Causal Inference in Recommender Systems:
A Survey and Future Directions

Generally, SCMs abstract the causal relationships between variables into causal graphs, build structural functions and then conduct causal inference to estimate the effects of interactions or counterfactuals (Pearl, 2009b).

Causal Models. Causal models involve two essential concepts: causal graphs and structural functions. Specifically, a causal graph describes the causal relationships via a Directed Acyclic Graph (DAG), in which the nodes denote variables and the edges indicate causal relationships. According to a causal graph, structural functions are used to model the relationships. For each variable, one structural function calculates its value based on its parent nodes.

Three Typical DAGs. As shown in Fig. 2, there are three classic structures in causal graphs: chain, fork, and collider, for each of which we give an example of recommender systems. In the chain structure, $X$ affects $Y$ via the mediator $Z$. For example, in Fig 2 (a), the user features affect the user preferences, and the user preferences affect the users’ click behavior. Besides, in the fork structure, $Z$ is a confounder, affecting both $X$ and $Y$. For example, as shown in In Fig. 2(b), an item’s quality can affect both its price and users’ preferences towards it. In such a fork structure, $Z$ is defined as confounder variable. Roughly ignoring confounder $Z$ leads to spurious correlation between $X$ and $Y$. That is, products with higher prices may have larger sales on an e-commerce platform, which does not mean users prefer to spend much money. In Fig. 2(c), differently, $Z$ represents a collider, which is affected by $X$ and $Z$. For example, the users’ click behavior is affected by user preference and item popularity. Conditioning on $Y$ will lead to correct correlation between $X$ and $Z$. That is, users’ behaviors on two items with the same popularity level are only affected by their preferences.

Intervention. Given the causal graph, a basic concept of intervention can be formally defined. Specifically, the intervention on a variable $X$ is formulated with $do$-calculus, $do(X=x)$ (Pearl, 2009b), which blocks the effect of $X$’s parents and set the value of $X$ as $x$. For example, $do(X=x)$ in Fig. 2(b) will rule out the path $Z\rightarrow X$ and force $X$ to be $x$ (Pearl and Mackenzie, 2018). That is, in our above-mentioned example, we set the item prices to a specific value.

Counterfactual. Another important concept is the counterfactual, which contrasts with the factual. It is used to address scenarios where the treatment variable’s value settings do not occur in the real world. In other words, counterfactual inference estimates what the outcome would have been if the treatment variable had taken on a different value compared to its observed value in the real world (Pearl, 2009b). For example, a bankrupted seller might wonder about potential sales in a counterfactual world where he/she had purchased advertisement services, setting the treatment variable $T_{\textbf{if\_ads}}=1$.

The potential outcome framework (Rubin, 1974) is another widely-used causal inference framework besides the structural causal model (Pearl, 2009b). It estimates the causal effect of a treatment variable on an outcome variable without the need for a causal graph.

Potential Outcome (Rubin, 1974). Given the treatment variable $T$ and the outcome variable $Y$, the potential outcome $Y_{t}^{i}$ denotes the value of $Y$ under the treatment $T=t$ for individual $i$. In the factual world, we can only observe the potential outcome of $Y$ under one treatment for each individual.

Treatment Effect (Rubin, 1974). Given binary treatments $T=0$ or $1$, the Individual Treatment Effect (ITE) for an individual $i$ is defined as $Y_{1}^{i}-Y_{0}^{i}$. However, ITE is impossible to calculate since we can only observe one potential outcome. Hence, ITE is extended to Average Treatment Effect (ATE) over a population. For a population $i=\{1,2,...,N\}$, ATE is calculated by $\mathbb{E}_{i}\left[Y_{1}^{i}-Y_{0}^{i}\right]=\frac{1}{N}\sum_{i=1}^{N}\left(Y_{1}^{i}-Y_{0}^{i}\right)$.

For estimating the causal effect, one golden rule is to conduct randomized experiments. Since individuals are divided into the treatment group and the control group randomly, there are no unobserved confounders. Under randomized experiments, some favorable properties of causal inference are guaranteed, such as covariate balance and exchangeability. Meanwhile, the causal effect can be estimated directly by comparing the two groups. For example, online A/B testing can be regarded as a kind of randomized experiment that divides users randomly into several groups and can obtain trustworthy evaluation results of recommendation performance.

However, randomized experiments can be expensive and sometimes impossible to conduct. For example, in recommender systems, experiments generating randomized recommendations can detrimentally affect user experiences and the platform’s profitability. Therefore, estimating the causal effect solely from observational data becomes critical. In general, a causal estimand is first transformed into a statistical estimand with a causal model like SCM. Then the statistical estimand is estimated with observed data. In other words, with the defined causal model, we can discern causal effects and non-causal effects, such as confounding associations between treatment and outcome. Subsequently, the causal effect is extrapolated by estimation using observed data in alignment with the identified causal mechanisms.

One classical method is backdoor adjustment (Pearl, 2009b). We say a set of variables $W$ satisfies backdoor criterion if $W$ contain no descendant of T and $W$ can block backdoor paths (which has arrow into $T$ rather than from $T$) between $T$ and $Y$. The causal effect of $T$ on $Y$ then can be obtained with backdoor adjustment as follows,

where $w\in W$ and the total causal effect is the weighted sum of the conditioned causal effect.

The above backdoor adjustment can address observed confounders, but not unobserved confounders, where frontdoor adjustment (Pearl, 2009b) comes to help. We say a set of variables $M$ satisfies frontdoor criterion if all the causal paths from treatment variable $T$ to the outcome variable $Y$ are through $M$, and there is no unblocked backdoor path from $T$ to $M$, as well as $M$ to $Y$ when conditioned on $T$. The causal effect of $T$ on $Y$ then can be obtained with frontdoor adjustment as follows,

where possible unobserved confounders are addressed.

With the sufficient adjustment set of variables $W$ in the high dimension, it is difficult to directly estimate the causal effect as the positivity property is hard to satisfy. Instead of modeling the whole set $W$, we can turn to the propensity score as follow,

which indicates the probability of receiving the treatment given $W$. Then the causal effect can be estimated by inverse propensity weighting (IPW) (Hirano et al., 2003) on the treatment and control group as follows,

All of the above causal effect estimations assume that we already have a causal graph. However, in the real world, we often lack prior knowledge about the causal relationships in collected data. This limitation gives rise to the problem of causal discovery, where the objective is to construct a causal graph from the existing data of a set of variables. Traditional approaches identify causal relations through conditional independence tests, bolstered by additional assumptions such as faithfulness (Spirtes et al., 2000). Score-based algorithms (Heckerman et al., 1995; Schwarz, 1978) have been also proposed to relax the strict assumptions for causal discovery. These methods utilize a score function to measure the quality of the discovered causal graph in comparison with observed data. Recently, various machine learning approaches have been developed to discover causal relations from large-scale data. For example, Zhu et al. (Zhu et al., 2019) utilize reinforcement learning method to find an optimal DAG with respect to a scoring function and penalties on acyclicity. There is a survey (Guo et al., 2020a) fully discusses different methods of causal discovery.

To summarize it, we have introduced the fundamental knowledge of causal inference, including two basic frameworks and two important research topics, causal effect estimation and causal discovery.

As an approach to information filtering, the recommender system has been widely deployed on various platforms in recent decades, such as TikTok, YouTube, Twitter, etc. In general, the modeling of user preferences based on historical interactions is the key point for the recommendation algorithm, and users’ future interactions are further predicted. In this way, the necessary data input of a recommendation task includes the records of user-item interactions, and the output is a model that can generate the interaction likelihood of a given user-item pair. This procedure can be formulated as,

where $\mathcal{U}$ and $\mathcal{I}$ denotes the user set and item set, respectively. $y_{u^{\prime},i^{\prime}}=1$ if user $u^{\prime}\in\mathcal{U}$ has interacted with item $i^{\prime}\in\mathcal{I}$; if not, $y_{u^{\prime},i^{\prime}}=0$; here the function $f(\cdot,\cdot)$ denotes the recommendation model. Furthermore, with different input data, there are two primary families of models in recommendation, i.e., collaborative filtering (CF) and click-through rate (CTR) prediction. Despite the vanilla CF which only considers user-item interaction data, some recommendation tasks enhance the behavioural data with auxiliary data, such as social network in social recommendation (Wu et al., 2019b; Fan et al., 2019b), behavioral sequences in sequential recommendation (Chang et al., 2021; Zhu et al., 2021), multiple-type behaviors in multi-behavior recommendation (Jin et al., 2020; Zhang et al., 2020b), multi-domain user behaviors in cross-domain recommendation (Hu et al., 2018; Gao et al., 2019a), etc. For CTR prediction problem, user and item features such as user profiles (occupation, age) and item attributes (category, brand) are also considered as input. The mainstream works of CTR prediction focus on extracting high-order cross-features with attention-based neural network (Guo et al., 2017), attention-based neural network (Guo et al., 2017), self-attentive layers (Song et al., 2019), etc.

Here we present two folds of design of recommendation models, collaborative filtering and click-through rate prediction.

The primary objective functions for optimization utilized in recommendation models are in two categories, i.e., point-wise and pair-wise. Specifically, the point-wise objective function focuses on the prediction of a user-item interaction of which the widely-used Logloss function is as follows,

where $\hat{y}_{u,i}=s(u,i)$ and $\mathcal{O}$ is the training set.

In terms of the pair-wise objective function, it encourages a larger disparity between positive ($y_{u,i}=1$) and negative ($y_{u,j}=0$) samples, and the widely-used BPR loss function (Rendle et al., 2009) is as follows,

where $\sigma(\cdot)$ denotes the sigmoid function, and $\mathcal{Q}$ denotes the training set.

Data bias refers to the uneven distribution of recommendation data that does not faithfully reflect user preference. Generally, there are two main types of data bias in recommendation over interactions and attributes.

Causal theory enables us to identify the root cause of data bias by scrutinizing the generation procedure of recommendation data and mitigating the impact of bias through causal recommendation modeling.

The data utilized in recommender systems is typically limited, which cannot cover all possible user-item feedback. For example, a user has only rated a small ratio of clicked movies; or the user purchasing the camera is not recorded as having bought a camera lens and a roll film, which is intuitively reasonable. Therefore, the obtained data cannot fully represent the users’ interest, leading to sub-optimal results for existing recommendation methods. First, the interaction data observed is constrained by the already-deployed recommendation policy of the recommender system (Saito, 2020a). Users can only interact with specific items if these items are exposed to them, which strongly correlates with the recommender system’s intrinsic strategy. In addition, users may refuse to give feedback (Wang et al., 2019b). For example, on movie rating websites such as IMDB or Douban²²2https://www.douban.com, users may only rate a few of the movies they have watched. Under this condition, it becomes more challenging to model users’ interests. Besides, features of users and items can also be missing in real-world recommender systems due to the high cost of feature collection.

Some earlier approaches (Steck, 2013; Schnabel et al., 2016a; Thomas and Brunskill, 2016) without causal inference were developed to address the data-missing problem. Steck (Steck, 2013) computes prediction errors for missing ratings. Schnabel et al. and Thomas et al. (Schnabel et al., 2016a; Thomas and Brunskill, 2016) consider weights for each observed rating based on the probability of collecting that record. However, these methods are limited by low accuracy and poor generalization ability. Causal inference actually provides the causal descriptions of how data is generated, which can serve as prior knowledge to data-driven models. As a result, the negative impact of data-missing issues can be alleviated, improving accuracy and generalization ability.

The recommender systems highly rely on the historical user-item interaction feedback to model users’ preferences and predict the interaction probability between the user and the unseen item; thus, the reliability of collected data is the basis of the effectiveness of recommender systems. However, the data collected in the real world may be noisy, i.e., incorrect. It is hard to detect and eliminate noisy interactions in traditional recommendation methods. Mahony et al. (O’Mahony et al., 2006) classified data noise into two categories: natural noises and malicious noises. Natural noise relates to the noise generated during the data-collection procedure by recommender systems, and malicious noise denotes the noise being deliberately inserted into the system.

As for the natural noise, Li et al. (Li et al., 2019) discussed various reasons that lead to the noisy data in recommender systems. The major reasons include the inaccurate impression of the users themself and the error in data collection. Jones et al. (Jones et al., 2011) points out that users can hardly accurately measure their preferences, thus leading to mismatch between their preferences and final ratings. Cosley et al. (Cosley et al., 2003) found that noisy data arises when users map their opinions into discrete ratings. Zhang er al. (Zhang et al., 2021b) argued that in some streaming applications, the conversion events may be delayed to the time when data is collected. Thus the feedback of users may have not yet occurred, resulting in a large number of incompletely labeled instances and introducing noise to data. Some existing work (Wang et al., 2021a; Hu et al., 2008; Lu et al., 2018; Wen et al., 2019) also pointed out the difference between the implicit feedback and users’ actual satisfaction because of noisy interactions. For example, in E-Commerce, many clicks do not lead to purchases, and a large portion of purchases finally get negative comments. Implicit interaction data widely used in recommender systems nowadays is easy to become noisy because of the inaccurate first impression of users. Since users are exposed to a flood of information in today’s online services, users are very likely to have accidentally triggered feedback such as click-by-mistake.

As for the malicious noise, it is produced by adversary attackers of recommender systems. For instance, on user-generated platforms such as TikTok³³3https://www.tiktok.com, some authors will create plenty of new accounts to rate their work with high scores, trying to earn over-exposure opportunities. In e-commerce websites such as Amazon, some adversary sellers may generate fake order records or positive comments on their products.

Many previous works have experimentally demonstrated the severity of data noise and its negative effects on recommender systems. Cosley et al. (Cosley et al., 2003) showed that only 60% of users will keep their rating to the same movie when they are asked to re-rate for it. Further experiments show that statistically significant MAE differences arise when exploiting CF models on the original rating data and new rating data. Amatriain et al. (Amatriain et al., 2009) showed that the recommendation performance will be significantly affected under noisy data compared to the noiseless data, with a difference of RMSE of about 40%. Wang et al. (Wang et al., 2021a) found through experiments on two representative datasets the performance of recommender system trained by noisy data experienced a performance drop of 9.56%-21.81% w.r.t. Recall@20 and drop of 3.92%-8.81% w.r.t. NDCG@20, compared with the recommender system trained over cleaned data. Although existing work has confirmed the widespread existence of data noise, which reveals that we need to consider its impact during training recommendation models, existing solutions are a few. Data noise can arise from various sources, such as limitations in data collection (e.g., inaccurate values in users’ questionnaire data) or during data preprocessing (e.g., crudely transforming feedback into simplified values, like converting continuous user watching durations into discrete positive/negative labels). Such noises present significant challenges in accurately discerning user preferences. Leveraging causal inference allows us to more effectively detect the presence of noise in interaction data or bridge the disparity between noisy training data and the expected clean testing data with the help of counterfactual learning and reasoning.

The requirement of explainability for recommender systems refers to the need that we should understand why some items are recommended while others are not. It helps build a bridge between users and recommendation lists for better transparency and trustworthiness. Specifically, it can be divided into two categories, explainable recommendation model and explainable recommendation results. Some existing work (Wang et al., 2018; Chen et al., 2020; Zhang et al., 2014) mainly took some item aspects to give explanations, which is concluded as the aspect-aware explainable recommendation. For example, Wang et al. (Wang et al., 2018) learned users’ preferences on given aspects by factorization method to get the aspect-aware explanations.

Filter bubble describes the phenomenon where people tend to be isolated from diverse content and information by online personalization (Pariser, 2011a). As a consequence, users are placed in a fixed environment where they can only encounter similar topics or information. Passe et al. (Passe et al., 2017) attribute this effect to homogenization, which means people’s behavior and interest show consistency and convergence.

The recommender system is one of the main causes of the filter bubble due to the principle of generating recommendation lists by learning the similarity between users or items (Pariser, 2011b), which inevitably leads to homogeneous recommendations. Gabriel Machado Lunardi et al. (Lunardi et al., 2020) empirically analyzed the filter-bubble formation based on popular CF methods and algorithms for diversified recommendation. In terms of human nature, researchers found that people tend to pursue a comfort zone and stay with the opinions they are interested in or agree with (Bozdag et al., 2014). In the long term, the filter bubble will narrow people’s views and radicalize their ideas. Thus, it is an urgent problem to break filter bubbles and improve recommendation heterogeneity.

Recently, the fairness of recommendations has gained significant attention. As we know, recommender systems operate as multi-stakeholder platforms, thus encompassing various aspects of fairness concerns, including both user-side and item-side (Burke, 2017).

The user-side fairness issue arises from the diverse fairness concerns among users. For instance, while some users may be predominantly worried about potential biases based on their gender, others might be more concerned about age-related biases (Li et al., 2021). To foster trust in the recommender system, it’s essential to address these user-side fairness concerns in a personalized manner. While some approaches (Hardt et al., 2016) have attempted to rectify these fairness challenges using association-based methods—aimed at eliminating statistical metric discrepancies between groups—research has shown these methods to be inadequate and lacking in certain areas (Khademi et al., 2019; Kusner et al., 2017). Notably, these association-based techniques often overlook the intricate relationship between objective features and model outputs. Conversely, a few studies have explored fairness through a causal lens, offering insights into how output variables evolve with changes in input (Junzhe Zhang and Elias Bareinboim, 2018; Junzhe Zhang and Elias Bareinboim, 2018).

Item-side fairness, on the other hand, evaluates the equity in treatment of each item during the recommendation process. Biases may emerge due to the oversight of particular items or their attributes. Several existing solutions (Gruson et al., 2019; Schnabel et al., 2016b) have ventured into unbiased learning or heuristic ranking adjustments to rectify these biases.

In short, we have systematically discussed the limitations of existing recommender systems and why causal inference is essential to address these limitations. In the following, we will introduce how these challenges can be addressed (at least partially addressed) by presenting the recent advances in the causality-enhanced recommendation.

In most cases, biases are caused by confounders, which lead to confounding effect in correlations estimated from the observations. To eliminate the confounding effect, there are mainly two lines of research regarding the causal inference frameworks adopted.

For example, Zhang et al.  (Zhang et al., 2021a) ascribed popularity bias to the confounding of item popularity, which affects both the item exposure and observed interactions. They then introduced backdoor adjustment to remove the confounding popularity bias during model training, and incorporated an inference strategy to mitigate popularity bias. Besides, Wang et al.  (Wang et al., 2021b) explored the bias amplification issue of recommender systems, i.e., over-recommending some majority item categories in users’ historical interactions. For instance, recommender systems tend to recommend more action movies to users if they have interacted with a large proportion of action movies before. To tackle this, Wang et al.  (Wang et al., 2021b) found that the imbalanced distribution of item categories is actually a confounder, affecting user representation and the interaction probability. Next, the authors proposed an approximation operator for backdoor adjustment, which can help alleviate the bias amplification.

However, the assumption of observed confounders might be infeasible in recommendation scenarios. To tackle the unobserved confounders (e.g., the temperature when users interact with items), frontdoor adjustment is a default choice (Pearl, 2009b). Xu et al.  (Xu et al., 2021) has made some initial attempts to address both global and personalized confounders via frontdoor adjustment. Zhu et al. (Zhu et al., 2022) gave a more detailed analysis of the conditions to apply the frontdoor adjustment in recommendation. Liu et al. (Liu et al., 2022) approached the selection bias challenge and proposed counterfactual learning-based method. Specifically, the authors focus on policy learning approaches for top-K recommendations in extensive item spaces, identifying key challenges like importance weight explosion and observation scarcity. A novel framework is introduced for efficient policy learning that addresses these complexities. Ding et al. (Ding et al., 2022) emphasizes the challenge of unmeasured confounders in recommender systems which can influence the accuracy of feedback predictions. The authors proposed Robust Deconfounder (RD) to consider the effects of these unmeasured confounders on propensities, using a bounded effect approach.

Nevertheless, estimating the proper propensity score is non-trivial and typically suffers from high variance. To address these issues, a line of research (Wang et al., 2019b; Saito, 2020b; Guo et al., 2021) pursues a doubly-robust model estimator by augmenting Equation 14 with an error imputation model, which is formulated as:

where $\hat{e}_{u,i}$ is the output of the imputation model with user-item features as inputs. To learn the parameter of the imputation model, a joint learning framework (Wang et al., 2019b) optimizes:

Undoubtedly, incorporating experimental data, i.e., interactions from randomized controlled trial (RCT) such as random exposure, can enhance the doubly-robust estimator. In this light, a line of research (Wang et al., 2021d; Chen et al., 2021) investigates data aggregation strategies, which largely focuses on tackling the sparsity issue of experimental data since RCT is costly. Recently, Li et al. (Li et al., 2023) considered that the exposure bias is largely depends on the socially-connected users, and proposed IPS-based methods with the auxiliary social network data.

Different with the existing works for specific kinds of bias, He et al. (He et al., 2023) studied how to address general feature biases. This work identified a challenge in recommender systems where some features, like video length, can bias user interaction data and misrepresent actual preferences. Approaching from a causal perspective, the study introduced the Deconfounding Causal Recommendation (DCR) framework to address this bias. The DCR used backdoor adjustment to counteract the effects of confounding features and combine it with the mixture-of-experts (MoE) model architecture.

We can discover many collider structures (cf. Fig 2(c)) in the interaction generation process by inspecting the causal relationships. A representative case is that many different variables affect the observed interactions, such as user interests and conformity. Conditioning on the collected user interactions will lead to the correlation between user interests and conformity: an interaction caused by user conformity has a higher probability of being uninterested. To mitigate the conformity bias, an existing work (Zheng et al., 2021) disentangles the interest and conformity representations by training over cause-specific data, which improves the robustness and interpretability of user representations.

Another SCM-based technique used for debiasing is counterfactual inference. In some SCMs of recommender systems, two causes (user and item features) lead to one effect (user behavior). If the user features or item features are significantly biased, this direct path (which we inaccurately referred to as a ”shortcut” in our original version) can result in biased interaction learning, especially when other unbiased features take a more indirect route. The counterfactual inference is able to estimate the path-specific causal effect and eliminate the causal effect of partial user/item features. Specifically, it first imagines a counterfactual world without these features along specific paths and then compares the factual and counterfactual worlds to estimate the path-specific causal effect. For example, Wang et al.  (Wang et al., 2021c) conducted counterfactual inference to remove the effect of exposure features (e.g., attractive titles) for mitigating clickbait issues. In addition, Wei et al.  (Wei et al., 2021) reduced the direct causal effect from the item node to the ranking score to alleviate popularity bias. Furthermore, Xu et al. (Xu et al., 2020) proposed an adversarial component to capture the counterfactual exposure mechanism and optimized the candidate model over the worst-case scenario with a min-max game between two recommendation models.

Interactions between users and items are the factual data, which expresses what really happens on the recommendation platforms and directly reflects user interest. However, factual data is usually scarce; thus, it is insufficient for recommender systems to accurately capture the user interest hidden in the data. The natural idea is to generate more samples that did not actually happen to augment the training data. Such data augmentation aims to answer a question in counterfactual world: “what would … if …”, which has been adopted in several research fields like computer vision (Fu et al., 2020), and natural language processing (Zmigrod et al., 2019). In terms of recommendation, counterfactual data augmentation aims to generate more interactions under situations that are different from the real cases when the factual data is collected.

Existing approaches answer counterfactual questions for the following recommendation scenarios,

• •

  Collaborative Filtering (Top-N Recommendation). In this scenario, users are provided with a ranked list of items, and they will interact with several items in the list. Data augmentation generates the feedback of unseen recommendation lists; thus the counterfactual question is “what would the given user’s feedback be if the system had provided a different recommendation list?” (Yang et al., 2021).

• •

  Sequential Recommendation. In this scenario, recommendation is made according to the historical interaction sequences of users. In other words, interactions of the same user are regarded as a sequence ordered by the timestamp of each interaction. Augmented data are interaction sequences that do not exist in the real scenario. Therefore, the counterfactual question is “what would users behave if their interaction sequences were different?”(Zhang et al., 2021c; Wang et al., 2021e).

• •

  Feature-based Recommendation. In this scenario, not only interactions but also features such as user profiles and item attributes are available for recommendation. In other words, user preference modeling can rely on the user/item features. The counterfactual question that data augmentation aims to answer is “what would the given user’s feedback be if his/her feature-level preference had been different?” (Xiong et al., 2021).

For all the above three scenarios, counterfactual data augmentation follows a similar paradigm of three steps, modeling, intervention, and inference. Fig. 6 provides a brief illustration of counterfactual data augmentation, and we will now introduce these three steps separately.

The modeling step captures the data generation process, which can be the recommendation model itself or another separate model. Specifically, it is usually a parametric model that is trained to fit factual data. In other words, given specific users and items that exist in factual data, the model serves as a simulator that is trained with the observed interactions and later generates unobserved interactions. For example, Yang et al. (Yang et al., 2021) first constructs a structural causal model to express the process of recommendation and then implements the SCM with an inner product between user and item embeddings. Xiong et al. (Xiong et al., 2021) utilizes a multi-layer neural network that takes feature vectors of users and items as input, then uses merging operators such as element-wise product or attention to fuse user and item feature-level properties. Zhang et al. (Zhang et al., 2021c) and Wang et al. (Wang et al., 2021e) propose model-agnostic counterfactual data augmentation thus the model can be off-the-shelf sequential recommendation models. The simulator is trained with existing factual data just as a normal recommendation task. After a well-trained simulator is obtained, the input is intervened to be different from factual cases, and the simulator is used to produce the counterfactual outcome. Gao (Gao et al., 2023) studied counterfactual interactive recommender system (CIRS), which combines offline reinforcement learning with causal inference. The authors used a causal user model derived from historical data to understand the overexposure effect on user satisfaction, with which model, the RL policy can be better planned.

In the intervention step, the input is set as different values from the factual data. Specifically, this step generates the counterfactual cases either by heuristic or another learning-based model. Heuristic-based counterfactual intervention is usually achieved by randomization. In (Wang et al., 2021e), a counterfactual interaction sequence can be generated by replacing an item at a random index with a random item. In (Zhang et al., 2021c), dispensable and indispensable items are replaced with random items to construct counterfactual positive and negative sequences, respectively. In contrast, the learning-based counterfactual intervention aims to construct more informative samples as data augmentation. In other words, it generates counterfactual data with higher importance for model optimization. For example, in (Yang et al., 2021), a counterfactual recommendation list is generated by selecting items with larger loss value i.e. the hard samples. In (Wang et al., 2021e) and (Xiong et al., 2021), items and feature-level preferences that are at the decision boundary are selected and then modified with minimal change to construct more effective counterfactual interaction sequences and input features, respectively.

In the inference step, counterfactual outputs are generated with the above counterfactual inputs and simulator. This step which uses the simulator to simulate the output of the intervened input, is usually straightforward. In (Yang et al., 2021), the counterfactual clicked items of the intervened recommendation list are generated by inferring according to the constructed SCM. In (Zhang et al., 2021c) and (Wang et al., 2021e), the intervened interaction sequences are fed into the sequential backbone model, and the obtained outputs can directly serve as the counterfactual user embeddings (Zhang et al., 2021c), or they can be used to derive counterfactual next items (Wang et al., 2021e).

Wang et al. (Wang et al., 2022b) further considered the problem of out-of-distribution recommendation, i.e., the data in another distribution is missing. The authors proposed to use a variational auto-encoder to help learn the user representations in the counterfactual distribution. Mu et al. (Mu et al., 2022) proposed to use counterfactual generator to obtain user-item interaction data with the item’s specific relation on the knowledge graph is changed. The counterfactual generator and recommender can be trained jointly to enhance each other.

A recent work (He et al., 2022) approaches the issue of data-missing from another perspective, causal discovery. Specifically, this work delves into the vulnerability of current recommender systems to distribution shifts (the missing of IID data). The authors propose a novel causal preference-based recommendation framework named CausPref, integrating a recommendation-specific DAG learner. With emphasizing causal learning of invariant user preference and anti-preference negative sampling, CausPref shows superiority and interpretability in varied OOD settings.

Interactions can be noisy or incorrect due to the tight time window of data collection. For example, users’ feedback can be delayed after the immediate interaction, such as purchasing an item a few days after adding it to the shopping cart. In real-time recommendation, these samples are used for model training before the complete reward is observed. Therefore, the reward at an early time is noisy, and whether the item will be purchased is unknown when it is added to the shopping cart. Zhang et al. (Zhang et al., 2021b) tackle the above problem of delayed feedback with the help of causal inference. Specifically, the authors utilize importance sampling (Bottou et al., 2013; Yasui et al., 2020) to re-weight the original reward and obtain the modified reward in counterfactual world.

In addition, noisy user feedback can be alleviated by incorporating reliable feedback (e.g., ratings). However, reliable feedback is usually sparse, leading to insufficient training samples. To solve the sparsity issue, Wang et al.  (Wang et al., 2021a) contributed a colliding inference strategy, which leverages the colliding effect (Pearl, 2009b) of reliable feedback on the predictions to facilitate the users with sparse reliable feedback.

The recommender systems impact data collection, resulting in the absence of real interaction data, as mentioned above. Existing recommendation approaches are primarily evaluated and trained using interaction data, where typically, more interactions with recommended items indicate a more successful recommendation. However, they overlook the fact that some items may be interacted with by users even without a recommendation. Take e-commerce recommendation as an example: users may have clear intentions and directly purchase the items they desire. On the contrary, some items are more effective in terms of recommendation, meaning that users will purchase these items if recommended but won’t purchase them if not recommended. Consequently, recommender systems boost the purchase probability of these effective items, referred to as uplift. These items reflect a stronger causal effect of recommendation, emphasizing the importance of recommending more items with a larger uplift.

Some studies (Sato et al., 2019; Sato et al., 2020; Sato, 2021; Xie et al., 2021) in recent years investigated the causal effect of recommender systems from the perspective of uplift. Sato et al. (Sato et al., 2019) applied the potential outcome framework to obtain the average treatment effect (ATE) of recommendation. Specifically, all the interactions are divided into four categories according to the treatment (recommendation) and the effect (user feedback), and then a sampling approach named ULO is proposed to learn the uplift of each sample. IPW was adopted to achieve unbiased offline learning (Sato et al., 2020) and online evaluation (Sato, 2021) on the causal effect estimation of recommendation. Xie et al. (Xie et al., 2021) proposed to estimate the uplift with tensor factorization by regarding treatment as an extra embedding, and they use regression discontinuity design (RDD) analysis to simulate randomized experiments. Xiao el al. (Xiao and Wang, 2022) proposed a doubly-robust estimator, along with which a deep variational information bottleneck method is proposed to aid the adjustment of causal effect estimation.

Other studies view the causal effect of the recommendation algorithm as a problem related to off-policy evaluation. In reinforcement learning, the policy determines how the agent behaves (i.e., selecting the action) given the environmental context and the current states. In response, the environment provides the corresponding reward (Levine et al., 2020). However, due to high costs and limitations in data collection, it is challenging to collect all possible rewards for every action. Consequently, researchers have proposed off-policy evaluation, aiming to estimate these rewards (Agarwal et al., 2017; Sachdeva et al., 2020). In the context of recommendation, the items recommended can be viewed as the policy, and the off-policy evaluation is understood as estimating the effect of the deployed algorithm. This is akin to uplift modeling but focuses more on a general framework that estimates rewards using historical data. To achieve off-policy evaluation, there are three major categories of estimators: model-based estimators (reward regression), model-free estimators like propensity score-based methods, and hybrid estimators using doubly robust methods (Saito and Joachims, 2021). Specifically, Swaminathan et al. (Swaminathan et al., 2017) tackled the problem of slate recommendation, where an ordered set of items is recommended. They built on techniques from combinatorial bandits to estimate a policy’s performance using logged data. Li et al. (Li et al., 2018) addressed a similar issue, aiming to estimate the number of clicks for a recommendation list. They introduced click models to construct estimators that learn with statistical efficiency, and the results showed the superior performance of these constructed estimators. Mcinerney et al. (McInerney et al., 2020) studied sequential recommendation and introduced a new counterfactual estimator that accounts for sequential interactions in the rewards, achieving lower variance. Specifically, they reweighted the rewards in the logging policy to approximate the expected sum of rewards under the target policy. Kiyohara et al. (Kiyohara et al., 2022) based their work on the assumption that users interact with items sequentially, starting from the top position in a ranking, leading them to propose a Cascade Doubly Robust estimator.

Causal inference naturally can improve the explainability of recommendation, since it captures how different factors (cause) leads to recommendation (effect) rather than only the correlations. To present the existing works, we divide them into three categories as follows.

• •

  Counterfactual learning. Tan et al. (Tan et al., 2021b) proposed CountER for explainable recommendation using counterfactual reasoning. CountER explained the recommendation by highlighting the distinctions between factual and counterfactual scenarios. Specifically, CountER included an optimization task with the goal of identifying an item that minimizes the difference to the original item, thereby reversing the recommendation outcome in the counterfactual world. CountER (Tan et al., 2021b) also used causal discovery techniques to extract causal relations from historical interactions and the recommended items to enhance the explanation.

• •

  Causal graph-guided representation learning. Zheng et al. (Zheng et al., 2022) built a recommendation model based on the causal graph. The authors pre-define the causal relationships that how user behaviors (effect) are generated from users’ two parts of preferences (causes), long-term preferences and short-term ones. Long-term preferences refer to those stable and intrinsic interests, while short-term preferences refer to dynamic and temporary interests. The evolution manner is also defined for these two kinds of preferences. Based on the pre-defined causal relations, the authors proposed to assign two disentangled embeddings for two parts of preferences, and the extracted self-supervised signals make the recommendation model explainable. Si et al. (Si et al., 2022) proposed to improve the recommendation model’s explainability by decomposing model parameters into two parts: causal part and non-causal part. Specifically, it built a model-agnostic framework by using users’ search behaviors as an instrumental variable.

• •

  Causal discovery. Xian et al. (Xian et al., 2019) proposed to make use of a knowledge graph for explainable recommendation, and the paths in the knowledge graph can be used for generating explanations. For example, the reason for purchasing AirPods may be that the user has purchased an iPhone before, and iPhone and AirPods are reachable in the knowledge graph via relation has_brand and node Apple Brand. Based on the knowledge graph and users’ interaction history, the authors (Xian et al., 2019) proposed to extract causal relations by a reinforcement learning method. Specifically, the policy function of reinforcement learning is optimized to explicitly select items via paths in knowledge graph, ensuring high performance of both accuracy and explanation. Tran et al. (Tran et al., 2021) approached the problem of explanable job-skill recommendation. Specifically, it is essential to know which skill to learn to meet the requirements of the job. The authors first proposed causal-discovery methods based on different features with the employment-status label. Then the authors proposed a counterfactual reasoning method that finds the most important feature, of which the modification can lead to employment, which served as the explanations.

As mentioned earlier, focusing solely on accuracy gives rise to the issue of overly homogeneous content, resulting in the phenomenon known as the filter bubble. By leveraging causal inference, which aids in gaining a deeper understanding and explicitly modeling the causal effects of user-decision factors, recommendations with improved diversity and the reduction of the filter bubble can be achieved.

• •

  Counterfactual learning. Wang et al. (Wang et al., 2022a) proposed a causal inference framework to alleviate the filter bubble with the help of user control. Specifically, the framework allows users’ active control commands with different granularity to seek out-of-bubble contents. Furthermore, the authors proposed a counterfactual learning method that generates new user embeddings in the counterfactual world to remove user representations of out-of-date features. By constructing counterfactual representations, the recommendation can keep both accurate and diverse.

• •

  Backdoor Adjustment. Wang et al. (Wang et al., 2021b) approached the problem of homogeneous recommendation, by regarding imbalanced item distribution as a confounder between user embedding and the prediction score. Specifically, the authors used the backdoor adjustment to block the effect of the imbalanced item-category distribution in training data, partly alleviating filter bubble. The proposed method is model agnostic and thus it can be adapted to different recommendation models, including both collaborative filtering and click-through rate prediction. Xu et al. (Xu et al., 2022) employed a causal graph with loops to represent the dynamic recommendation process which leads to the filter bubble. A Dynamic Causal Collaborative Filtering ($\partial$CCF) model is proposed, which leverages back-door adjustment to estimate post-intervention user preferences and employs counterfactual reasoning to alleviate the echo chamber effect. Real-world dataset experiments validate the efficacy of the model in mitigating echo chambers, while maintaining strong recommendation performance.

The concept of achieving fairness naturally aligns with the counterfactual world in causal inference. For instance, when evaluating the fairness of a recommender system for a specific user profile, one can pose a counterfactual question: Would the recommendation results change if the user profile were altered? Li et al. (Li et al., 2021) introduced the notion of counterfactual fairness in recommendation, where modifying the value of a given feature ensures that the distribution of recommendation probabilities remains unchanged. The authors address this issue by introducing personalized fairness criteria for users. The core idea is to acquire user embeddings that are independent of specific features. To accomplish this, they propose a filtering module positioned after the embedding layer, which eliminates information relevant to sensitive features and generates filtered embeddings. Subsequently, the authors introduce a prediction module that utilizes these filtered embeddings to predict sensitive features, employing an adversarial learning approach in conjunction with the primary recommendation loss functions.