Knowledge Graph Assisted Personalized Recommendation Systems

In today's booming internet industry, personalized recommendation systems are a core technology for any user‑facing platform. Traditional recommender models suffer from sparsity and cold‑start problems, making it difficult to compute reliable user‑item similarities.

Collaborative Filtering (CF) is a classic method that assumes similar users have similar preferences, but it relies solely on historical rating matrices, leading to limited performance when interactions are sparse.

Knowledge Graphs (KG) serve as a rich side‑information source. A KG is a heterogeneous directed graph where nodes represent entities and edges represent relations, typically expressed as triples (head, relation, tail). By linking items (e.g., movies) to entities (actors, genres, directors) and further to non‑item entities, KGs provide additional semantic connections that can alleviate sparsity and cold‑start issues.

Using KG in recommendation brings three main benefits: higher precision through richer similarity signals, increased diversity by exploring multiple semantic paths, and better explainability by tracing recommendation paths in the graph.

KG processing methods include Knowledge Graph Embedding (KGE) techniques such as TransE, TransH, and TransR, which learn low‑dimensional vector representations for entities and relations by enforcing translation‑based constraints (e.g., h + r ≈ t).

KG‑aware recommender approaches can be grouped into three categories:

Embedding‑based methods: DKN (Deep Knowledge‑Aware Network) for news recommendation and MKR (Multi‑Task Feature Learning for KG‑Enhanced Recommendation) which jointly optimizes a recommendation task and a KG embedding task via cross‑compress units.

Path‑based methods: (not covered in this summary).

Hybrid methods: RippleNet propagates user preferences over multi‑hop KG neighborhoods; KGCN and KGNN‑LS apply Graph Convolutional Networks on weighted KG graphs, introducing relation‑specific scoring functions and label‑smoothness regularization to prevent over‑fitting.

Hybrid models such as RippleNet construct user‑specific weighted adjacency matrices by scoring relations per user, then perform multi‑hop message passing to obtain a user embedding that is combined with candidate item embeddings for click‑through prediction.

KGCN transforms the heterogeneous KG into a weighted graph where edge weights are learned via a relation scoring function s_u(r). A user‑specific adjacency matrix A_u is fed into a GNN (A_u H^{(l)} W^{(l)}) to aggregate multi‑hop entity features, producing final entity embeddings used for recommendation.

KGNN‑LS further introduces a label‑propagation regularizer: known user‑item interactions serve as binary labels on item nodes, and a label‑propagation algorithm spreads these labels over the weighted KG. The resulting smoothness loss encourages neighboring nodes with strong edges to have similar predictions, acting as an additional regularization term.

Overall, the lecture covered the fundamentals of recommendation systems, the challenges they face, how knowledge graphs can be leveraged as auxiliary information, and a suite of state‑of‑the‑art KG‑aware recommendation models (DKN, MKR, RippleNet, KGCN, KGNN‑LS) along with their underlying embedding and graph‑neural techniques.