TL;DR
This paper introduces a novel framework called Disentangled Counterfactual Reasoning (DCR) to tackle two major biases in recommender systems: item popularity bias and user conformity bias. By disentangling user and item embeddings and employing counterfactual reasoning, DCR effectively mitigates both the indirect and direct effects of these biases, leading to more accurate and fair recommendations.
What Problem Does This Paper Address?
Recommender systems are crucial for personalizing user experiences, but they often suffer from biases present in the observational data they learn from. Two prominent biases are:
- Popularity Bias: Popular items tend to be recommended more frequently, not necessarily because they are better, but simply because they are popular. This can lead to a "rich get richer" scenario, where less popular but potentially relevant items are overlooked.
- Conformity Bias: Users might interact with items based on what others are doing (social influence or conformity), rather than their genuine preferences. This can skew the understanding of individual user tastes.
While previous work has addressed these biases individually or conflated their effects, this paper highlights the need for a comprehensive approach that simultaneously tackles both popularity and conformity biases, and crucially, distinguishes between their direct and indirect influences on recommendations.
Methods
The core of the proposed solution is the Disentangled Counterfactual Reasoning (DCR) framework. It operates in two main stages:
- Disentangled Embeddings (Training):
- The framework learns to disentangle user and item embeddings into "biased" and "unbiased" components.
- Mutual Information (MI) Minimization: To achieve disentanglement, the authors use MI to ensure that the unbiased embeddings have minimal shared information with the biased embeddings. This forces the unbiased components to capture true user preferences and item characteristics, independent of popularity or conformity.
- Objective Functions: The biased embeddings are trained to capture item popularity and user conformity, while the unbiased embeddings are trained to accurately estimate these same factors. This separation is key to isolating the indirect effects of biases.
- Prediction Score: A recommendation score is calculated using a combination of biased embeddings, propensity scores (derived from popularity and conformity), and a matching function.
- Counterfactual Inference (Inference):
- Mitigating Indirect Effects: During inference, the biased embeddings are adjusted using inverse propensity weights (derived from popularity and conformity). This step aims to suppress the influence of popularity and conformity on the user-item matching process, thereby mitigating their indirect effects.
- Mitigating Direct Effects: The authors then imagine a counterfactual world where user-item matching is disregarded, and interactions are solely driven by popularity and conformity. The prediction score is adjusted by subtracting a term related to these bias factors, effectively removing their direct impact on the final recommendation score.
Datasets
The proposed DCR framework was evaluated on four publicly available datasets:
- Adressa
- MovieLens-1M
- Gowalla
- Yelp2018
The authors also mention constructing "unbiased test sets" for evaluation, ensuring that all items in these test sets receive an equal number of interactions. This is a crucial step for robustly evaluating debiasing methods, as it removes the inherent popularity bias from the evaluation itself.
Key Insights and Ideas
- Dual Bias is Crucial: The paper strongly advocates for addressing both popularity and conformity biases simultaneously, as focusing on only one can lead to suboptimal performance.
- Disentanglement is Key: Separating user and item embeddings into biased and unbiased components is fundamental to isolating and mitigating the different ways biases manifest.
- Indirect vs. Direct Effects: The distinction between indirect (embedding-level) and direct (score-level) effects of biases is critical for a comprehensive debiasing strategy.
- Counterfactual Reasoning for Debias: Applying counterfactual inference at the inference stage is a novel way to adjust predictions by simulating worlds with and without bias influences.
Key Results
The experimental results, summarized in Table 1, demonstrate the effectiveness of DCR:
- Superior Performance: DCR consistently outperforms state-of-the-art methods that focus on single biases (UDIPS, DICE, PPAC) and even methods that attempt to address dual biases (MACR).
- Significant Improvements: DCR achieves notable improvements in Recall and NDCG scores across all datasets. For instance, on Adressa, DCR improves Recall by 12.1% and NDCG by 15.9% over the best existing method.
- Addressing MACR's Weakness: While MACR also targets dual biases, DCR shows better performance, suggesting that MACR might conflate true user interests with bias-influenced behaviors, a problem DCR's disentanglement aims to solve.
Limitations and Future Work
- Limitations: The paper acknowledges that while MACR attempts to address dual biases, it struggles to distinguish true user interests from bias-influenced behaviors. This implies that even DCR might have room for improvement in perfectly separating these aspects. The paper also mentions that MACR conflates biased and unbiased representations, which DCR explicitly addresses.
- Future Work: Not explicitly detailed in the student abstract, but the success of DCR suggests future work could explore more sophisticated disentanglement techniques, investigate other types of biases, or apply the framework to different recommendation scenarios (e.g., sequential recommendations). The tunable hyperparameter 'X' in the inference formulation also presents an avenue for further optimization.
Summary updated