Heterophily is a prevalent property of real-world graphs and is well known to impair the performance of homophilic Graph Neural Networks (GNNs). Prior work has attempted to adapt GNNs to heterophilic graphs through non-local neighbor extension or architecture refinement. However, the fundamental reasons behind misclassifications remain poorly understood. In this work, we take a novel perspective by examining recurring inductive subgraphs, empirically and theoretically showing that they act as spurious shortcuts that mislead GNNs and reinforce non-causal correlations in heterophilic graphs. To address this, we adopt a causal inference perspective to analyze and correct the biased learning behavior induced by shortcut inductive subgraphs. We propose a debiased causal graph that explicitly blocks confounding and spillover paths responsible for these shortcuts. Guided by this causal graph, we introduce Causal Disentangled GNN (CD-GNN), a principled framework that disentangles spurious inductive subgraphs from true causal subgraphs by explicitly blocking non-causal paths. By focusing on genuine causal signals, CD-GNN substantially improves the robustness and accuracy of node classification in heterophilic graphs. Extensive experiments on real-world datasets not only validate our theoretical findings but also demonstrate that our proposed CD-GNN outperforms state-of-the-art heterophily-aware baselines.

Sequential recommender systems (SRSs) are designed to suggest relevant items to users by analyzing their interaction sequences. However, SRSs often suffer from exposure bias in these sequences due to imbalanced item exposure and varied user activity levels, creating a self-reinforcing loop favoring popular items regardless of their true relevance. Most SRSs only focus on item dependencies to address exposure bias, while overlooking user-side exposure bias and the rich semantics behind interactions. These oversights result in a limited understanding of less active users' preferences and inaccurate preference capture for less exposed items, exacerbating exposure biases. Towards this end, we propose a novel methodLLM-enhancedDualPropensity ScoreEstimation (LDPE), which synergistically integrates Large Language Models (LLMs) and causal inference. First, LDPE leverages LLMs' superior ability in capturing rich semantics from textual data and then integrates collaborative information to generate debiased semantic-rich LLM-based user/item embeddings. With these debiased item/user embeddings, LDPE estimates time-aware debiased propensity scores from both the item and user sides. These dual propensity scores can fully mitigate exposure bias by considering item popularity, user activity levels, and temporal dynamics. Lastly, LDPE employs the transformer as the backbone of our method, incorporating estimated dual propensity scores for accurately predicting users' true preferences. Extensive experiments show that our LDPE outperforms state-of-the-art baselines in terms of recommendation performance.

Conversational recommender systems (CRSs) provide personalised recommendations by strategically querying attributes matching users’ preferences. However, this process suffers from confounding effects of time and user attributes, as users’ preferences naturally evolve over time and differ among similar users due to their unique attributes. These confounding effects distort user behaviors’ causal drivers, challenging CRSs in learning users’ true preferences and generalizable patterns. Recently, causal inference provides principled tools to clarify cause-effect relations in data, offering a promising way to address such confounding effects. In this context, we introduceCausalConversationalRecommender (CCR), which applies causal inference to model the causality between user behaviors and time/user attribute, enabling deeper understanding of user behaviors’ causal drivers. First, CCR employs stratification and matching to ensure attribute asked per round is independent from time and user attributes, mitigating their confounding effects. Following that, we apply the Average Treatment Effect (ATE) to quantify the unbiased causal impact of each unasked attribute on user preferences, identifying the attribute with the highest ATE per round as the causal-based attribute, i.e., causal driver of user behaviour. Finally, CCR iteratively refines user preferences through feedback on causal-based attributes. Extensive experiments verified CCR's robustness and personalization.

Predicting user satisfaction over time is crucial in news recommendations, as users’ preferences are significantly influenced by various time-variant factors. Traditional correlation-based recommenders often suffer from redundant relationships, which can undermine their effectiveness over time. This work takes a time-aware causal approach to news recommendations, treating exposed news at a predicted time as the treatment variable and the resulting user satisfaction as the outcome variable. Capturing the evolving causal effects of exposed news items on user satisfaction poses significant challenges, particularly stemming from the need to model complex dependencies among time-variant covariates, such as news popularity and recency, as well as to effectively leverage the inherent user preferences embedded in time-invariant covariates. To these ends, we propose the CA u S al T ime-aware Rec ommender, named CAST-Rec , which accounts for the causal influences of both time-variant and time-invariant covariates. Specifically, we model the intricate causal dependencies among time-variant covariates through a series of transformer-based causal blocks. For time-invariant covariates, we utilize the semantic understanding and generative capabilities of Large Language Models (LLMs) to infer inherent user preferences while mitigating potential confounding effects. Extensive experiments demonstrate the superior performance of CAST-Rec compared to various news recommendation models and across multiple LLM implementations.

Multimodal recommendation, which utilizes rich multimodal information to learn user preferences, has attracted significant attention. Most works focus on designing powerful encoders for extracting multimodal features, and simply aggregate the learned features together to make prediction. Consequently, they have a limited capacity to learn the inter-modality knowledge including the modality-shared and modality-unique knowledge. In fact, learning the modality-shared knowledge enables us to align cross-modality data for fusing heterogeneous modality features. Learning the modality-unique knowledge is equally important when recommendation tasks only involve a small amount of shared features and the necessary information is contained within specific modality. In this article, we propose Contrastive Modality-Disentangled Learning (CMDL) to overcome this critical limitation. CMDL exactly captures the inter-modality knowledge by achieving modality disentanglement. Specifically, CMDL first disentangles the initial representation into the modality-invariant and modality-specific representations. Afterwards, CMDL introduces a novel manner of contrastive learning to approximate the MI upper bounds for achieving disentanglement regularization. Building upon the proposed regularization, CMDL encourages the modality-invariant and modality-specific representations to capture the modality-shared and modality-unique knowledge respectively and to be statistically independent to each other. Empirically, extensive experiments are conducted on benchmark datasets, demonstrating the superior performance of CMDL compared with strong multimodal recommenders.

In social networks, echo chambers form when users primarily encounter information that reinforces their existing views with limited exposure to different perspectives. This self-reinforcing isolation worsens societal issues such as division and declining public discourse. Traditional approaches attempt to mitigate echo chambers by analyzing observable interaction patterns to identify their formative mechanisms. However, they overlook unobserved implicit factors, called hidden confounders in causal inference, that significantly influence content exposure and user behaviors despite not being directly captured in the data. To address this, we propose Causal Echo Diffusion Attenuator (CEDA) , a novel framework that integrates causal learning with sequential recommendations to detect and adjust for hidden confounders in social networks. Generally, CEDA comprises four key components: (1) User Dual Modelling builds comprehensive user embeddings by combining users’ attributes and structural information to fully capture behavior patterns. (2) Causal Transformer then estimates residual embeddings that account for hidden confounders, incorporating them into the Transformer as causal adjustments for unbiased user embeddings. (3) Social Diffusion Predictor uses unbiased user embeddings to jointly optimize diffusion prediction accuracy and information diversity. (4) Targeted Interventions strategically reshapes information flows to disrupt echo chambers based on the generated prediction and diversity insights. Extensive experiments demonstrate CEDA’s superior performance in both predicting information diffusion patterns and mitigating echo chambers.

Graph collaborative filtering (GCF) has emerged as a prominent method in recommendation systems, leveraging the power of graph learning to enhance traditional collaborative filtering (CF). One common approach in GCF involves employing Graph Convolutional Networks (GCN) to learn user and item embeddings and utilize these embeddings to optimize CF models. However, existing GCN-based methods often fall short of generating satisfactory embeddings, mainly due to their limitations in capturing node dependencies and variable dependencies within the graph. Consequently, the learned embeddings are fragile in uncovering the root causes of user preferences, leading to sub-optimal performance of GCF models. In this work, we propose integrating causal modeling with the learning process of GCN-based GCF models, leveraging causality-aware graph embeddings to capture complex dependencies in recommendations. Our methodology encompasses three key designs: 1) Causal Graph conceptualization, 2) Neural Causal Model parameterization, and 3) Variational inference for the Neural Causal Model. We define a Causal Graph to model genuine dependencies in GCF models and utilize this Causal Graph to parameterize a Neural Causal Model. The proposed framework, termed Neural Causal Graph Collaborative Filtering (NCGCF), uses variational inference to approximate neural networks under the Neural Causal Model. As a result, NCGCF is able to leverage the expressive causal effects from the Causal Graph to enhance graph representation learning. Extensive experimentation on four datasets demonstrates NCGCF's ability to deliver precise recommendations consistent with user preferences.

Counterfactual explanations interpret the recommendation mechanism by exploring how minimal alterations on items or users affect recommendation decisions. Existing counterfactual explainable approaches face huge search space, and their explanations are either action-based (e.g., user click) or aspect-based (i.e., item description). We believe item attribute-based explanations are more intuitive and persuadable for users since they explain by fine-grained demographic features, e.g., brand. Moreover, counterfactual explanations could enhance recommendations by filtering out negative items. In this work, we propose a novelCounterfactual Explainable Recommendation (CERec)to generate item attribute-based counterfactual explanations meanwhile to boost recommendation performance. OurCERecoptimizes an explanation policy upon uniformly searching candidate counterfactuals within a reinforcement learning environment. We reduce the huge search space with an adaptive path sampler by using rich context information of a given knowledge graph. We also deploy the explanation policy to a recommendation model to enhance the recommendation. Extensive explainability and recommendation evaluations demonstrateCERec's ability to provide explanations consistent with user preferences and maintain improved recommendations. We release our code and processed datasets athttps://github.com/Chrystalii/CERec.

Conversational Recommender Systems (CRSs) fundamentally differ from traditional recommender systems by interacting with users in a conversational session to accurately predict their current preferences and provide personalized recommendations. Although current CRSs have achieved favorable recommendation performance, the explainability is still in its infancy stage. Most of the CRSs tend to provide coarse explanations and fail to explore the impact of minimal alterations on the recommendation decisions on items. In this paper, we are the first to incorporate the counterfactual techniques into CRS and propose a Counterfactual Explainable Conversational Recommender (CECR) to enhance the recommendation model from a counterfactual perspective. Counterfactual explanations can offer fine-grained reasons to explain users' real-time intentions, meanwhile generating counterfactual samples for augmenting the training dataset to enhance recommendation performance. Specifically, CECR adaptively learns users' preferences based on the conversation context and effectively responds to users' real-time feedback during multiple rounds of conversation. Furthermore, CECR actively generates counterfactual samples to augment the training set and thus leading to a constant improvement in recommendation performance. Empirical experiments carried out on three benchmark datasets show that our CECR outperforms state-of-the-art CRSs in terms of recommendation performance and explainability

Fairness-aware recommendation alleviates discrimination issues to build trustworthy recommendation systems. Explaining the causes of unfair recommendations is critical, as it promotes fairness diagnostics, and thus secures users’ trust in recommendation models. Existing fairness explanation methods suffer high computation burdens due to the large-scale search space and the greedy nature of the explanation search process. Besides, they perform feature-level optimizations with continuous values, which are not applicable to discrete attributes such as gender and age. In this work, we adopt counterfactual explanations from causal inference and propose to generate attribute-level counterfactual explanations, adapting to discrete attributes in recommendation models. We use real-world attributes from Heterogeneous Information Networks (HINs) to empower counterfactual reasoning on discrete attributes. We propose a Counterfactual Explanation for Fairness (CFairER) that generates attribute-level counterfactual explanations from HINs for item exposure fairness. Our CFairER conducts off-policy reinforcement learning to seek high-quality counterfactual explanations, with attentive action pruning reducing the search space of candidate counterfactuals. The counterfactual explanations help to provide rational and proximate explanations for model fairness, while the attentive action pruning narrows the search space of attributes. Extensive experiments demonstrate our proposed model can generate faithful explanations while maintaining favorable recommendation performance.

Fairness-aware recommendation eliminates discrimination issues to build trustworthy recommendation systems. Existing fairness-aware approaches ignore accounting for rich user and item attributes and thus cannot capture the impact of attributes on affecting recommendation fairness. These real-world attributes severely cause unfair recommendations by favoring items with popular attributes, leading to item exposure unfairness in recommendations. Moreover, existing approaches mostly mitigate unfairness for static recommendation models, e.g., collaborative filtering. Static models can not handle dynamic user interactions with the system that reflect users’ preferences shift through time. Thus, static models are limited in their ability to adapt to user behavior shifts to gain long-run user satisfaction. As user and item attributes are largely involved in modern recommenders and user interactions are naturally dynamic, it is essential to develop a novel method that eliminates unfairness caused by attributes meanwhile embrace the dynamic modeling of user behavior shifts. In this article, we propose Constrained Off-policy Learning over Heterogeneous Information for Fairness-aware Recommendation (Fair-HINpolicy) , which uses recent advances in context-aware off-policy learning to produce fairness-aware recommendations with rich attributes from a Heterogeneous Information Network. In particular, we formulate the off-policy learning as a Constrained Markov Decision Process (CMDP) by dynamically constraining the fairness of item exposure at each iteration. We also design an attentive action sampling to reduce the search space for off-policy learning. Our solution adaptively receives HIN-augmented corrections for counterfactual risk minimization, and ultimately yields an effective policy that maximizes long-term user satisfaction. We extensively evaluate our method through simulations on large-scale real-world datasets, obtaining favorable results compared with state-of-the-art methods.

Contrastive-learning-based neural networks have recently been introduced to recommender systems, due to their unique advantage of injecting collaborative signals to model deep representations, and the self-supervision nature in the learning process. Existing contrastive learning methods for recommendations are mainly proposed through introducing augmentations to the user-item (U-I) bipartite graphs. Such a contrastive learning process, however, is susceptible to bias towards popular items and users, because higher-degree users/items are subject to more augmentations and their correlations are more captured. In this paper, we advocate a Siamese Graph Contrastive Consensus Learning (SGCCL) framework, to explore intrinsic correlations and alleviate the bias effects for personalized recommendation. Instead of augmenting original U-I networks, we introduce siamese graphs, which are homogeneous relations of user-user (U-U) similarity and item-item (I-I) correlations. A contrastive consensus optimization process is also adopted to learn effective features for user-item ratings, user-user similarity, and item-item correlation. Finally, we employ the self-supervised learning coupled with the siamese item-item/user-user graph relationships, which ensures unpopular users/items are well preserved in the embedding space. Different from existing studies, SGCCL performs well on both overall and debiasing recommendation tasks resulting in a balanced recommender. Experiments on four benchmark datasets demonstrate that SGCCL outperforms state-of-the-art methods with higher accuracy and greater long-tail item/user exposure.

Graph contrastive learning has emerged as a powerful unsupervised graph representation learning tool. The key to the success of graph contrastive learning is to acquire high-quality positive and negative samples as contrasting pairs to learn the underlying structural semantics of the input graph. Recent works usually sample negative samples from the same training batch with the positive samples or from an external irrelevant graph. However, a significant limitation lies in such strategies: the unavoidable problem of sampling false negative samples. In this paper, we propose a novel method to utilize Counterfactual mechanism to generate artificial hard negative samples for Graph Contrastive learning, namely CGC. We utilize a counterfactual mechanism to produce hard negative samples, ensuring that the generated samples are similar but have labels that differ from the positive sample. The proposed method achieves satisfying results on several datasets. It outperforms some traditional unsupervised graph learning methods and some SOTA graph contrastive learning methods. We also conducted some supplementary experiments to illustrate the proposed method, including the performances of CGC with different hard negative samples and evaluations for hard negative samples generated with different similarity measurements. The implementation code is available online to ease reproducibility1.

In recommendation systems, the existence of the missing-not-at-random (MNAR) problem results in the selection bias issue, degrading the recommendation performance ultimately. A common practice to address MNAR is to treat missing entries from the so-called “exposure” perspective, i.e., modeling how an item is exposed (provided) to a user. Most of the existing approaches use heuristic models or re-weighting strategy on observed ratings to mimic the missing-at-random setting. However, little research has been done to reveal how the ratings are missing from a causal perspective. To bridge the gap, we propose an unbiased and robust method called DENC ( De-Bias Network Confounding in Recommendation ), inspired by confounder analysis in causal inference. In general, DENC provides a causal analysis on MNAR from both the inherent factors (e.g., latent user or item factors) and auxiliary network’s perspective. Particularly, the proposed exposure model in DENC can control the social network confounder meanwhile preserve the observed exposure information. We also develop a deconfounding model through the balanced representation learning to retain the primary user and item features, which enables DENC generalize well on the rating prediction. Extensive experiments on three datasets validate that our proposed model outperforms the state-of-the-art baselines.

Traditional recommendation models trained on observational interaction data have generated large impacts in a wide range of applications, it faces bias problems that cover users’ true intent and thus deteriorate the recommendation effectiveness. Existing methods track this problem as eliminating bias for the robust recommendation, e.g., by re-weighting training samples or learning disentangled representations. The disentangled representation methods as the state-of-the-art eliminate bias by revealing cause-effect of the bias generation. However, how to design the semantic-aware and unbiased representations for users’ true intents is largely unexplored. To bridge the gap, we are the first to propose an unbiased and semantic-aware disentanglement learning calledCaDSI(CausalDisentanglement forSemantics-AwareIntent Learning) from a causal perspective. Particularly, CaDSI explicitly models the causal relations underlying recommendation task, and thus produces semantic-aware representations via disentangling users’ true intents aware of specific item context. Moreover, the causal intervention mechanism is designed to eliminate confounding bias stemming from context information, which further aligns the semantic-aware representation with users’ true intent. Extensive experiments and case studies both validate the robustness and interpretability of our proposed model.

Recommender systems which captures dynamic user interest based on time-ordered user-item interactions plays a critical role in the real-world. Although existing deep learning-based recommendation systems show good performances, these methods have two main drawbacks. Firstly, user interest is the consequence of the coaction of many factors. However, existing methods do not fully explore potential influence factors and ignore the user-item interaction formation process. The coarse-grained modeling patterns cannot accurately reflect complex user interest and leads to suboptimal recommendation results. Furthermore, these methods are implicit and largely operate in a black-box fashion. It is difficult to interpret their modeling processes and recommendation results. Secondly, recommendation datasets usually exhibit scale-free distributions and some existing recommender systems take advantage of hyperbolic space to match the data distribution. But they ignore that the operations in hyperbolic space are more complex than that in Euclidean space which further increases the difficulty of model interpretation. To tackle the above shortcomings, we propose an E xplainable H yperbolic T emporal P oint P rocess for User-Item Interaction Sequence Generation (EHTPP) . Specifically, EHTPP regards each user-item interaction as an event in hyperbolic space and employs a temporal point process framework to model the probability of event occurrence. Considering that the complexity of user interest and the interpretability of the model,EHTPP explores four potential influence factors related to user interest and uses them to explicitly guide the probability calculation in the temporal point process. In order to validate the effectiveness of EHTPP, we carry out a comprehensive evaluation of EHTPP on three datasets compared with a few competitive baselines. Experimental results demonstrate the state-of-the-art performances of EHTPP.

Graph meta-learning which is used to deal with graph few-shot learning attracts more and more research interests. Existing graph meta-learning methods mainly focus on capturing node-level relations, but they ignore task-level relations which are beneficial for improving the performance of few-shot node classification. Furthermore, contrastive learning which can learn knowledge without labeled data is suitable for few-shot scenario, but existing graph few-shot learning methods have never exploited it. To tackle above problems, in this paper, we combine conventional graph meta-learning framework with graph contrastive learning and propose a novel joint model named -${\underline T}$asklevel -${\underline R}$elations Modelling for -${\underline G}$raph ${\underline M}$eta-learning (TRGM). By constructing auxiliary contrastive pretext tasks, TRGM can fully capture the inter-task relations (task correlation and task discrepancy) and promote the primary few-shot learning. Finally, we conduct extensive experiments on six benchmark datasets to validate the effectiveness and efficiency of TRGM. Experimental results show that our model outperforms several strong baselines and achieves the new state-of-the-art.

Off-policy learning has drawn huge attention in recommender systems (RS), which provides an opportunity for reinforcement learning to abandon the expensive online training. However, off-policy learning from logged data suffers biases caused by the policy shift between the target policy and the logging policy. Consequently, most off-policy learning resorts to inverse propensity scoring (IPS) which however tends to be over-fitted over exposed (or recommended) items and thus fails to explore unexposed items.

Reinforcement learning has recently become an active topic in recommender system research, where the logged data that records interactions between items and users feedback is used to discover the policy. Much off-policy learning, referring to the procedure of policy optimization with access only to logged feedback data, has been a popular research topic in reinforcement learning. However, the log entries are biased in that the logs over-represent actions favored by the recommender system, as the user feedback contains only partial information limited to the particular items exposed to the user. As a result, the policy learned from such off-line logged data tends to be biased from the true behaviour policy.

Abstract Causal inference is capable of estimating the treatment effect (i.e., the causal effect oftreatmenton theoutcome) to benefit the decision making in various domains. One fundamental challenge in this research is that the treatment assignment bias in observational data. To increase the validity of observational studies on causal inference, representation-based methods as the state-of-the-art have demonstrated the superior performance of treatment effect estimation. Most representation-based methods assume all observed covariates are pre-treatment (i.e., not affected by the treatment) and learn a balanced representation from these observed covariates for estimating treatment effect. Unfortunately, this assumption is often too strict a requirement in practice, as some covariates are changed by doing an intervention on treatment (i.e., post-treatment). By contrast, the balanced representation learned from unchanged covariates thus biases the treatment effect estimation. In light of this, we propose a deep treatment-adaptive architecture (DTANet) that can address the post-treatment covariates and provide a unbiased treatment effect estimation. Generally speaking, the contributions of this work are threefold. First, our theoretical results guarantee DTANet can identify treatment effect from observations. Second, we introduce a novel regularization of orthogonality projection to ensure that the learned confounding representation is invariant and not being contaminated by the treatment, meanwhile mediate variable representation is informative and discriminative for predicting the outcome. Finally, we build on the optimal transport and learn a treatment-invariant representation for the unobserved confounders to alleviate the confounding bias.

Underwriting is an important process in insurance and is concerned with accepting individuals into insurance policy with tolerable claim risk. Underwriting is a tedious and labor intensive process relying on underwriters' domain knowledge and experience, thus is labor intensive and prone to error. Machine learning models are recently applied to automate the underwriting process and thus to ease the burden on the underwriters as well as improve underwriting accuracy. However, observational data used for underwriting modelling is high dimensional, sparse and incomplete, due to the dynamic evolving nature (e.g., upgrade) of business information systems. Simply applying traditional supervised learning methods e.g., logistic regression or Gradient boosting on such highly incomplete data usually leads to the unsatisfactory underwriting result, thus requiring practical data imputation for training quality improvement. In this paper, rather than choosing off-the-shelf solutions tackling the complex data missing problem, we propose an innovative Generative Adversarial Nets (GAN) framework that can capture the missing pattern from a causal perspective. Specifically, we design a structural causal model to learn the causal relations underlying the missing pattern of data. Then, we devise a Causality-aware Generative network (CaGen) using the learned causal relationship prior to generating missing values, and correct the imputed values via the adversarial learning. We also show that CaGen significantly improves the underwriting prediction in real-world insurance applications.

Entity linking is the task of aligning mentions to corresponding entities in a given knowledge base. Previous studies have highlighted the necessity for entity linking systems to capture the global coherence. However, there are two common weaknesses in previous global models. First, most of them calculate the pairwise scores between all candidate entities and select the most relevant group of entities as the final result. In this process, the consistency among wrong entities as well as that among right ones are involved, which may introduce noise data and increase the model complexity. Second, the cues of previously disambiguated entities, which could contribute to the disambiguation of the subsequent mentions, are usually ignored by previous models. To address these problems, we convert the global linking into a sequence decision problem and propose a reinforcement learning model which makes decisions from a global perspective. Our model makes full use of the previous referred entities and explores the long-term influence of current selection on subsequent decisions. We conduct experiments on different types of datasets, the results show that our model outperforms state-of-the-art systems and has better generalization performance.

As a popular heuristic to the matrix rank minimization problem, nuclear norm minimization attracts intensive research attentions. Matrix factorization based algorithms can reduce the expensive computation cost of SVD for nuclear norm minimization. However, most matrix factorization based algorithms fail to provide the theoretical guarantee for convergence caused by their non-unique factorizations. This paper proposes an efficient and accurate Linearized Grassmannian Optimization (Lingo) algorithm, which adopts matrix factorization and Grassmann manifold structure to alternatively minimize the subproblems. More specially, linearization strategy makes the auxiliary variables unnecessary and guarantees the close-form solution for low per-iteration complexity. Lingo then converts linearized objective function into a nuclear norm minimization over Grassmannian manifold, which could remedy the non-unique of solution for the low-rank matrix factorization. Extensive comparison experiments demonstrate the accuracy and efficiency of Lingo algorithm. The global convergence of Lingo is guaranteed with theoretical proof, which also verifies the effectiveness of Lingo.