TL;DR
This paper introduces ESCOFILT, a novel unsupervised approach to collaborative filtering that generates extractive summaries from user and item reviews. These summaries serve a dual purpose: they represent user/item preferences and provide natural, human-like explanations for recommendations, improving both prediction accuracy and explainability.
What Problem Does This Paper Address?
Recommender systems, particularly those using collaborative filtering (CF), are powerful but often act as "black boxes." While they can predict what users might like, they struggle to explain why. Existing explanation methods, like highlighting single reviews or specific words, don't always feel natural or comprehensive. The authors argue that these methods are often byproducts of the underlying recommendation mechanisms rather than genuine explanations. They aim to bridge this gap by creating explanations that closely mimic how humans naturally explain their preferences and decisions, using extractive summaries derived from multiple reviews.
Methods
The core of ESCOFILT lies in its ability to unify representation learning and explanation generation. The model employs a multi-stage process:
- Sentence Extraction and BERT Encoding: Reviews for each item (or user) are first segmented into sentences. These sentences are then fed into a pre-trained BERT-LARGE model to obtain contextualized word embeddings. The word embeddings for each sentence are averaged to produce sentence embeddings.
- Embedding Clustering: K-Means clustering is applied to these sentence embeddings to group similar sentences. This step aims to reduce redundancy and identify the most representative sentences for each cluster. The number of clusters ($K$) is determined by an "item summary ratio" ($i$), which dictates the desired length of the summary.
- Fusion Layers: The clustered sentence embeddings (which form the "representation-explanation") are then fused with traditional latent factor embeddings (UserIV and ItemIV). This fusion combines the learned representations with the generated explanations.
- Multilayer Perceptron (MLP) and Rating Prediction: The fused representations are passed through a tower-structured MLP to learn complex user-item interactions. The final layer of the MLP outputs the predicted rating.
The model is trained using the Mean Squared Error (MSE) objective function to minimize the difference between predicted and actual ratings.
Datasets
The authors evaluated ESCOFILT on four public datasets:
- Amazon 5-core (users and items with at least five reviews)
- Digital Music
- Instant Video
- Patio, Lawn, & Garden
These datasets were split into training (80%), validation (10%), and test (10%) sets.
Key Insights and Ideas
- Unifying Representation and Explanation: The most significant insight is that extractive summaries can serve both as a representation of user/item preferences and as a natural explanation. This dual functionality is a key innovation.
- Mimicking Human Explanations: The paper emphasizes that summary-level explanations, derived from multiple reviews, are more akin to how humans naturally explain their choices than single-review or word-level explanations.
- BERT's Power for Context: Utilizing BERT's contextual embeddings is crucial for capturing nuanced meaning in reviews, leading to better representations and explanations.
- Clustering for Summarization: K-Means clustering effectively groups similar sentences, helping to distill the most important information for summary generation and reducing redundancy.
- Items Benefit More from Clustering: The experiments suggest that items benefit more from embedding clustering than users, likely because items tend to have a larger volume of reviews compared to the number of reviews written by a single user.
Key Results
- Prediction Accuracy: ESCOFILT consistently outperformed all baselines across all datasets in terms of Root Mean Square Error (RMSE). For example, on the Automotive dataset, ESCOFILT achieved an RMSE of 0.8968, outperforming the next best model (BENEFICT) at 0.9023.
- Explainability: Human evaluations showed that ESCOFILT's summary-level explanations were rated highest for coherence, completeness, novelty, and perceived truth. They were also ranked as the most helpful explanations by judges for a significant majority of items (83%).
- Effectiveness of Clustering: Variants of ESCOFILT without embedding clustering (ESCOFILT-N) performed significantly worse, demonstrating the crucial role of clustering in achieving high accuracy and explainability.
Limitations and Future Work
- Unsupervised Nature: The model performs unsupervised summarization, which is a strength as obtaining ground-truth summaries is unrealistic. However, this also means the "quality" of the summary is inherently tied to the quality and informativeness of the available reviews.
- Future Work: The authors suggest expanding the model's explainability capabilities by incorporating other Natural Language Processing (NLP) principles, such as abstractive summarization and natural language generation. This could lead to even more fluid and human-like explanations.
Summary updated