The Multi-Armed Bandit (MAB) problem has been extensively studied in order to address real-world challenges related to sequential decision making. In this setting, an agent selects the best action to be performed at time-step t, based on the past rewards received by the environment. This formulation implicitly assumes that the expected payoff for each action is kept stationary by the environment through time. Nevertheless, in many real-world applications this assumption does not hold and the agent has to face a non-stationary environment, that is, with a changing reward distribution. Thus, we present a new MAB algorithm, named f-Discounted-Sliding-Window Thompson Sampling (f-dsw TS), for non-stationary environments, that is, when the data streaming is affected by concept drift. The f-dsw TS algorithm is based on Thompson Sampling (TS) and exploits a discount factor on the reward history and an arm-related sliding window to contrast concept drift in non-stationary environments. We investigate how to combine these two sources of information, namely the discount factor and the sliding window, by means of an aggregation function $f(.)$. In particular, we proposed a pessimistic ($f=min$), an optimistic ($f=max$), as well as an averaged ($f=mean$) version of the f-dsw TS algorithm. A rich set of numerical experiments is performed to evaluate the f-dsw TS algorithm compared to both stationary and non-stationary state-of-the-art TS baselines. We exploited synthetic environments (both randomly-generated and controlled) to test the MAB algorithms under different types of drift, that is, sudden/abrupt, incremental, gradual and increasing/decreasing drift. Furthermore, we adapt four real-world active learning tasks to our framework—a prediction task on crimes in the city of Baltimore, a classification task on insects species, a recommendation task on local web-news, and a time-series analysis on microbial organisms in the tropical air ecosystem. The f-dsw TS approach emerges as the best performing MAB algorithm. At least one of the versions of f-dsw TS performs better than the baselines in synthetic environments, proving the robustness of f-dsw TS under different concept drift types. Moreover, the pessimistic version ($f=min$) results as the most effective in all real-world tasks.

Reproducibility is a main principle in science and fundamental to ensure scientific progress. However, many recent works point out that there are widespread deficiencies for this aspect in the AI field, making the reproducibility of results impractical or even impossible. We therefore studied the state of reproducibility support on the topic of Reinforcement Learning & Recommender Systems to analyse the situation in this context. We collected a total of 60 papers and analysed them by defining a set of variables to inspect the most important aspects that enable reproducibility, such as dataset, pre-processing code, hardware specifications, software dependencies, algorithm implementation, algorithm hyperparameters, and experiment code. Furthermore, we used the ACM Badges definitions assigning them to the selected papers. We discovered that, like in many other AI domains, the Reinforcement Learning & Recommender Systems field is grappling with a reproducibility crisis, as none of the selected papers were reproducible when strictly applying the ACM Badges definitions according to our analysis.

Causality is gaining more and more attention in the machine learning community and consequently also in recommender systems research. The limitations of learning offline from observed data are widely recognized, however, applying debiasing strategies like Inverse Propensity Weighting does not always solve the problem of making wrong estimates. This concept paper contributes a summary of debiasing strategies in recommender systems and the design of several toy examples demonstrating the limits of these commonly applied approaches. Therefore, we propose to map the causality frameworks of potential outcomes and structural causal models onto the recommender systems domain in order to foster future research and development. For instance, applying causal discovery strategies on offline data to learn the causal graph in order to compute counterfactuals or improve debiasing strategies.