Introduction to the Special Issue on Causal Inference for Recommender Systems

Contemporary recommendation systems predominantly utilize advanced machine learning techniques and latent representation frameworks, such as matrix factorization and deep neural networks, to process diverse information sources, including but not limited to ratings, text, images, and audio or video signals. While these systems employ meticulously designed models tailored to these tasks, the foundational design of most models remains centered on extracting correlative signals from the data, encompassing techniques such as representation learning for feature extraction and neural function learning for similarity matching. Nonetheless, the shift from correlation-based learning to causal learning in the domain of recommendation systems presents a critical area for research. This is because causal models provide the capability to extend beyond mere observational data in the processes of representation learning and ranking, offering a broader understanding and enhanced accuracy in recommendations.

Causal learning offers multifaceted benefits to the Recommender System community: (1) Explainable Recommendation Models–utilizing counterfactual reasoning, causal learning facilitates the development of models that enhance transparency and explainability in recommendations; (2) Unbiased Recommendation Models–through interventionist and counterfactual approaches, it aids in mitigating biases within the data or models, fostering unbiased recommendations; (3) Fairness-aware Recommendation Models–counterfactual fairness ensures that recommendation models do not rely on sensitive feature profiling, thus promoting fairness in ranking and recommendations; (4) Robust Recommendation Models–leveraging both factual and counterfactual data, causal learning enables the creation of recommendation models that are resilient to manipulations such as shilling attacks, spam, and echo chambers; (5) Cognitive Reasoning Models–combining causal and logical reasoning, causal learning is pivotal in transitioning from perceptual AI to cognitive AI, thereby supporting the development of cognitively-aware recommendation systems, such as those used in conversational interfaces that interact dynamically with users; (6) New Evaluation Methods for Recommendation–innovative counterfactual evaluation methods improve the reliability of evaluation protocols within recommender systems. This special issue underscores the importance of causal modeling in various recommendation tasks and promotes the integration of AI Ethics in recommender systems through causal approaches. As highlighted, causal modeling enhances recommender systems not only in traditional metrics such as ranking accuracy but also in ethical dimensions including explainability, fairness, robustness, and privacy. More broadly, the AI community recognizes the shift from correlative to causal learning as crucial for addressing a wide array of challenges in deep learning, computer vision, and natural language processing.

Towards the above goal, the special issue has included papers on both the methodology and the application of causal inference for recommendation. Cavenaghi et al. [1] propose a unified causal decision-making framework for recommender systems, which emphasizes the growing importance of causality in machine learning, particularly in the context of recommender systems due to the requirement of personalization and de-biasing. de Villa et al. [2] propose an approach to rank the causal impact of recommendations under collider bias in k-spots recommender systems, which examines how insights from recommendation data can inform decisions about item catalog management, such as identifying highly impactful recommendations or underperforming items. Li et al. [3] propose an explicitly weighted GCN aggregator for recommendation, which explores the enhancement of graph convolutional networks (GCN) for recommender systems by using temporal evolution and popularity bias as key features. Furthermore, Pargal et al. [4] propose a personalized guideline recommendation system based on explainable domain adaptation for driving through new cities, which addresses the challenges that drivers face when navigating unfamiliar city environments, contributing to the novel applications of causal-based recommendation methods.

Overall, the special issue covered research areas on both causal inference methods for recommendation and domain-specific applications of causal-based recommendation.

In the paper “Towards a Causal Decision-Making Framework for Recommender Systems,” Cavenaghi et al. [1] emphasize the growing importance of causality in machine learning, particularly in the context of recommender systems. The paper examined the limitations of the traditional offline learning approach from observational data, due to the inefficacy of common debiasing techniques such as Inverse Propensity Weighting, which often lead to inaccurate estimations. To address these challenges, the paper provides a review of existing debiasing strategies and illustrates their limitations through several toy examples. It further advocates for integrating causal frameworks, specifically potential outcomes and structural causal models, into recommender systems, which enhances the research and development of recommender systems by enabling causal discovery from offline data. This helps to improve the construction of causal graphs, computation of counterfactuals, and refinement of debiasing strategies in recommender systems.

In the paper “Ranking the Causal Impact of Recommendations under Collider Bias in k-Spots Recommender Systems,” de Villa et al. [2] discuss an advanced perspective on the role of recommender systems, moving beyond mere personalization to analyze the broader impact of past recommendations on user behavior. The paper examines how insights from recommendation data can inform decisions about item catalog management, such as identifying highly impactful recommendations or underperforming items. A key challenge addressed in the research is the need to evaluate the causal impact of item recommendations on user outcomes. The paper advances from existing studies that focus on confounding bias to the study of collider bias, which is significant in scenarios where multiple items are recommended simultaneously. To overcome this, the paper introduces a new method for estimating the differences in impact among items, which is applicable regardless of the outcome or type of recommender system, provided there is some level of randomization. The approach is validated through simulations and real-world application using data from a digital healthcare app.

In the paper “An Explicitly Weighted GCN Aggregator based on Temporal and Popularity Features for Recommendation,” Li et al. [3] explore the enhancement of graph convolutional networks (GCN) for recommender systems (RS). GCNs traditionally aggregate neighbor information in a network, which models user-item relevancy by assigning uniform or trainable weights, often based on implicit features such as user preferences. This approach, however, typically overlooks explicit features such as temporal aspects, which can be crucial for enhancing a model’s representational ability and explainability. To address this limitation, the paper proposes a GCN-based framework that incorporates explicit features derived with specific intentions into the recommendation process. The framework specifically focuses on temporal evolution and popularity bias as key features. It introduces a Temporal and Popularity weighted Aggregator (TPA) that uses an Interest-Forgetting Curve to adjust weights over time and to manage weights based on item popularity.

In the paper “GRIDS: Personalized Guideline Recommendations while Driving Through a New City,” Pargal et al. [4] address the challenges drivers face when navigating unfamiliar city environments, which can vary significantly in terms of road conditions and traffic rules. The paper introduces an innovative solution called GRIDS (Guidelines Recommendation for Inter-domain Driving Safety). This model leverages explainable domain adaptation techniques to understand different driving environments and provides drivers with personalized driving guidelines. These recommendations are categorized into four major feature categories relevant to driving conditions. The effectiveness of GRIDS is validated using the Carla driving simulator, which demonstrates that the model’s recommendations significantly enhance driving safety.

The special issue serves as one of the initial attempts to promote the research and development of causal inference for recommender systems. The future of causal inference in recommender systems presents numerous promising directions for enhancing system robustness and accuracy. Research can delve into counterfactual reasoning to enable robust prediction and evaluation, develop debiasing techniques to correct systemic biases, and explore causal inference methods for large language model-based recommendation. Additionally, the fields of causal discovery and integration with reinforcement learning offer potential for automated identification of causal relationships and dynamic decision-making. Emphasizing causal interpretability and explainability will ensure that recommender systems remain transparent and trustworthy to users. Besides, personalizing recommendations through heterogeneous treatment effects and ensuring ethical considerations through causal fairness in algorithmic recommendations are essential. Further, causal and robust evaluation methods and experimental design protocols are crucial for accurately measuring algorithmic impacts of recommender systems. Altogether, these research avenues not only advance the technical frontiers of recommender systems but also address critical aspects of user engagement, fairness, and ethical implications.