Research Of Recommendation Algorithms Based On Convolutional Neural Network

New generation of information technologies such as big data and mobile Internet are used wildly and deep.They dominate the way people obtain information and bring a surge of it which people acquire.People have gradually increased the efficiency and individual needs for getting information.Exactly the recommendation service fits the users' requirements in these two dimensions.Recommendation tasks often have to handle the data of users' interactions,and sequential recommendation is suitable for this scenario.Sequential recommendation models the interaction between user and item as a dynamic sequence,and uses the sequence in chronological order to capture the user's preferences for recent period.The computing power of hardware and theories of neural network are also constantly developing and progressing.Data-driven deep learning is sweeping over.Combining it with recommendation algorithms has gradually become a hot topic.Among them,convolutional neural networks are used in sequential recommendation problems dramatically.This paper uses convolutional neural networks as the technical background,combining with recommendation algorithms,focusing on sequential recommendation problems,mining patterns in users' behavior,capturing interaction patterns of sequential data,and painting users' portraits.The main work is as follows:1.This thesis analyzes the existing recommendation algorithms,removes its shortcomings,and proposes Daser algorithm together with atrous convolution to jointly model the users' recent and enduring interests.This algorithm maps the user-item interaction sequences in the embedding space to building an ”image”,which provides convenience for the application of convolutional networks.In the meanwhile,aiming to extract different types of sequential patterns,the algorithm designs multiple types of convolutional kernels,which can be used to capture single-point,joint,discrete,and jump-type behaviors.2.This thesis combines deep neural network and sequential recommendation,extracts deep features of interactive data,and proposes a recommendation algorithm Conv FC.The algorithm transforms the embedding space dimension into the channel dimension by converting the perspective on the embedding matrix,and thus constructs a convolutional structure which is common in computer vision.It uses layer-by-layer multiplication of the number of channels to extract more abstract features and atrous convolution to enlarge receptive field,then realizes deep digging of interactive information.3.This thesis uses residual structure to deepen the network,and then proposes two different residual blocks,DBlock and VBlock,corresponding to two residual networks.The residual structure can simplify the network's learning process,redouble the gradient information,and enhance the adaptation of the network.The convolutional layer extracts the cross features and abstract relationships of items layer by layer,then enhances the feature expression through skip connections.This thesis conducts experiments on two public datasets,and the results show that the algorithms proposed in this thesis are better than other popular recommendation algorithms,which verifies the effectiveness of the algorithms.The datasets come from real life,and the results also reflect that the algorithms have wide application prospect in sequential recommendation.