PaperlayerView exact revision on arXiv

Source-linked AI summary

Heterogeneous Information Network Embedding for Recommendation

Chuan Shi, Binbin Hu, Wayne Xin Zhao, Philip S. Yu

arXiv:1711.10730v1cs.SI

TL;DR

HIN-based recommendation needs better ways to extract latent structure and use heterogeneous information than path-based similarity alone. HERec learns meta-path-guided network embeddings, transforms them with fusion functions, and integrates them into an extended matrix factorization model; experiments report effectiveness and cold-start capability. The paper limits its meta-paths to those with the same starting and ending types and identifies broader embedding and explainability extensions as future work.

  • Problem

    HIN-based recommendation methods often rely on path-based similarities that cannot fully extract latent user and item structure or map HIN information effectively to recommendation.

  • Method

    HERec uses meta-path-based random walks for network embedding, transforms embeddings with fusion functions, and jointly optimizes them with an extended matrix factorization model.

  • Results

    Experiments on three real-world datasets demonstrate HERec's effectiveness, including capability for the cold-start prediction problem.

  • Takeaways & Limitations

    Transformed embedding information from HINs can improve recommendation performance within the evaluated setting.

  • Takeaways & Limitations

    The method uses only meta-paths with the same starting and ending types, while arbitrary meta-paths and improved embedding fusion remain future directions.

Abstract

from arXiv · show

Due to the flexibility in modelling data heterogeneity, heterogeneous information network (HIN) has been adopted to characterize complex and heterogeneous auxiliary data in recommender systems, called HIN based recommendation. It is challenging to develop effective methods for HIN based recommendation in both extraction and exploitation of the information from HINs. Most of HIN based recommendation methods rely on path based similarity, which cannot fully mine latent structure features of users and items. In this paper, we propose a novel heterogeneous network embedding based approach for HIN based recommendation, called HERec. To embed HINs, we design a meta-path based random walk strategy to generate meaningful node sequences for network embedding. The learned node embeddings are first transformed by a set of fusion functions, and subsequently integrated into an extended matrix factorization (MF) model. The extended MF model together with fusion functions are jointly optimized for the rating prediction task. Extensive experiments on three real-world datasets demonstrate the effectiveness of the HERec model. Moreover, we show the capability of the HERec model for the cold-start problem, and reveal that the transformed embedding information from HINs can improve the recommendation performance.

1 INTRODUCTION

HINs can represent heterogeneous auxiliary information for recommendation, but path-based methods have difficulty extracting latent structure and mapping it effectively to prediction. HERec addresses these issues with meta-path-guided embeddings, fusion functions, and an extended matrix factorization model.

  • HINs model heterogeneous auxiliary information in recommender systems, but extracting and using that information effectively remains challenging.
  • Meta-path similarities can be unreliable with sparse or noisy paths and may not directly improve recommendation performance.
  • HERec learns heterogeneous network representations using meta-path-guided node sequences rather than relying solely on explicit path connections.
  • Three fusion functions transform embeddings from different meta-paths into representations used for recommendation.
  • The transformed HIN information is integrated with an extended matrix factorization model, whose prediction and fusion components are jointly optimized.
  • Experiments on three real-world datasets evaluate effectiveness, cold-start capability, and the impact of meta-paths on performance.

2 RELATED WORK

Prior work uses collaborative filtering, HIN-based similarities, and network embedding to improve recommendation, but existing approaches do not fully combine heterogeneous structure representation with rating prediction. HERec connects HIN embedding with recommendation through a transformation mechanism.

  • Matrix factorization is an effective collaborative-filtering framework, but collaborative-filtering methods commonly suffer from cold-start problems.
  • HIN-based recommendation methods commonly use meta-path similarities to represent complex heterogeneous relations.
  • Network embedding extracts structural features, but many methods focus on homogeneous networks and cannot directly handle heterogeneous networks.
  • HERec combines HIN flexibility with network embedding to extract information for rating prediction.

3 PRELIMINARY

The paper defines HINs, network schemas, meta-paths, and HIN-based recommendation as foundations for modeling heterogeneous users, items, relations, and auxiliary information.

  • A HIN contains multiple object or link types, represented through object-type and link-type mappings.
  • A network schema is a directed meta-template over object types, with edges representing relations between those types.
  • Meta-paths connect objects through sequences of relations and can encode different semantics, such as friendship or co-watch relations.
  • HIN-based recommendation methods use heterogeneous network information to enhance recommendation performance.
  • In recommendation-oriented HINs, users, items, and rating relations are the primary entities and links of interest.

4 THE PROPOSED APPROACH

HERec has two major components: HIN embedding and recommendation. It learns user and item embeddings from HINs, then incorporates them into an extended matrix factorization model through flexible fusion functions.

  • HIN embedding: HERec first learns user and item embeddings from heterogeneous information networks.
  • Recommendation: It then extends matrix factorization by incorporating the learned embeddings with flexible fusion functions.
  • The overall pipeline constructs HINs before performing HIN embedding and recommendation.

4.1 Heterogeneous Network Embedding

HERec learns HIN node representations by guiding random walks with meta-paths, retaining homogeneous target-type sequences, and optimizing embeddings from their co-occurring neighbors. It then fuses the embeddings from multiple meta-paths into representations intended for recommendation.

  • Meta-path based Random Walk: Meta-path based random walks generate sequences that capture heterogeneous-network semantics for embedding.The walk follows a meta-path pattern through typed neighbors until a predefined length.
  • Homogeneous Sequence Construction: HERec retains only nodes matching the starting user or item type, producing homogeneous sequences from heterogeneous walks.This lets same-type nodes share an embedding space and use more relevant homogeneous neighbors within a fixed window.
  • Embedding Learning: For each meta-path, node embeddings are learned by optimizing co-occurrence with neighbors sampled in fixed-length windows.The neighborhood is constructed from homogeneous nodes encountered by meta-path based random walks, and SGD optimizes the embedding mapping.
  • Embedding Fusion: The method learns separate d-dimensional representations for nodes with respect to each meta-path, then fuses them through a task-specific function g(·).The fusion function is left to the recommendation model so it can transform embeddings into a form useful for prediction.
  • Algorithm: Algorithm 1 generates target-type walks and applies SGD to return embeddings for a single meta-path.It specifies the target node type, embedding dimension, walk length, neighborhood size, and number of walks per node.

4.2 Integrating Matrix Factorization with Fused HIN Embedding for Recommendation

HERec integrates transformed HIN embeddings with matrix factorization by jointly learning fusion-function and rating-prediction parameters. The model supports several fusion choices and adds latent factors to pair with transformed embeddings.

  • Extended Matrix Factorization: HERec extends matrix factorization by incorporating transformed user and item HIN embeddings into the rating predictor.The model retains conventional user/item latent factors and introduces additional factors to pair with the transformed embeddings.
  • Fusion Functions: Simple linear fusion assigns every user a unified weight for each meta-path and linearly transforms embeddings into the target space.The transformation uses a meta-path-specific matrix and bias vector.
  • Fusion Functions: Personalized linear fusion assigns each user a meta-path weight vector to represent individual preferences over meta-paths.This addresses the limitation of treating all users as having the same meta-path preferences.
  • Fusion Functions: Personalized non-linear fusion uses a non-linear function to increase expressive power for complex data relations.The paper uses a sigmoid transformation and notes that additional layers could be added.
  • Model Learning: The fusion functions and matrix-factorization parameters are jointly optimized with SGD under regularization.The objective includes rating-prediction error and regularization for latent factors, fusion parameters, and pairing factors.
  • Complexity: HIN embedding costs O(|P|·d·(|U|+|I|)), while per-triplet updates remain efficient when the meta-path set and embedding dimensions are small.The paper states that embedding can be trained in parallel and that updating transformation matrices costs O(|P| · D · d).

5 EXPERIMENTS

The experiments evaluate HERec on three real-world recommendation datasets from movie, book, and business domains. Recommendation quality is measured with MAE and RMSE, where smaller values indicate better performance.

  • Evaluation Datasets: HERec is evaluated on Douban Movie, Douban Book, and Yelp datasets spanning movie, book, and local-business recommendation.The datasets include user-item ratings together with social or attribute information.
  • Evaluation Datasets: Douban Movie contains 13,367 users, 12,677 movies, and 1,068,278 ratings, while Douban Book contains 13,024 users, 22,347 books, and 792,026 ratings.Both rating ranges are 1 to 5.
  • Evaluation Metrics: Recommendation quality is measured using mean absolute error (MAE) and root mean square error (RMSE).The metrics compare actual and predicted ratings on the test set.
  • Evaluation Metrics: Smaller MAE or RMSE values indicate better recommendation performance.The metrics are computed over test-set rating records.

5.3 Methods to Compare

The experiments compare HERec with conventional matrix-factorization, HIN-based, and network-embedding baselines for rating prediction.

  • The comparison includes PMF and SoMF as matrix-factorization baselines, alongside FMHIN, HeteMF, SemRec, and DSR as HIN-based methods.
  • The selected HIN-based methods cover similarity-based recommendation, context-aware factorization, weighted heterogeneous-network collaborative filtering, and dual similarity regularization.
  • HERecdw replaces HERec’s embedding component with homogeneous DeepWalk, while HERecmp uses the heterogeneous metapath2vec++ embedding method.
  • Table 4 reports MAE, RMSE, and improvement ratios relative to PMF across three datasets; smaller MAE or RMSE and larger improvement ratios indicate better performance.
  • HERec uses heterogeneous information network embedding with a personalized non-linear fusion function in the effectiveness experiments.

5.4 Effectiveness Experiments

Across three datasets, the experiments find that HERec consistently outperforms the compared baselines, with larger gains when training data are limited and on sparse data.

  • HIN-based baselines outperform traditional PMF and SoMF, suggesting that heterogeneous information provides useful evidence for recommendation.
  • HERec is consistently better than all baselines, from PMF through DSR, and its superiority becomes more significant with less training data.
  • Up to 40% improvement over PMF occurs with 20% training data on the Douban Book dataset.
  • With 60% training data on Yelp, HERec improves over PMF by about 26%, indicating effectiveness on a sparse dataset.
  • HERec outperforms its variants, while HERecdw performs much worse than HERecmp, highlighting the importance of the network-embedding component and task-specific HIN embedding.

5.5 Detailed Analysis of The Proposed Approach

Detailed analyses examine fusion functions, cold-start prediction, meta-path selection, latent-factor tuning, parameter weights, and convergence to characterize HERec’s behavior.

  • Selection of Different Fusion Functions: Across three datasets, the fusion-function ranking is HERecpnl > HERecpl > HERecsl, with personalized non-linear fusion performing best.
  • Cold-start Prediction: HERec performs best among the cold-start comparison methods, and its improvement over PMF increases for users with fewer rating records.
  • Impact of Different Meta-Paths: Performance generally improves as more meta-paths are incorporated, although additional paths can cause slight fluctuations because some contain noisy or conflicting information.
  • Impact of Different Meta-Paths: The performance reaches a relatively good level with only a few meta-paths, supporting complexity control through selecting a small number of high-quality paths.
  • Parameter Sensitivity: Ten latent factors yield the best performance when the number of factors is varied from 5 to 40.
  • Parameter Sensitivity: The optimal α and β values are near 1, and about 40–60 iterations suffice for dense datasets versus about 20 for sparse Yelp.

6 CONCLUSION

HERec integrates meta-path-guided heterogeneous network embeddings with recommendation-specific fusion and extended matrix factorization. Experiments evaluate its effectiveness, cold-start capability, and the impact of meta-paths, while identifying extensions for embedding fusion and broader meta-path coverage.

  • Conclusion: HERec uses meta-path-based random walks to generate meaningful node sequences for heterogeneous network embedding.The approach targets useful structural information from heterogeneous information networks.
  • Conclusion: Embeddings from different meta-paths are fused and jointly optimized with an extended matrix factorization model for rating prediction.The fusion functions and matrix factorization model are optimized together for the recommendation task.
  • Conclusion: Experiments on three real datasets demonstrate HERec’s effectiveness and evaluate its ability to alleviate cold-start problems and assess meta-path impact.The conclusion reports these evaluations without specifying quantitative results here.
  • Conclusion: Future work includes deep-learning-based fusion, arbitrary meta-paths for any node types, and improved recommendation explainability.The current work restricts meta-paths to those with the same starting and ending types.

RMSE

The supplied figures examine RMSE performance under changes in latent-factor dimension, parameters α and β, and iteration count across three datasets.

  • RMSE: Fig. 7 examines performance as the dimension of latent factors changes on three datasets.The supplied caption does not state the observed RMSE comparison.
  • RMSE: Fig. 8 examines performance while varying parameters α and β on three datasets.The supplied caption does not state the observed RMSE comparison.
  • RMSE: Fig. 9 examines performance as the number of iterations changes on three datasets.The supplied caption does not state the observed RMSE comparison.
Loading 1711.10730v1…