Research On Attribute Based Cold-start Recommendation Problems

When recommendation system is in operation,new users and new items are added continually.Owing to the absence of historical information,traditional recommen-dation algorithms can not produce a valid recommendation directly to new users or with new items.This challenge is the cold-start issue.Cold start problems can be divided into three scenarios:user cold-start-recommending existing items to new users;item cold-start-recommending new items to existing users;system cold-start-recommending new items to new users.In order to solve the cold-start problems,an effective strategy is to consider user attributes(such as gender,occupation and age)and item properties(such as categories and manufacturers)as auxiliary information.This article focuses on attribute-based cold-start recommendations,and proposes the following four recommendation models.For the problem of adaptability of attribute measurement function,we propose an attribute metric based cold-start recommendation model(MetricRec),to deal with user cold-start scenario.Considering that users with similar attributes will show a similar interest(such as women would buy cosmetics).We search users with similar interest based on the similarity of attributes.To solve the adaptability problem of traditional similarity function(cosine and Jaccard for instances),we propose MetricRec model that learns a general similarity function such that users with similar interests are as close as possible.To efficiently optimize MetricRec,we propose an interior point based stochastic gradient descent optimization method(ISGD).In the iterating process,ISGD can adaptively adjust the learning rate such that the parameters are updated in the feasible region.Experimental results on two movie datasets show that MetricRec can effectively deal with the user cold-start problems,and obtain a similarity function that is superior to cosine similarity and Jaccard similarity.For the problem of mapping from attributes to latent profiles,we propose n-dimensional Markov random field prior constrained matrix factorization model(MRF-MF),to deal with three types of cold-start.Using the MRF constraint,MRF-MF can make users/items with similar attributes have similar latent profiles.According to Markovianity,new users/items can consider neighbors in the attribute space to ap-proximate their latent profiles.We apply alternating least squares method to optimize MRF-MF model and analyze the corresponding time complexity.Experimental results on two movie datasets show that MRF-MF is stable to parameters,can simultaneously effectively tackle three cold-start problems with a single training.Based on the observation that users/items with the same attribute will show similar behaviors,we propose a model of learning profile of attributes(LPA)to deal with three cold-start scenarios.LPA divides the latent profile of a user into two parts:an attribute profile and an individual profile.The attribute profile represents the preference of a class of users with the same attribute.The individual profile can describe the different preference between considering user and its class.Similarly,item latent profile is also composed of an attribute profile and an individual profile.In the cold-start scenarios,although the individual profile is not available,the effective recommendations can be made by attribute profiles.We apply block coordinates descent algorithm to optimize the LPA,and analyze the time complexity of optimization algorithm.Experimental results show that LPA can obtain high performance in each of three cold-start scenarios by appropriately adjusting the regulization of the attribute profile and the individual profile.Compared to MRF-MF,LPA can obtain better performance on a very sparse data set Bookcrossing.For the problem of too many parameters of LPA,we propose an attribute based multi-level preference regression model(MPR),to address three kinds of cold-start problem.MPR considers ratings as the combination of three correlations:1)corre-lation between user attributes and item attributes;2)correlation between individual user and item attributes;3)correlation between user attributes and individual item.Compared to the previous three models,MPR is a linear regression model and has a globally optimal solution.Because it is time consuming to directly calculate the global solution,we apply alternating least squares method to optimize the MPR.In addition,we analyze the internal relations,advantages and disadvantages of LPA and MPR.Experimental results on Movielens and Bookcrossing datasets show that MPR can get better performance than the LPA.Finally,we summarize the relevance,applicability,advantages and disadvantages of these four models.Based on these analyzes,we introduce several possible expansions.