[Paper Review] Factorization Machines for Data with Implicit Feedback

Babak Loni, Martha A. Larson+1 more|arXiv (Cornell University)|Dec 19, 2018

Recommender Systems and Techniques18 references4 citations

TL;DR

This paper proposes FM-Pair, a pairwise learning-to-rank adaptation of Factorization Machines optimized for implicit feedback datasets. By replacing the standard point-wise loss with a BPR-inspired pairwise objective, FM-Pair improves ranking performance over standard FMs and state-of-the-art methods like BPR-MF, while retaining linear complexity and scalability.

ABSTRACT

In this work, we propose FM-Pair, an adaptation of Factorization Machines with a pairwise loss function, making them effective for datasets with implicit feedback. The optimization model in FM-Pair is based on the BPR (Bayesian Personalized Ranking) criterion, which is a well-established pairwise optimization model. FM-Pair retains the advantages of FMs on generality, expressiveness and performance and yet it can be used for datasets with implicit feedback. We also propose how to apply FM-Pair effectively on two collaborative filtering problems, namely, context-aware recommendation and cross-domain collaborative filtering. By performing experiments on different datasets with explicit or implicit feedback we empirically show that in most of the tested datasets, FM-Pair beats state-of-the-art learning-to-rank methods such as BPR-MF (BPR with Matrix Factorization model). We also show that FM-Pair is significantly more effective for ranking, compared to the standard FMs model. Moreover, we show that FM-Pair can utilize context or cross-domain information effectively as the accuracy of recommendations would always improve with the right auxiliary features. Finally we show that FM-Pair has a linear time complexity and scales linearly by exploiting additional features.

Motivation & Objective

• Address the limitation of standard Factorization Machines (FMs) in handling implicit feedback data, which is common in real-world recommender systems.
• Overcome the shortcomings of trivial implicit-to-explicit mapping, which fails to optimize for ranking due to point-wise optimization and unreliable negative sampling.
• Develop a scalable, general-purpose model that maintains FMs’ expressiveness and linear complexity while being effective for ranking tasks with implicit feedback.
• Demonstrate the effectiveness of FM-Pair in context-aware and cross-domain collaborative filtering by integrating auxiliary features such as context and cross-domain item features.
• Show that FM-Pair scales linearly with feature count and factorization dimensionality, making it suitable for large-scale applications.

Proposed method

• Adapt Factorization Machines to use a pairwise ranking loss function inspired by the Bayesian Personalized Ranking (BPR) criterion, replacing the standard point-wise regression loss.
• Formulate the optimization objective to maximize the predicted score of positive user-item interactions over negative (sampled) interactions, improving ranking performance.
• Retain the original FM structure, including factorization of user, item, and auxiliary features into a shared latent space, preserving model generality and expressiveness.
• Integrate auxiliary features such as context (e.g., time, location) and cross-domain item features by mapping them into the same latent space as users and items.
• Train the model using stochastic gradient descent with pairwise ranking objective, ensuring linear time complexity O(k) per sample, where k is the number of non-zero features.
• Leverage the WrapRec toolkit for reproducible evaluation across multiple datasets and configurations, enabling consistent comparison with baselines like BPR-MF and standard FMs.

Experimental results

Research questions

• RQ1 Can a pairwise optimization objective improve the ranking performance of Factorization Machines on implicit feedback datasets compared to standard point-wise training?
• RQ2 How does FM-Pair compare to state-of-the-art learning-to-rank methods such as BPR-MF in terms of accuracy and scalability on implicit feedback data?
• RQ3 To what extent can FM-Pair effectively exploit auxiliary features such as context or cross-domain information in collaborative filtering tasks?
• RQ4 Does the trivial mapping of implicit feedback to explicit labels (e.g., +1 for clicks, 0 for non-interactions) lead to suboptimal ranking performance in FMs?
• RQ5 How does FM-Pair scale with increasing dimensionality of factorization and number of auxiliary features, and does it maintain linear time complexity?

Key findings

• FM-Pair significantly outperforms standard Factorization Machines on ranking tasks, even when trained on explicit feedback data, due to its pairwise optimization objective.
• On four real-world datasets with implicit feedback, FM-Pair achieves higher accuracy than BPR-MF, the state-of-the-art method for implicit feedback ranking.
• The model demonstrates linear scalability with respect to both the number of latent factors and the number of auxiliary features, maintaining low training time growth.
• Incorporating item biases significantly improves performance on popularity-skewed datasets like Frappe, indicating the importance of bias modeling in implicit feedback settings.
• FM-Pair effectively leverages auxiliary features—such as context and cross-domain item features—resulting in consistent performance gains without requiring model reconfiguration.
• The pairwise training objective in FM-Pair leads to better convergence and ranking quality compared to the point-wise loss used in standard FMs, especially when negative samples are unreliable.