Research On User Behavior Recognition Problem Based On CNN And LSTM

Time series classification has always been one of the hot issues in the field of data min-ing.However,with the development of information technology,time series data increasingly tends to be high-dimensional and large-scale,traditional methods based on distance and feature have been difficult to get the desired results.Deep learning models are beginning to be applied to time series classification problems due to their powerful feature extraction capabilities and end-to-end characteristics.Among them,the one-dimensional fully convolutional neural net-work has been proved to have incomparable efficiency and precision in dealing with sequence problems,and it is recognized as a relatively strong end-to-end time series classification algo-rithm.However,subject to the one-dimensional convolution kernel,the one-dimensional fully convolutional neural network cannot extract the cross-combination information between differ-ent variables,which limits the accuracy of the model in dealing with multivariate time series problems.Aiming at the problems of the one-dimensional convolution kernel,this paper combines the long and short-term memory unit with the two-dimensional convolutional neural network module and proposes the long and short term convolutional neural network model(LS-CNN)model.This model improves the generalization ability of the model by combining the time se-ries features extracted from the sample data by the long-term short-term memory unit and the joint features between different variables extracted by the convolutional neural network mod-ule.Relevant experiments prove that the model proposed in this paper has achieved a certain effect on the data set of user behavior recognition.The specific work of this article is summarized as follows:(1)Aiming at the shortcomings of the one-dimensional convolution kernel,this paper uses a two-dimensional convolution to extract the combination information between different features,and merges it with the feature information extracted by the long short term memory module,and proposes an end-to-end deep learning model LS-CNN,which significantly improves the accuracy of the model.(2)Aiming at the common over-fitting problems of deep learning models,this article has made improvements from three aspects:model structure,model parameters and data.Including the improvement of the loss function,replacing the softmax loss function with the L-softmax,A-softmax and AM-softmax loss functions;for the improvement of the activation function,use mish,swish and leakyrelu activation functions instead of relu activation functions;as well as kaming parameter initialization and label smoothing for model parameters and data labels.(3)In order to reduce the size of the model and improve the inference efficiency of the model,this paper uses different compression methods to compress the model.In the end,when the model accuracy is lost by 5%,the model size is half of the original.