Research Of Recommendation Algorithm Based On Implicit Vector Representation And Domain Adaptation

Recommendation system has a wide range of application scenarios in various online services.Its role is to recommend matching projects based on users' interests,improve user experience and bring more benefits to the system,so as to achieve a win-win situation between users and the system.Recommendation algorithms based on deep learning are a popular research direction.These methods map features into hidden vectors of low latitude,but they cannot effectively measure the similarity between hidden vectors.The information in the recommendation system is related to each other,according to which a graph network can be constructed.The network representation learning method can extract the unstructured information in the graph and enrich the expression of feature hidden vectors.Text information exists in some recommendation scenarios,and effective use of auxiliary information such as text category and emotion can improve recommendation performance.Complex recommendation system has multi-domain text information,which has the problem of domain adaptation and the difficulty of extracting auxiliary information.In order to solve the above problems,Deep Neural Factorization Machine for Recommendation Systems,Recommendation algorithm based on Heterogeneous Graph and Distant Domain Adaption for Text Classification in the recommendation scenario are studied.The specific work is as follows:(1)Deep Neural Factorization Machine for Recommendation Systems.In this paper,a recommendation model,Deep Neural Factorization Machine DNFM based on "wide and deep" structure is proposed.The "wide" part is an improved model,Dimension-weighted Factorization Machine DwFM based on FM.DwFM improves the calculation method of feature similarity and learns different weight vectors for different cross fields,so as to measure feature similarity more accurately.The "deep" part is a neural network that can capture the high-order nonlinear information of the data,which improves the generalization ability of the model.(2)Recommendation algorithm based on Heterogeneous Graph.A recommendation model,Personalized Taxi Demand Prediction TDP based on heterogeneous graph embedding is proposed to recommend the starting point and ending point pairs to users.Firstly,multiple heterogeneous maps were constructed based on the co-occurrence information and time attribute information of the region,and the vectorization representation of the region was obtained by combining multiple heterogeneous maps.Then vectorization recall technique is used to obtain the starting point and the end point set of users.Finally,the sorting model based on deep neural network is used to get the pairs of Top K starting point and ending point that users are interested in.(3)Distant Domain Adaption for Text Classification in the recommendation scenario.Selective Domain Adaption Algorithm SDAA based on selection iteration is proposed to provide good multi-source information assistance for recommendation.Text is an important auxiliary information of recommended applications,which has the problem of distant domain adaptation.In SDAA iteration,meaningful data are selected from the source domain and the intermediate domain based on the similarity of category and structure,then knowledge transfer is conducted with the target domain,and manifold loss is used to control the process of iterative optimization.