Universidade de São Paulo Biblioteca Digital da Produção Intelectual - BDPI Departamento de Ciências de Computação - ICMC/SCC Comunicações em Eventos - ICMC/SCC 2014-11 Combining multiple metadata types in movies recommendation using ensemble algorithms Brazilian Symposium on Multimedia and the Web, 20th, 2014, João Pessoa. http://www.producao.usp.br/handle/BDPI/48641 Downloaded from: Biblioteca Digital da Produção Intelectual - BDPI, Universidade de São Paulo Combining Multiple Metadata Types in Movies Recommendation Using Ensemble Algorithms Bruno Cabral Renato D. Beltrão Marcelo G. Manzato Department of Computer Mathematics and Computing Mathematics and Computing Science Institute - University of São Institute - University of São Federal University of Bahia Paulo Paulo Salvador, Brazil São Carlos, SP – Brazil São Carlos, SP – Brazil [email protected] [email protected] [email protected] Frederico Araújo Durão Department of Computer Science Federal University of Bahia Salvador, Brazil [email protected] ABSTRACT important tools in assisting users to filter what is relevant In this paper, we analyze the application of ensemble al- in this complex information world. There are a number of gorithms to improve the ranking recommendation problem ways to build recommender systems; they are classified as with multiple metadata. We propose three generic ensemble content-based filtering, collaborative filtering or the hybrid strategies that do not require modification of the recom- approach, which combines both filtering strategies [1, 5]. mender algorithm. They combine predictions from a recom- Content-based filtering recommends multimedia content mender trained with distinct metadata into a unified rank to the user based on a profile containing information re- of recommended items. The proposed strategies are Most garding the content, such as genre, keywords, subject, etc. Pleasure, Best of All and Genetic Algorithm Weighting. These metadata are weighted according to past ratings, in The evaluation using the HetRec 2011 MovieLens 2k dataset order to characterize the user's main interests. However, with five different metadata (genres, tags, directors, actors this approach has problems such as over-specialization [1] and countries) shows that our proposed ensemble algorithms and limited performance due to metadata scarcity or quality. achieve a considerable 7% improvement in the Mean Aver- An alternative to this problem is the collaborative filtering, age Precision even with state-of-art collaborative filtering which is based on clusters of similar users or items. One dis- algorithms. advantage of collaborative filtering is the computational ef- fort spent to calculate similarity between users and/or items in a vectorial space composed of user ratings in a user-item Categories and Subject Descriptors matrix. H.3.3 [Information Search and Retrieval]: Information Such limitations have inspired researchers to use matrix Filtering factorization techniques, such as Singular Value Decomposi- tion (SVD), in order to extract latent semantic relationships General Terms between users and items, transforming the vectorial space into a feature space containing topics of interest [20, 11, 17, Design, Algorithms 10]. Nevertheless, other challenges have to be dealt with, such as sparsity, overfitting and data distortion caused by Keywords imputation methods [10]. recommendation; ensemble; metadata; movie; collaborative Considering the limitations and challenges depicted above, filtering hybrid recommenders play an important role because they group together the benefits of content based and collabo- 1. INTRODUCTION rative filtering. It is known that limitations of both ap- proaches, such as the cold start problem, overspecialization Recommender systems have become increasingly popular and limited content analysis, can be reduced when combin- and widely adopted by many sites and services. They are ing both strategies into a unified model [1]. However, most Permission to make digital or hard copies of all or part of this work for recent systems which exploit latent factor models do not con- personal or classroom use is granted without fee provided that copies are not sider the metadata associated to the content, which could made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components provide significant and meaningful information about the of this work owned by others than ACM must be honored. Abstracting with user's interests. Another issue of current metadata aware credit is permitted. To copy otherwise, or republish, to post on servers or to recommenders is that usually they support only one type redistribute to lists, requires prior specific permission and/or a fee. Request of item attribute at a time. To overcome this issue, Bel- permissions from [email protected]. tr~ao et al. [3] analyzed the performance of a recommender WebMedia’14, November 18–21, 2014, João Pessoa, Brazil. using multiple types of metadata, by concatenating the dif- Copyright 2014 ACM 978-1-4503-3230-9/14/11 ...$15.00. http://dx.doi.org/10.1145/2664551.2664569. 231 ferent pieces of information, and although the performance a set of metadata aware algorithms which use the Bayesian improved, the results were still modest. Personalized Ranking (BPR) framework [6] to personalize a Similarly to Beltr~ao et al. [3], this paper proposes a differ- ranking of items using only implicit feedback. These tech- ent approach for handling multiple metadata, using ensem- niques will be considered in our evaluation in the context of ble algorithms. We use three different ensemble strategies movies recommendation. to combine different metadata, but with the advantage that it does not require the algorithm to be modified, or to be 3.1 Notation trained multiple times with the same dataset, and therefore, Following the same notation in [10, 12], we use special it can be used in all current Recommender Systems. indexing letters to distinguish users, items and attributes: This work is structured as follows: in Section 2 we review a user is indicated as u, an item is referred as i; j; k and an related works that use ensemble algorithms; in Section 3 we item's attribute as g. The notation rui is used to refer to briefly describe the models considered in this evaluation; in explicit or implicit feedback from a user u to an item i. In the Section 4 we detail our proposed Ensemble framework and first case, it is an integer provided by the user indicating how strategies; Section 5 presents the evaluation and validation much he liked the content; in the second, it is just a boolean of the approach with HetRec dataset with 855598 ratings, indicating whether the user consumed or visited the content and analysis of the performance of the three proposed strate- or not. The prediction of the system about the preference gies; and finally, in Section 6 we discuss the final remarks, of user u to item i is represented byr ^ui, which is a floating future work and acknowledgments. point value calculated by the recommender algorithm. The set of pairs (u; i) for which rui is known is represented by the set K = f(u; i)jrui is knowng. 2. RELATED WORK Additional sets used in this paper are: N(u) to indicate An ensemble method combines the predictions of different the set of items for which user u provided an implicit feed- algorithms, or the same algorithm with different parameters back, and N¯(u) to indicate the set of items that is unknown to obtain a final prediction. Ensemble algorithms have been to user u. successfully used, for instance, in the Netflix Prize contest consisting of the majority of the top performing solutions. 3.2 BPR-Linear [23, 18]. The BPR-Linear [6] is an algorithm based on the Bayesian Most of the related works in the literature point out that Personalized Ranking (BPR) framework, which uses item ensemble learning has been used in recommender system as attributes in a linear mapping for score estimation. The a way of combining the prediction of multiple algorithms prediction rule is defined as: (heterogeneous ensemble) to create a stronger rank [9], in a technique known as blending. They have been also used n X with a single collaborative filtering algorithm (single-model r^ui = φf (~ai) = wugaig ; (1) or homogeneous ensemble), with methods as Bagging and g=1 Boosting [2]. However, those solutions do not consider the n where φf : R ! R is a function that maps the item at- multiple metadata present in the items, and are often not tributes to the general preferencesr ^ui and ~ai is a boolean practical to implement in a production scenario because of vector of size n where each element aig represents the oc- the computational cost and complexity. In the case of het- currence or not of an attribute, and wug is a weight matrix erogeneous ensemble, it needs to train all models in parallel learned using LearnBPR, which is variation of the stochastic and treat the ensemble as one big model, but unfortunately gradient descent technique [7]. This way, we first compute training 100+ models in parallel and tuning all parameters the relative importance between two items: simultaneously is computationally not feasible [23]. In con- trast, the homogeneous ensemble demands the same model s^uij =r ^ui − r^uj to be trained multiple times, and some methods such as n n Boosting requires that the underlying algorithm be modi- X X = wugaig − wugajg fied to handle the weighted samples. Beltr~ao et al. [3] tried g=1 g=1 (2) n a different approach and combined multiple metadata by X concatenating them, with a modest performance increase. = wug(aig − ajg) : In comparison to the above approaches, our method uses g=1 three different ensemble strategies to combine distinct meta- Finally, the partial derivative with respect to wug is taken: data, but with the advantage that it does not require the al- gorithm to be modified, or to be trained multiple times with @ s^ = (a − a ) ; (3) the same dataset, and therefore, it can be used in all of the @w uij ig jg current Recommender Systems.