Unified Embedding Model over Heterogeneous Information Network for Personalized Recommendation

Zekai Wang, Hongzhi Liu, Yingpeng Du, Zhonghai Wu, Xing ZhangIJCAI 2019 · OpenReview

Abstract

Most heterogeneous information network (HIN) based recommendation models rely on user and item modeling with meta-paths. However, they often model users and items in isolation under each meta-path, which can lead to misleading information extraction. Furthermore, they typically only consider structural features, potentially losing useful recommendation signals. To address these issues, we propose HueRec, a HIN-based unified embedding model. HueRec assumes common characteristics exist across different meta-paths for each user or item and learns unified representations to alleviate data sparsity and noise. By combining graph exploration and recommendation into an end-to-end framework, HueRec avoids information loss and utilizes an attention mechanism to measure personalized preferences on meta-paths. Extensive experiments demonstrate that HueRec consistently outperforms state-of-the-art methods.

Introduction

Personalized recommendation systems increasingly leverage Heterogeneous Information Networks (HINs)—complex graphs where nodes (e.g., users, movies, directors) and edges (e.g., watched, directed) have diverse types. To navigate these networks, researchers use meta-paths (such as User → Movie → Director → Movie) to uncover latent similarities between entities.

Existing HIN-based models suffer from two primary limitations. First, they often analyze meta-paths independently (Isolation & Noise). If a specific path is sparse or noisy, the resulting user representation becomes skewed. Second, they typically follow a "Two-Step" process where features are extracted first and then fed into a recommender. This separation is inherently lossy; signals that are structurally minor but critical for recommendation may be discarded prematurely.

The HueRec Framework

HueRec addresses these flaws by treating HIN exploration as a unified embedding problem within an end-to-end architecture. The model bridges the gap between graph structure and recommendation utility.

The HueRec framework: (a) HIN example, (b) Information extraction via meta-paths, and (c) Recommendation via attention.

Unified Embeddings and Path Estimation

Instead of learning separate vectors for a user under every meta-path, HueRec learns one unified embedding for each user and item. This allows information from dense meta-paths to regularize sparse ones, making the model robust to noise. Inspired by knowledge graph models like TransE, HueRec represents meta-paths as vectors in the same latent space, estimating connection strength by calculating interactions between user, item, and path embeddings.

Personalized Attention Mechanism

Not all users value relationships equally; one might prioritize directors while another follows genres. HueRec employs an attention mechanism to calculate personalized weights for each meta-path. This mechanism combines a global importance score with a user-specific preference, ensuring the model captures individual nuances.

Detailed supervision and embedding architecture of HueRec, illustrating the joint optimization of HIN structure and recommendation loss.

Joint Optimization

The model is trained using a multi-task objective that simultaneously optimizes for two losses:

  1. $Loss^{HIN}$: Ensures embeddings accurately reflect the HIN's structural properties.
  2. $Loss^{Rec}$: Optimizes for the actual recommendation task (e.g., predicting ratings or clicks).

Experimental Evaluation

The model was evaluated on three real-world datasets: Yelp (business/social), MovieLens (movies/directors/genres), and Douban Movie (movies/user groups).

Key Results

HueRec consistently outperformed state-of-the-art baselines like FMG and HeteRec:

  • Significant Performance Gains: On MovieLens, HueRec improved MAP@10 by 24.76% and NDCG@10 by 17.71% over the strongest baseline.
  • Stability: While traditional Factorization Machines (FM) struggled with "feature overload" on complex datasets like Douban, HueRec remained stable and accurate.
  • Ablation Studies: Removing either the unified embedding or the attention mechanism led to significant performance drops, validating the necessity of both components.

Conclusion and Future Work

HueRec demonstrates that interrelation between meta-paths is key to filtering noise and that end-to-end training prevents irreversible information loss. Future work may explore more sophisticated non-linear fusion methods, such as deep neural networks, to combine meta-path scores and automate the selection of optimal meta-paths to reduce manual feature engineering.

Summary updated

Abstract

Most of heterogeneous information network (HIN) based recommendation models are based on the user and item modeling with meta-paths. However, they always model users and items in isolation under each meta-path, which may lead to information extraction misled. In addition, they only consider structural features of HINs when modeling users and items during exploring HINs, which may lead to useful information for recommendation lost irreversibly. To address these problems, we propose a HIN based unified embedding model for recommendation, called HueRec. We assume there exist some common characteristics under different meta-paths for each user or item, and use data from all meta-paths to learn unified users’ and items’ representations. So the interrelation between meta-paths are utilized to alleviate the problems of data sparsity and noises on one meta-path. Different from existing models which first explore HINs then make recommendations, we combine these two parts into an end-to-end model to avoid useful information lost in initial phases. In addition, we embed all users, items and meta-paths into related latent spaces. Therefore, we can measure users’ preferences on meta-paths to improve the performances of personalized recommendation. Extensive experiments show HueRec consistently outperforms state-of-the-art methods.
BibTeX
@inproceedings{wang2019unified,
  author = {Zekai Wang and Hongzhi Liu and Yingpeng Du and Zhonghai Wu and Xing Zhang},
  title = {{Unified Embedding Model over Heterogeneous Information Network for Personalized Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=FwkBknQ_0Tw},
}
Figure