The author of more than three years of business experience which has accumulated a lot of experience with time-series data modeling predicted, because of the nature of work, exposed to more types of data are time-series data, will be greater use when dealing with this type of data to the regression model, RNN LSTM or model, so this is mainly based on past experience to share some of the timing of households inside the field of modeling common practice.
Now comes LSTM, will brief look RNN (Recurrent Neural Network, RNN) a recurrent neural network, LSTM neural network can be seen as a model of RNN, RNN is a type of nerve designed to handle time-series data samples network, it is not only the output of each layer to the next layer, and also outputs a hidden state, the processing of the current layer in a sample when used.
RNN can be inferred based on the information of the current information before the appearance, especially in the language processing, RNN can be used to predict the next word according to a above will appear. But it can only process information at regular intervals, and if the interval is too far above, there may arise where it is difficult Lenovo. This time LSTM came into being. LSTM deployed configuration with different RNN in the presence of mainly a control configuration of the storage state, the following diagram is a schematic expanded RNN classical models and structural models LSTM:
Want in-depth understanding of the mechanism of LSTM mode, clear three kinds of door LSTM is very important, LSTM main models include: Forgot door, gate input layer and output layer door. Each door briefly described in the following table:
