PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-view Intent Disentangle Graph Networks for Bundle Recommendation

Sen Zhao, Wei Wei, Ding Zou, Xianling Mao

arXiv:2202.11425v3cs.IRcs.AI

TL;DR

Bundle recommendation methods can overlook users’ diverse intents and item associations within bundles. MIDGN disentangles intents across global and local views and contrasts those representations; experiments report gains over state-of-the-art methods on two benchmark datasets. The paper assumes each user has K intents and slices user and bundle embeddings into K chunks.

Abstract

from arXiv · show

Bundle recommendation aims to recommend the user a bundle of items as a whole. Nevertheless, they usually neglect the diversity of the user's intents on adopting items and fail to disentangle the user's intents in representations. In the real scenario of bundle recommendation, a user's intent may be naturally distributed in the different bundles of that user (Global view), while a bundle may contain multiple intents of a user (Local view). Each view has its advantages for intent disentangling: 1) From the global view, more items are involved to present each intent, which can demonstrate the user's preference under each intent more clearly. 2) From the local view, it can reveal the association among items under each intent since items within the same bundle are highly correlated to each other. To this end, we propose a novel model named Multi-view Intent Disentangle Graph Networks (MIDGN), which is capable of precisely and comprehensively capturing the diversity of the user's intent and items' associations at the finer granularity. Specifically, MIDGN disentangles the user's intents from two different perspectives, respectively: 1) In the global level, MIDGN disentangles the user's intent coupled with inter-bundle items; 2) In the Local level, MIDGN disentangles the user's intent coupled with items within each bundle. Meanwhile, we compare the user's intents disentangled from different views under the contrast learning framework to improve the learned intents. Extensive experiments conducted on two benchmark datasets demonstrate that MIDGN outperforms the state-of-the-art methods by over 10.7% and 26.8%, respectively.

Introduction

Bundle recommendation predicts bundles as a whole, but existing methods can miss users’ diverse intents and item associations. MIDGN disentangles intents from global and local views and outperforms state-of-the-art methods by over 10%.

Introduction

  • Bundle recommendation recommends bundles of items that a user is interested in purchasing together.Unlike conventional recommendation of individual items, bundling can expose items seldom purchased in isolation.Source: Introduction · p. 1
  • Existing bundle recommendation methods can neglect users’ intent diversity and the associations among items within bundles.The global view captures intents distributed across a user’s bundles; the local view captures item associations within each bundle.Source: Introduction · pp. 1–2
  • MIDGN outperforms state-of-the-art methods by over 10% on NetEase and Youshu.It models and contrasts user intents under global and local views.Source: Introduction · p. 2
  • MIDGN disentangles user and bundle representations into intent-specific chunks using neighbor routing on user-item and bundle-item graphs.A contrastive module compares intents from the two views to enhance learned intent representations.Source: Introduction · p. 2

Related Work

Prior bundle recommendation work ranges from item-set construction and basket completion to graph-based methods. The reviewed approaches leave gaps in personalized preference modeling, item associations, or intent disentanglement.

Related Work

  • Early approaches mined common item groups or formulated bundle selection as a knapsack problem, often without modeling item-to-item dependency.Other work treated cross-item dependencies as hard constraints, while the review notes difficulty capturing personalized preference.Source: Related Work · p. 2
  • Intra-basket methods model item correlations or pairwise associations, while inter-basket methods rank bundles for recommendation.The review says later methods still miss multi-hop connectivity or user preferences over within-bundle item associations.Source: Related Work · p. 2
  • Graph convolution methods encode collaborative or user-bundle-item graph structure, but may neglect the diversity of users’ intents.The cited bundle graph work learns user and bundle representations with item-level semantics.Source: Related Work · pp. 2–3
  • Intent-disentangled representations have been studied for item recommendation, but bundle recommendation distributes intents both across and within bundles.The review characterizes bundle intent disentanglement as more complicated and in need of further methods.Source: Related Work · p. 3

Preliminaries

The paper represents users, bundles, and items as sets linked by interaction and affiliation matrices. Bundle recommendation is framed as predicting a user’s probability of interacting with a previously unseen bundle.

Preliminaries

  • U, B, and I denote the user, bundle, and item sets, with M, O, and N their respective sizes.Source: Preliminaries · p. 3
  • Y, H, and R encode user-bundle interactions, bundle-item affiliations, and user-item interactions, respectively.An entry of 1 indicates that the corresponding interaction or item inclusion occurred.Source: Preliminaries · p. 3
  • The task is to predict the probability that a user will interact with a previously unseen bundle.Source: Preliminaries · p. 3

The Proposed Model

MIDGN combines intent disentangling across global and local views with cross-view propagation, intent contrast, and bundle prediction. Its modules represent user and bundle interactions at intent-level granularity and exchange information between the views.

The Proposed Model

  • MIDGN combines graph disentangling, cross-view propagation, intent contrasting, and prediction to model user and bundle interactions across global and local views.The model predicts ratings from the learned user and bundle embeddings.Source: The Proposed Model · p. 3, Prediction and Optimization · p. 5

Graph Disentangling Module

  • User and bundle embeddings are divided into K intent chunks, and neighbor routing refines their intent-aware interaction graphs and representations.The user-item graph captures intents distributed across bundles, while the bundle-item graph captures multiple intents within each bundle.Source: Graph Disentangling Module · pp. 3–4

Cross-View Propagating Module

  • LightGCN propagates signals over the user-bundle graph to exchange intent information between representations learned from the two views.After this exchange, user and bundle chunks represent intents from the local and global views, respectively.Source: Cross-View Propagating Module · p. 4

Intent Contrasting Module

  • InfoNCE contrasts user and bundle chunks across views, pairing same-intent chunks as positives and treating other chunks as negatives.The paper presents this contrast as a way for the views to complement one another in representing user preferences and item associations.Source: Intent Contrasting Module · p. 5

Prediction and Optimization

  • The model concatenates representations from both views and uses their inner product to estimate interaction likelihood, optimizing with BPR and alternating prediction and contrast losses.The training data pairs observed interactions with unobserved ones, and the objective includes L2 regularization.Source: Prediction and Optimization · p. 5

Experiments

Experiments compare MIDGN with prior methods on two public datasets and examine its components and parameters. MIDGN leads the baselines on both datasets, while ablations and parameter studies probe the roles of its views, contrast module, layer depth, and intent count.

Experiments

  • The evaluation uses NetEase music playlists and songs and Youshu book bundles, with Recall@K and NDCG@K as ranking metrics.NetEase includes bundles with at least 10 songs appearing in at least 5 bundles; its users consume at least 10 songs and 10 playlists.Source: Experiments · p. 5
  • Comparisons include matrix-factorization and graph-based recommendation baselines, including MFBPR, DAM, RGCN, BasConv, and BGCN.The experiment uses a shared BPR loss and Adam optimizer across methods.Source: Experiments · pp. 5–6

Performance Comparison(RQ1)

Study of MIDGN (RQ2&RQ3)

  • Removing the global-view module, local-view module, or intent contrast module lowers performance; all three ablated models still significantly outperform the baselines.The variants remove contrast, replace local disentangling with GCN, or replace global disentangling with GCN.Source: Study of MIDGN (RQ2&RQ3) · p. 6
  • Increasing layer depth can help: MIDGN-2 highly outperforms MIDGN-1, but MIDGN-3 outperforms MIDGN-4 on Youshu, where added-hop noise may matter.MIDGN-1 uses only one-hop information and misses higher-order collaborative information.Source: Study of MIDGN (RQ2&RQ3) · p. 7, Table 4 · p. 7, Figure 3 · p. 7
  • MIDGN performs worst at K = 1, while increasing K from 4 to 8 sharply reduces performance, suggesting too few or too-fine-grained intents are inadequate.The authors attribute the latter drop to item groups being too small to represent their associated intents well.Source: Study of MIDGN (RQ2&RQ3) · p. 7, Table 5 · p. 7

Conclusion

MIDGN disentangles user intents through global and local views, using neighbor-routing GNNs and cross-view comparison; experiments on Yoush and NetEase demonstrate its superiority.

  • Experiments on Yoush and NetEase demonstrate the superiority of MIDGN.Source: Conclusion · p. 7
  • MIDGN uses neighbor-routing GNNs to disentangle user-item and bundle-item graphs under global and local views.Source: Conclusion · p. 7
  • MIDGN compares intents learned from different views to better represent user preferences and item associations at finer intent granularity.Source: Conclusion · p. 7

Related papers

Loading 2202.11425v3…