Multiple Traditional and Deep Learning-based Prediction Models for Quantity of Domestic Mail

Abstract

As the main role of domestic mail is gradually changing from the field of communication to the field of economy and culture, it is very important to provide relatively accurate volume prediction and feedback through the visualization process. This paper focuses on finding models with a high matching ability to achieve the purpose of prediction, and selects the best option by comparing multiple models. Eight different models ranging from simple to complex will be used in the following steps. More specifically, the predictive power of the models has steadily improved, from linear to curved, from constant to trend, from a single dimension to adding gradient factors to combining seasonal periods. The prediction principle is also gradually complicated. In the beginning, it focuses on the average, then weights the values in different stages to obtain the weighted average, and then makes a more refined prediction by means of single-layer convolution. To achieve the purpose of obtaining better results, it is of great necessity to introduce a self-learning prediction model, so Recurrent Neural Network (RNN) model and Long Short-term Memory (LSTM) model are chosen. In addition to obtaining the input data of this time, the RNN model will also combine the input data and output data of previous times to make the comprehensive prediction. The LSTM model further solves the situation that the RNN model cannot predict in some cases, by setting the limit to determine whether to keep the data better to fit the data with the trend. The data visualization results show that the performance of LSTM model is the most outstanding among all models, which can provide learners with more accurate prediction information.