Distributionally Robust Policy Learning under Concept Drifts

Distributionally robust policy learning aims to find a policy that performs well under the worst-case distributional shift, and yet most existing methods for robust policy learning consider the worst-case joint distribution of the covariate and the outcome. The joint-modeling strategy can be unnecessarily conservative when we have more information on the source of distributional shifts. This paper studiesa more nuanced problem -- robust policy learning under the concept drift, when only the conditional relationship between the outcome and the covariate changes. To this end, we first provide a doubly-robust estimator for evaluating the worst-case average reward of a given policy under a set of perturbed conditional distributions. We show that the policy value estimator enjoys asymptotic normality even if the nuisance parameters are estimated with a slower-than-root-$n$ rate. We then propose a learning algorithm that outputs the policy maximizing the estimated policy value within a given policy class $\Pi$, and show that the sub-optimality gap of the proposed algorithm is of the order $\kappa(\Pi)n^{-1/2}$, with $\kappa(\Pi)$ is the entropy integral of $\Pi$ under the Hamming distance and $n$ is the sample size. A matching lower bound is provided to show the optimality of the rate. The proposed methods are implemented and evaluated in numerical studies, demonstrating substantial improvement compared with existing benchmarks.

Adaptively Learning to Select-Rank in Online Platforms

Ranking algorithms are fundamental to various online platforms across e-commerce sites to content streaming services. Our research addresses the challenge of adaptively ranking items from a candidate pool for heterogeneous users, a key component in personalizing user experience. We develop a user response model that considers diverse user preferences and the varying effects of item positions, aiming to optimize overall user satisfaction with the ranked list. We frame this problem within a contextual bandits framework, with each ranked list as an action. Our approach incorporates an upper confidence bound to adjust predicted user satisfaction scores and selects the ranking action that maximizes these adjusted scores, efficiently solved via maximum weight imperfect matching. We demonstrate that our algorithm achieves a cumulative regret bound of $O(d\sqrt{NKT})$ for ranking $K$ out of $N$ items in a $d$-dimensional context space over $T$ rounds, under the assumption that user responses follow a generalized linear model. This regret alleviates dependence on the ambient action space, whose cardinality grows exponentially with $N$ and $K$ (thus rendering direct application of existing adaptive learning algorithms -- such as UCB or Thompson sampling -- infeasible). Experiments conducted on both simulated and real-world datasets demonstrate our algorithm outperforms the baseline.

Statistical Properties of Robust Satisficing

The Robust Satisficing (RS) model is an emerging approach to robust optimization, offering streamlined procedures and robust generalization across various applications. However, the statistical theory of RS remains unexplored in the literature. This paper fills in the gap by comprehensively analyzing the theoretical properties of the RS model. Notably, the RS structure offers a more straightforward path to deriving statistical guarantees compared to the seminal Distributionally Robust Optimization (DRO), resulting in a richer set of results. In particular, we establish two-sided confidence intervals for the optimal loss without the need to solve a minimax optimization problem explicitly. We further provide finite-sample generalization error bounds for the RS optimizer. Importantly, our results extend to scenarios involving distribution shifts, where discrepancies exist between the sampling and target distributions. Our numerical experiments show that the RS model consistently outperforms the baseline empirical risk minimization in small-sample regimes and under distribution shifts. Furthermore, compared to the DRO model, the RS model exhibits lower sensitivity to hyperparameter tuning, highlighting its practicability for robustness considerations.

Proportional Response: Contextual Bandits for Simple and Cumulative Regret Minimization

Simple regret minimization is a critical problem in learning optimal treatment assignment policies across various domains, including healthcare and e-commerce. However, it remains understudied in the contextual bandit setting. We propose a new family of computationally efficient bandit algorithms for the stochastic contextual bandit settings, with the flexibility to be adapted for cumulative regret minimization (with near-optimal minimax guarantees) and simple regret minimization (with SOTA guarantees). Furthermore, our algorithms adapt to model misspecification and extend to the continuous arm settings. These advantages come from constructing and relying on "conformal arm sets" (CASs), which provide a set of arms at every context that encompass the context-specific optimal arm with some probability across the context distribution. Our positive results on simple and cumulative regret guarantees are contrasted by a negative result, which shows that an algorithm can't achieve instance-dependent simple regret guarantees while simultaneously achieving minimax optimal cumulative regret guarantees.

Post-Episodic Reinforcement Learning Inference

We consider estimation and inference with data collected from episodic reinforcement learning (RL) algorithms; i.e. adaptive experimentation algorithms that at each period (aka episode) interact multiple times in a sequential manner with a single treated unit. Our goal is to be able to evaluate counterfactual adaptive policies after data collection and to estimate structural parameters such as dynamic treatment effects, which can be used for credit assignment (e.g. what was the effect of the first period action on the final outcome). Such parameters of interest can be framed as solutions to moment equations, but not minimizers of a population loss function, leading to Z-estimation approaches in the case of static data. However, such estimators fail to be asymptotically normal in the case of adaptive data collection. We propose a re-weighted Z-estimation approach with carefully designed adaptive weights to stabilize the episode-varying estimation variance, which results from the nonstationary policy that typical episodic RL algorithms invoke. We identify proper weighting schemes to restore the consistency and asymptotic normality of the re-weighted Z-estimators for target parameters, which allows for hypothesis testing and constructing reliable confidence regions for target parameters of interest. Primary applications include dynamic treatment effect estimation and dynamic off-policy evaluation.

Two-Stage Constrained Actor-Critic for Short Video Recommendation

The wide popularity of short videos on social media poses new opportunities and challenges to optimize recommender systems on the video-sharing platforms. Users sequentially interact with the system and provide complex and multi-faceted responses, including watch time and various types of interactions with multiple videos. One the one hand, the platforms aims at optimizing the users' cumulative watch time (main goal) in long term, which can be effectively optimized by Reinforcement Learning. On the other hand, the platforms also needs to satisfy the constraint of accommodating the responses of multiple user interactions (auxiliary goals) such like, follow, share etc. In this paper, we formulate the problem of short video recommendation as a Constrained Markov Decision Process (CMDP). We find that traditional constrained reinforcement learning algorithms can not work well in this setting. We propose a novel two-stage constrained actor-critic method: At stage one, we learn individual policies to optimize each auxiliary signal. At stage two, we learn a policy to (i) optimize the main signal and (ii) stay close to policies learned at the first stage, which effectively guarantees the performance of this main policy on the auxiliaries. Through extensive offline evaluations, we demonstrate effectiveness of our method over alternatives in both optimizing the main goal as well as balancing the others. We further show the advantage of our method in live experiments of short video recommendations, where it significantly outperforms other baselines in terms of both watch time and interactions. Our approach has been fully launched in the production system to optimize user experiences on the platform.

Deconfounding Duration Bias in Watch-time Prediction for Video Recommendation

Watch-time prediction remains to be a key factor in reinforcing user engagement via video recommendations. It has become increasingly important given the ever-growing popularity of online videos. However, prediction of watch time not only depends on the match between the user and the video but is often mislead by the duration of the video itself. With the goal of improving watch time, recommendation is always biased towards videos with long duration. Models trained on this imbalanced data face the risk of bias amplification, which misguides platforms to over-recommend videos with long duration but overlook the underlying user interests. This paper presents the first work to study duration bias in watch-time prediction for video recommendation. We employ a causal graph illuminating that duration is a confounding factor that concurrently affects video exposure and watch-time prediction -- the first effect on video causes the bias issue and should be eliminated, while the second effect on watch time originates from video intrinsic characteristics and should be preserved. To remove the undesired bias but leverage the natural effect, we propose a Duration Deconfounded Quantile-based (D2Q) watch-time prediction framework, which allows for scalability to perform on industry production systems. Through extensive offline evaluation and live experiments, we showcase the effectiveness of this duration-deconfounding framework by significantly outperforming the state-of-the-art baselines. We have fully launched our approach on Kuaishou App, which has substantially improved real-time video consumption due to more accurate watch-time predictions.

ResAct: Reinforcing Long-term Engagement in Sequential Recommendation with Residual Actor

Long-term engagement is preferred over immediate engagement in sequential recommendation as it directly affects product operational metrics such as daily active users (DAUs) and dwell time. Meanwhile, reinforcement learning (RL) is widely regarded as a promising framework for optimizing long-term engagement in sequential recommendation. However, due to expensive online interactions, it is very difficult for RL algorithms to perform state-action value estimation, exploration and feature extraction when optimizing long-term engagement. In this paper, we propose ResAct which seeks a policy that is close to, but better than, the online-serving policy. In this way, we can collect sufficient data near the learned policy so that state-action values can be properly estimated, and there is no need to perform online exploration. Directly optimizing this policy is difficult due to the huge policy space. ResAct instead solves it by first reconstructing the online behaviors and then improving it. Our main contributions are fourfold. First, we design a generative model which reconstructs behaviors of the online-serving policy by sampling multiple action estimators. Second, we design an effective learning paradigm to train the residual actor which can output the residual for action improvement. Third, we facilitate the extraction of features with two information theoretical regularizers to confirm the expressiveness and conciseness of features. Fourth, we conduct extensive experiments on a real world dataset consisting of millions of sessions, and our method significantly outperforms the state-of-the-art baselines in various of long term engagement optimization tasks.

Constrained Reinforcement Learning for Short Video Recommendation

The wide popularity of short videos on social media poses new opportunities and challenges to optimize recommender systems on the video-sharing platforms. Users provide complex and multi-faceted responses towards recommendations, including watch time and various types of interactions with videos. As a result, established recommendation algorithms that concern a single objective are not adequate to meet this new demand of optimizing comprehensive user experiences. In this paper, we formulate the problem of short video recommendation as a constrained Markov Decision Process (MDP), where platforms want to optimize the main goal of user watch time in long term, with the constraint of accommodating the auxiliary responses of user interactions such as sharing/downloading videos. To solve the constrained MDP, we propose a two-stage reinforcement learning approach based on actor-critic framework. At stage one, we learn individual policies to optimize each auxiliary response. At stage two, we learn a policy to (i) optimize the main response and (ii) stay close to policies learned at the first stage, which effectively guarantees the performance of this main policy on the auxiliaries. Through extensive simulations, we demonstrate effectiveness of our approach over alternatives in both optimizing the main goal as well as balancing the others. We further show the advantage of our approach in live experiments of short video recommendations, where it significantly outperforms other baselines in terms of watch time and interactions from video views. Our approach has been fully launched in the production system to optimize user experiences on the platform.

Off-Policy Evaluation via Adaptive Weighting with Data from Contextual Bandits

It has become increasingly common for data to be collected adaptively, for example using contextual bandits. Historical data of this type can be used to evaluate other treatment assignment policies to guide future innovation or experiments. However, policy evaluation is challenging if the target policy differs from the one used to collect data, and popular estimators, including doubly robust (DR) estimators, can be plagued by bias, excessive variance, or both. In particular, when the pattern of treatment assignment in the collected data looks little like the pattern generated by the policy to be evaluated, the importance weights used in DR estimators explode, leading to excessive variance. In this paper, we improve the DR estimator by adaptively weighting observations to control its variance. We show that a t-statistic based on our improved estimator is asymptotically normal under certain conditions, allowing us to form confidence intervals and test hypotheses. Using synthetic data and public benchmarks, we provide empirical evidence for our estimator's improved accuracy and inferential properties relative to existing alternatives.