Informative household recommendation with feature-based matrix factorization

Qiuxia Lu, Diyi Yang, Tianqi Chen, Weinan Zhang, Yong Yu2011 · DOI

TL;DR

This paper tackles the challenge of recommending movies to entire households, not just individuals, by developing an "informative" feature-based matrix factorization model. They introduce a novel ranking-oriented matrix factorization approach and demonstrate the crucial role of incorporating contextual information like household relationships, item neighborhoods, and implicit user feedback for improved recommendation accuracy.

What Problem Does This Paper Address?

The paper addresses the problem of generating personalized movie recommendations for households, a task that differs significantly from traditional individual-user recommendation systems. The CAMRa2011 challenge, which this paper participated in, specifically aimed to create movie ranking lists for households. This presents several challenges:

  • Group Recommendation: Recommending for a household means considering the preferences of multiple individuals and aggregating them into a single, coherent recommendation. This is a less explored area than individual recommendations.
  • Ranking vs. Rating Prediction: The goal is to produce a ranked list of movies, not just predict a user's rating for a movie. This requires models that are optimized for ordering items.
  • Rich Contextual Information: The dataset provided a wealth of contextual data beyond simple user-item interactions, including user-household relationships, temporal information, and implicit feedback. Effectively leveraging this information is key to improving recommendation quality.

Methods

The authors propose a multi-faceted approach built upon a feature-based matrix factorization (FBMF) framework. This framework is designed to be flexible, allowing various types of contextual information (features) to be incorporated into the model.

Key techniques include:

  • Ranking-Oriented Matrix Factorization (RMF): Instead of predicting explicit ratings, this method focuses on learning pairwise preferences (e.g., user u prefers item i over item j). They introduce a refined pairwise preference generation rule (Equation 7) that is stricter than previous methods, leading to more confident preferences.
  • Matrix Factorization with Negative Examples Sampling: Recognizing that users rarely rate movies they dislike, the authors address the "negative sample missing problem" by strategically sampling negative examples (movies a user likely doesn't like) to augment the training data. This is shown to be crucial for improving performance.
  • Informative Models via FBMF: The core innovation lies in using the FBMF framework to integrate diverse contextual information. This includes:
  • Item Neighborhood: Incorporating information about similar items.
  • User-Household Hierarchy: Modeling relationships within households.
  • User Implicit Feedback: Utilizing signals like whether a user has rated a movie or not. The FBMF framework allows these features to be seamlessly integrated, creating "informative models."
  • Household Recommendation Strategies: Two methods are proposed for generating household recommendations:
  • User-Oriented Approach: Predicts item rankings for individual users and then aggregates these preferences for the household (e.g., using a weighted average).
  • Household-Oriented Approach: Treats households as entities and directly learns their preferences using matrix factorization.

Datasets

The primary dataset used is from the CAMRa2011 challenge. This dataset contains approximately 4.5 million ratings from 17,000 users for 2,400 items, along with contextual information such as user-household relationships, time stamps, and implicit feedback. The paper does not introduce a new benchmark dataset but rather uses this existing one for evaluation.

Key Insights and Ideas

  • Context is King (for Households): The paper strongly emphasizes that incorporating contextual information is not just beneficial but essential for accurate household recommendations. Models that integrate item neighborhood, household hierarchy, and implicit feedback significantly outperform basic matrix factorization.
  • Ranking Matters: Directly optimizing for ranking (as in RMF) is more effective for the task of generating ranked lists than relying on models that primarily predict ratings.
  • Negative Sampling is Crucial: The "negative sample missing problem" is a significant hurdle. Strategically sampling negative examples dramatically improves performance, highlighting the importance of understanding what users don't like.
  • Feature-Based Matrix Factorization is a Powerful Abstraction: The FBMF framework provides a flexible and unified way to incorporate diverse features, making it easier to build complex, informative models without reinventing the wheel for each new piece of information.
  • User-Oriented Household Recommendation is Sufficient (for now): While a household-oriented approach is explored, the user-oriented approach (aggregating individual preferences) proved to be effective and yielded similar results for this specific task.

Key Results

The paper demonstrates significant improvements over baseline methods.

  • RMF vs. BMF: The ranking matrix factorization (RMF-S2) achieved approximately a 60% improvement over basic matrix factorization (BMF) in metrics like P@10 and MAP.
  • Negative Sampling Impact: BMF with negative sampling (BMF-3N) showed a ~60% improvement over basic BMF, underscoring its importance.
  • Informative Models: Integrating contextual information (item neighborhood, household hierarchy, implicit feedback) into RMF-S2 led to substantial gains. For instance, incorporating neighborhood information alone improved MAP and P@5 by ~25%. The best model, RMF-S2-ALL (integrating all contextual info), showed a ~30% improvement over RMF-S2.
  • Best Models: The top-performing models (e.g., RMF-S2-ALL) achieved significantly better performance across various metrics (MAP, AUC, P@5, P@10) compared to basic models. For instance, the best user-based model achieved a MAP of 0.3096 and a rank of 111, while the best household-based model achieved a MAP of 0.3186 and a rank of 104.

Limitations and Future Work

  • Monotony of Recommendations: The authors acknowledge that while accuracy is improved, recommendations might become monotonous.
  • Future Work: They suggest exploring other recommendation characteristics like explanation, novelty, and diversity. Further research into group recommendation, especially concerning evaluation and dataset suitability, is also advocated. The household-oriented approach, though not superior in this study, is noted as an interesting direction for future exploration in other scenarios.

Summary updated

Abstract

In this paper, we describe our solutions to the first track of CAMRa2011 challenge. The goal of this track is to generate a movie ranking list for each household. To achieve this goal, we propose to use the ranking oriented matrix factorization and the matrix factorization with negative examples sampling. We also adopt feature-based matrix factorization framework to incorporate various contextual information to our model, including user-household relations, item neighborhood, user implicit feedback, etc. Finally, we elaborate two kinds of methods to recommend movies for each household based on our models. Experimental results show that our proposed approaches achieve significant improvement over baseline methods.
BibTeX
@misc{lu2011informative,
  author = {Qiuxia Lu and Diyi Yang and Tianqi Chen and Weinan Zhang and Yong Yu},
  title = {{Informative household recommendation with feature-based matrix factorization}},
  year = {2011},
  doi = {10.1145/2096112.2096116},
}
Figure