Abstract:Sequential recommendation seeks to predict users' next behaviors and recommend related items over time. Existing research has mainly focused on modeling users' dynamic preferences from their sequential behaviors. However, most of these studies have ignored the negative effects of noise behaviors in the given sequences, which may mislead the recommender. In addition, users' behavior data is always sparse, which makes it difficult to effectively learn users' preferences purely from their historical behaviors. Most recently, knowledge graphs (KGs) have been exploited by few researchers for sequential recommendation. However, they always assume all information in KGs or KG paths with limited length are useful for recommendation, which may bring irrelevant information from KGs into the recommender and further mislead the recommender. To address these issues, we propose a novel KG-based behavior denoising and preference learning model named KGDPL for sequential recommendation. We argue that the paths in KGs that reflect semantic relations between entities can not only help to remove noise behaviors and recommend successive items for users, but also provide relevant explanations. Therefore, we first devise a supervised knowledge path selection module to select effective paths between items from KGs for behavior prediction, which aims to filter out irrelevant information from KGs for the given recommendation task. Then, we design a knowledge-enhanced behavior denoising module to mitigate the negative effects of the noise behaviors contained in historical sequences by using the knowledge path information. After that, we propose a knowledge-enhanced preference learning module to better learn users' personalized and dynamic preferences from their historical behavior sequences and related knowledge information, which can also help tag users and provide explanations for recommendation results. Experimental results on four real-world datasets demonstrate the effectiveness and interpretability of the proposed model KGDPL.

KERL: A Knowledge-Guided Reinforcement Learning Model for Sequential Recommendation

Knowledge-guided Deep Reinforcement Learning for Interactive Recommendation

Knowledge-Enhanced Causal Reinforcement Learning Model for Interactive Recommendation

Multi-modal Knowledge-aware Reinforcement Learning Network for Explainable Recommendation

Interactive Recommender System via Knowledge Graph-enhanced Reinforcement Learning

Model-enhanced Contrastive Reinforcement Learning for Sequential Recommendation

KASR: Knowledge-Aware Sequential Recommendation

Teach and Explore: A Multiplex Information-guided Effective and Efficient Reinforcement Learning for Sequential Recommendation

Multi-level Recommendation Reasoning over Knowledge Graphs with Reinforcement Learning

CADRL: Category-aware Dual-agent Reinforcement Learning for Explainable Recommendations over Knowledge Graphs

Knowledge Graph-Based Behavior Denoising and Preference Learning for Sequential Recommendation

A Knowledge Graph-based Interactive Recommender System Using Reinforcement Learning

User Behavior Enriched Temporal Knowledge Graphs for Sequential Recommendation.

Reinforcement Learning based Path Exploration for Sequential Explainable Recommendation

KPRLN: deep knowledge preference-aware reinforcement learning network for recommendation

Empowering recommender systems using automatically generated Knowledge Graphs and Reinforcement Learning

Ekar: an Explainable Method for Knowledge Aware Recommendation

Enhancing User Interest Modeling with Knowledge-Enriched Itemsets for Sequential Recommendation

A Joint Framework for Explainable Recommendation with Knowledge Reasoning and Graph Representation

AdaRec: Adaptive Sequential Recommendation for Reinforcing Long-term User Engagement

DIRS-KG: a KG-enhanced interactive recommender system based on deep reinforcement learning