Abstract
Collaborative filtering plays an important role in promoting the service recommendation ecosystem, and matrix decomposition technology has been proven to be one of the most effective recommendation methods. However, traditional collaborative filtering algorithms struggle with "cold start" items—new entries that lack historical rating data. This neglect reduces recommendation diversity and hinders the growth of new services. To address this, we propose Attribute-Fused Singular Value Decomposition (FASVD). This method combines item attribute information and historical rating data into a unified matrix decomposition model, incorporating a time-decay penalty. Experimental results on MovieLens and JD datasets demonstrate that FASVD significantly improves recommendation accuracy for new items compared to baseline methods.
The Challenge: The Cold Start Problem
Modern recommendation systems primarily rely on Collaborative Filtering (CF) to predict user interests based on the behavior of similar peers. However, CF faces a critical bottleneck known as the Cold Start Problem. When a new movie, product, or mobile application enters the system, it has no ratings. Traditional models often ignore these items because they lack the data necessary to calculate similarity.
In high-velocity environments like mobile app stores or e-commerce platforms, failing to recommend new content leads to stagnant catalogs and poor user retention. A robust system must be able to "warm up" these new items immediately upon entry.
Methodology: Attribute-Fused SVD (FASVD)
The FASVD approach bridges the gap between existing user data and "cold" items through a three-stage fusion process:
- Attribute-Based KNN with Time Penalty: For items without ratings, the model calculates similarity using metadata (e.g., genre, publisher). A time penalty factor is introduced to prioritize recent items, acknowledging that user preferences shift over time.
- SVD for Existing Items: Standard Singular Value Decomposition (SVD) is applied to items with sufficient historical data to map users and items into a latent factor space.
- The FASVD Fusion: The core innovation lies in replacing the static item bias term ($b_i$) in the SVD objective function with the preliminary rating predicted by the attribute-based KNN. This allows the model to inherit knowledge from item metadata to inform the matrix factorization.
Experimental Evaluation
The researchers evaluated FASVD across three datasets to test scalability and real-world applicability:
- MovieLens (100K & 20M): Standard benchmarks for movie recommendations.
- JD Dataset: Real-world data from the Jingdong e-commerce platform, specifically captured during the "Double Eleventh" shopping festival to simulate a high-velocity "new product" environment.
Key Results
- Superior Accuracy: FASVD consistently achieved lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) than baselines like Item-KNN and Funk-SVD.
- Resilience to Sparsity: The model maintained high performance even when training data density was as low as 30%, proving its effectiveness in data-poor scenarios.
- Dynamic Bias: By treating attributes as a proxy for history, FASVD creates a dynamic bias that reflects how a user’s specific history aligns with an item's metadata.
Conclusion and Future Directions
FASVD provides a practical solution for the item cold start problem by fusing content-based attributes with collaborative filtering. While highly effective for new items, the model does not yet address the "user cold start" problem (new users with no history). Future research may explore using deep learning architectures, such as Autoencoders, to extract even richer features from item descriptions and images to further refine the recommendation engine.
Summary updated