Research On Sequence Recommendation Based On Generative Adversarial Network

Nowadays,the application range of recommendation systems is becoming more and more extensive which has become the infrastructure of e-commerce platforms,video playback platforms,entertainment and social platforms.Sequence recommendation is a branch of recommendation system.It mine hidden information from user-item interaction records and makes predictions based on the information.Implied information includes connections between items,user preferences,commodities popularity,etc.Dependency refers to the connection between products that users interact with at different times.For example,there is a sequential dependency between commodities in a transaction sequence.There are many problems with sequence recommendation,such as the sparseness of sequence data,the problem of long-term dependence mining.To solve these problems,this paper studies sequence recommendation,proposes two algorithms and conducts experiments on actual data.The main research work of this article is as follows:(1)For the sparseness of sequence data,this paper first proposes a collaborative filtering algorithm based on generative adversarial networks.The algorithm expresses the user's sequence data with vectors where learns implicit feedback information.It learns the characteristics of users and items,uses the generator to generate fake user interaction data,and makes discriminator to discriminate between the real data and the generated data.To solve the sparseness of the sequence data,the algorithm learns the sequence information and uses the learned features to generate data.Generating adversarial network model is prone to mode collapse that the generator deceives the discriminator by learning a certain distribution feature of real data which results in the problem that the loss value oscillation is too small to improve the recommended precision.This paper uses the method of weight clipping and adds gradient penalty on discriminator to improve the original algorithm.The method of weight clipping improves the gradient disappearance in training,but it will cause weight-polarization,training instability and slow the model convergence.Therefore,this paper adds a gradient penalty term to the loss function,which meets the Lipschitz limit of the discriminator gradient.Experimental results show that the algorithm can solve the data sparsity problem of sequence recommendation.The precision of the model improves 61.1%for the neural collaborative filtering method.(2)Aiming at the problem of long-term dependence mining in sequence recommendation,this paper proposes a generative adversarial network algorithm embedded with bidirectional encoder representation on Transformer.The algorithm uses bidirectional encoder representation model to interpret the user-item binary relationship sequence,and uses adversarial training of generate adversarial network model to generate more sample data fitted the real data distribution to guide the model and obtain better prediction results.The algorithm includes the following steps.Firstly,the bidirectional encoder representation model is used to interpret user sequence data.It considers the context information of any item in the sequence.The multi-head self-attention mechanism is used to capture the dependence and information transfer between the item and its context.Then,we train the generative adversarial network model by the characteristics of users and items,use the generator generate sample data that approximates the distribution of real data,and use the discriminator to distinguish the real data from the generated data.Finally,this paper uses the training results of the collaborative filtering algorithm based on generative adversarial networks and conducts experiments with the generative adversarial network algorithm embedded with bidirectional encoder representation.The experimental results show that the algorithm also solves the problem of sequence data sparsity.The NDCG@10 obtained by the algorithm in real datasets up to 93.0%.