Movie Recommender System for Profit Maximization (Short LBP)

In this paper we provide an algorithm for utility maximization of a movie supplier service, in two different settings, one with prices and the other without. This algorithm is provided along with an extensive experiment demonstrating its performance. We also uncover a phenomenon where movie consumers prefer watching and even paying for movies that they have already seen in the past than movies that are new to them. Introduction & Related Work The main goal in designing recommender systems is usually to predict the user’s wish list and to supply her with the best list of recommendations. However, in most cases, the engineers that design the recommender system are hired by the business which provides the suggestions. The business’ end goal is usually to increase sales, revenues, user engagement, or some other metric. In that sense, the user is not the end customer of the recommendation system, although she sees the recommendations (Lewis 2010). Still, one could argue that it is better for the business to give the user the best possible recommendations, as it will also maximize the business’ profit, either in the short run (no point in giving recommendations which are not followed by the users) or at least in the long run (good recommendations make users happy). In this paper, we provide evidence that a business may gain significantly (with little or no long-term loss) by providing users with recommendations that may not be best from the users point of view but serve the business’ needs. We provide an algorithm which uses a general recommender system as a black-box and increases the utility of the business. We perform extensive experiments with it in various cases. In particular, we consider two settings: 1. The Hidden Agenda setting: In this setting, the business has items that it wants to promote, in a way which is opaque to the user. For example, a movie supplier which provides movies on a monthly fee basis but has different costs for different movies. 2. The Revenue Maximizing setting: In this case the goal of the recommender system is to maximize the expected revenue, e.g. by recommending expensive items. In this Copyright c © 2013, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. setting, there is an inherent conflict between the user and the business. An interesting phenomenon that we uncover is that subjects are more willing to pay for movies that they’ve already seen. While a similar phenomena is known for other types of consumer goods, coming across it with regards to movies is new and somewhat counter-intuitive. We supply some explanation for this phenomenon, and explain how it can improve the design of recommender systems for movies. However, further research is required. Chen et al. (2008) develop a recommender system which tries to maximize product profitability. Chen et al. assume the usage of a collaborative filtering recommender system which, as part of its construction, provides a theoreticallybased probability that a user will purchase each item. They multiply this probability by the revenue from each item and recommend the items which yield the highest expected revenue. However, in practice, many recommender systems do not rely only upon collaborative filtering, but also rely on different engines (such as popularity, semantic similarity, etc.). Therefore, we assume a generic recommender system which is treated as a black-box component, and dedicate most of our work to building a human model in order to predict the acceptance rate of a given item using a generic recommender system. In (Azaria et al. 2012) we model the long-term affect of advice given by a self-interested system on the users in path selection problems. However, in (Azaria et al. 2012) we assume that the user must select his action among a limited number of options and the system merely recommends a certain action. Therefore the system does not act as a classic recommender system, which recommends a limited number of items from a very large corpus. Still, this work may be found useful if combined with the approach given in this paper, when considering repeated interactions scenarios. Profit and Utility Maximizing Algorithm In this section we present the Profit and Utility Maximizer Algorithm (PUMA) for the revenue maximizing setting. PUMA mounts a black-boxed recommender system which supplies a ranked list of movies. In order to learn the impact of the price on the likelihood of the users buying a movie, we use the recommender system as is, providing recommendations from 1 to n. We 5 Late-Breaking Developments in the Field of Artificial Intelligence Papers Presented at the Twenty-Seventh AAAI Conference on Artificial Intelligence