Similarity Neural Network Based Collaborative Annotation Model

As a mature technology,collaborative annotation is often used to solve problems such as indexing,management,and retrieval of information resources,etc.However,there are cold start problems caused by insufficient or no interactive information between data.At present,the commonly used solution is to use content information to enhance and improve annotation algorithm,especially text content-based collaborative filtering.However,how to better extract text features and measure item similarity still needs to be further explored.In view of the above problems,this thesis proposes a Similarity Neural Network based Collaborative Annotation model(SNNCA)which combine Sparse Linear Method(SLIM)with Siamese Convolutional Neural Network(SCNN).This model uses the SCNN component for feature extraction and item similarity calculation of content infor-mation,and its output can be regarded as the nonlinear similarity measurement function of two input contents.To solve the parameter training problem of SCNN,this thesis uses SLIM to learn the similarity relationship sparse coefficient matrix W between existing item-annotation data,and use the obtained results to train the parameters in the SCNN component.Based on this idea,this thesis specially designs the model learning method,and realizes the joint optimization of model parameters through cross-iterative method,that is,in each iteration,it fixes SCNN component parameters,solves SLIM parameter W by coordinate descent method;Then,W is fixed as the training mark of SCNN component,and a least squares problem is solved by the back-propagation method to update the parameters of SCNN component.The innovation of this thesis lies in the introduction of neural network components to measure the similarity of item content,and the organic combination with item-based collaborative filtering model,meanwhile,it solves the problems of feature extraction and similarity measure learning,etc.This thesis examines the performance of the proposed model using three different datasets respectively collected from CiteULike and Programmable Web.After a thorough experiment,the validity of the proposed model is verified.Through quantitative comparison,it is found that the proposed model is 10%and 6%higher than the baseline models in the accuracy of recommendations.After the comparative analysis of different dimensions,it is found that the richer the representation of content information is,the better the performance of the model will be,and the sparser the data set is,the more prominent performance will be achieved.