Olivier Jeunen

Olivier Jeunen

Approximate Nearest Neighbour search indices form the backbone of real-world recommender systems, enabling real-time candidate retrieval over million-item catalogues. Typically, a single point estimate embedding is learnt for every user and every item. At serving time, the user embedding queries the index for relevant items. Since these representations are learnt from sparse interaction data, they are noisy and might fail to capture all the nuances that contribute to ``relevance'' -- ignoring the fundamental uncertainty that is inherent to them. The result is a retrieval pipeline that is systematically biased toward the small minority of popular head items with well-estimated embeddings, at the expense of the long-tail majority of niche, diverse, and serendipitous content. We propose DINOSAUR (Distributional Approximate Nearest Neighbour Search for Uncertainty-Aware Retrieval): a simple and infrastructure-compatible framework to incorporate embedding uncertainty into candidate generation. Rather than indexing point estimates, DINOSAUR samples $S_i$ embeddings per item and constructs an index on this augmented set. Analogously, at query time, a user embedding is sampled. This two-sided stochastic retrieval process implicitly marginalises over embedding uncertainty, without requiring changes to model architecture or ANN index infrastructure. On the analytical side, we show that DINOSAUR recovers standard point-estimate retrieval as uncertainty vanishes, and we characterise how increased embedding variance expands the regions of latent space in which uncertain items are retrievable. Reproducible empirical observations align with these expectations, showing large coverage gains with small losses in offline recall.

Neeti Pokharna, Olivier Jeunen, Yatharth Saraf et al.

Online evaluation of ranking and retrieval systems often relies on downstream monetization metrics such as app revenue or creator earnings. These metrics are typically heavy-tailed, with a small fraction of users dominating both mean and variance, leading to low statistical power and unreliable conclusions in A/B experiments -- especially under limited traffic. We present a practical framework for variance reduction in online experiments by combining post-stratification with CUPED. Our approach leverages pre-experiment covariates to improve the sensitivity of monetization experiments without requiring additional traffic. Deployed at ShareChat across ranking-driven monetization experiments, the method substantially reduces variance and improves decision stability, achieving equivalent statistical confidence with ~45\% less traffic than standard metrics. We further discuss practical design choices, guardrails, and limitations, providing guidance on when post-stratification is appropriate for real-world information retrieval and Recommendation systems.

Gabriel Bénédict, Olivier Jeunen, Samuele Papa et al.

In this paper we propose RecFusion, which comprise a set of diffusion models for recommendation. Unlike image data which contain spatial correlations, a user-item interaction matrix, commonly utilized in recommendation, lacks spatial relationships between users and items. We formulate diffusion on a 1D vector and propose binomial diffusion, which explicitly models binary user-item interactions with a Bernoulli process. We show that RecFusion approaches the performance of complex VAE baselines on the core recommendation setting (top-n recommendation for binary non-sequential feedback) and the most common datasets (MovieLens and Netflix). Our proposed diffusion models that are specialized for 1D and/or binary setups have implications beyond recommendation systems, such as in the medical domain with MRI and CT scans.

Olivier Jeunen, Ciarán M. Gilligan-Lee, Rishabh Mehrotra et al.

The ability to answer causal questions is crucial in many domains, as causal inference allows one to understand the impact of interventions. In many applications, only a single intervention is possible at a given time. However, in some important areas, multiple interventions are concurrently applied. Disentangling the effects of single interventions from jointly applied interventions is a challenging task -- especially as simultaneously applied interventions can interact. This problem is made harder still by unobserved confounders, which influence both treatments and outcome. We address this challenge by aiming to learn the effect of a single-intervention from both observational data and sets of interventions. We prove that this is not generally possible, but provide identification proofs demonstrating that it can be achieved under non-linear continuous structural causal models with additive, multivariate Gaussian noise -- even when unobserved confounders are present. Importantly, we show how to incorporate observed covariates and learn heterogeneous treatment effects. Based on the identifiability proofs, we provide an algorithm that learns the causal model parameters by pooling data from different regimes and jointly maximizing the combined likelihood. The effectiveness of our method is empirically demonstrated on both synthetic and real-world data.

Olivier Jeunen, Ivan Potapov, Aleksei Ustimenko

Approaches to recommendation are typically evaluated in one of two ways: (1) via a (simulated) online experiment, often seen as the gold standard, or (2) via some offline evaluation procedure, where the goal is to approximate the outcome of an online experiment. Several offline evaluation metrics have been adopted in the literature, inspired by ranking metrics prevalent in the field of Information Retrieval. (Normalised) Discounted Cumulative Gain (nDCG) is one such metric that has seen widespread adoption in empirical studies, and higher (n)DCG values have been used to present new methods as the state-of-the-art in top-$n$ recommendation for many years. Our work takes a critical look at this approach, and investigates when we can expect such metrics to approximate the gold standard outcome of an online experiment. We formally present the assumptions that are necessary to consider DCG an unbiased estimator of online reward and provide a derivation for this metric from first principles, highlighting where we deviate from its traditional uses in IR. Importantly, we show that normalising the metric renders it inconsistent, in that even when DCG is unbiased, ranking competing methods by their normalised DCG can invert their relative order. Through a correlation analysis between off- and on-line experiments conducted on a large-scale recommendation platform, we show that our unbiased DCG estimates strongly correlate with online reward, even when some of the metric's inherent assumptions are violated. This statement no longer holds for its normalised variant, suggesting that nDCG's practical utility may be limited.

Hitesh Sagtani, Madan Jhawar, Rishabh Mehrotra et al.

Ad-load balancing is a critical challenge in online advertising systems, particularly in the context of social media platforms, where the goal is to maximize user engagement and revenue while maintaining a satisfactory user experience. This requires the optimization of conflicting objectives, such as user satisfaction and ads revenue. Traditional approaches to ad-load balancing rely on static allocation policies, which fail to adapt to changing user preferences and contextual factors. In this paper, we present an approach that leverages off-policy learning and evaluation from logged bandit feedback. We start by presenting a motivating analysis of the ad-load balancing problem, highlighting the conflicting objectives between user satisfaction and ads revenue. We emphasize the nuances that arise due to user heterogeneity and the dependence on the user's position within a session. Based on this analysis, we define the problem as determining the optimal ad-load for a particular feed fetch. To tackle this problem, we propose an off-policy learning framework that leverages unbiased estimators such as Inverse Propensity Scoring (IPS) and Doubly Robust (DR) to learn and estimate the policy values using offline collected stochastic data. We present insights from online A/B experiments deployed at scale across over 80 million users generating over 200 million sessions, where we find statistically significant improvements in both user satisfaction metrics and ads revenue for the platform.

Bram van den Akker, Olivier Jeunen, Ying Li et al.

The bandit paradigm provides a unified modeling framework for problems that require decision-making under uncertainty. Because many business metrics can be viewed as rewards (a.k.a. utilities) that result from actions, bandit algorithms have seen a large and growing interest from industrial applications, such as search, recommendation and advertising. Indeed, with the bandit lens comes the promise of direct optimisation for the metrics we care about. Nevertheless, the road to successfully applying bandits in production is not an easy one. Even when the action space and rewards are well-defined, practitioners still need to make decisions regarding multi-arm or contextual approaches, on- or off-policy setups, delayed or immediate feedback, myopic or long-term optimisation, etc. To make matters worse, industrial platforms typically give rise to large action spaces in which existing approaches tend to break down. The research literature on these topics is broad and vast, but this can overwhelm practitioners, whose primary aim is to solve practical problems, and therefore need to decide on a specific instantiation or approach for each project. This tutorial will take a step towards filling that gap between the theory and practice of bandits. Our goal is to present a unified overview of the field and its existing terminology, concepts and algorithms -- with a focus on problems relevant to industry. We hope our industrial perspective will help future practitioners who wish to leverage the bandit paradigm for their application.

Olivier Jeunen

Online experiments such as Randomised Controlled Trials (RCTs) or A/B-tests are the bread and butter of modern platforms on the web. They are conducted continuously to allow platforms to estimate the causal effect of replacing system variant "A" with variant "B", on some metric of interest. These variants can differ in many aspects. In this paper, we focus on the common use-case where they correspond to machine learning models. The online experiment then serves as the final arbiter to decide which model is superior, and should thus be shipped. The statistical literature on causal effect estimation from RCTs has a substantial history, which contributes deservedly to the level of trust researchers and practitioners have in this "gold standard" of evaluation practices. Nevertheless, in the particular case of machine learning experiments, we remark that certain critical issues remain. Specifically, the assumptions that are required to ascertain that A/B-tests yield unbiased estimates of the causal effect, are seldom met in practical applications. We argue that, because variants typically learn using pooled data, a lack of model interference cannot be guaranteed. This undermines the conclusions we can draw from online experiments with machine learning models. We discuss the implications this has for practitioners, and for the research literature.

Olivier Jeunen, Ben London

Off-Policy Estimation (OPE) methods allow us to learn and evaluate decision-making policies from logged data. This makes them an attractive choice for the offline evaluation of recommender systems, and several recent works have reported successful adoption of OPE methods to this end. An important assumption that makes this work is the absence of unobserved confounders: random variables that influence both actions and rewards at data collection time. Because the data collection policy is typically under the practitioner's control, the unconfoundedness assumption is often left implicit, and its violations are rarely dealt with in the existing literature. This work aims to highlight the problems that arise when performing off-policy estimation in the presence of unobserved confounders, specifically focusing on a recommendation use-case. We focus on policy-based estimators, where the logging propensities are learned from logged data. We characterise the statistical bias that arises due to confounding, and show how existing diagnostics are unable to uncover such cases. Because the bias depends directly on the true and unobserved logging propensities, it is non-identifiable. As the unconfoundedness assumption is famously untestable, this becomes especially problematic. This paper emphasises this common, yet often overlooked issue. Through synthetic data, we empirically show how naïve propensity estimation under confounding can lead to severely biased metric estimates that are allowed to fly under the radar. We aim to cultivate an awareness among researchers and practitioners of this important problem, and touch upon potential research directions towards mitigating its effects.

Imad Aouali, Amine Benhalloum, Martin Bompaire et al.

Both in academic and industry-based research, online evaluation methods are seen as the golden standard for interactive applications like recommendation systems. Naturally, the reason for this is that we can directly measure utility metrics that rely on interventions, being the recommendations that are being shown to users. Nevertheless, online evaluation methods are costly for a number of reasons, and a clear need remains for reliable offline evaluation procedures. In industry, offline metrics are often used as a first-line evaluation to generate promising candidate models to evaluate online. In academic work, limited access to online systems makes offline metrics the de facto approach to validating novel methods. Two classes of offline metrics exist: proxy-based methods, and counterfactual methods. The first class is often poorly correlated with the online metrics we care about, and the latter class only provides theoretical guarantees under assumptions that cannot be fulfilled in real-world environments. Here, we make the case that simulation-based comparisons provide ways forward beyond offline metrics, and argue that they are a preferable means of evaluation.

Olivier Jeunen, Schaun Wheeler

Marketing and product personalisation provide a prominent and visible use-case for the application of Information Retrieval methods across several business domains. Recently, agentic approaches to these problems have been gaining traction. This work evaluates the behavioural and retention effects of agentic personalisation on a financial service application's customer communication system during a 2025 national tax filing period. Through a two month-long randomised controlled trial, we compare an agentic messaging approach against a business-as-usual (BAU) rule-based campaign system, focusing on two primary outcomes: unsubscribe behaviour and conversion timing. Empirical results show that agent-led messaging reduced unsubscribe events by 21\% ($\pm 0.01$) relative to BAU and increased early filing behaviour in the weeks preceding the national deadline. These findings demonstrate how adaptive, user-level decision-making systems can modulate engagement intensity whilst improving long-term retention indicators.

Olivier Jeunen, Eleanor Hanna, Schaun Wheeler

In consumer applications, Customer Relationship Management (CRM) has traditionally relied on the manual optimisation of static, rule-based messaging strategies. While adaptive and autonomous learning systems offer the promise of scalable personalisation, it remains unclear to what extent ``human-in-the-loop'' oversight is required to sustain performance uplift over time. This paper presents a longitudinal case study analysing a real-world consumer application that leverages agentic infrastructure to personalise marketing messaging for a large-scale user base over an 11-month period. We compare two distinct periods: an active phase where marketers directly curated content, audiences, and strategies -- followed immediately by a passive phase where agents operated autonomously from a fixed library of components. Our results demonstrate that whilst active human management generates the highest relative lift in engagement metrics, the autonomous agents successfully sustained a positive lift during the passive period. These findings suggest a symbiotic model where human intervention drives strategic initialisation and discovery, yet autonomous agents can ensure the scalable retention and preservation of performance gains.

Nilavra Pathak, Olivier Jeunen, Eric Lambert

Organizations routinely make strategic budget allocations under operational constraints, but often lack a principled way to assess whether realized allocations were close to the best feasible choices in hindsight. We present a retrospective auditing framework based on hindsight regret, defined as the opportunity cost of the realized allocation relative to a constraint-faithful benchmark under the same budget and stability guardrails. The framework estimates regime-specific spend--response functions from historical logs, computes feasible hindsight allocations via constrained optimization, and propagates uncertainty through Monte Carlo evaluation to produce regret distributions, expected lift, and probability-of-improvement summaries. This separates allocation inefficiency from uncertainty in the estimated response surfaces. Experiments on real marketing allocation logs show that the framework yields interpretable post-hoc diagnostics and reveals a practical trade-off between allocation flexibility and detectability: moderate feasible reallocations often capture most measurable gain, while larger shifts move into weak-support regions with higher uncertainty. The result is a practical method for auditing historical budget decisions when online experimentation is costly or infeasible.

Olivier Jeunen, Jatin Mandav, Ivan Potapov et al.

Real-world recommender systems often need to balance multiple objectives when deciding which recommendations to present to users. These include behavioural signals (e.g. clicks, shares, dwell time), as well as broader objectives (e.g. diversity, fairness). Scalarisation methods are commonly used to handle this balancing task, where a weighted average of per-objective reward signals determines the final score used for ranking. Naturally, how these weights are computed exactly, is key to success for any online platform. We frame this as a decision-making task, where the scalarisation weights are actions taken to maximise an overall North Star reward (e.g. long-term user retention or growth). We extend existing policy learning methods to the continuous multivariate action domain, proposing to maximise a pessimistic lower bound on the North Star reward that the learnt policy will yield. Typical lower bounds based on normal approximations suffer from insufficient coverage, and we propose an efficient and effective policy-dependent correction for this. We provide guidance to design stochastic data collection policies, as well as highly sensitive reward signals. Empirical observations from simulations, offline and online experiments highlight the efficacy of our deployed approach.

Olivier Jeunen, Shashank Gupta

Off-policy evaluation (OPE) is essential for assessing ranking and recommendation systems without costly online interventions. Self-Normalised Inverse Propensity Scoring (SNIPS) is a standard tool for variance reduction in OPE, leveraging a multiplicative control variate. Recent advances in off-policy learning suggest that additive control variates (baseline corrections) may offer superior performance, yet theoretical guarantees for evaluation are lacking. This paper provides a definitive answer: we prove that $β^\star$-IPS, an estimator with an optimal additive baseline, asymptotically dominates SNIPS in Mean Squared Error. By analytically decomposing the variance gap, we show that SNIPS is asymptotically equivalent to using a specific -- but generally sub-optimal -- additive baseline. Our results theoretically justify shifting from self-normalisation to optimal baseline corrections for both ranking and recommendation.

Olivier Jeunen, Aleksei Ustimenko

Online controlled experiments are a crucial tool to allow for confident decision-making in technology companies. A North Star metric is defined (such as long-term revenue or user retention), and system variants that statistically significantly improve on this metric in an A/B-test can be considered superior. North Star metrics are typically delayed and insensitive. As a result, the cost of experimentation is high: experiments need to run for a long time, and even then, type-II errors (i.e. false negatives) are prevalent. We propose to tackle this by learning metrics from short-term signals that directly maximise the statistical power they harness with respect to the North Star. We show that existing approaches are prone to overfitting, in that higher average metric sensitivity does not imply improved type-II errors, and propose to instead minimise the $p$-values a metric would have produced on a log of past experiments. We collect such datasets from two social media applications with over 160 million Monthly Active Users each, totalling over 153 A/B-pairs. Empirical results show that we are able to increase statistical power by up to 78% when using our learnt metrics stand-alone, and by up to 210% when used in tandem with the North Star. Alternatively, we can obtain constant statistical power at a sample size that is down to 12% of what the North Star requires, significantly reducing the cost of experimentation.

Shubham Baweja, Neeti Pokharna, Aleksei Ustimenko et al.

Online controlled experiments, such as A/B-tests, are commonly used by modern tech companies to enable continuous system improvements. Despite their paramount importance, A/B-tests are expensive: by their very definition, a percentage of traffic is assigned an inferior system variant. To ensure statistical significance on top-level metrics, online experiments typically run for several weeks. Even then, a considerable amount of experiments will lead to inconclusive results (i.e. false negatives, or type-II error). The main culprit for this inefficiency is the variance of the online metrics. Variance reduction techniques have been proposed in the literature, but their direct applicability to commonly used ratio metrics (e.g. click-through rate or user retention) is limited. In this work, we successfully apply variance reduction techniques to ratio metrics on a large-scale short-video platform: ShareChat. Our empirical results show that we can either improve A/B-test confidence in 77% of cases, or can retain the same level of confidence with 30% fewer data points. Importantly, we show that the common approach of including as many covariates as possible in regression is counter-productive, highlighting that control variates based on Gradient-Boosted Decision Tree predictors are most effective. We discuss the practicalities of implementing these methods at scale and showcase the cost reduction they beget.

Schaun Wheeler, Olivier Jeunen

Large Language Models (LLMs) represent a landmark achievement in Artificial Intelligence (AI), demonstrating unprecedented proficiency in procedural tasks such as text generation, code completion, and conversational coherence. These capabilities stem from their architecture, which mirrors human procedural memory -- the brain's ability to automate repetitive, pattern-driven tasks through practice. However, as LLMs are increasingly deployed in real-world applications, it becomes impossible to ignore their limitations operating in complex, unpredictable environments. This paper argues that LLMs, while transformative, are fundamentally constrained by their reliance on procedural memory. To create agents capable of navigating ``wicked'' learning environments -- where rules shift, feedback is ambiguous, and novelty is the norm -- we must augment LLMs with semantic memory and associative learning systems. By adopting a modular architecture that decouples these cognitive functions, we can bridge the gap between narrow procedural expertise and the adaptive intelligence required for real-world problem-solving.

Olivier Jeunen, Aleksei Ustimenko

The off-policy paradigm casts recommendation as a counterfactual decision-making task, allowing practitioners to unbiasedly estimate online metrics using offline data. This leads to effective evaluation metrics, as well as learning procedures that directly optimise online success. Nevertheless, the high variance that comes with unbiasedness is typically the crux that complicates practical applications. An important insight is that the difference between policy values can often be estimated with significantly reduced variance, if said policies have positive covariance. This allows us to formulate a pairwise off-policy estimation task: $Δ\text{-}{\rm OPE}$. $Δ\text{-}{\rm OPE}$ subsumes the common use-case of estimating improvements of a learnt policy over a production policy, using data collected by a stochastic logging policy. We introduce $Δ\text{-}{\rm OPE}$ methods based on the widely used Inverse Propensity Scoring estimator and its extensions. Moreover, we characterise a variance-optimal additive control variate that further enhances efficiency. Simulated, offline, and online experiments show that our methods significantly improve performance for both evaluation and learning tasks.

Sami Abboud, Eleanor Hanna, Olivier Jeunen et al.

Consumer applications provide ample opportunities to surface and communicate various forms of content to users. From promotional campaigns for new features or subscriptions, to evergreen nudges for engagement, or personalised recommendations; across e-mails, push notifications, and in-app surfaces. The conventional approach to orchestration for communication relies heavily on labour-intensive manual marketer work, and inhibits effective personalisation of content, timing, frequency, and copy-writing. We formulate this task under a sequential decision-making framework, where we aim to optimise a modular decision-making policy that maximises incremental engagement for any funnel event. Our approach leverages a Difference-in-Differences design for Individual Treatment Effect estimation, and Thompson sampling to balance the explore-exploit trade-off. We present results from a multi-service application, where our methodology has resulted in significant increases to a variety of goal events across several product features, and is currently deployed across 150 million users.

Olivier Jeunen

Recommender systems exemplify sequential decision-making under uncertainty, strategically deciding what content to serve to users, to optimise a range of potential objectives. To balance the explore-exploit trade-off successfully, Thompson sampling provides a natural and widespread paradigm to probabilistically select which action to take. Questions of causal and counterfactual inference, which underpin use-cases like offline evaluation, are not straightforward to answer in these contexts. Specifically, whilst most existing estimators rely on action propensities, these are not readily available under Thompson sampling procedures. We derive exact and efficiently computable expressions for action propensities under a variety of parameter and outcome distributions, enabling the use of off-policy estimators in Thompson sampling scenarios. This opens up a range of practical use-cases where counterfactual inference is crucial, including unbiased offline evaluation of recommender systems, as well as general applications of causal inference in online advertising, personalisation, and beyond.

Olivier Jeunen

Off-policy estimation (OPE) methods enable unbiased offline evaluation of recommender systems, directly estimating the online reward some target policy would have obtained, from offline data and with statistical guarantees. The theoretical elegance of the framework combined with practical successes have led to a surge of interest, with many competing estimators now available to practitioners and researchers. Among these, Doubly Robust methods provide a prominent strategy to combine value- and policy-based estimators. In this work, we take an alternative perspective to combine a set of OPE estimators and their associated confidence intervals into a single, more accurate estimate. Our approach leverages a correlated fixed-effects meta-analysis framework, explicitly accounting for dependencies among estimators that arise due to shared data. This yields a best linear unbiased estimate (BLUE) of the target policy's value, along with an appropriately conservative confidence interval that reflects inter-estimator correlation. We validate our method on both simulated and real-world data, demonstrating improved statistical efficiency over existing individual estimators.

Olivier Jeunen

A/B-tests are a cornerstone of experimental design on the web, with wide-ranging applications and use-cases. The statistical $t$-test comparing differences in means is the most commonly used method for assessing treatment effects, often justified through the Central Limit Theorem (CLT). The CLT ascertains that, as the sample size grows, the sampling distribution of the Average Treatment Effect converges to normality, making the $t$-test valid for sufficiently large sample sizes. When outcome measures are skewed or non-normal, quantifying what "sufficiently large" entails is not straightforward. To ensure that confidence intervals maintain proper coverage and that $p$-values accurately reflect the false positive rate, it is critical to validate this normality assumption. We propose a practical method to test this, by analysing repeatedly resampled A/A-tests. When the normality assumption holds, the resulting $p$-value distribution should be uniform, and this property can be tested using the Kolmogorov-Smirnov test. This provides an efficient and effective way to empirically assess whether the $t$-test's assumptions are met, and the A/B-test is valid. We demonstrate our methodology and highlight how it helps to identify scenarios prone to inflated Type-I errors. Our approach provides a practical framework to ensure and improve the reliability and robustness of A/B-testing practices.

Olivier Jeunen

Randomised Controlled Trials (RCTs) are the gold standard for estimating treatment effects across many fields of science. Technology companies have adopted A/B-testing methods as a modern RCT counterpart, where end-users are randomly assigned various system variants and user behaviour is tracked continuously. The objective is then to estimate the causal effect that the treatment variant would have on certain metrics of interest to the business. When the outcomes for randomisation units -- end-users in this case -- are not statistically independent, this obfuscates identifiability of treatment effects, and harms decision-makers' observability of the system. Social networks exemplify this, as they are designed to promote inter-user interactions. This interference by design notoriously complicates measurement of, e.g., the effects of sharing. In this work, we propose a simple Markov Decision Process (MDP)-based model describing user sharing behaviour in social networks. We derive an unbiased estimator for treatment effects under this model, and demonstrate through reproducible synthetic experiments that it outperforms existing methods by a significant margin.

Shashank Gupta, Olivier Jeunen, Harrie Oosterhuis et al.

The off-policy learning paradigm allows for recommender systems and general ranking applications to be framed as decision-making problems, where we aim to learn decision policies that optimize an unbiased offline estimate of an online reward metric. With unbiasedness comes potentially high variance, and prevalent methods exist to reduce estimation variance. These methods typically make use of control variates, either additive (i.e., baseline corrections or doubly robust methods) or multiplicative (i.e., self-normalisation). Our work unifies these approaches by proposing a single framework built on their equivalence in learning scenarios. The foundation of our framework is the derivation of an equivalent baseline correction for all of the existing control variates. Consequently, our framework enables us to characterize the variance-optimal unbiased estimator and provide a closed-form solution for it. This optimal estimator brings significantly improved performance in both evaluation and learning, and minimizes data requirements. Empirical observations corroborate our theoretical findings.

Olivier Jeunen, Dmytro Mykhaylov, David Rohde et al.

In machine learning we often try to optimise a decision rule that would have worked well over a historical dataset; this is the so called empirical risk minimisation principle. In the context of learning from recommender system logs, applying this principle becomes a problem because we do not have available the reward of decisions we did not do. In order to handle this "bandit-feedback" setting, several Counterfactual Risk Minimisation (CRM) methods have been proposed in recent years, that attempt to estimate the performance of different policies on historical data. Through importance sampling and various variance reduction techniques, these methods allow more robust learning and inference than classical approaches. It is difficult to accurately estimate the performance of policies that frequently perform actions that were infrequently done in the past and a number of different types of estimators have been proposed. In this paper, we review several methods, based on different off-policy estimators, for learning from bandit feedback. We discuss key differences and commonalities among existing approaches, and compare their empirical performance on the RecoGym simulation environment. To the best of our knowledge, this work is the first comparison study for bandit algorithms in a recommender system setting.

Olivier Jeunen, David Rohde, Flavian Vasile

In academic literature, recommender systems are often evaluated on the task of next-item prediction. The procedure aims to give an answer to the question: "Given the natural sequence of user-item interactions up to time t, can we predict which item the user will interact with at time t+1?". Evaluation results obtained through said methodology are then used as a proxy to predict which system will perform better in an online setting. The online setting, however, poses a subtly different question: "Given the natural sequence of user-item interactions up to time t, can we get the user to interact with a recommended item at time t+1?". From a causal perspective, the system performs an intervention, and we want to measure its effect. Next-item prediction is often used as a fall-back objective when information about interventions and their effects (shown recommendations and whether they received a click) is unavailable. When this type of data is available, however, it can provide great value for reliably estimating online recommender system performance. Through a series of simulated experiments with the RecoGym environment, we show where traditional offline evaluation schemes fall short. Additionally, we show how so-called bandit feedback can be exploited for effective offline evaluation that more accurately reflects online performance.

Dmytro Mykhaylov, David Rohde, Flavian Vasile et al.

There are three quite distinct ways to train a machine learning model on recommender system logs. The first method is to model the reward prediction for each possible recommendation to the user, at the scoring time the best recommendation is found by computing an argmax over the personalized recommendations. This method obeys principles such as the conditionality principle and the likelihood principle. A second method is useful when the model does not fit reality and underfits. In this case, we can use the fact that we know the distribution of historical recommendations (concentrated on previously identified good actions with some exploration) to adjust the errors in the fit to be evenly distributed over all actions. Finally, the inverse propensity score can be used to produce an estimate of the decision rules expected performance. The latter two methods violate the conditionality and likelihood principle but are shown to have good performance in certain settings. In this paper we review the literature around this fundamental, yet often overlooked choice and do some experiments using the RecoGym simulation environment.