A Survey on Causal Inference for Recommendation

The Potential Outcomes Framework (aka the Neyman-Rubin Causal Model)  (Splawa-Neyman et al., 1990; Rubin, 1974; Imbens and Rubin, 2015) is the most widely used framework across many disciplines. With a hypothetical treatment (or manipulation, intervention), the causal effect, i.e., treatment effect, is defined as the difference between the potential outcomes under treatment and control for the same unit (Imbens and Rubin, 2015).

A unit can be a physical object, an individual, or a collection of objects or persons, such as a classroom or a market, at a particular point in time (Imbens and Rubin, 2015). In recommendation research, a user-item pair will usually be defined as a unit. It should be noticed that the same physical object or person at a different time is a different unit. This is a reasonable restriction, considering the same user will make different decisions at a different time even if exposed to the same item due to factors like preference shift, mood, occasion and so on.

This paper focuses on binary treatment (e.g., recommend or not), the most common setting in the recommendation field. In practice, we refer to the more active treatment simply as the “treatment” $T=1$ and the other treatment as the “control” $T=0$.

Potential Outcome. For each treatment-unit pair, the potential outcome is the outcome that the treatment is applied to the unit, denoted as $Y(T=t)$ (ignoring unit). For a unit, only the potential outcome corresponding to the treatment actually taken will be observed, denominated as observed outcome, while others are referred to as counterfactual outcomes. The fundamental problem of causal inference in PO framework is that we can never obtain both observed and counterfactual outcomes for a unit: it is impossible to realize all treatments and observe the corresponding outcomes.

Treatment Effect/ Causal effect. Treatment effect is represented by the difference between the potential outcomes under treatment and control for the same unit, formulated as:

where ${Y}(T=1)$ and ${Y}(T=0)$ are the potential treated and control outcome of the unit, respectively. Treatment effect like Equation 1 is also called Individual Treatment Effect. Furthermore, the treatment effect can be defined at the population and subpopulation levels. At the population level, Average Treatment Effect (ATE) is the expectation of ITE over the whole population (Guo et al., 2020), denoted as:

The ATE on the subpopulation level is often of particular interest; thus we define Conditional Average Treatment Effect (CATE) on the units with the same features $X=x$ as:

Assumptions. Despite the simple definition of the causal effect, the fundamental problem in causal inference, i.e., the missing data problem, appear to be a major obstacle to the estimation of the causal effect. Therefore, it is critical to make additional assumptions.

The stable unit treatment value assumption, or SUTVA (Imbens and Rubin, 2015) is the most fundamental assumption in causal inference, incorporating both the No Interference idea that treatments applied to one unit do not affect the outcome for another unit and the No Hidden Variations of Treatments concept that for each unit there is only a single version of each treatment level. The second assumption, ignorability or unconfoundedness (Rubin, 1990), states that treatment assignment is free from dependence on the potential outcomes.

In other words, within subpopulations defined by the values of observed background variables, or covariates, the treatment assignment is random. The ignorability assumption rules out unmeasured confounders, which causally influences both the treatment $T$ and the outcome $Y(T)$. $\textrm{Pr}(T=1|X)$ is called the propensity score (Rosenbaum and Rubin, 1983). The last assumption is positivity, or overlap:

In large data samples, positivity requires that there are both treated and control units for all values of the covariates. In contrast to the untestable ignorability assumption (Imbens and Rubin, 2015), positivity can be tested from observed data. The combination of unconfoundedness and positivity is referred to as “strong ignorability (Rosenbaum and Rubin, 1983).”

Structural causal models (SCM) (Pearl, 1995, 1988, 2009) serve as a comprehensive causality framework, which unifies graphical models, nonparametric structural equations, and counterfactual and interventional logic. The most significant advantage of SCM is its intuitive structure of real-world causal dependencies based on graphical models as well as the wise and friendly symbiosis between counterfactual and graphical methods.

Causal Graph. A causal graph, or a causal diagram, is usually a Bayesian network, which describes the causal relations between variables by a Directed Acyclic Graph (DAG), where the nodes represent the variables and the edges record the causal relations. Causal graphs play an essential role in the SCM framework, for they provide a vivid representation of sets of variables that are relevant to each other in any given state of knowledge, and serves as a carrier of conditional independence relationships along the order of construction, through which we can confirm whether it satisfies the criteria such that certain causal inference methods can be applied (Pearl, 2009).

d-Separation. We first review the concept of dependency-separation (d-Separation) as the knowledge base for conditional independence. There are three typical causal graphs of three disjoint sets of variables, shown in Fig. 4, with the help of which we can characterize any pattern of arrows in the network. In the chain (Fig. 4(a)), $B$ is the $mediator$ that transmits the effect of $A$ to $C$. In the fork (Fig. 4(b)), $B$ is often called a common cause or confounder of $A$ and $C$. A confounder will make $A$ and $C$ statistically correlated even though there is no direct causal link between them, which may give rise to a so-called spurious correlation in the application. In the collider (Fig. 4(c)), though $A$ and $C$ are independent to begin with, conditioning on (i.e., knowing the value of) $B$ will make them dependent. A good example is three features of Hollywood actors: Talent $\rightarrow$ Celebrity $\leftarrow$ Beauty  (Elwert and Winship, 2014). Although beauty and talent are completely unrelated to one another in the general population, an unanticipated negative correlation is found between talent and beauty if we only focus on famous actors: a celebrity is unattractive increases our belief that he or she is talented (Pearl and Mackenzie, 2018). This negative correlation is sometimes called collider bias or the “explain-away” effect. In recommendation systems, a similar example can be found in analyzing popularity bias from the user’s perspective. The factors influencing user interaction can be summarized as: Conformity $\rightarrow$ Interaction $\leftarrow$ Inherent Interest (Zheng et al., 2021). When a user interacts with a popular item, it does not necessarily indicate his or her true preference for it; such interaction may be driven by the desire to conform to prevailing trends. Conversely, if a user engages with an unpopular item — where the influence of conformity is significantly reduced — it is more probable that the item is in close alignment with the user’s inherent interests.

A path means a sequence of consecutive edges (of any directionality) in the graph, and we regard stopping the flow of dependency between the variables that are connected by such paths as blocked. In the chain and fork, the path between $A$ and $C$ will be blocked by conditioning on $B$, while in the collider, any conditioning on $B$ will introduce a correlation between them. The formal definition of d-separation or blocking is defined as follows.

Structural Equations. Beside causal graph, structural equation is another representation of causal information, where the former is an abstraction of the latter. In its general form, a structural equation of a variable $Y$ is defined as:

where $Pa$ (connoting parents) stands for the set of variables that directly determine the value of $Y$ and where $U$ represents exogenous variables or errors (or “disturbances”) due to omitted factors. For example, the causal graph in Fig. 5(a) is associated with the structural model as follows. In this section, uppercase letters are used to represent variables, while their corresponding lowercase counterparts denote the values of these variables. Thus, we have:

where $U_{A}$, $U_{B}$ and $U_{C}$ represent exogenous variables. A set of equations in the form of Equation 4 is called a structural model; if each variable has a distinct equation in which it appears on the left-hand side, then the model is called a structural causal model.

Intervention. The do-calculus allows researchers to complete intervention, interpreted as controlling the value of a variable, by purely mathematical means instead of by carrying out a physical experiment, which is one of the outstanding contributions of Pearl’s SCM framework. The do-calculus involves the do-operation, like $do(T=t)$, which denotes the intervention of setting the variable $T$ to $t$, realizing by blocking the effect of $T$’s parents on $T$ and set the value of $T$ as $t$. For example, if we $do(B=b_{0})$ on the model in Fig. 5(a), Equations 5 will be modified as:

the graphical description of which is shown in Fig. 5(b).

It is crucial to note that $\textrm{Pr}(Y=y|do(T=t))$ and $\textrm{Pr}(Y=y|T=t)$ are not the same. For example. the fork structure (Fig.4(b))might represent the causal mechanism that connects the number of sales at a local ice cream shop on that day ($A$), a day’s temperature in a city ($B$), and the number of violent crimes in the city on that day ($C$) (Glymour et al., 2016). Because both ice cream sales and violent crime are more common in hot weather, a positive correlation might be found when estimate $P(C=c|A=a)$. However, as illustrated in manipulated graphical model of Fig. 5, crime rates $C$ are independent of ice cream sales $B$, which results in a different $\textrm{Pr}(C=c|do(A=a))$ from $\textrm{Pr}(C=c|A=a)$.

Although causal graph manipulation is the most fundamentalist approach to calculating $\textrm{Pr}(Y=y|do(T=t))$, it can be challenging and even impossible in reality. Fortunately, we can estimate $\textrm{Pr}(Y=y|do(T=t))$ from observed data with the following causal effect rule:

The most important benefit brought by the rule is that it enables us to finish the do-calculus purely on passive observational data (Xu et al., 2023a). The factor $\textrm{Pr}(T=t\mid PA=x)$ is the propensity score, and Equation 7 is named inverse propensity score (Section 4.1.2) in PO framework, which partly reflects the unity of the two frameworks.

Counterfactuals. Counterfactuals are employed to emphasize our wish to compare two outcomes under the exact same conditions, differing only in one aspect: the antecedent, or hypothetical condition (Glymour et al., 2016). For example, in the counterfactual question “What would be the user’s interaction if the recommended items had been different?” mentioned above in the Section 1, we would like to compare the user’s interaction under the same conditions except for the recommended item. Counterfactuals, situations which are non-existent in reality, cannot be inferred by do-calculus. Fortunately, Pearl (Pearl, 2009) proposed a new set of notations: $\textrm{Pr}(Y(T=1)|T=0,Y=Y(T=0))$ indicates the probability of the outcome $Y(T=1)$ would be if the observed treatment value is $T=0$, given the fact that we observe $Y=Y(T=0)$ in the data.

As mentioned above, the two frameworks are equivalent logically: an assumption or a theorem can be translated to its counterpart in the other, and a problem solved in one framework would yield the same solution in another (Pearl, 2009; Glymour et al., 2016). For example, $\textrm{Pr}(Y=y|do(T=t))$ in the SCM is equivalent to $\textrm{Pr}(Y(T=t)=y)$ in the PO, which the regular assessment in a controlled experiment, in which the distribution of $Y$ is estimated for each level $w$ of a random variable $T$. Causal effects that are measured between the results of the counterfactual world and the real world can be estimated conveniently in both frameworks. However, there are several important differences between PO and SCM. The most significant difference is that PO does not assume the causal relations between concerned variables, while SCM makes assumptions of causal mechanisms among a set of variables or searches for ones based on some assumptions. In other words, any given PO model corresponds to multiple causal graphs in SCM. For PO, it can be a strength for PO that causal effects can be reasoned without knowing the causal model, and be a weakness either. According to the unconfoundednes assumption, all confounders should be observed to infer a correct treatment effect since the mechanism is unknown, almost impossible in practice (Aliprantis, 2015). In contrast, in SCM, causal diagrams allow us to work with causal effects by interventions on the fewest number of variables or the observed variables as much as possible.

RSs are usually classified into the following three categories (Adomavicius and Tuzhilin, 2005; Zhang et al., 2019): content-based, collaborative filtering (CF), and hybrid. Content-based recommendation learns to recommend primarily based on comparisons across items’ and users’ auxiliary information (Zhang et al., 2019), such as items’ human-set tags, images, texts, and users’ sex. Collaborative filtering recommender systems recommend items according to user/item historical interactions, i.e., explicit (e.g., user’s previous ratings) or implicit feedback (e.g., click behavior) (Zhang et al., 2019). Hybrid approaches are those that combine collaborative filtering and content-based methods.

If we review the model structure, recommender systems can generally be divided into shallow models and neural network models. Shallow models involve methods that directly calculate the similarity of interactions and CF methods with matrix factorization (MF) (Koren et al., 2009) or factorization machine (FM)  (Rendle, 2010), but suffer from insufficient learning of users’ complicated interest. Neural network models are proposed to solve this issue, with the advantage of high-order feature interactions (Guo et al., 2017). For example, Wide & Deep (Cheng et al., 2016) jointly trains linear models and deep neural networks to combine the benefits of memorization and generalization. Deep factorization machine (DeepFM) (Guo et al., 2017) combines traditional factorization machine (FM) with multi-layer perceptrons (MLP) in parallel. Graph neural networks (GNN)-based methods adopt embedding propagation to iteratively aggregate neighborhood embedding, thereby more effectively exploring structural information (Gao et al., 2022b).

Considering a general recommender system, we assume $\mathcal{O}$ and $\mathcal{O}^{-}$ denote the observed dataset and unobserved dataset. Each observed sample includes the treatment $T$, background features $X$, and an interaction label $Y$. Background features $X$, aka., covariates are usually formulated as a high dimensional sparse vector containing information such as user ID, item ID, user profile, item category, etc. The interaction label $Y$, or the outcome, can be explicit feedback (e.g., rating) or implicit feedback (such as click and watch behavior).

In normal circumstances, researchers prefer choosing whether to recommend as the treatment. Therefore, the observed dataset can be denoted as ${\mathcal{O}}=\{(T=1,X,Y)\}|_{1}^{|\mathcal{O}|}\in\mathcal{T}\times\mathcal{X}\times\mathcal{Y}$, where $\mathcal{T}$ means the treatment space, $\mathcal{X}$ is the feature spaces, and $\mathcal{Y}$ is the label space. In general, the observed dataset is obtained with the deployed recommender policy $\pi$; thus ${\mathcal{O}}$ will be specifically expressed as ${\mathcal{O}_{\pi}}$ if we are concerned about the policy. Note that settings of $T,X,Y$ vary slightly according to specific work. It would be better to understand with reference to the context.

In this part, we will introduce the phenomena and factors of missing not at random, to provide explanations and conclusions of challenges in recommender systems in a causal language to understand existing work better.

Recommendation algorithms often obey the missing at random (MAR) (Rubin, 1976) assumption but may lead to biased prediction and suboptimal policy (Little and Rubin, 2019; Marlin and Zemel, 2009). The MAR condition essentially states that the probability that a potential outcome is missing does not depend on the value of that potential outcome and can be easily violated in recommender systems (Marlin and Zemel, 2009). For example, on movie rating websites, movies with high ratings are less likely to be missing compared to movies with low ratings (Pradel et al., 2012). The issue of missing not at random (MNAR) has been demonstrated by Marlin and Zemel  (Marlin and Zemel, 2009) and it is a phenomena stemming from selection bias and confounding bias (Correa et al., 2019; Wu et al., 2022).

Selection bias, or sampling bias, is usually discussed in the prediction task and can be further classified into model selection bias and user self-selection bias (Wu et al., 2022). For example, the case that the platform may systematically recommend pop music to younger users who may be more active on the service regardless of genre preferences (McInerney et al., 2020) will be regarded as model selection bias (Yuan et al., 2019a; McInerney et al., 2020) and can be eliminated by random recommendation. User self-selection bias (Bareinboim and Pearl, 2012; Elwert and Winship, 2014), on the contrary, can not be removed by randomization of recommendation (Correa et al., 2019). It is caused by preferential exclusion of samples from the data (Bareinboim and Pearl, 2012). A typical example is a song recommender system, in which users usually rate songs they like or dislike and seldom rate what they feel neutral about (Saito, 2020). Some of the most frequently discussed biases like popularity bias (Zhang et al., 2021b; Wei et al., 2021) and exposure bias (Liang et al., 2016; Wang et al., 2018) will lead to model selection bias, while conformity bias (Zhang et al., 2021b; Zheng et al., 2021) and clickbait bias (Wang et al., 2021b) fall under user self-selection bias as a result of user preference.

Confounding bias (Hernán et al., 2002; Pearl, 2009) arises from the confounder described in Section 3.1.2, which affects both the treatment and the outcome, illuminated in Fig.6(b). Alternatively, it can be identified if the probabilistic distribution representing the statistical association is not always equivalent to the interventional distribution, i.e., $\textrm{Pr}(y\mid t)\neq\textrm{Pr}(y\mid do(t))$ (Guo et al., 2020). A notable example of confounding bias is that a system trained with historical user interactions may over recommend items that the user used to like, and the user’s decision (i.e., outcome) is also affected by historical interactions (Wang et al., 2021a).

Both biases can lead to invalid estimates of causality from the data, and they are not mutually exclusive because selection bias does not explicitly involve causality. Many model selection biases, including popularity bias and exposure bias, are also confounding biases. As for user self-selection bias, the model in Fig. 6 (a) gives an illustration of its causal nature in which $S$ is a variable affected by both $T$ (treatment) and $Y$ (outcome), indicating entry into the data pool (Bareinboim and Pearl, 2012). Therefore, confounding bias is significantly different from user self-selection bias from the causal perspective. The former originates from common causes, whereas the latter originates from common outcomes (Elwert and Winship, 2014). The former stems from the systematic bias introduced during the treatment assignment, while the latter comes from the systematic bias during the collection of units into the sample (Correa et al., 2019).

Inverse Propensity Score (IPS) (Horvitz and Thompson, 1952; Rosenbaum, 1987; Rosenbaum and Rubin, 1983; Little and Rubin, 2019), also named as inverse propensity weighting (IPW), or inverse propensity of treatment weighting (IPTW), is one of the favorite counterfactual techniques and has inspired a lot of causal inference methods in RS, especially for unbiased learning (Joachims et al., 2017). Propensity score is the probability of receiving the treatment given covariates $X$, formulated as:

IPS assigns a weight $w$ to each sample:

which indicates the inverse probability of receiving the observed treatment and control. The unbiasedness of IPS can be proven (Rosenbaum, 1987). More specifically, for the reward estimation of recommendation policy, IPS adjusts the distribution of background features in the logged dataset to be consistent with that during $\pi$ tests online, formulated as:

where we assume that only positive feedback is taken into account, and $w=\frac{e_{\pi}(X)}{e_{\pi_{0}}(X)}$ is the ratio of the evaluation and logged policies. Note that in most applications in RS, IPS is model-agnostic, applied to the training objective function for policy evaluation directly or for policy learning indirectly.

Much IPS-based recommendation focuses on data debiasing in user interactions, mainly selection bias  (Schnabel et al., 2016; Saito et al., 2020; Sato et al., 2020; Zhang et al., 2021e; Sato, 2021; Zhang et al., 2021g; Wu et al., 2021; Li et al., 2023a). For example,  (Schnabel et al., 2016) is a representative work adopting IPS to recommender system for the elimination of selection bias, in which the recommendation algorithm is based on matrix factorization and propensity scores are estimated via naive Bayes or logistic regression. Similarly, Saito et al. (Saito et al., 2020) estimate the exposure propensity for each user-item pair and Sato et al. (Sato et al., 2020) propose the DLCE (Debiased Learning for the Causal Effect) model with IPS-based estimators to evaluating unbiased ranking uplift. Unbiased IPS-based uplift is also concerned by  (Sato et al., 2019). In addition,  (Zhang et al., 2021e) proposes UR-IPW (User Retention Modeling with Inverse Propensity Weighting) to model revisit rate estimation accounting for the selection bias problem and  (Li et al., 2021b) adjusts domain weights based on IPS to reduce domain bias. Though IPS-based methods do not require an explicit analysis of the causal correlation between variables, some works (Christakopoulou et al., 2020; McInerney et al., 2020; Ding et al., 2022b) still discuss causal graphs as an excellent guide to accurate model. For example, Ding et al. (Ding et al., 2022b) leverage a causal graph to explain the risk of unmeasurable confounders on the accuracy of propensity estimation and propose RD (Robust Deconfounder) with the sensitivity analysis, obtaining the bound of propensity score to enhance the robustness of methods against unmeasured confounders. Li et al. (Li et al., 2023a) construct the DENC (De-bias Network Confounding in Recommendation). This causal graph-based recommendation framework disentangles three determinants for the outcomes, including inherent factors, social network-based confounder and exposure, and estimates each of them with a specific component, respectively. By the way, there are some works (Christakopoulou et al., 2020; Cai et al., 2022; Zhang et al., 2021e) integrate multi-task models with IPS to learn propensity scores and user interactions simultaneously.

In addition to debiasing, some IPS-based methods are dedicated to addressing other issues that abound in RS (Mehrotra et al., 2018; Zhang et al., 2021c; Krauth et al., 2022). For example, Mehrotra et al. (Mehrotra et al., 2018) proposes an unbiased estimator of user satisfaction based on IPS to jointly optimize for supplier fairness and consumer relevance. Besides, the CBDF (Counterfactual Bandit with Delayed Feedback) algorithm  (Zhang et al., 2021c) re-weights the observed feedback with importance sampling, which is determined by a survival model to deal with delayed feedbacks. The CAFL (causal adjustment for feedback loops)  (Krauth et al., 2022) extends the IPS estimator to break feedback loops.

Despite the unbiasedness strength of IPS, the inaccurate estimation of the unknown propensity $e(x)$ or sample weight, which results in high variance (Gilotte et al., 2018), becomes the biggest obstacle to achieving it. To alleviate this problem, modified versions of IPS have been proposed to control variance and applied to RS, including Self Normalized IPS  (Schnabel et al., 2016; Zhang et al., 2021e), Clipped IPS (Saito et al., 2020; Sato et al., 2020), Reward interaction IPS (McInerney et al., 2020), and Regularized per-Item IPS (Liu et al., 2022b). Self Normalized Inverse Propensity Scoring (SNIPS) (Swaminathan and Joachims, 2015) rescales the estimate of the original IPS without any parameters to reduce the high variance, which is:

and is introduced to RS by works like  (Schnabel et al., 2016) and  (Zhang et al., 2021e) to alleviate selection bias. Clipped IPS (CIPS) (Bottou et al., 2013; Saito et al., 2020; Sato et al., 2020), or Capped IPS, tightens the bound of the sample weight by introducing a scalar hyperparameter $\lambda_{\mathrm{CIPS}}$, formulated as:

which has a lower variance but gives away its unbiasedness. Expanding upon the groundwork established by NCIPS (Swaminathan and Joachims, 2015), which amalgamated SNIPS and CIPS, the study by Gilotte et al. (Gilotte et al., 2018) advances PieceNCIS and PointNCIS as enhancements that utilize contextual information to refine bias modeling. McInerney et al. (McInerney et al., 2020) loosen the SUTVA assumption and propose Reward interaction IPS (RIPS) for sequential recommendations, which assumes a causal model in which users interact with a list of items from the top to the bottom. RIPS uses iterative normalization and lookback to estimate the average reward and achieves a better bias-variance trade-off than IPS. In addition to high variance, violation of the Unconfoundedness assumption is another challenge of utilizing IPS in RS. That is, the treatment mechanism is identifiable (Glymour et al., 2016; Mohan and Pearl, 2021) from observed covariates due to the existence of unobserved ones, which leads to the inaccurate estimate of propensity score and the disagreement between the online and offline evaluations. To address the uncertainty brought by the identifiability issue,  (Xu et al., 2020) proposes minimax empirical risk formulation, which can be converted to an adversarial game between two recommendation models via duality arguments and relaxations.

More recently, Liu et al. (Liu et al., 2022b) propose Regularized per-item IPS (RIIPS) with an additional penalty function that constrains the difference in recommended outcomes between the deployed system and the new system so that the explosion of propensity scores can be avoided.

Doubly Robust (DR) (Funk et al., 2011; Dudík et al., 2014; Jiang and Li, 2016; Wang et al., 2019) is another powerful and effective causal method account for the MNAR issue. To understand DR, let us consider the two common-used approaches to mitigate against MNAR: direct method (DM) (Beygelzimer and Langford, 2009) and IPS (Saito et al., 2021). The former designs a model (linear regression, deep neural network, etc.) to directly learn the missing outcomes based on the observed data, which has low variance due to the advantage of supervised learning but suffers from high bias caused by unmet IID assumptions, denoted as (Saito et al., 2021):

where $\hat{y}\left(x,t\right)$ is the estimated outcomes. The latter, though unbiased theoretically, often causes training losses to oscillate stemming from the inverse of propensity with high variance (Thomas and Brunskill, 2016). What DR does is to combine the direct method and IPS, which takes advantage of both and overcomes their limitations:

DR uses the estimated outcomes to decrease the variance of IPS. It is also doubly robust in that it is consistent with the policy reward value if either the propensity scores or the imputed outcomes are accurate for all user-item pairs (Wang et al., 2019; Saito et al., 2021). By the way, advanced versions like Switch-DR (Wang et al., 2017) and DRos (Doubly Robust with Optimistic Shrinkage) (Su et al., 2020) are proposed to further control the variance.

Based on the above advantages, DR has found an increasingly wide utilization in RSs (Yuan et al., 2019a; Wang et al., 2019; Zhang et al., 2020; Guo et al., 2021; Kiyohara et al., 2022; Mondal et al., 2022; Xiao and Wang, 2022; Dai et al., 2022; Song et al., 2023a; Li et al., 2023b). Wang et al. (Wang et al., 2019) utilize DR for unbiased RS prediction and further propose a joint learning approach that simultaneously learns rating prediction and propensity to guarantee a low prediction inaccuracy at inference time. Yuan et al. (Yuan et al., 2019a) propose a propensity-free doubly robust method to address the issue that samples with low propensity scores are absent in the observed dataset. Zhang et al. (Zhang et al., 2020) propose Multi-DR based on a multi-task learning framework to address selection bias and data sparsity issues in CVR estimation. Gun et al. (Guo et al., 2021) propose the MRDR (more robust doubly robust) estimator to further reduce the variance caused by inaccurate imputed outcomes in DR while retaining its double robustness. In addition, Kiyohara et al. (Kiyohara et al., 2022) expand previous RIPS to Cascade Doubly Robust estimator, which has the same user interaction assumption as RIPS. Xiao et al. (Xiao and Wang, 2022) propose an information bottleneck-based approach to effectively learn the DR estimator for the estimation of recommendation uplift, with the hope of a better trade-off between the bias and variance of propensity scores. Dai et al. (Dai et al., 2022) learns imputation with balancing the variance and bias of DR loss. More recently, Song et al. (Song et al., 2023a) filter imputation data through examination of their mean and variance, in order to reduce poisonous imputations that significantly deviate from the truth and impair the debiasing performance.

Uplift, denoting the causal effect of recommendations, refers to the increase in user interactions purely caused by recommendations. Typical evaluations of recommender systems regard positive user interactions as a success. However, a subset of these interactions might persist even in the absence of recommendations. This assertion is substantiated by the conclusion of Sharma et al. (Sharma et al., 2015), which indicates that more than 75% of click-throughs would still occur in the absence of recommendations. For marketing campaigns where Return on Investment (ROI) is paramount, targeting ’voluntary buyers’ — individuals who would interact with or without any recommendations — is deemed unnecessary. Therefore, the industry regards uplift as a valuable metric for recommendations in expectation of higher rewards.

It is a natural application to introduce the causality concepts such as ATE and CATE for uplift modeling since the definition of uplift is a counterfactual problem and consistent with the objective of causal effect estimation (Yamane et al., 2018; Zhang et al., 2021d; Gutierrez and Gérardy, 2017). Causal approaches with traditional machine learning methods for uplift estimation include two-model approach (Radcliffe, 2007; Nassif et al., 2013), transformed outcome (Jaskowski and Jaroszewicz, 2012) and uplift trees (Radcliffe and Surry, 2011; Rzepakowski and Jaroszewicz, 2012). Regarding recommender systems, uplift estimation on online A/B testing suffers from the high expense and large fluctuations due to user self-selection bias (Sato, 2021), while uplift estimated offline is bedeviled by a wide variety of biases that could lead to MNAR. In order to deal with these issues, much of the literature has been published. Sato et al. (Sato et al., 2019) utilize SNIPS-based ATE to accomplish offline uplift-based evaluation. Goldenberg et al. (Goldenberg et al., 2020) leverage the Retrospective Estimation technique that relies solely on data with positive outcomes for CATE-based uplift modeling, which makes it especially suited for many recommendation scenarios where only the treatment outcomes are observable.  (Betlei et al., 2021) learns a model that directly optimizes an upper bound on AUUC, a popular uplift metric based on the uplift curves and unified with ATE (Yamane et al., 2018). In addition, CausCF (Xie et al., 2021) extends the classical MF to the tensor factorization with three dimensions—user, item, and treatment effect for better uplift performance. CF-MTL (Li et al., 2023b) accounts for whether users actively accept the treatment, leading to a more granular classification of users, and then estimates the probability for each user type within a multi-task learning framework. It is worth mentioning that in the uplift modeling literature (Diemert et al., 2018; Gutierrez and Gérardy, 2017; Zhang et al., 2021d), there are two closely related metrics for uplift modeling, uplift and Qini curves, the latter of which is evaluated based on the ranking of conditional treatment effect estimations.

There are some other impressive recommendation works with causal effect (Mehrotra et al., 2020; Zhang et al., 2021a; Rosenfeld et al., 2017; Bonner and Vasile, 2018; Yao et al., 2022; Tran et al., 2022). For example,  (Mehrotra et al., 2020) adapts a Bayesian model to infer the causal impact of new track releases, which may be an essential consideration in the design of music recommendation platforms.  (Zhang et al., 2021a) minimizes the distance between the traditional attention weights in the recommendation method and the ITE to reflect the true impact of the features on the interactions.  (Rosenfeld et al., 2017) and  (Bonner and Vasile, 2018) frames causal inference as a domain adaptation problem and leverages ITE with a large sample of biased data and a small sample of unbiased data to eliminate the bias problems, which are described in more detail in 6.1.