Specializing Joint Representations for the task of Product Recommendation

Thomas Nedelec, Elena Smirnova, Flavian Vasile

TL;DR

This paper introduces Content2Vec, a novel deep learning architecture that unifies various product information sources (text, images, collaborative filtering) into a single, specialized representation for product recommendation. This approach significantly improves performance, especially in challenging "cold-start" and cross-category scenarios, by learning a more robust and expressive product embedding.

What Problem Does This Paper Address?

Online product recommendation systems are crucial for e-commerce, but a key challenge is effectively recommending new or niche products (cold-start problem) and suggesting items across different categories. Traditional recommendation systems often struggle with these scenarios because they rely heavily on past user-item interactions, which are scarce for new items or when bridging category gaps. While combining different data sources like product descriptions, images, and user behavior (collaborative filtering) can help, doing so in a unified, optimized representation for the recommendation task itself, particularly for candidate generation, remains an open problem. This paper aims to bridge this gap by creating a specialized joint representation that leverages all available product information to improve recommendation relevance and robustness.

Methods

The core of the proposed solution is the Content2Vec architecture, designed for modularity and specialization. It comprises three main components:

  • Content-specific Embedding Modules: These modules independently process different types of product data (text, images, collaborative filtering signals) to generate modality-specific embeddings. For text, they use a combination of Word2Vec for word embeddings and a TextCNN for sentence-level representation. For images, they leverage a pre-trained AlexNet, fine-tuned for product similarity. Collaborative filtering signals are captured using Prod2Vec.
  • Joint Product Embedding Modules: These modules fuse the modality-specific embeddings into a single, unified product representation. The paper proposes two architectures for this:
  • Content2Vec-perf: This approach uses a "Cross Interaction Unit" (CIU) to model complex, non-linear interactions between modalities, going beyond simple linear combinations. It aims to capture residual similarities not explained by individual modalities.
  • Content2Vec-compressed: This architecture prioritizes retrieval efficiency by compressing the fused representation into a fixed, smaller dimension (e.g., 200), making it suitable for large-scale indexing and searching.
  • Output Layer: This layer computes the probability of two products being co-purchased, typically using a sigmoid function over their similarity.

The training is performed module-by-module, allowing for efficient updates and leveraging pre-trained models. The entire system is optimized for the product link prediction task, aiming to maximize the AUC (Area Under the Curve) of predicting co-purchased product pairs.

Datasets

The evaluation was conducted on the publicly available Amazon dataset, which contains product images, text descriptions, and co-purchase information. The authors focused on the "Book" and "Movie" subgraphs to create datasets for evaluating different recommendation scenarios:

  • Hard Cold Start Dataset: This dataset simulates a scenario where all product pairs in the validation and testing sets consist of products unseen during training. This is a particularly challenging benchmark.
  • Soft Cold Start Dataset: This dataset includes some products from the test set that were also present during training, representing a more common cold-start scenario.

Key Insights and Ideas

  • Specialization is Key: Simply combining embeddings from different modalities isn't enough; they need to be specialized for the recommendation task. Content2Vec achieves this by fine-tuning modality-specific modules.
  • Beyond Linear Combinations: The "Cross Interaction Unit" (CIU) is a novel way to capture complex, non-linear relationships between product features, leading to better performance than simple linear fusion.
  • Modularity for Real-World Deployment: The modular design of Content2Vec makes it practical for large-scale systems, allowing individual modules to be updated or replaced without retraining the entire architecture.
  • Robustness in Challenging Scenarios: The proposed method shows significant improvements in hard cold-start and cross-category recommendations, areas where traditional methods often falter.
  • Trade-off Between Performance and Efficiency: Content2Vec-perf offers higher accuracy, while Content2Vec-compressed provides a more efficient, smaller representation suitable for retrieval systems, highlighting a practical design choice.

Key Results

The paper reports significant improvements in AUC across various datasets and scenarios:

  • Hard Cold Start: Content2Vec-perf achieved an AUC of 89% on the Books dataset and 83% on the Movies dataset, outperforming baselines like ImageCNN (81%, 92%), TextCNN (72%, 90%), and Fusion-linear (83%, 94%).
  • Soft Cold Start: Content2Vec-perf+ (an ensemble of Prod2Vec and Content2Vec) achieved an AUC of 92% on the Movies dataset, demonstrating the benefit of combining content and collaborative filtering signals.
  • Cross-Category Performance: The Content2Vec-perf model showed better generalization on mixed-category datasets compared to Content2Vec-compressed, indicating its ability to learn more robust, category-agnostic representations.

Limitations and Future Work

The authors acknowledge that while Content2Vec-perf performs well, there is still room for improvement in cross-category tasks. They also suggest exploring sparsity constraints on product representations to further enhance the performance of the final product retrieval system. Future work could also involve investigating more advanced fusion techniques and exploring other modalities for product representation.

Summary updated

BibTeX
@inproceedings{nedelec2017specializing,
  author = {Thomas Nedelec and Elena Smirnova and Flavian Vasile},
  title = {{Specializing Joint Representations for the task of Product Recommendation}},
  booktitle = {DLRS@RecSys},
  year = {2017},
  doi = {10.1145/3125486.3125489},
}
Figure