Diversity in recommendation has been studied extensively. It has been shown that maximizing diversity subject to constrained relevance yields high user engagement over time. Existing work largely relies on setting some attributes that are used to craft an item similarity function and diversify results. In this paper, we examine the question of learning diversity attributes. That is particularly important when users receive recommendations over multiple sessions. We devise two main approaches to look for the best diversity attribute in each session: the first is a generalization of traditional diversity algorithms and the second is based on reinforcement learning. We implement both approaches and run extensive experiments on a semi-synthetic dataset. Our results demonstrate that learning diversity attributes yields a higher overall diversity than traditional diversity algorithms. We also find that training policies using reinforcement learning is more efficient in terms of response time, in particular for high dimensional data.

We tackle the longstanding question of checking hypotheses on the social Web. In particular, we address the challenges that arise in the context of testing an input hypothesis on many data samples, in our case, user groups. This is referred to as Multiple Hypothesis Testing, a method of choice for data-driven discoveries. Ensuring sound discoveries in large datasets poses two challenges: the likelihood of accepting a hypothesis by chance, i.e., returning false discoveries, and the pitfall of not being representative of the input data. We develop GroupTest, a framework for group testing that addresses both challenges. We formulate CoverTest, a generic top-n problem that seeks n user groups satisfying one-sample, two-sample, or multiple-sample tests, and maximizing data coverage. We show the hardness of CoverTest and develop a greedy algorithm with a provable approximation guarantee as well as a faster heuristic-based algorithm based on α-investing. Our extensive experiments on four real-world datasets demonstrate the necessity to optimize coverage for sound data-driven discoveries, and the efficiency of our heuristic-based algorithm.

Despite the proliferation of recommendation algorithms, the question of which recommender works best for which user-item instance remains widely open. In this paper, we develop a meta-learning approach that chooses among several recommendation algorithms, which one is best suited for predicting the preference of a user for an item. We propose an empirical investigation of the meta-learner when applied to implicit and explicit datasets. The meta-learner is trained using four classifiers/regressors: logistic regression, decision trees, stochastic gradient descent, and gradient boosting. We run extensive experiments on four real datasets: RETAIL, a proprietary implicit dataset provided by our industrial partner, TAFENG, a publicly available grocery shopping dataset and two publicly available AMAZON datasets with explicit preferences. Results show that using a meta-learner yields higher accuracy than single recommendation algorithms for explicit datasets when compared to state-of-the-art ensemble-learned models and factorization machines. This work is an ongoing collaboration with the marketing department of a major industrial partner to test promotional offers for different customer segments.

Industrial applications of recommendation systems aim at recommending top-N products that are the most appealing to their customers, often focusing on those products that customers are likely to purchase in the near future. In this experiments and analyses paper, we present an extensive experimental evaluation of various top-N collaborative filtering recommendation algorithms based on a real-world dataset of customer's purchase history provided by our business partners at TOTAL. Our study aims to compare representative collaborative filtering approaches in practice and study the ones yielding the highest recommendation accuracy, with respect to well-established evaluation measures. These experiments are part of the development of a promotional offers campaign for TOTAL customers owning a loyalty card. We show how different settings for training and applying the selected algorithms influence their absolute and relative performances. The results are valuable to our TOTAL partners as they constitute the first large-scale analysis of recommendation algorithms in the context of their datasets. In particular, the study of the impact of recency in the training set and the role of customer activity and of context in recommendation shed light on a finer design of promotional product campaigns.

We propose a bi-objective formulation for product recommendations. Our formulation goes beyond traditional recommendations by capturing two conflicting objectives: utility that serves customers' interests, and profit margin, a business-oriented goal. To satisfy the needs of our business partners, we formulate a new problem, namely generating a result containing all sets of k products such that there does not exist any other set of k products that dominates the returned sets, i.e., whose cumulative values for each objective is higher than a set of k products in the result. We study properties of k-Pareto sets that enable us to reduce the number of candidates, as well as the number of dominance tests between candidate sets. We develop a dynamic programming algorithm that leverages those properties to prune the space of solutions. We generalize traditional measures of recommendation accuracy to be applicable to sets of k products. Our experiments on a large set of real customer transactions validate the need for a bi-objective optimization to reconcile customer and business interests, and the scalability of our solution.

Classical recommender systems provide users with ranked lists of recommendations, where each one consists of a single item. However, these ranked lists are not suitable for applications such as trip planning, which deal with heterogeneous items. In this paper, we focus on the problem of recommending a set of packages to the user, where each package is constituted with a set of different Points of Interest that may constitute a tour. Given a collection of POIs, our goal is to recommend the most interesting packages for the user, where each package satisfies the budget constraints. We formally define the problem and we present a novel composite recommendation system, inspired from composite retrieval. Experimental evaluation of our proposed system, using a real-world dataset demonstrates its quality and its ability to improve both diversity and relevance of recommendations.

Classical recommender systems provide users with ranked lists of recommendations that are relevant to their preferences. Each recommendation consists of a single item, e.g., a movie or a book. However, these ranked lists are not suitable for applications such as travel planning, which deal with heterogeneous items. In fact, in such applications, there is a need to recommend packages the user can choose from, each package being a set of Points of Interest (POIs), e.g., museums, parks, monuments, etc. In this paper, we focus on the problem of recommending a set of packages to the user, where each package is constituted with a set of POIs that may constitute a tour. Given a collection of POIs, where each POI has a cost and a time associated with it, and the user specifying a maximum total value for both the cost and the time (budgets), our goal is to recommend the most interesting packages for the user, where each package satisfies the budget constraints. We formally define the problem and we present a novel composite recommendation system, inspired from composite retrieval. We introduce a scoring function and propose a ranking algorithm that takes into account the preferences of the user, the diversity of POIs included in the package, as well as the popularity of POIs in the package. Extensive experimental evaluation of our proposed system, using a real dataset demonstrates its quality and its ability to improve both diversity and relevance of recommendations.