MODELO DE ACONDICIONAMIENTO DE UN LABORATORIO DE METROLOGÍA

3.1

Introduction

Social recommender systems (SRS) are important to help users to find relevant content. This is in part because of social media contents now account for the majority of content published on web. Typical social recommender systems assume a social network among users and makes recommendations based on the ratings of the users that have direct or indirect social relations with the target user (JAMALI; ESTER, 2010).

However, explicit user’s ratings suffer from two known drawbacks: (i) The problems of calibration (consistency), which consists in incompatible users ratings on same scale, for example, on 1 to 5 star ratings scale, a rating of 4 for user X might be comparable to a rating of 5 for user Y. (ii) Resolution (granularity), this problem states that any numeric scale for ratings, say 1 to 5 stars, may be insufficient to capture all the users interests without loss of information (BALAKRISHNAN; CHOPRA, 2012) (AMO; RAMOS, 2014).

Thus, building on PrefRec (AMO; OLIVEIRA, 2014), we propose So cial Pr e fRec a social recommender that applies user preference mining and clustering techniques to incorporate social information on the pairwise preference recommender system. Besides the advantage of the pairwise preference model, our focus is mitigate the user cold-start problem and sparsity.

Researches related to collaborative social recommendation argue that social informa­ tion can easily deal with cold users and data sparsity, because instead of relying only in user’s preferences they use available ratings from users whose hold a relationship with the target user (MA et al., 2011; WANG et al., 2014). In this work, we propose an approach to incorporate social network to provide recommendations. To leverage social influence in our model, we exploit several well know social network metrics.

In addition, social recommender systems in general make use of social information to build prediction models (recommendation models). Thus, for each cold user a new model must be built. In comparison, our approach harnessing pre-existent models. Instead of building a new model from scratch for each cold user, we cluster existent users and generate prediction models for each group. Through social information we select among existent models the most appropriated for a cold user.

Different factors of social relationships have influence on cold users. Some of these factors contribute or even harm social recommender systems (YUAN et al., 2015). Un­ derstanding the extent to which these factors impact SR systems provides valuable insights for building recommenders. We investigated the role of several social metrics on pairwise preference recommendations. Given that user’s preference is similar to or influenced by their connected friends (TANG; HU; LIU, 2013), we also studied how to apply social similarities in a pairwise preference recommender. So cial Pr e fRec is evaluated on two datasets, named Facebook and Flixster, to verify the integrity of our results. Focusing on social pairwise preference recommendation, our study addresses six questions:

Q1: How accurately does social information help on item recommendation?

We assessed the accurateness of So cial Pr e fRec by comparing it to Pr e fRe c. This is the key to determine whether a pairwise preference recommender can beneht from social information.

Q2: How relevant are the recommendations made by a social pairwise preference rec- ommender?

One of the main reasons for the relevance of So cial Pr e fRec is to mitigate the cold start problem for users through social information. To further assess our model, we compare So cial Pr e fRec to three state-of-art social recommenders.

Q3: Which social metrics are the most important for item recommendation?

The previous questions focus on understanding whether pairwise recommenders could beneht from contextual social information. Here, we want to evaluate the overall perfor­ mance of each social metric: friendship, mutual friends, similarity, centrality and interac­ tion.

Q4: How effective is So cial Pr e fRec to mitigate data sparsity problems?

In social recommender systems there is a common assumption that contextual social information mitigates data sparsity problems. To assess our model in this context, we evaluate the effectiveness of So c ia l Pr e fRec with regards to Pr e fRec against hve data sparsity levels.

Q5: Does social degree affect So c ia l Pr e fRec as much as profile length affects Pre­ fRec?

To achieve high-quality personalization, recommender systems must maximize the information gained about users from item ratings. The more ratings a user’s prohle has, the merrier will be. We want to check whether increasing the number of friends impacts

3.2. Background 37

our approach.

Q6: Are there major differences between recommendations quality of popular and un­ popular users?

Here we further investigate social popularity effects on recommender systems. This question complements Q5, offering valuable insights into when and which social metric impacts the predictions.

This chapter is organized as follows. Section 3.2 presents the background knowledge undertaking in this chapter. Section 3.3 describes our proposed framework the Social Pr e fRe c, as well as the applied social metrics and recommender model selection strate­ gies. Section 3.4 describes our experimental settings and Section 3.5 presents the results. Finally, Section 3.6 concludes the chapter.

3.2

Background

In this section, we introduce the main concept underlying this chapter: pairwise pref­ erence recommender systems.