PaperlayerView exact revision on arXiv

Source-linked AI summary

Joint Neural Collaborative Filtering for Recommender Systems

Wanyu Chen, Fei Cai, Honghui Chen, Maarten de Rijke

arXiv:1907.03459v2cs.IR

TL;DR

Traditional collaborative filtering may miss complex user-item interactions when it predicts preferences with linear dot products. J-NCF jointly trains neural feature learning and interaction modeling; across MovieLens and Amazon Movies, it outperforms state-of-the-art baselines in prediction accuracy.

  • Problem

    Traditional collaborative filtering often predicts preferences with linear dot products of user and item factors, which may not capture complex interaction structure.

    Source: §1 · p. 2

  • Method

    J-NCF jointly trains neural feature learning from rating information and neural modeling of user-item interactions using the learned feature vectors.

    Source: §1 · p. 2, §7 · p. 27

  • Results

    J-NCF outperforms state-of-the-art baselines in prediction accuracy across MovieLens and Amazon Movies; HR@10 improves by up to 8.24% on MovieLens 100K.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Joint training lets feature learning and interaction modeling optimize each other, improving recommendation performance.

    Source: §1 · p. 2

    J-NCF's future work includes adding auxiliary information such as item content and reviews, session context, and attention mechanisms.

    Source: §7 · p. 28

Abstract

from arXiv · show

We propose a J-NCF method for recommender systems. The J-NCF model applies a joint neural network that couples deep feature learning and deep interaction modeling with a rating matrix. Deep feature learning extracts feature representations of users and items with a deep learning architecture based on a user-item rating matrix. Deep interaction modeling captures non-linear user-item interactions with a deep neural network using the feature representations generated by the deep feature learning process as input. J-NCF enables the deep feature learning and deep interaction modeling processes to optimize each other through joint training, which leads to improved recommendation performance. In addition, we design a new loss function for optimization, which takes both implicit and explicit feedback, point-wise and pair-wise loss into account. Experiments on several real-word datasets show significant improvements of J-NCF over state-of-the-art methods, with improvements of up to 8.24% on the MovieLens 100K dataset, 10.81% on the MovieLens 1M dataset, and 10.21% on the Amazon Movies dataset in terms of HR@10. NDCG@10 improvements are 12.42%, 14.24% and 15.06%, respectively. We also conduct experiments to evaluate the scalability and sensitivity of J-NCF. Our experiments show that the J-NCF model has a competitive recommendation performance with inactive users and different degrees of data sparsity when compared to state-of-the-art baselines.

1 INTRODUCTION

J-NCF addresses limits of linear interaction models by jointly learning user and item features and their interactions from rating data. The paper reports stronger recommendation performance than state-of-the-art baselines and competitive performance across inactive users and data sparsity levels.

1 INTRODUCTION

  • The introduction identifies linear dot-product interaction models as a limitation because they may not capture complex user-item relationships well.It situates J-NCF among deep-learning approaches that often use auxiliary information or retain matrix-factorization approaches for rating matrices.Source: §1 · p. 2
  • The model jointly trains deep feature learning from rating information and nonlinear user-item interaction modeling, allowing the processes to optimize each other.The interaction network uses the feature representations produced by the feature-learning network.Source: §1 · p. 2
  • Its loss combines implicit and explicit feedback with point-wise and pair-wise loss information.The cited prior neural approaches CDAE, NCF, and DMF use point-wise loss functions and leave pair-wise loss for future work.Source: §1 · pp. 2–3
  • HR@10 improvements reach 8.24% on MovieLens 100K, 10.81% on MovieLens 1M, and 10.21% on Amazon Movies over the best baseline.NDCG@10 improvements are 12.42%, 14.24%, and 15.06% on those datasets, respectively.Source: §1 · p. 3
  • J-NCF also achieves competitive recommendation performance with inactive users and different degrees of data sparsity.The paper investigates scalability and sensitivity alongside these conditions.Source: §1 · p. 3

2 RELATED WORK

Related work moves from traditional collaborative filtering to deep-learning recommender models, distinguishing approaches by how they represent users and items and model their interactions. J-NCF is positioned as a rating-matrix-based model that jointly learns representations and nonlinear interactions.

2.1 Traditional recommender systems

  • Traditional top-N systems include neighborhood methods based on user or item similarity and latent-factor models that represent users and items as vectors.The paper focuses on latent-factor modeling rather than neighborhood-based approaches.Source: §2.1 · pp. 3–4
  • SVD-based methods factor the rating matrix, but missing ratings complicate conventional SVD and filling them can create a computationally infeasible dense matrix.Later methods learn factors from known ratings with regularized prediction-error objectives; SVD++ incorporates explicit and implicit feedback.Source: §2.1 · p. 4
  • Traditional inner-product models use a linear kernel, which may describe user characteristics and interactions inaccurately compared with nonlinear approaches.Source: §2.1 · p. 4

2.2 Deep learning-based recommender system

  • Deep-learning recommenders include models using auxiliary content and models applied to rating matrices; RBM, AutoRec, and CDAE have differing feedback and recommendation-task limitations.RBM targets rating prediction and uses observed ratings, while CDAE uses implicit feedback and an item-item personalization model.Source: §2.2 · pp. 4–5
  • NCF models nonlinear user-item interactions with multilayer perceptrons, but initializes representations from identifiers and uses implicit feedback only.J-NCF instead learns user and item features jointly and combines implicit and explicit feedback.Source: §2.2 · p. 5
  • J-NCF differs from DMF’s inner-product interaction and content-based integration models by using rating information to learn features and multilayer perceptrons to model interactions.Its end-to-end network jointly optimizes feature learning and interaction modeling, with a loss combining point-wise and pair-wise losses and both feedback types.Source: §2.2 · pp. 5–6

3 APPROACH

J-NCF frames top-N recommendation as returning user-specific item lists and jointly learns user and item features with their interactions from rating information. Its objective combines point-wise and pair-wise losses to use explicit and implicit feedback.

3.1 Problem formulation and notation

  • Top-N recommendation returns a list of items for an individual user to maximize that user's satisfaction.The paper denotes the user-item rating information by R.Source: §3.1 · p. 6

3.2 Joint Neural Collaborative Filtering

  • The DI network takes feature representations from the DF network and transforms user-item interactions through hidden layers into predicted scores.Intermediate layers provide a multi-layer non-linear projection of the interactions.Source: §3.2 · pp. 6, 9
  • The DF network maps user and item rating vectors through parallel MLPs into deep feature representations.The user and item networks are coupled in their last layer.Source: §3.2 · pp. 6–8
  • J-NCF fuses user and item vectors either by concatenation for non-linear fusion or element-wise product for a linear kernel.These alternatives form two versions of J-NCF.Source: §3.2 · p. 9

3.3 Loss function

  • J-NCF combines point-wise and pair-wise losses, using log loss with implicit and explicit feedback and testing three pair-wise variants.The variants are point-wise+TOP1, point-wise+BPR-max, and point-wise+TOP1-max; user ratings are scaled by that user's maximum rating so different ratings influence the loss differently.Source: §3.3 · pp. 10–11

4 EXPERIMENTAL SETUP

The experiments evaluate J-NCF against collaborative-filtering baselines across several datasets and settings, using top-N accuracy metrics and a leave-one-out evaluation protocol. The setup also varies model choices, user activity, and dataset sparsity.

4.1 Model summary and research questions

  • The study tests nine research questions spanning baseline performance, loss and network choices, feedback types, user activity, sparsity, large datasets, and runtime.Comparisons include traditional collaborative-filtering methods and deep-learning baselines; J-NCFm combines user and item vectors element-wise, while J-NCFc concatenates them.Source: §4.1 · p. 11
  • The algorithm forms positive interaction pairs from the rating matrix, samples negatives, computes user and item representations and interactions, then updates parameters by backpropagation.The pseudocode returns the user, item, and interaction-network weights and biases.Source: §4.1 · p. 12

4.2.1 Datesets.

  • The core evaluation uses MovieLens 100K, MovieLens 1M, and Amazon Movies; Amazon Movies retains users with at least 20 interactions and items with at least 5.Amazon Electronics, with 7,824,482 ratings, is included for large, sparse dataset tests; it retains users with at least 2 interactions and items with at least 5.Source: §4.2.1 · p. 13
  • For sparsity tests, each dataset is subsampled and ratings are iteratively removed while keeping the user and item counts unchanged, producing successively sparser versions.Dataset statistics for the sparsity variants are summarized in Table 4.Source: §4.2.1 · p. 14, Table 4 · p. 15

4.2.2 Experimental setup.

  • Leave-one-out testing trains on all but each user's latest interaction and ranks that test item among 100 randomly sampled unseen items.The authors note that this sampling may overestimate all algorithms' performance, while comparisons remain fair.Source: §4.2.2 · p. 14
  • HR and NDCG assess top-N accuracy, while hyperparameters are selected through 100 random search trials on a validation set.Unless otherwise specified, N is 10; neural models use mini-batch Adam, batch size 256, and learning rate 0.0001.Source: §4.2.2 · p. 15

5 RESULTS AND DISCUSSION

Across the evaluated settings, J-NCF variants generally outperform the baselines, while results also vary with loss choice, network depth, and feedback type. The experiments examine overall performance and these design factors across recommendation metrics and datasets.

5.1 Overall performance

  • J-NCF variants beat DMF by up to 8.24% in HR@10 and 12.42% in NDCG@10 on ML100K; the table reports model performance and significance tests.The largest gains are among the reported ranges for the three datasets.Source: Table 5 · p. 16, §5.1 · p. 16

5.2 Impact of different loss functions

  • List-wise J-NCFXE performs worst; point-wise loss favors HR@10, while pair-wise loss favors NDCG@10, with loss-model trends differing somewhat on AMovies.On AMovies, J-NCFBMP is slightly better on HR@10, while its NDCG@10 is similar to J-NCFTP and better than J-NCFTMP.Source: §5.2 · p. 17, §5.3 · p. 18

5.3 Utility of hybrid loss function

  • On ML100K, hybrid loss beats point-only and pair-only J-NCF by 2.68% and 7.61% in HR@10, and 3.99% and 2.36% in NDCG@10, respectively.Across the recommendation-list sizes examined, hybrid J-NCF also consistently improves over DMF and the single-loss models.Source: §5.3 · pp. 17, 19

5.4 Number of of layers in the networks

  • On ML100K, adding a DF layer improves HR@10 and NDCG@10 by 2.82% and 4.31%, versus 1.05% and 2.62% when adding a DI layer.The best results on each dataset use DF-5, DI-4; performance no longer increases beyond four DI layers, while deeper DF networks continue to help.Source: Table 6 · p. 20, §5.4 · p. 21
  • The best J-NCF performance improves over the shallowest configuration by 20.52%, 25.37%, and 34.52% in HR@10, and 28.96%, 63.05%, and 53.37% in NDCG@10 across the three datasets.Deeper layers yield more obvious improvements in NDCG@10 than HR@10.Source: §5.4 · p. 21

5.5 Impact of feedback

  • Compared with implicit-only J-NCF, J-NCFex improves HR@10 by 3.72%, 5.22%, and 4.89%, and NDCG@10 by 4.61%, 5.58%, and 5.31% across ML100K, ML1M, and AMovies.J-NCFex uses both feedback types in its input and loss; J-NCF converges to its best performance faster than DMF and NCF, while additional iterations can degrade performance after a peak.Source: §5.5 · pp. 21, 23

6 SCALABILITY AND SENSITIVITY

The experiments assess J-NCF across user activity levels, datasets with different sparsity, a large sparse dataset, and training and inference time. J-NCF outperforms DMF across the reported activity and sparsity settings, while timing comparisons distinguish training cost from prediction time.

6.1 Model scalability with user ratings

  • J-NCFc outperforms DMF across user activity levels and is best on HR@10 and NDCG@10; J-NCF gains are larger for inactive than very active users.On ML100K, gains fall from 11.08% to 7.85% in HR@10 and from 9.57% to 7.32% in NDCG@10 when comparing the 50% and 90% user groups.Source: §6.1 · pp. 23–24

6.2 Sensitivity to data sparsity

  • J-NCF beats DMF across sparsity levels; on smaller ML100K subsets gains are larger on sparser data, while on larger ML1M and AMovies subsets they are larger on denser data.For ML100K-3, J-NCFc gains 7.77% HR@10 and 12.02% NDCG@10 over DMF; on ML1M, its HR@10 gain is 11.13% on the denser ML1M-1 subset versus 6.53% on ML1M-3.Source: §6.2 · pp. 24–25

6.3 Performance with a large and sparse dataset

  • On the large, sparse AEle dataset, J-NCF outperforms DMF and NCF across top-N list sizes; at N = 10, gains over DMF are 4.62% in HR and 5.82% in NDCG.At N = 5, the corresponding gains are 5.88% in HR and 6.12% in NDCG.Source: §6.2 · p. 26, §6.3 · p. 26

6.4 Training and inference time

  • J-NCF takes slightly longer per training epoch than DMF but less total time to reach best performance; all three models take roughly a few milliseconds per user for top-10 ranking.NCF has the lowest training and prediction time overall.Source: §6.4 · pp. 26–27, Table 8 · p. 27

7 CONCLUSIONS AND FUTURE WORK

J-NCF jointly trains deep user/item feature learning and interaction modeling, using a hybrid loss for top-N recommendation. The reported results include stronger performance across user activity levels and dataset sparsity settings, while future work includes addressing its higher computational cost than NCF.

7 CONCLUSIONS AND FUTURE WORK

  • J-NCF jointly trains deep feature learning for users and items with neural modeling of their interactions, using explicit and implicit feedback as input.Its top-N recommendation loss incorporates both pair-wise and point-wise loss.Source: §7 · p. 27
  • Experiments support the hybrid loss, deeper networks, and combined implicit and explicit feedback as effective settings for J-NCF.Source: §7 · p. 27
  • J-NCF outperforms the best baseline across user activity levels and shows competitive performance across sparsity settings, including on the large, sparse AEle dataset.The authors report particular strength for inactive users, who constitute the majority of users in the datasets.Source: §7 · pp. 27–28
  • J-NCF is computationally more expensive than NCF, and planned work includes improving its efficiency and adding auxiliary, contextual, and attention-based information.The authors also propose exploring recurrent neural networks for session context.Source: §7 · p. 28
Loading 1907.03459v2…