This work presents an extension of Thompson Sampling bandit policy for orchestrating the collection of base recommendation algorithms for e-commerce. We focus on the problem of item-to-item recommendations, for which multiple behavioral and attribute-based predictors are provided to an ensemble learner. In addition, we detail the construction of a personalized predictor based on k-Nearest Neighbors (kNN), with temporal decay capabilities and event weighting. We show how to adapt Thompson Sampling to realistic situations when neither action availability nor reward stationarity is guaranteed. Furthermore, we investigate the effects of priming the sampler with pre-set parameters of reward probability distributions by utilizing the product catalog and/or event history, when such information is available. We report our experimental results based on the analysis of three real-world e-commerce datasets.

Despite the increasing interest towards session-based and streaming recommender systems, there is still a lack of publicly available evaluation frameworks supporting both these paradigms. To address the gap, we propose FlowRec — an extension of the streaming framework Scikit-Multiflow, which opens plentiful possibilities for prototyping recommender systems operating on sessionized data streams, thanks to the underlying collection of incremental learners and support for real-time performance tracking. We describe the extended functionalities of the adapted prequential evaluation protocol, and develop a competitive recommendation algorithm on top of Scikit-Multiflow’s implementation of a Hoeffding Tree. We compare our algorithm to other known baselines for the next-item prediction task across three different domains.

Reported evidence of biased matchmaking calls into question the ethicality of recommendations generated by a machine learning algorithm. In the context of dating services, the failure of an automated matchmaker to respect the user’s expressed sensitive preferences (racial, religious, etc.) may lead to biased decisions perceived by users as unfair. To address the issue, we introduce the notion of preferential fairness, and propose two algorithmic approaches for re-ranking the recommendations under preferential fairness constraints. Our experimental results demonstrate that the state of fairness can be reached with minimal accuracy compromises for both binary and non-binary attributes.

Recommender systems are algorithmic tools that assist users in discovering relevant items from a wide range of available options. Along with the apparent user value in mitigating the choice overload, they have an important business value in boosting sales and customer retention. Last, but not least, they have brought a substantial research value to the algorithm developments of the past two decades, mainly in the academic community. This thesis aims to address some of the aspects that are important to consider when recommender systems pave their way towards real-life applications.

Recommender systems have become an integral part of virtually every e-commerce application on the web. These systems enable users to quickly discover relevant products, at the same time increasing business value. Over the past decades, recommender systems have been modeled using numerous machine learning techniques. However, the adoptability of these models by commercial applications remains unclear. We assess the receptiveness of the industrial sector to algorithmic contributions of the research community by surveying more than 30 e-commerce platforms, and experimenting with various recommenders on proprietary e-commerce datasets. Another overlooked but important factor that complicates the design and use of recommender systems is their ethical implications. We identify and summarize these issues in our ethical recommendation framework, which also enables users to control sensitive moral aspects of recommendations via the proposed “ethical toolbox”. The feasibility of this tool is supported by the results of our user study. Because of moral implications associated with user profiling, we investigate algorithms capable of generating user-agnostic recommendations. We propose an ensemble learning scheme based on Thompson Sampling bandit policy, which models arms as base recommendation functions. We show how to adapt this algorithm to realistic situations when neither arm availability nor reward stationarity is guaranteed.

This work presents an extension of Thompson Sampling bandit policy for orchestrating the collection of base recommendation algorithms for e-commerce. We focus on the problem of item-to-item recommendations, for which multiple behavioral and attribute-based predictors are provided to an ensemble learner. We show how to adapt Thompson Sampling to realistic situations when neither action availability nor reward stationarity is guaranteed. Furthermore, we investigate the effects of priming the sampler with pre-set parameters of reward probability distributions by utilizing the product catalog and/or event history, when such information is available. We report our experimental results based on the analysis of three real-world e-commerce datasets.

In our ongoing work we extend the Thompson Sampling (TS) bandit policy for orchestrating the collection of base recommendation algorithms for e-Commerce. We focus on the problem of item-to-item recommendations, for which multiple behavioral and content-based predictors are provided to an ensemble learner. The extended TS-based policy must be able to handle situations when bandit arms are non-stationary and non-answering. Furthermore, we investigate the effects of priming the sampler with pre-set parameters of reward distributions by analyzing the product catalog and/or event history, when such information is available. We report our preliminary results based on the analysis of two real-world e-Commerce datasets.

We study bandit algorithms for e-commerce recommender systems. The question we pose is whether it is necessary to consider reinforcement learning effects in recommender systems. A key reason to introduce a recommender system for a product page on an e-commerce site is to increase the order value by improving the chance of making an upsale. If the recommender system merely predicts the next purchase, there might be no positive effect at all on the order value, since the recommender system predicts sales that would have happened independent of the recommender system. What we really are looking for are the false negatives, i.e., purchases that happen as a consequence of the recommender system. These purchases entail the entire uplift and should be present as reinforcement learning effects. This effect cannot be displayed in a simulation of the site, since there are no reinforcement learning effects present in a simulation. The attribution model must capture the uplift to guarantee an increased order value. However, such an attribution model is not practical, due to data sparsity. Given this starting point, we study some standard attribution models for e-commerce recommender systems, and describe how these fare when applied in a reinforcement learning algorithm, both in a simulation and on live sites.

The goal of our study is to provide a holistic view on various ethical challenges that complicate the design and use of recommender systems (RS). Our findings materialize into an ethical recommendation framework, which maps RS development stages to the corresponding ethical concerns, and further down to known solutions and the proposed user-adjustable controls. The need for such a framework is dictated by the apparent lack of research in this particular direction and the severity of consequences stemming from the neglect of the code of ethics in recommendations. The framework aims to aid RS practitioners in staying ethically alert while taking morally charged design decisions. At the same time, it would give users the desired control over the sensitive moral aspects of recommendations via the proposed “ethical toolbox”. The idea is embraced by the participants of our feasibility study.