Research On Factorization Machine And Deep Learning For Recommendation

In today's Internet era,the growth rate of information is very amazing.The rec-ommendation system has become more and more important in our lives and has been extensively researched in recent years.As an upgraded version of matrix factorization,the factorization machine has been extensively studied.However,it still belongs to linear models and cannot learn the com-plex non-linear relationships between users and items.As deep variants of the factoriza-tion machine,NFM and DeepFM introduce deep neural networks into the factorization machine,effectively solving the above problems.Nevertheless,they still have several key problems: 1)no distinction is made between various the input features? 2)the user's historical record cannot be effectively utilized? 3)the importance of auxiliary information cannot be adaptively weighed.In view of the above problems,this thesis conducts in-depth research,proposes a memory-aware collaborative filtering algorithm,and further proposes a memory-aware gated factorization machine algorithm.Experimental results on multiple real datasets verify the effectiveness of the algorithm proposed in this thesis.The main research work of this thesis is as follows:1.This thesis conducts in-depth research on hybrid recommendation algorithms that consider user historical interaction records and introduce additional auxiliary information,and analyzes the advantages and disadvantages of existing research.Based on the idea of matrix factorization and memory network,this thesis proposes a memory-aware collabo-rative filtering algorithm(MACF),which uses a deep neural network to learn the user's recent preferences from the user's recent interacted items.And then the user's recent pref-erences are combined with the user's long-term preferences,it can make MACF more accurately predict the items that the user is interested in.2.Inspired by the gated filtering unit in the LSTM network,this thesis designs a gated filtering unit that can eliminate the negatively affected features in the auxiliary informa-tion.In the situation where auxiliary information can be obtained,this thesis proposes a memory-aware gated factorization machine algorithm(MAGFM),which can effectively use the user's historical interaction items and the auxiliary information about the user and item.It improves the factorization machine method in the following ways: 1)Introduce an external user storage matrix to each user,and enrich the expressive power of the model by using the user's historical interaction items and auxiliary information related to the his-torical interaction items.2)The gated filtering unit is applied to the user/item auxiliary information,and can adaptively filter out features with negative effects to achieve higher accuracy? 3)Design a simple and effective calculation method to distinguish the input features,avoiding many redundant second-order interactions in the factorization machine and its deep variants.3.For the proposed algorithms in this thesisi,extensive experiments and analyses are performed on multiple real datasets,and the effectiveness of the proposed algorithms is verified.