PaperlayerView exact revision on arXiv

Source-linked AI summary

CDRNP: Cross-Domain Recommendation to Cold-Start Users via Neural Process

Xiaodong Li, Jiawei Sheng, Jiangxia Cao, Wenyuan Zhang, Quangang Li, Tingwen Liu

arXiv:2401.12732v3cs.IRcs.SI

TL;DR

Cold-start recommendation in a target domain is challenging when users have few interactions, and existing cross-domain mapping approaches leave gaps in modeling user-specific preferences and preference correlations. CDRNP combines meta-learning with a Neural Process to model both, and the paper reports gains over state-of-the-art methods across three CDR scenarios.

  • Problem

    Cross-domain recommendation seeks to support target-domain recommendations for cold-start users using source-domain interactions, while existing mapping approaches neglect user-specific preference, user correlations, or explicit interactions from both domains.

    Source: §1 · pp. 1–2

  • Method

    CDRNP combines meta-learning and a Neural Process to model user-specific preferences and common preference correlations, with a preference remainer and adaptive conditional decoder for target-domain prediction.

    Source: §1 · p. 2

  • Results

    30.68%, 11.27%, and 39.34% MAE improvements over PTUPCDR at α=20% across three CDR scenarios; CDRNP is reported to consistently outperform state-of-the-art methods.

    Source: §4.3.1 · p. 6, §6 · p. 8

  • Takeaways & Limitations

    The reported results support CDRNP for cold-start recommendation across the three evaluated CDR scenarios.

    Source: §6 · p. 8

    The model assumes a standard normal prior for its latent variable for simplicity in the variational framework.

    Source: §2.2 · p. 3

Abstract

from arXiv · show

Cross-domain recommendation (CDR) has been proven as a promising way to tackle the user cold-start problem, which aims to make recommendations for users in the target domain by transferring the user preference derived from the source domain. Traditional CDR studies follow the embedding and mapping (EMCDR) paradigm, which transfers user representations from the source to target domain by learning a user-shared mapping function, neglecting the user-specific preference. Recent CDR studies attempt to learn user-specific mapping functions in meta-learning paradigm, which regards each user's CDR as an individual task, but neglects the preference correlations among users, limiting the beneficial information for user representations. Moreover, both of the paradigms neglect the explicit user-item interactions from both domains during the mapping process. To address the above issues, this paper proposes a novel CDR framework with neural process (NP), termed as CDRNP. Particularly, it develops the meta-learning paradigm to leverage user-specific preference, and further introduces a stochastic process by NP to capture the preference correlations among the overlapping and cold-start users, thus generating more powerful mapping functions by mapping the user-specific preference and common preference correlations to a predictive probability distribution. In addition, we also introduce a preference remainer to enhance the common preference from the overlapping users, and finally devises an adaptive conditional decoder with preference modulation to make prediction for cold-start users with items in the target domain. Experimental results demonstrate that CDRNP outperforms previous SOTA methods in three real-world CDR scenarios.

1 INTRODUCTION

The introduction frames cross-domain recommendation as a way to address cold-start users, while identifying limits in shared and user-specific mapping approaches. CDRNP combines meta-learning and Neural Processes to model user preferences and their correlations, and is reported to outperform state-of-the-art methods in three scenarios.

1 INTRODUCTION

  • Cross-domain recommendation uses source-domain interactions to support target-domain recommendations for users with few interactions.Cold-start users may lack enough user-item interactions to learn effective representations.Source: §1 · p. 1
  • Shared mappings overlook user-specific preferences, while user-specific meta-learning overlooks correlations among users and explicit interactions from both domains.The introduction describes both approaches as having limitations in preference modeling.Source: §1 · p. 2
  • CDRNP combines meta-learning with Neural Processes to model user-specific preferences and preference correlations among overlapping and cold-start users.Its characteristic embedding uses source-domain interactions, and its preference aggregator models users within a stochastic process.Source: §1 · p. 2
  • Experiments across three CDR scenarios show that CDRNP outperforms state-of-the-art methods.The framework also includes a preference remainer and an adaptive conditional decoder for target-domain rating prediction.Source: §1 · p. 2

2 PRELIMINARY

The preliminary section introduces the CDR setting and the Neural Process principle. It formulates target-domain rating prediction for cold-start users from source-domain interactions, then describes a stochastic-process model that predicts from support-set information and captures preference correlations across users.

2 PRELIMINARY

  • The preliminary section introduces the CDR problem setting and Neural Process background used in the paper.The following subsections specify the task and explain the Neural Process principle.Source: §2 · p. 2

2.1 Problem Setting

  • For each cold-start source-domain user, the task is to predict ratings for candidate target-domain items using that user's source-domain interactions.The problem setting distinguishes overlapping users, present in both domains, from cold-start users, present only in the source domain.Source: §2.1 · p. 2

2.2 Neural Process Principle

  • The model treats each meta-learning task as samples from a stochastic process and approximates that process with a neural latent-variable model.The variational posterior is learned with amortized variational inference; the prior over the latent variable is assumed to be standard normal for simplicity.Source: §2.2 · p. 3
  • At test time, the support-set latent representation is used to predict cold-start query users, encouraging user tasks to follow a shared process.Training aligns the latent distributions inferred from support and query sets, which the paper says captures preference correlations among overlapping and cold-start users.Source: §2.2 · p. 3

3 MODEL

CDRNP combines characteristic embeddings, a preference encoder, and an adaptive conditional decoder to predict target-domain ratings for cold-start users. Its encoder represents both user-specific preferences and common preference correlations.

3 MODEL

  • CDRNP uses a characteristic embedding layer, preference encoder, and adaptive conditional decoder to predict target-domain item ratings for cold-start users.The encoder includes a preference aggregator and a preference remainer.Source: §3 · pp. 3–4

3.1 Characteristic Embedding Layer

  • Source-domain interactions are attended to form user-specific preferences, which are fused with user features and a candidate target item into rating-observed information.The resulting representation combines the source-domain user information with the target-domain candidate item.Source: §3.1 · p. 4

3.2 Preference Encoder

  • The preference encoder uses support and query sets to approximate a stochastic process and capture preference correlations among users.It comprises a preference aggregator and a preference remainer.Source: §3.2 · p. 4, §3.2.1 · p. 4

3.2.1 Preference Aggregator.

  • The preference aggregator forms a permutation-invariant task representation from user representations using a commutative operation, with time complexity O(n + m).The representation supports the stochastic-process approximation for the support and query sets.Source: §3.2.1 · pp. 4–5

3.2.2 Preference Remainer.

  • The preference remainer fuses support-set rating-observed information with target-domain ratings, then aggregates user representations into a common preference representation.This retains original overlapping-user preference features from the support set.Source: §3.2.2 · p. 5

3.3 Adaptive Conditional Decoder

  • The adaptive conditional decoder uses the preference aggregator and remainer features to modulate decoding across layers and predict target-domain ratings.Its inputs include the candidate-item representation and sampled preference representation; the remainer representation generates modulation coefficients.Source: §3.3 · p. 5

3.4 Loss Function

  • Training combines a rating-prediction likelihood loss with KL-divergence regularization, while testing predicts held-out query ratings using support-set information and learned global information.At test time, query ground-truth ratings are removed and the decoder uses a latent variable sampled from the support-set distribution.Source: §3.4 · p. 5

4 EXPERIMENTS

The experiments evaluate CDRNP on Amazon review data across three cross-domain recommendation scenarios. They use all available data to simulate real-world settings.

4 EXPERIMENTS

  • The experiments compare the proposed model with previous models to evaluate CDRNP.Table 1 summarizes statistics for the three scenarios, with overlap denoting the number of overlapping users.Source: §4 · p. 5, Table 1 · p. 6

4.1 Datasets

  • The experiments use Amazon-5scores data, where each user and item has at least five ratings.The evaluation follows existing cross-domain recommendation methods.Source: §4.1 · p. 6
  • The evaluation covers Movie→Music, Book→Movie, and Book→Music scenarios using all available data rather than a subset.The categories are movies_and_tv, cds_and_vinyl, and books.Source: §4.1 · p. 6

4.2 Experimental Setting

The experiments evaluate recommendation with MAE and RMSE, compare CDRNP with target-only, cross-domain, and meta-learning baselines, and use shared tuning settings.

4.2.1 Evaluation Metrics.

  • MAE and RMSE are averaged over five random runs, with predictions limited to candidate items in each user’s query set.Source: §4.2.1 · p. 6

4.2.2 Compared Baselines.

  • Baselines span target-only MF (TGT), source-feature CMF, mapping-based EMCDR, DCDCSR [38], SSCDR, and personalized meta-learning PTUPCDR.Source: §4.2.2 · p. 6

4.2.3 Implementation and Hyperparameter Setting.

  • All methods use the same hyperparameter settings; embedding dimensions, learning rates, hidden sizes, and λ are selected from specified candidate sets.The test-user proportion α is set to 20%, 50%, or 80%, and sequential timestamps are used to avoid information leakage.Source: §4.2.3 · p. 6

4.3 Performance Comparisons on CDR

CDRNP is compared with existing methods across three CDR scenarios and in bidirectional cold-start recommendation. The reported results favor CDRNP across these evaluations.

4.3 Performance Comparisons on CDR

  • 30.68%, 11.27%, and 39.34% lower MAE than PTUPCDR are reported for CDRNP across three CDR scenarios at α=20%, where CDRNP performs best overall.The comparison uses support-set size 40 and source interaction-history length 20.Source: §4.3.1 · p. 6, Table 2 · p. 7

4.3.1 Performance Comparisons.

  • CDRNP achieves the best bidirectional cold-start recommendation performance in both source and target domains.Source: §4.3.2 · p. 6, Table 3 · p. 7

4.3.2 Bi-directional Recommendation Performance.

  • Training CDRNP only on Music-domain→Movie-domain yields outstanding bidirectional cold-start performance, and it also performs outstandingly for target-domain cold-start users versus EMCDR and PTUPCDR.Source: §4.3.2 · p. 7

4.4 Ablation Study

The ablation study compares CDRNP with versions lacking its preference remainer and attentive characteristic preference modules, and reports the importance of preference correlations across layer depths.

  • CDRNP significantly improves over CDRNP (w/o PRM), supporting the preference remainer’s contribution.Source: §4.4 · p. 7
  • CDRNP (w/o PRM) significantly outperforms CDRNP (w/o PRM+ACP), indicating the value of ACP for using source-domain auxiliary information.Source: §4.4 · p. 7
  • The results demonstrate the importance of capturing preference correlations among users at different layer depths.Source: §4.4 · p. 7

4.5 Impact of |V𝑠(𝑢𝑗)| and |C𝑖|

The experiments vary source interaction-history length and support-set size separately across three CDR scenarios, assessing CDRNP’s performance under reduced information.

  • The study changes history length and support-set size separately in three CDR scenarios, using α=20%.Source: §4.5 · p. 7

4.5.1 Impact of |V𝑠(𝑢𝑗)|.

  • With reduced source interaction histories, CDRNP outperforms CDRNP (w/o PRM) and PTUPCDR and is less affected by history length.The evaluated history lengths are 10, 30, 50, 70, and 90 items.Source: §4.5 · p. 7

4.5.2 Impact of |C𝑖|.

  • CDRNP and CDRNP (w/o PRM) remain robust as support-set users decrease, while CDRNP outperforms the ablated model at every tested size.The tested support-set sizes are 20, 30, 40, 50, and 60; the authors attribute stability partly to random function realizations.Source: §4.5 · p. 8

4.6 Hyper-parameter Analysis

CDRNP remains ahead of other baselines even at its worst tested λ, and its performance is stable as λ changes. The preferred λ varies with the interaction scenario.

4.6 Hyper-parameter Analysis

  • Even at its worst λ, CDRNP outperforms other baselines, and its results remain stable as λ changes.The analysis evaluates three CDR scenarios with α=20%.Source: §4.6 · p. 8
  • Larger interaction scenarios such as Book-domain→Movie-domain may benefit from a larger λ; smaller λ works well in the other two scenarios.The authors suggest the larger scenario has more complex preference correlations to capture.Source: §4.6 · p. 8
  • The sensitivity analysis varies λ in Eq. (18) and compares CDRNP with CDRNP(w/o PRM) across three scenarios at α=20%.Source: §4.6 · p. 8

5 RELATED WORKS

Related work spans cross-domain methods for cold-start recommendation, meta-learning approaches, and Neural Process methods. CDRNP is situated among approaches that model user-specific tasks and preference correlations.

5 RELATED WORKS

  • CDR research addresses cold-start recommendation, while the reviewed scope excludes the separate data-sparsity issue.The section focuses on methods recommending target-domain items to cold-start users.Source: §5 · p. 8
  • CDRNP differs from prior approaches by capturing preference correlations among overlapping and cold-start users and using explicit interactions from both domains.Source: §5 · p. 8
  • Meta-learning work spans optimization-, gradient-, and model-based methods; CDRNP is model-based and uses NP to capture user preference correlations.Source: §5 · p. 8
  • Neural Process models distributions over functions while using neural networks for efficient prediction; TANP applies NP to single-domain, rather than cross-domain, cold-start recommendation.The section also notes recommender-system methods using NP.Source: §5 · p. 8

6 CONCLUSION

CDRNP uses Neural Process to capture preference correlations among overlapping and cold-start users, with a preference remainer and adaptive conditional decoder. It consistently outperforms state-of-the-art methods in three CDR scenarios.

6 CONCLUSION

  • CDRNP consistently outperforms state-of-the-art methods in three CDR scenarios.Source: §6 · p. 8
  • CDRNP captures preference correlations among overlapping and cold-start users through an associated stochastic process.Source: §6 · p. 8
  • The method adds a preference remainer to enhance overlapping users’ common preference and an adaptive conditional decoder for prediction.The authors identify multi-domain recommendation with Neural Process as future work.Source: §6 · p. 8
Loading 2401.12732v3…