Exploiting User and Business Attributes for Personalized Business Recommendation

Kai Lu, Yi Zhang, Lanbo Zhang, Shuxin WangSIGIR 2015 · DOI

TL;DR

This paper introduces an Integrated Bias and Factorization Model (IBFM) to tackle the severe data sparsity and cold-start problems in business recommendation. By effectively leveraging rich user and business attributes, IBFM significantly outperforms existing state-of-the-art methods, achieving top performance on the RecSys 2013 Yelp business challenge.

What Problem Does This Paper Address?

Recommender systems are crucial for navigating information overload, but they face significant challenges, particularly in domains like business recommendations. Two major hurdles are data sparsity (users interact with only a tiny fraction of available items) and the cold-start problem (difficulty recommending for new users or new businesses with no historical data). Business recommendations are especially prone to these issues because transactions are often offline, and users rarely provide explicit ratings. Traditional matrix factorization (MF) models, while effective in general, struggle here because they tend to overfit with limited data and produce unreliable predictions for new entities. This paper aims to build a more robust recommendation system that can handle these challenges by incorporating diverse contextual information.

Methods

The core of the proposed solution is the Integrated Bias and Factorization Model (IBFM). It builds upon a baseline rating bias model ($r_{ui} = \mu + b_u + b_i$) by extending it with rich user and business attributes.

  • User Bias Model: User bias ($b_u$) is modeled using demographic information (gender), user voting patterns (e.g., how often a user's reviews are marked as "funny," "useful," or "cool"), and user's points-of-interest derived from the locations of businesses they've interacted with.
  • Business Bias Model: Business bias ($b_i$) is enriched by considering:
  • Check-in Information: How a business's rating bias is influenced by businesses with similar check-in patterns.
  • Location Information: Incorporating zip code, street, city, longitude, and latitude, and considering the bias of nearby businesses.
  • Name and Category: Leveraging the fact that businesses with similar names or belonging to the same category might share biases.
  • Unified Bias Model: These individual bias components are integrated into a comprehensive business bias term ($b_i = b_i + c_i + l_i + t_i$).
  • Integrated Bias and Factorization Model (IBFM): The full model combines the unified bias with latent factor representations of users ($p_u$) and businesses ($q_i$), predicting a rating as $r_{ui} = \mu + b_u + b_i + q_i^T p_u$.
  • Handling Cold-Start: For new users or businesses, IBFM employs a novel sampling strategy. Instead of random initialization, it identifies similar existing users/businesses based on their attributes and then samples their latent factor vectors from a Gaussian distribution fitted to these neighbors. This provides a more informed starting point for new entities.

Datasets

The primary dataset used is the RecSys 2013 Yelp business rating prediction challenge dataset. This dataset is publicly available and includes:

  • Rating data
  • Business check-in data
  • User profile data
  • Business profile data

The dataset is characterized by extremely low rating density (0.03%), making it a challenging benchmark for sparsity and cold-start issues. A significant portion of users have only one rating, and a large percentage of users and businesses in the test set are new. The authors also created a validation set from the original training data, carefully matching the cold-start rating distributions of the test set for hyper-parameter tuning and early stopping.

Key Insights and Ideas

  • Attribute Richness is Key: The paper strongly demonstrates that incorporating diverse user and business attributes is crucial for overcoming sparsity and cold-start problems, especially in domains like business recommendations where explicit feedback is scarce.
  • Location Matters Most: Among the various business attributes explored, location information proved to be the most impactful feature for improving recommendation accuracy. This highlights the spatial nature of business discovery.
  • Complementarity of Features: Combining user and business features in a unified model (Unified Bias Model) significantly boosts performance, indicating that these different sources of information are complementary and provide a more holistic view.
  • Intelligent Cold-Start Initialization: The sampling strategy for new users/businesses, based on attribute-driven similarity to existing entities, is a critical innovation. It provides a more grounded starting point than random initialization, leading to better predictions for unseen entities.

Key Results

The experimental results, measured by Root Mean Squared Error (RMSE), show that IBFM significantly outperforms several state-of-the-art baseline models:

  • IBFM achieved the lowest RMSE on both the public and private leaderboards of the RecSys 2013 Yelp business challenge.
  • Compared to traditional MF models like ALS-SVD and BasicMF, IBFM showed substantial improvements.
  • Even compared to advanced models like SVD++ and Factorization Machines (FM), IBFM increased prediction accuracy by 5.15%.
  • The Unified Bias Model, which incorporates attributes but not the sampling strategy, also showed significant improvements over simpler bias models and MF, demonstrating the value of attribute integration.

Limitations and Future Work

  • Limitations: The paper focuses on rating prediction. While IBFM shows strong performance, the authors don't explicitly discuss limitations related to scalability for extremely large datasets or potential biases introduced by specific attributes. The sampling strategy for cold-start relies on finding "similar" existing entities, which might still be challenging if truly similar entities don't exist.
  • Future Work: The authors suggest that their sampling method can effectively generate reasonable initial hidden representations for new users and new businesses. Future work could explore more sophisticated sampling techniques or alternative methods for cold-start scenarios. Further investigation into how to best combine different types of user and business attributes and their interactions could also be a fruitful direction.

Summary updated

BibTeX
@inproceedings{lu2015exploiting,
  author = {Kai Lu and Yi Zhang and Lanbo Zhang and Shuxin Wang},
  title = {{Exploiting User and Business Attributes for Personalized Business Recommendation}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2015},
  doi = {10.1145/2766462.2767806},
}
Figure