Local Item-Item Models for Top-N Recommendation

Several algorithms for the top-N recommendation problem have been developed. The latent space methods factorize the user-item matrix into lower rank user factor and item factor matrices and the neighborhood-based methods identify similar users or items. The item-based methods, which include item k-NN and Sparse Linear Methods (SLIM) have been shown to outperform the user-based schemes for the top-N recommendation task. However, item-based methods have the drawback of estimating

only a single model for all users.

In this paper they present a top-N recommendation method that extends the SLIM model in order to capture the differences

in the preferences between different user subsets. Their method, which is called GLSLIM (Global and Local SLIM), combines global and local SLIM models in a personalized way and automatically identifies the appropriate user subsets.

GLSLIM computes a global model and has a personalization factor for each user determining the interplay between the global and the local information. GLSLIM updates the assignment of the users to subsets, allowing better local models to be estimated.

Previous works use user and item latent factors, while GLSLIM focuses on item-item models.

A global item-item model may not be sufficient to capture the preferences of a set of users, especially when there are user subsets with diverse and sometimes opposing preferences.

GLSLIM computes top-N recommendations that utilize user-subset specific models and a global model. These models are jointly optimized along with computing the user assignments for them. The estimation of the item-item coefficient matrices, the user assignments and the personalized weight is done with alternating minimization.

To estimate item-item model, they separate the users into subsets with either a clustering algorithm (CLUTO) or randomly. Then

initially set the parameters in order to have equal contribution of the global and the local part and estimate the coefficient matrices. Following SLIM, the item-item coefficient matrices can be calculated per column, which allows for the different columns (of both the global and the local coefficient matrices) to be estimated in parallel. GLSLIM solves an optimization problem using coordinate descent and soft thresholding. After estimating the local models (and the global model), GLSLIM fixes them and proceeds with the second part of the optimization: updating the user subsets. While doing that, GLSLIM also determines the personalized weight.

They employed leave-one-out cross-validation to evaluate the performance of the proposed model. For each user, we randomly selected an item, which we placed in the test set. The rest of the data comprised the training set.

A Survey on Transfer Learning

This survey categorizes and reviews the latest progress on transfer learning in different data mining and machine learning areas. Transfer learning emerged to overcome the problem that assumes training and test set data can’t be in different feature space and should have the same distribution. Therefore, models don’t need to be rebuilt after using new data which is expensive and sometimes impossible. The main focus of the paper is on transfer learning for classification, regression and clustering problems.

Traditional supervised and semi-supervised methods can remove noisy data or use cost-sensitive learning but most of them assume that the labeled and unlabeled data have the same distributions. However, transfer learning lets the training and test data distribution to be different. Traditional machine learning techniques try to learn each task from scratch, while transfer learning techniques try to transfer the knowledge from some previous task to a target task when the latter has fewer high-quality training data.

In transfer learning, we have the following three main research issues: 1) what to transfer, 2) how to transfer, 3) when to transfer. “What to transfer” asks which part of knowledge can be

transferred across domains or tasks. “how to transfer “corresponds to learning algorithms need to be developed to transfer the knowledge “When to transfer” asks in which situations, transferring skills should be done.

categorize transfer learning under three subsettings, inductive transfer learning, transductive transfer learning, and unsupervised transfer learning, based on different situations between the source and target domains and tasks.

Matrix factorization techniques for recommender systems

Recommender systems recognize patterns of user preference in products and provide personalized recommendations for users and are used in a variety of online shopping websites. Content filtering approach and collaborative filtering are two main methods used in recommendation systems. In former method, the characteristics of users and products are stored but the latter, needs the user behavior in the past. Collaborative filtering is more accurate than content based methods but doesn’t work very accurate for new products and users. Collaborative filtering approach is based on neighborhood methods and latent factor models. Neighborhood methods finds the relation between both users and items. In other words, the interest of a user to and item, could be based on the rating of the similar items. But in latent factor models, both items and users are characterized. Matrix factorization methods that are concerned in this paper are the base of latent factor models which characterize users and items by vectors. More interaction between user and item factors recommends items for the user. In these systems usually input data are derived from explicit feedback and marked on the dimensions of a matrix which are users and items. The matrix is likely sparse but there are methods that can overcome this issue.

Matrix factorization models are modeled as inner product of users and items on a joint latent factor space. Due to sparsity of the matrix, this method has difficulties. There are methods that can dominate this problem and add data to the matrix but are expensive and create massive amount of data. Therefore, methods using observed data are recommended.

The factor vectors of users and items are learned by Minimizing the regularized square error on the set of known ratings with stochastic gradient descent and alternating least squares which are learning algorithms. The extent of the regularization is determined by cross-validation.

Various data aspects and other application-specific requirements can be considered in matrix factorization approach. As we know, the variation in ratings is because of effects associated with either users or items. This variation is called biases or intercepts and are independent of interactions. To consider the biases, observed rating is broken down into global average, item bias, user bias and user-item interaction. The square error function should be minimized in order to make the system learn.

When there isn’t enough rating from the users, we can incorporate additional sources of information like implicit feedback or user attributes. The matrix factorization model integrates all signal sources so items can get a similar treatment when necessary.

In reality product perceptions, popularity and customer’s inclination evolve over time. This temporal effects should be considered in the system. The matrix factorization approach assumes these terms vary over time: item biases, user biases and user preferences. Temporal dynamics also affect the interaction between users and items.

In another situation, the ratings don’t have same weight or confidence. The matrix factorization model can easily accept varying confidence levels by giving less weight to less meaningful observations.

The more complex factor models with different amount of details, are more accurate. Since there are temporal effects in the data, the temporal components are very important.

The experiments in this paper are done on Netflix dataset. The results are superior to classical nearest-neighbor techniques and their model is memory-efficient. Also many crucial aspects of the data in forms of feedback, temporal dynamics and confidence levels are integrated to make the model accurate.