Training method and system of neural network model and prediction method and system

The training method of claim 63, wherein said determining of the dimension of the each embedding layer proportionally based on a value of the information entropy of the feature input to the each embedding layer comprises:

• setting a maximum dimension and a minimum dimension;
• determining the embedding layer corresponding to a first predetermined number of features with the lowest information entropy to be assigned the minimum dimension, and determining the embedding layer corresponding to a second predetermined number of features with the highest information entropy to be assigned the maximum dimension;
• for a remaining feature except the first predetermined number of features and the second predetermined number of features, the dimension of the embedding layer respectively corresponding to the remaining feature is determined proportionally according to the values of the information entropy of the remaining feature between the minimum dimension and the maximum dimension,
• wherein an optimal solution regarding the first predetermined number and the second predetermined number is determined by enumerating the first predetermined number and the second predetermined number, wherein the optimal solution corresponds to minimizing or maximizing a variance value of the dimension assigned to the each embedding layer.

A system of training a neural network model comprising a three-level model, comprising at least one computing device and at least one storage device storing instructions, wherein the instructions, when executed by the at least one computing device, cause the at least one computing device to:

• acquire a training data record;
• generate features of a training sample based on attribute information of the training data record, and use a label of the training data record as the label of the training sample;
• and
• train the neural network model using a set of the training samples,
• wherein, in training the neural network model, learning a feature information representation of each feature itself respectively by a plurality of under-layer neural network models comprised in a first-level model of the neural network model, learning an interaction representation between corresponding input items respectively by a plurality of intermediate models comprised in a second-level model of the neural network model, learning a prediction result at least based on the interaction representations output by the second-level model by a top-layer neural network model as a third-level model of the neural network model, and adjusting the neural network model comprising the three-level model at least based on a difference between the prediction result and the label, wherein the corresponding input items comprise at least one of the feature information representations output by the first-level model.