A Stacked Denoising Autoencoders Based Collaborative Approach For Recommender System

The rapid development of Internet technology has brought human beings into an new information era.People create,store and inquire information at an exponentially growing speed.In particular,the continuous expansion in e-commerce has a huge impact on the life of individuals.Electronic retailers and content providers offer huge selections of products and colorful information to meet a variety of special needs and tastes of customers.Immersed in the ocean of unprecedented opportunities,Customers are often getting lost and have to waste a lot of time and energy to find their own personalized needs of the information or merchandise.As an effective tool to alleviate the status of information overload,personalized recommendation technologies can help users find their preferred items easily and quickly and have been widely used in the industry.However,there are still many problems to be solved,and recommendation technology still maintains a continuous research heat in the academia.Although the deep learning methods have achieved excellent results in various fields in recent years,the fields of recommendation system rarely use the deep learning method.Firstly,this paper analyzes the advantages and disadvantages of various methods in the existing recommendation technology,and then deeply studies the principle and scope of each method.Secondly,to handle the problem of unreliable similarity calculation between users or items leading to the poor recommendation quality caused by the sparseness of the rating matrix,this thesis proposes a Stacked Denoising Auto-encoder based Collaborative Filtering algorithm,referred to as SDAE-CF algorithm.Aiming to introduce the auto-encoders into the recommendation system to extract the user's preference feature from the rating matrix to assist the collaborative filtering recommendation task,further to improve the recommendation quality.The main innovation works of this thesis has two points,summarized as follows:(1)Introduce the stacked de-noising auto-encoders as a user preference learning component in collaborative filtering recommendation task.In this thesis,a method named SDAE-CF was proposed.Compared with the traditional recommendation algorithms taking the whole products in a system as the dimension of the user preference model,the proposed method reduces the dimension of the user preference feature representation and obtains the relatively dense user preference vector.In order to adapt to the recommended task,we reconstruct the structure of prototype auto-encoders and also following the new training strategy.As for structure,the noise model is introduced in the prototype auto-encoders and then stacking them together,enhancing the ability to extract the features of auto-encoders.Meanwhile,as for the training strategy,the tightly-coupled synchronous training method was employed,reducing the number of super-parameters and decreasing the training difficulty for the single user's personalized auto-encoders.(2)With certain compromising in accuracy,the modified Hamming distance is used as the similarity calculation method of the users' preference vectors,so that the similarity calculation between users becomes simple and efficient.Compared with the other similarity standards,the experiments results show that the proposed method still can achieve good recommended performance.In conclusion,this thesis introduces stacked de-noising auto-encoders for collaborative filtering task to build the SDAE-CF model.This model alleviates the unreliable similarity calculation and low recommendation accuracy caused by the sparseness of rating matrix in the recommendation system.In addition,the SDAE-CF algorithm has better expansibility for Top-N recommendation task on the implicit feedback dataset.Experiments on MovieLens datasets and the result confirmed the effectiveness of SDAE-CF model.