Research On Super-Parameter Adaptive Neural Network Model For E-Commerce Sales Forecasting

With the improvement of logistics system and e-commerce platform,more and more enterprises join the e-commerce industry,and a large number of user purchase data are generated in the background of the sales system.These sales data can be used to predict the future sales of goods in advance,so as to improve the inventory turnover rate of enterprises,reduce investment costs,and further maximize the profits of enterprises.The model based on traditional statistical methods is limited by the periodicity of data,which leads to low prediction accuracy,while deep learning has strong learning ability,which can automatically extract features from data and predict based on these features,so it is often used in time series prediction.However,some hyperparameters of deep learning model need to be set by experimenters before training,which directly affects the prediction performance of the model.Therefore,based on the above problems,the following research is carried out on the model superparameter optimization in e-commerce sales forecast:(1)A forecasting method based on ant colony algorithm and long-term and short-term memory network(ACO-LSTM)is proposed.Ant Colony Algorithm(ACO)is used to optimize the number of neurons,the number of iterations and the learning rate of the LSTM model,so as to solve the problem that the prediction performance of the model is affected by the inaccurate setting of the parameters.(2)A prediction method based on improved ant colony algorithm and bidirectional gated cyclic neural network(IACO-BiGRU)is proposed.Because of the slow convergence speed of ACO algorithm,this paper proposes an improved ant colony optimization algorithm(IACO)to improve the updating mode of ant pheromones in the algorithm.The experimental results show that the improved algorithm can efficiently adapt to the super-parameter optimization process,and thus improve the prediction performance of BiGRU model.(3)A prediction method based on Q-Learning and bidirectional gated cyclic neural network(QLBiGRU)is proposed.Although the improved ant colony algorithm effectively improves the efficiency of super-parameter optimization,the heuristic algorithm has the disadvantage of easily falling into local optimal solution,and its output is the solution of a problem.When the environment is disturbed,it needs to be optimized from scratch,while the output of reinforcement learning is a strategy,and it does not need to be retrained when the environment is disturbed.Therefore,in this thesis Q-Learning algorithm is used to optimize the superparameter of BiGRU model,which can automatically search the optimal strategy efficiently,improve the efficiency of superparameter optimization,and then reduce the prediction error of BiGRU model.