Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation

Wenbo Wang, Bingquan Liu, Lili Shan, Chengjie Sun, Ben Chen, Jian GuanAAAI 2024 · DOI

TL;DR

This paper introduces PAD-CLRec, a novel recommendation system that tackles the "cold-start" problem for new items by leveraging user preferences and a dual contrastive learning approach. By aligning item content with aggregated user preferences, PAD-CLRec significantly improves recommendations for items with no prior interaction history.

What Problem Does This Paper Address?

Recommender systems, particularly those based on collaborative filtering (CF), excel at suggesting items based on past user interactions. However, they struggle when new items (cold-start items) appear, as these items lack the interaction data needed to learn collaborative features. Existing solutions often try to align an item's content features (like images or attributes) with its collaborative features. While this helps, it often overlooks the crucial aspect of how a user might perceive or prefer a cold item, especially if its content is only subtly similar to items they've interacted with before. This paper aims to bridge this gap by incorporating user preferences into the alignment process, leading to better recommendations for cold items.

Methods

The proposed method, PAD-CLRec, consists of two main stages:

  1. Feature Encoding:
    • Warm Branch: Learns collaborative features for users and existing (warm) items using standard embedding matrices based on their interaction history.
  2. Cold Branch: Extracts content features for cold items, primarily from their image inputs, using a pre-trained model and a trainable feature extractor.
  3. Joint Objective Function Optimization: This is where the core innovation lies, using a combination of contrastive learning losses:
  4. Dual Recommendation Contrastive Loss (LUI): Captures the relationship between users and items based on their collaborative features, ensuring that interacted user-item pairs are more similar than non-interacted ones.
  5. Dual Item-Level Aligning Contrastive Loss (LA): Aligns an item's collaborative feature with its content feature. This is a standard approach but is enhanced here by considering both directions (item's collaborative to content, and content to collaborative).
  6. Preference Aware Contrastive Loss (LP): This is the key novelty. It addresses the cold-start problem by:
  7. Aggregating the collaborative features of warm items a user has interacted with to form a "user preference" representation.
  8. Treating a randomly selected warm item in the training set as a simulated cold item.
  9. Aligning this simulated cold item's content feature with the aggregated user preference. This allows the model to learn how cold items might fit into a user's broader preferences, even without direct interaction data.

A joint objective function combines these losses to balance performance across both warm and cold items.

Datasets

The paper evaluates PAD-CLRec on two real-world datasets:

  • Amazon Rec dataset: A general recommendation dataset.
  • Amazon Fashion dataset: A dataset specifically focused on fashion items.

Both datasets provide user-item interaction records and item images, which are crucial for the content-based aspects of the proposed method. The authors define three tasks: "Cold" (focusing on new items), "Warm" (focusing on existing items), and "All" (a mix of both).

Key Insights and Ideas

  • User Preference is Key for Cold Items: The most significant insight is that simply aligning content and collaborative features isn't enough for cold items. Understanding a user's aggregated preference from their past interactions with warm items provides a much richer signal for recommending new, unseen items.
  • Simulating Cold Items in Training: The clever use of randomly selected warm items as "cold items" during training allows the model to learn the preference-aware alignment without needing actual cold items in the training set. This is a practical way to address the cold-start challenge during model development.
  • Dual Contrastive Learning for Comprehensive Alignment: Employing both item-level and group-level (preference-aware) contrastive learning ensures that the model captures relationships at multiple granularities, leading to a better overall recommendation performance.
  • Balancing Warm and Cold Item Performance: The joint objective function is designed to strike a crucial trade-off between recommending items users have interacted with before and recommending new items, a common challenge in cold-start scenarios.

Key Results

  • Superior Performance: PAD-CLRec consistently outperforms state-of-the-art baseline methods across all three tasks (All, Warm, Cold) on both Amazon Rec and Amazon Fashion datasets.
  • Significant Improvement for Cold Items: The method shows particularly strong gains on the "Cold" task, demonstrating its effectiveness in handling new items. For instance, on Amazon Rec, it achieved a 10.4% improvement in Recall@10 and 13.0% in NDCG@10 compared to the best baseline.
  • Visualization Confirms Effectiveness: t-SNE visualizations show that PAD-CLRec clusters users, warm items, and cold items more cohesively than baseline methods, indicating better learned representations. Cold items are shown to be closer to users and warm items, suggesting better integration into user profiles.
  • Ablation Studies Validate Components: Experiments confirm the importance of the preference-aware contrastive loss (LP) and the dual contrastive losses (LUI and LA), with PAD-CLRec achieving the best results when all components are present.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper notes that the parameter η (controlling the weight of the preference-aware loss) can significantly impact performance, especially on the cold task. Finding the optimal balance is crucial.
  • Preference Calculation Methods: While the "Average" method for calculating user preference was found to be the best compromise, other methods like "Weighted Sum" and "Sampling" were explored, suggesting potential for further research into more sophisticated preference aggregation techniques.
  • Future Work: The authors suggest that their method can be extended to other recommendation scenarios and that further exploration of preference calculation methods could yield even better results. The focus on image features could also be expanded to include other multimedia information.

Summary updated

BibTeX
@inproceedings{wang2024preference,
  author = {Wenbo Wang and Bingquan Liu and Lili Shan and Chengjie Sun and Ben Chen and Jian Guan},
  title = {{Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i8.28763},
}
Figure