knowledge between the RS and KGE tasks, which are interconnected but traditionally treated separately in the literature. The paper's theoretical analysis demonstrates that these units can approximate high-order interactions to an exponential degree, providing a robust framework for feature sharing.

Empirical results further validate the effectiveness of MKR across various domains—including movie, book, music, and news recommendation—yielding substantial improvements over state-of-the-art methods in both click-through rate prediction (with average AUC improvement ranging from 0.3% to 66.4%) and top-K recommendation scenarios.

Implications and Future Directions

Practically, MKR's ability to maintain strong recommendation performance even in sparse user-item interaction scenarios suggests it is well-suited for real-world applications where data incompleteness is a significant barrier. Theoretical implications include deeper explorations into the nature of shared representations between RS and KG tasks, which could spur further research in more sophisticated multi-task and transfer learning approaches.

Looking forward, the MKR framework can be enhanced by exploring alternative neuronal architectures, like Convolutional Neural Networks (CNNs), to enrich the feature extraction process. Additionally, integrating different KGE methods by redesigning the cross-compress unit would likely broaden its applicability across domains outside of straightforward item recommendations. Furthermore, as multi-tasking continues to evolve, the balance of task-specific and shared layers remains an open question, offering an avenue for further empirical validation and theoretical refinement.

In sum, this paper poses a comprehensive model that not only leverages the abundant relational information within KGs but also challenges existing paradigms of isolated task processing, underscoring the potential for MKR to reshape approaches in recommendation scenarios laden with sparse data challenges.