Decision-Guided Recommenders With Composite Alternatives

Abstract

Recommender systems aim to support users in their decision-making process while interacting with large information spaces and recommend items of interest to users based on preferences they have expressed, either explicitly or implicitly. Recommender systems are increasingly used with product and service selection over the Internet. Although technology has made it easy to search for and interact with most information types, the volume surge in data presented is overwhelming and hard to filter. While state-of-the-art recommender systems focus on atomic products or services, this research focuses on developing a framework, models and algorithms for recommending composite services and products based on decision optimization. Composite services are characterized by a set of sub-services, which, in turn, can be composite or atomic and make the recommendation space very large (or infinite, for continuous case). Complex recommendation models involving composite alternatives, such as product configurations and service packages, have not been addressed. The proposed framework contains models that allow for fast and easy user preference elicitation that can be captured in a utility function, and provides algorithms for diversifying a recommendation set. Such recommendations will be dynamically defined using database views extended with decision optimization based on mathematical programming.

A key challenge addressed in this research is combining the flexibility of diversity ranking functionality with the capabilities of information processing to learn and capture users' preferences through an iterative learning process.

The proposed framework presents a method for utility function elicitation, which is based on iteratively refining a set of axes in the n-dimensional utility space. User preferences are initially learned using regression analysis or Collaborative Filtering (CF) techniques. At every step, the user is asked to rank a set of recommendations, each being optimal for one of the current axes. Based on the user feedback, utility axes are adaptively adjusted based on a confidence degree. Consequently, the utility function is constructed. In addition, the framework proposes a new approach to diversify a set of recommendations, which is based on constructing and using an m-dimensional diversity feature space, which is separate from the utility space used for utility elicitation. Furthermore, the framework presents a diversity algorithm to address the Maximum Diversity Problem (MDP) of recommendations, which is randomized and based on iterative relaxation of selections by the Greedy algorithm with an exponential probability distribution.

Finally, the proposed framework is validated with several experimental studies using publically available datasets, such as MovieLens and Yahoo, to measure the efficacy and efficiency against state-of-the art algorithms and techniques. The results show that the proposed algorithm is highly efficient computationally, and consistently outperforms competing algorithms and systems in terms of precision, recall, diversity measures, and MAE (mean absolute error). In addition, the proposed framework converges to the optimal or near-optimal solutions in under 100 ms using a machine with Intel Core 2 Duo CPU 2.53GHz and 3GB RAM. The proposed collaborative filtering technique achieved a precision of 90% on average.