Multi-Perspective Neural Architecture for Recommendation System

Han Xiao, Yidong Chen, Xiaodong ShiarXiv 2018 · arXiv

TL;DR

This paper introduces a hierarchical, multi-perspective neural architecture for top-N recommendation systems. By modeling user preferences and item characteristics through sequential stages of "perspectives" (e.g., moving from specific traits like actors to abstract concepts like acting quality) and employing a novel cross-attention mechanism, the model significantly outperforms state-of-the-art baselines across multiple domains.

What Problem Does This Paper Address?

Traditional recommendation systems, even recent neural models like NeuMF or DMF, often treat user-item interactions as a black box. They typically represent users and items as simple latent vectors (embeddings) and use a metric like cosine similarity or a Multi-Layer Perceptron (MLP) to predict a rating.

The authors argue this is too simplistic. In reality, a user’s preference is multi-faceted and hierarchical. For example, when watching a movie, a user might care about specific elements (the lead actor, the director) which then inform more abstract judgments (acting quality, overall style). Furthermore, user preferences aren't static; they are subjective and can be slightly adjusted based on the specific item being considered. Existing models fail to capture this hierarchical reasoning and the fine-grained correlation between a specific user and a specific item during the encoding process.

Methods

The proposed architecture (shown in Figure 2) moves away from independent encoding.

Figure 2: Proposed Neural Architecture. We leverage the corresponding row/column of interaction matrix as the in- put of user/item. To characterize hierarchically organized perspectives, we employ several sequential stages to encode the input. In each stage, there exist several perspectives. In each perspective, the input of this perspective will be en- coded into hidden representations by linear transformation with ReLU activation function and then attention mecha- nism addresses the correlations for the encoded representa- tions of user/item to generate the output of this perspective. Furthermore, the outputs of all the perspectives in one stage are respectively concatenated as the output representation of user and item for this stage. Finally, the representation of user and item would be metric by cosine similarity to pre- dict the user’s preference.
  1. Hierarchical Sequential Stages: The model processes inputs through multiple sequential stages. Each stage contains several "perspectives." The idea is that lower stages capture specific features, while higher stages aggregate these into abstract representations, mimicking human decision-making (see Figure 1).
  2. Multi-Perspective Encoding: In each stage, the user and item are processed through $M$ parallel perspectives. The outputs of these perspectives are concatenated to form the input for the next stage.
  3. Cross-Attention Mechanism: This is the core innovation. Within each perspective, the model uses an attention mechanism so that the user representation "attends" to the item and vice versa. The authors propose two versions:
  4. Softmax-ATT: A standard approach where the attention signal for the user is generated from the item's representation.
  5. Correlated-ATT: A more advanced version that constructs a correlation matrix (the outer product of user and item attention vectors) to explicitly model the interaction between every dimension of the user and item representations.
  6. Final Prediction: The final stage's output representations are compared using cosine similarity to predict the preference score. The model is trained using a cross-entropy loss function.

Datasets

The authors evaluated their model on six widely used datasets across five distinct domains:

  • Movie: MovieLens 100K and MovieLens 1M.
  • Music: Amazon Music.
  • Book: Amazon Kindle Books.
  • Office: Amazon Office Products.
  • Baby: Amazon Baby Products.

The datasets were filtered to ensure users had at least 20 interactions and items had at least 5, following standard preprocessing practices in the field.

Key Insights and Ideas

  • Hierarchical Abstraction Matters: The authors found that a 3-layer (stage) architecture performed best. This supports the hypothesis that recommendation benefits from a "Specific Level" to "Abstract Level" flow of information.
  • Mutual Subjectivity: The success of the attention mechanism suggests that encoding a user in the context of the item (and vice versa) is more effective than encoding them independently. A user's "actor preference" might shift slightly depending on the "director style" of the movie they are currently looking at.
  • Correlation is Key: The "Correlated-ATT" mechanism consistently outperformed the simpler "Softmax-ATT." By explicitly calculating a correlation matrix, the model captures deeper, non-linear dependencies between user preferences and item features.

Key Results

The multi-perspective model consistently outperformed baselines like NeuMF and DMF across all datasets in terms of Hit Ratio (HR@10) and Normalized Discounted Cumulative Gain (NDCG@10).

  • Significant Gains in Sparse Domains: The improvements were particularly notable in sparser datasets. For the Amazon Music dataset, the "Correlated-ATT" model showed a 13.90% improvement in NDCG and a 16.49% improvement in HR over the best baseline.
  • Ablation Studies: The researchers demonstrated that increasing the number of perspectives (up to 6) and the final latent dimension (up to 128) generally improved performance, suggesting that more complex preferences require higher-capacity representations.
NDCG@K for Music Dataset
Figure 6(d) shows the consistent performance lead of the proposed method over DMF and NeuMF as the rank list length (K) increases.

Limitations and Future Work

While the paper demonstrates strong results, a few areas remain for exploration:

  • Computational Complexity: The use of multiple stages, parallel perspectives, and correlation matrices increases the number of parameters and computational overhead compared to simpler models. The authors do not provide a detailed analysis of training or inference time.
  • Cold Start: Like most collaborative filtering models based on interaction matrices, this architecture likely struggles with the "cold start" problem for entirely new users or items with no history.
  • Interpretability: Although the model is inspired by human-like hierarchical perspectives, the "perspectives" themselves remain hidden latent spaces. Future work could involve trying to map these latent perspectives to actual metadata (like genre or price) to make the recommendations more explainable.

Summary updated

Abstract

Currently, there starts a research trend to leverage neural architecture for recommendation systems. Though several deep recommender models are proposed, most methods are too simple to characterize users' complex preference. In this paper, for a fine-grain analysis, users' ratings are explained from multiple perspectives, based on which, we propose our neural architecture. Specifically, our model employs several sequential stages to encode the user and item into hidden representations. In one stage, the user and item are represented from multiple perspectives and in each perspective, the representations of user and item put attentions to each other. Last, we metric the output representations of final stage to approach the users' rating. Extensive experiments demonstrate that our method achieves substantial improvements against baselines.
BibTeX
@misc{xiao2018multi,
  author = {Han Xiao and Yidong Chen and Xiaodong Shi},
  title = {{Multi-Perspective Neural Architecture for Recommendation System}},
  year = {2018},
  eprint = {1807.09751},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.09751},
}
Figure