PaperlayerView exact revision on arXiv

Source-linked AI summary

An Efficient Neighborhood-based Interaction Model for Recommendation on Heterogeneous Graph

Jiarui Jin, Jiarui Qin, Yuchen Fang, Kounianhua Du, Weinan Zhang, Yong Yu, Zheng Zhang, Alexander J. Smola

arXiv:2007.00216v1cs.IR

TL;DR

HIN recommenders may rely on sparse or noisy path connections or compress neighborhoods before prediction, leaving neighborhood interactions underused. NIRec models interactions through metapath-guided neighborhoods and learns them convolutionally with FFT; it outperforms state-of-the-art baselines on four public datasets.

  • Problem

    HIN recommenders may rely on sparse or noisy path connections or summarize neighborhoods before prediction, while rich neighborhood interactions remain underused.

    Source: §1 · p. 2

  • Method

    NIRec captures interactions between metapath-guided neighborhoods using a convolutional formulation and FFT for efficient learning.

    Source: §1 · p. 2, §4.3 · pp. 4–5

  • Results

    NIRec outperforms state-of-the-art baselines on four public datasets.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper reports gains in recommendation effectiveness and interpretability, with learned node- and path-level importance indicating contributions of interactions and metapaths.

    Source: §4.6 · p. 6, §6 · p. 8

    NIRec currently captures interactive information in only one direction; extending interaction and aggregation to both sides and broader scenarios remains future work.

    Source: §6 · pp. 8–9

Abstract

from arXiv · show

There is an influx of heterogeneous information network (HIN) based recommender systems in recent years since HIN is capable of characterizing complex graphs and contains rich semantics. Although the existing approaches have achieved performance improvement, while practical, they still face the following problems. On one hand, most existing HIN-based methods rely on explicit path reachability to leverage path-based semantic relatedness between users and items, e.g., metapath-based similarities. These methods are hard to use and integrate since path connections are sparse or noisy, and are often of different lengths. On the other hand, other graph-based methods aim to learn effective heterogeneous network representations by compressing node together with its neighborhood information into single embedding before prediction. This weakly coupled manner in modeling overlooks the rich interactions among nodes, which introduces an early summarization issue. In this paper, we propose an end-to-end Neighborhood-based Interaction Model for Recommendation (NIRec) to address the above problems. Specifically, we first analyze the significance of learning interactions in HINs and then propose a novel formulation to capture the interactive patterns between each pair of nodes through their metapath-guided neighborhoods. Then, to explore complex interactions between metapaths and deal with the learning complexity on large-scale networks, we formulate interaction in a convolutional way and learn efficiently with fast Fourier transform. The extensive experiments on four different types of heterogeneous graphs demonstrate the performance gains of NIRec comparing with state-of-the-arts. To the best of our knowledge, this is the first work providing an efficient neighborhood-based interaction model in the HIN-based recommendations.

1 INTRODUCTION

HIN recommenders can lose neighborhood interactions through early summarization or struggle with sparse or noisy path connections. NIRec models interactions between metapath-guided neighborhoods and reports better performance than state-of-the-art baselines on four public datasets.

1 INTRODUCTION

  • Graph-based methods can compress a node and its neighborhood before prediction, while metapath-based methods may perform poorly when path connections are sparse or noisy.Neighborhood structure outside explicit metapaths may also be omitted.Source: §1 · pp. 1–2
  • The paper frames three challenges: capturing neighborhood interactions, aggregating information across paths and nodes, and learning efficiently on large-scale HINs.The motivating example uses co-movies or co-directors as interactions between users’ neighborhoods.Source: §1 · p. 2
  • NIRec uses metapath-guided neighborhoods, an interaction module, node- and path-level attention, and convolutional learning with FFT.The interaction module captures patterns between neighborhoods; the aggregation module weighs nodes and combines neighborhood content.Source: §1 · p. 2
  • NIRec outperforms state-of-the-art baselines on four public datasets.Source: §1 · p. 2

2 RELATED WORK

Related work covers heterogeneous-network methods for recommendation and graph representation learning. NIRec is positioned as an interaction model that uses information in neighborhoods.

2 RELATED WORK

  • Prior HIN recommendation work uses heterogeneous relations and metapath contexts across recommendation settings,,, [14], [24],.Source: §2 · p. 2
  • Many HIN recommenders rely on path-based similarity, which may not fully mine latent user and item features [22].NIRec instead introduces metapath-guided neighborhoods to capture interactions hidden in neighborhoods.Source: §2 · p. 3
  • Graph representation methods include unsupervised approaches that preserve graph structure [5] [17] [21] [28] and semi-supervised approaches using labeled vertices [11] [29].The paper distinguishes NIRec as an efficient neighborhood-based interaction method for heterogeneous graphs.Source: §2 · p. 3

3 PRELIMINARY

This section defines recommendation over HINs and introduces metapath-guided neighborhoods as semantic context for interactions between source and target neighbors.

3 PRELIMINARY

  • The recommendation task predicts an interaction, such as a click-through rate or link, between a source user and target item, using interactions between their neighbors.Source: §3 · p. 3
  • An HIN contains multiple node or link types, while a metapath is a relation sequence connecting two objects.The network schema represents the network’s object types and interaction relations.Source: §3 · p. 3
  • A metapath-guided neighborhood is the set of objects visited from an object along a given metapath, preserving heterogeneous semantic content.For example, a User-Movie-Director path yields user, movie, and director objects in the neighborhood.Source: §3 · p. 3
  • The setup motivates NIRec because existing HIN recommenders may overlook early summarization and rich interaction information.Source: §3 · p. 3

4 METHODOLOGY

NIRec uses metapath-guided neighborhoods to model interactions between source and target nodes before aggregating their information. Its pipeline samples neighbors, constructs interactions, attends to node- and path-level information, and predicts recommendations.

4.1 Overview

  • NIRec samples metapath-guided neighborhoods, constructs interactive information, aggregates it, and produces a final prediction.The framework uses an interaction layer followed by an aggregation layer.Source: §4.1 · p. 4, Figure 2 · p. 4

4.2 Neighborhood Sampling

  • Metapath-guided random walks generate node sequences that follow the metapath pattern, without requiring a complete source-to-target metapath sample.The walk repeats the metapath pattern until reaching a predefined length.Source: §4.2 · p. 4

4.3 Interaction Module

  • The model groups neighbors by distance from source and target nodes, interacting only corresponding groups; singleton groups use element-wise products, while larger groups use products followed by summation.The neighborhood can be divided into 2I − 1 groups, where I is metapath length.Source: §4.3 · pp. 4–5
  • NIRec formulates neighborhood interaction as convolution: repeated shifts, products, and sums capture both node similarity and co-ratings, with FFT reducing convolution complexity from O(I^2) to O(I log(I)).The operation combines interactions across aligned neighborhood nodes.Source: §4.3 · p. 5

4.4 Aggregation Module

  • Node-level attention selects important interactions within metapath neighborhoods, and multi-head attention aggregates semantic-specific interaction embeddings for each metapath.Attention weights express the importance of interactions relative to other elements in the neighborhood.Source: §4.4 · pp. 5–6
  • Path-level attention estimates each metapath’s task-specific contribution and uses the normalized weights to fuse semantic-specific embeddings into final embedding Z.The resulting aggregation contains interaction information at both node and path levels.Source: §4.4 · p. 6

4.5 Optimization Objective

  • A nonlinear projection such as an MLP derives the final prediction, trained with log loss over positive and negative instances.Source: §4.5 · p. 6

4.6 Model Analysis

  • NIRec’s complexity is linear in nodes and metapath-based node pairs, its node- and path-level attention can be parallelized, and learned attention offers interpretability of interactions and metapaths.The model analysis also describes potential interpretation through neighbor similarity and co-ratings.Source: §4.6 · p. 6

5 EXPERIMENTS

NIRec is evaluated for CTR prediction on four heterogeneous datasets against graph, recommendation, and ablation baselines. It outperforms all baselines across the datasets, with analyses examining component contributions, attention patterns, and metapath-neighborhood choices.

5 EXPERIMENTS

  • The experiments ask whether NIRec improves recommendation performance, whether its components matter, what patterns it captures, and how neighborhood choices affect results.These are the four stated research questions.Source: §5 · p. 6

5.1 Dataset Description

  • The evaluation uses four datasets—Movielens, LastFM, AMiner, and Amazon—with ratings treated as user–item interaction records.Table 2 summarizes dataset statistics and selected metapaths.Source: §5.1 · pp. 6–7

5.2 Compared Methods

  • The comparison includes six graph-embedding and recommendation baselines, alongside NIRecCNN and NIRecGCN component variants.The evaluation measures CTR prediction using AUC and ACC; the variants test interaction and aggregation components.Source: §5.2 · p. 7, §5.3 · p. 7

5.3 Result Analysis

  • NIRec consistently outperforms all baselines on four datasets in CTR prediction.Table 1 reports AUC and ACC; the caption defines asterisked improvements over the best baseline as statistically significant at p < 10^-6.Source: §5.3 · p. 7, Table 1 · p. 7
  • NIRec exceeds NIRecCNN and NIRecGCN, supporting the convolutional interaction strategy and attention-based aggregation over the tested alternatives.The authors attribute the NIRec–NIRecCNN difference to better interaction capture and the NIRec–NIRecGCN difference to better use of metapath-based interactive information.Source: §5.3 · pp. 7–8
  • For user u692, different user–movie pairs emphasize different metapaths, and node-level attention highlights neighborhood similarity and co-ratings.The u692–m805 pair relies mainly on UMUM and UMMM, while u692–m920 relies mainly on UMGM; the attention distributions are skewed.Source: §5.3 · p. 8, Figure 6 · p. 8
  • Adding metapaths generally improves NIRec, but their effects vary; adding UMGM boosts Movielens and adding UIVI boosts Amazon.The analysis evaluates metapaths incrementally against NeuMF.Source: §5.3 · p. 8
  • Performance rises and then declines as neighborhood length increases, with length 4 performing best in the reported analysis.The authors suggest longer neighborhoods can introduce noisy information that harms recommendation performance.Source: §5.3 · p. 8

6 CONCLUSION AND FUTURE WORK

NIRec addresses early summarization by modeling interactions between nodes through metapath-guided neighborhoods. The authors report effectiveness and interpretability gains, while noting that current interaction capture covers the graph side only.

6 CONCLUSION AND FUTURE WORK

  • NIRec is proposed to address the early summarization issue in HIN-based recommendation.The paper describes this as a neighborhood-based interaction-enhanced recommendation model.Source: §6 · p. 8
  • Metapath-guided neighborhoods preserve HIN heterogeneity, while interaction and node- and path-level attention capture pairwise interactions and fuse key interactions and metapaths.The interaction module captures similarity between each source–target pair through their neighborhoods.Source: §6 · p. 8
  • Experiments demonstrate NIRec’s superiority in recommendation effectiveness and interpretability.Source: §6 · p. 8
  • Current interaction capture is effective only on the graph side; future work would extend interaction and aggregation to capture information from both graph and node sides.Source: §6 · pp. 8–9

A.1 Pseudocode of NIRec Training Procedure

NIRec training takes a heterogeneous graph, node features, metapaths, and a source–target pair, then predicts their link or recommendation outcome. It generates metapath-guided neighborhoods, computes and attends over interactions, and predicts through an MLP.

A.1 Pseudocode of NIRec Training Procedure

  • The procedure takes a HIN, node features, a metapath set, and source and target nodes, and predicts their link; the task can also be stated as CTR prediction.Source: §A.1 · p. 10
  • For each metapath, NIRec generates source and target neighborhoods and computes their interaction result.Source: §A.1 · p. 10
  • Node-level attention fuses interaction results into metapath embeddings, and path-level attention combines those embeddings into a final representation.Source: §A.1 · p. 10
  • An MLP produces the final prediction, after which the procedure calculates loss and backpropagates.Source: §A.1 · p. 10

A.2 Data Preproccess

Ratings are converted into binary feedback before the data are divided into training, validation, and test sets.

A.2 Data Preproccess

  • Ratings of 4–5 stars are converted to positive feedback labeled 1.Source: §A.2 · p. 10
  • Ratings below 4 stars are converted to negative feedback labeled 0.Source: §A.2 · p. 10
  • Each processed dataset is split into training, validation, and test sets at a 6:2:2 ratio.Source: §A.2 · p. 10

A.3 Model Variants Configuration

The configurations compare NIRec variants that replace its interaction or aggregation design. The supplied table caption identifies dataset statistics and selected metapaths.

A.3 Model Variants Configuration

  • NIRecCNN replaces the interaction module with direct CNN fusion of source and target neighborhoods while keeping NIRec’s aggregation module.The variant is intended to show the performance gain from NIRec’s interaction module.Source: §A.3 · p. 10
  • NIRecGCN replaces NIRec’s attention mechanism with GCN layers for node- and path-level features while retaining the interaction module.The variant is intended to show the performance gain from NIRec’s aggregation module.Source: §A.3 · p. 10
  • Table 2 reports statistics for four datasets and identifies the selected metapaths for each.Source: Table 2 · p. 10

A.4 Implementation Details

NIRec uses fixed embedding and metapath-guided neighborhood settings, while the comparison setup standardizes path sampling and baseline inputs. The neighborhood length follows the metapath length.

A.4 Implementation Details

  • NIRec uses 128-dimensional embeddings, samples 20 neighbors per node under the current metapath, and sets neighborhood length to metapath length.The paper also investigates different neighborhood lengths, including n-hop neighborhoods, in Section 5.3.Source: §A.4 · p. 10
  • The unified sampling framework stores metapath-guided paths and neighborhoods from the same source node, but paths end at the target while neighborhoods may end at any node of the target type.For the example, the neighborhoods may end at any movie node.Source: §A.4 · p. 10
  • Metapath-based baselines receive the same paths, GNN-based baselines receive n-hop nodes generated from sampled paths, and all baseline embeddings are 128-dimensional.Source: §A.4 · p. 10
Loading 2007.00216v1…