Proposing a Highly Accurate Hybrid Component-Based Factorised Preference Model in Recommender Systems

Farhad Zafari, Rasoul Rahmani, Irene MoserIJCAI 2017 · DOI

TL;DR

This paper introduces a novel hybrid recommender system model that integrates multiple preference components (user/item biases, feature preferences, temporal dynamics, social influence, and conditional dependencies) into a unified framework. By optimizing these components and their interactions, the model achieves higher accuracy than existing state-of-the-art methods, highlighting the critical role of understanding these interdependencies for effective recommendation.

What Problem Does This Paper Address?

In today's vast online marketplaces, recommender systems are crucial for helping users navigate an overwhelming number of products. While existing recommender systems, particularly those based on collaborative filtering and latent factor models, have achieved significant success, they often model only a subset of user preferences. The authors identify five key preference components: user/item biases, feature value preferences, conditional dependencies, temporal preference drifts, and social influence. The critical gap this paper addresses is the lack of a unified model that integrates all these components and, more importantly, investigates their interactions. Previous work has largely disregarded these interactions, potentially limiting the accuracy and adaptability of recommender systems across different domains and datasets. Understanding these interactions is key to designing more accurate and tailored recommendation engines.

Methods

The paper proposes a new hybrid latent factor model called TCSFVSVD (Time Conditional Social Feature Value Singular Value Decomposition). This model extends traditional matrix factorization by incorporating the five identified preference components. To optimize the model, a two-pronged approach is used:

  1. Gradient Descent: This standard optimization technique is employed to fine-tune the model parameters (user and item latent features, biases, etc.) based on the observed ratings.
  2. Evolutionary Algorithm (EA): A novel aspect is the use of an EA to optimize the hyper-parameters and learning rates for each of the preference components. This adaptive approach, termed ATCSFVSVD, is crucial because the optimal settings for these components can vary significantly across different datasets. The EA iteratively refines these parameters, feeding back the model's error to guide the optimization process.

The authors also explore the interactions between components by systematically switching them on and off and observing the impact on model performance.

Datasets

The proposed model was evaluated on three publicly available datasets:

  • Ciao: A dataset of movie ratings with associated trust relationships.
  • Epinions: A larger dataset of product ratings with explicit trust statements.
  • Flixster: A social movie site dataset including movie ratings and friendship links.

The authors note that for computational efficiency, they used uniformly randomly selected subsets of users and items from these datasets, while preserving the original density and relationships.

Key Insights and Ideas

  • Interactions Matter: The most significant insight is that the performance of recommender systems is not just about including various preference components, but critically about how these components interact with each other. These interactions are highly dataset-dependent.
  • No One-Size-Fits-All: Different combinations of preference components yield varying levels of accuracy. What works best for one dataset might not for another, underscoring the need for adaptive modeling.
  • Component Contributions Vary: Some components, like user and item biases (B), can even have a deteriorating effect on accuracy in certain datasets (e.g., Epinions), while others, like Feature Value Preferences (FVP), can be highly beneficial.
  • Social Influence Nuances: Social influence (SI) can be helpful, but its effectiveness can depend on the presence of other components, such as temporal dynamics (T).
  • Conditional Preferences are Key: Conditional preferences (CP) are shown to be particularly helpful when user and item biases (B) are also modeled.

Key Results

The proposed ATCSFVSVD model consistently outperformed TrustSVD, a strong baseline, across the evaluated datasets in terms of Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).

  • Ciao Dataset: The model including all five components achieved the lowest MAE.
  • Epinions Dataset: The model with only FVP performed best for MAE, while a combination of CP, T, and B yielded the best RMSE.
  • Flixster Dataset: A combination of FVP, SI, and B achieved the highest accuracy for MAE, while CP, T, and B led to the best RMSE.

The paper also presents detailed interaction plots illustrating how the inclusion or exclusion of specific components affects model accuracy, visually demonstrating the complex interdependencies.

Limitations and Future Work

The authors acknowledge that while their model integrates multiple preference components, the computational cost of optimizing all possible combinations can be high. They also note that some datasets may not contain all the necessary information (e.g., time-stamped social relationships).

Future work suggested includes:

  • Quantifying the significance of individual components and their interactions across different datasets and domains.
  • Exploring more sophisticated methods for modeling temporal dynamics and social influence.
  • Investigating how to efficiently incorporate these components when certain data types are missing.

Summary updated

BibTeX
@inproceedings{zafari2017proposing,
  author = {Farhad Zafari and Rasoul Rahmani and Irene Moser},
  title = {{Proposing a Highly Accurate Hybrid Component-Based Factorised Preference Model in Recommender Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/185},
}
Figure