Knowledge Graph Enhanced Recommender Systems: Methods, Models, and Experiments

In the era of information overload, recommender systems help users discover personalized items, but traditional collaborative filtering suffers from data sparsity and cold‑start issues.

Knowledge graphs (KGs), represented as heterogeneous directed graphs of entities and relations, provide auxiliary information that can link items and enrich their representations.

1. Collaborative Filtering (CF) predicts ratings by aggregating similar users' feedback, yet it faces sparsity and cold‑start problems.

2. CF + Side Information incorporates additional signals such as social networks, item attributes, multimedia content, and contextual data to mitigate CF drawbacks.

Knowledge Graph Embedding (KGE) learns low‑dimensional vectors for entities and relations. Common translational models include:

TransE: enforces head + relation ≈ tail.

TransH: projects entities onto relation‑specific hyperplanes to handle 1‑N, N‑1, and N‑N relations.

TransR: maps entities into relation‑specific spaces before applying translational constraints.

3. Deep Knowledge‑Aware Networks (e.g., DKN) extract entity embeddings from KG subgraphs, combine them with word embeddings via Knowledge‑Aware CNN (KCNN), and use attention mechanisms to weight historical user interactions for click‑through prediction.

4. Multi‑Task Feature Learning for KG‑Enhanced RS jointly optimizes recommendation loss, KGE loss, and regularization using cross‑&‑compress units to model interactions between item and entity embeddings.

5. RippleNet (Structure‑Based) propagates user preferences over KG hops, computing relevance scores between candidate items and multi‑hop neighbor entities, and aggregates them to form the final user representation.

6. Knowledge Graph Convolutional Networks (KGCN) treat KG edges as unweighted, introduce edge weights via user‑relation inner products, and perform layer‑wise graph convolutions to obtain item embeddings; the final click probability is derived from the inner product of user and item embeddings.

7. Comparative Analysis experiments on multiple datasets show that KGCN achieves the best accuracy, while DKN performs worst; embedding‑based methods scale better than structure‑based ones, but the latter offer more interpretable graph‑based reasoning.

In conclusion, incorporating knowledge graphs into recommender systems can improve accuracy, diversity, and explainability, effectively alleviating sparsity and cold‑start challenges.