Causality is receiving increasing attention from the artificial intelligence and machine learning communities. Similarly, growing attention to causality is currently going on in the Recommendation Systems (RSs) community, which has realised that RSs could greatly benefit from causality to transform accurate predictions into efective and explainable decisions. Indeed, the RS literature has repeatedly highlighted that, in real-world scenarios, recommendation algorithms sufer many types of biases since assumptions ensuring unbiasedness are likely not met. In this discussion paper, we formulate the RS problem in terms of causality, using potential outcomes and structural causal models, by giving formal definitions of the causal quantities to be estimated and a general causal graph to serve as a reference to foster future research and development.
Predicting and deciding are two fundamentally diferent tasks. As described by the Ladder of Causation
[
Furthermore, human beings are not interested in mere correlations but in understanding the actual
causes of the efects, manipulating the world to achieve the desired outcome. In fact, scientists are
familiar with the phrase: “Correlation is not causation”, that is, for example, “the rooster’s crow is highly
correlated with the sunrise; yet it does not cause the sunrise” [
This is why causality becomes important: we need a way to translate cause-and-efect relations and
interventions on a system using a mathematical formulation. To this end, in [
The first step is to have a CG that describes the data-generating process of the system under study. The
CG can be learned by combining observational data with experts’ knowledge through a process called
causal discovery, which is enabled by several causal discovery algorithms [
U
I
C
The problem of recommending a single item with features I to a user U in context C is described by the CG of Figure 2 where the node represents the action of recommending an item whose domain corresponds to the item set, i.e., ∈ (). For example, in the context of film recommendation, () is the catalogue of the films and our recommendation is one of the films in the catalogue. To decide which item to recommend to the current user U in context C, we use a policy based on user and context features. Once we decide which item to recommend, the corresponding item’s features I are fixed and they mediate the efect of our recommendation on the user feedback through the path → I → . For example, once we decide to recommend a film, its genre is fixed ( → ), and the genre (likely) afects a user’s feedback ( → ). It is worth noticing that not all the item’s features have to influence the user’s feedback, i.e., some item’s features are not taken into account by the users. On the other hand, part of the efect of the recommendation on the user’s feedback may not be captured by the modelled features I and flows directly through the edge → . For example, if we are not able to model the film’s popularity, since it is dificult to know or to model it, the efect of the film’s popularity will flow through the edge → as this feature is not included in our model.
To exploit the potential of causality, we should frame the quantity to estimate as a causal estimand that
encodes the notion of the causal efect of a variable (the cause) on another (the efect). Generally, we
can define it, using the POs framework and the do-operator [
However, expressions with the do-operator can only be estimated in controlled experiments where
the variables in the do-terms can be appropriately controlled. To estimate a causal estimand using only
observational data, it is necessary to remove the do-terms and obtain an equivalent expression. To this
end, the adjustment formula estimator [
To identify the variables that must be included in Z, we can query the CG by evaluating an
identification criterion, such as the backdoor criterion [
Once we have proved the identifiability of the causal estimand, i.e., once we have shown that the causal estimand is equal to a statistical estimand, this can be estimated using classical statistical estimators. To this end, any model that is compatible with the type of the outcome variable, e.g. linear regression for a continuous outcome or neural networks for non-linear relations, is suitable for this estimation. Clearly, the model should be chosen carefully for each problem by considering the available data characteristics to avoid estimation errors.
Finally, with the estimated causal efects, e.g., the efect of our recommendation on the user’s propensity
to click on the recommended item(s), we can decide which items to recommend. This could be done
in diferent ways: (i) greedy, (ii) -greedy and (iii) more sophisticated policies. To this end, in recent
years, several works have exploited causality by linking it to Multi-Armed Bandit (MAB) [
In this paper, we proposed a causal view of the RS problem and highlighted the importance of framing the recommendation problem in terms of causality. The causality framework can, in our view, be considered as a single framework allowing researchers to wholistically define and address several problems widely acknowledged in the RSs community to bridge the gaps in future works.
However, we would like to stress that causality is not magic but ruthlessly honest and, diferently from other approaches, it makes explicit assumptions, such as ignorability and unconfoundedness, leaving us with the burden of judging whether they are likely to be satisfied for the addressed context. Indeed, causality is not the sole ingredient to solve the RS problem while we are fully convinced that exploiting the body of knowledge generated over more than 30 years of research in RSs and users’ behaviour remains fundamental. [22] J. Zhang, E. Bareinboim, Near-optimal reinforcement learning in dynamic treatment regimes, in: H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, R. Garnett (Eds.), Advances in Neural Information Processing Systems, volume 32, Curran Associates, Inc., Vancouver, Canada, 2019.