Personalized Sets Recommendation System

With the rapid growth of fashion-focused social networks and online shopping,intelligent personalized fashion recommendation is now in great need.Different to traditional recommendation system which recommend single item,fashion outfits recommendation consider the combination of items from different category.Due to the tremendous number of oufits in combination space,different users have rarely posted the same outfits.That makes the collaborative filtering can hardly applied into the problme of set recommendation.On the other hand,factorization models have been extensively used for recommendation system for recovering the missing entries of a matrix or tensor.By learning the feature vectors of items and users,the value in target tensor can be computed based on a decomposition model.However,direct computing all entries using the learned factorization models is prohibitive when the size of matrix/tensor is large.We propose two ways to imporve the efficiency the recommender system.Firstly,we propose a high-performance,sampling-based algorithm for finding the top entries of a tensor which is decomposed by the CANDECOMP/PARAFAC model.We develop an algorithm to sample the entries with probabilities proportional to their values for the general problem.We further extend it to make the sampling proportional to the 6)-th power of the values,amplifying the focus on the top ones.We provide theoretical analysis of the sampling algorithm and evaluate its performance on several real-world data sets.One the other hand,a hashing tecnique is developed to make the recommendation system more efficient.We design algorithms to learn a compact binary codes for items and users,which suggest users outfits and fit their personal fashion preferences.Personalized sets recommendation system is relatively new to recommender systems.The object is a set which is composed of multiple interacted items.We explore the use of deep neural networks for this challenging task.Our proposed model consists of two components,a feature network for feature extraction and matching networks based on tensor decomposition to model the score of outfits.We introduce two kinds of binary codes and evaluate our algorithm on a large scale data set.We run comprehensive experiments to show the efficiency of our approaches.