Causal relationship mining method based on deep learning

Abstract

The invention discloses a causal relationship mining method based on deep learning. The method comprises the following steps: firstly, preprocessing data by using technologies such as missing data supplement, data normalization and independent hot coding; then, based on a Keras deep learning framework, carrying out univariate time sequence prediction on the target characteristics by using an LSTM;adjusting the model structure and a series of hyper-parameters to obtain an optimal model; recording the R2 score of the model on the test set; and then, predicting all the alternative features by using the model to obtain R2 scores of the alternative features on the test set, subtracting the two scores to obtain a Granger causal relationship score of the alternative features and the target feature, and obtaining a quantitative number for describing the Granger causal relationship between the alternative features and the target feature. The method is suitable for the problem of influence factor analysis of other time sequences. In conclusion, the Granger causal relationship mining method based on deep learning has the advantages of mining much data in various fields.

Description

technical field [0001] The invention belongs to the technical field of data mining, and in particular relates to mining Granger causality from multivariate time series data based on deep learning technology. Background technique [0002] Time series data is a set of data points observed at constant time intervals. It has two main characteristics. One is that it depends on time, and there is correlation between the observed values. Most time series also have a seasonal trend, that is, a specific change in a specific time window. As the cost of data storage has been greatly reduced, a large amount of time series data generated in real life has also been completely recorded. This kind of data is commonly found in fields such as finance, commodity prices, and transportation. For example, in the field of air quality, in recent years, atmospheric visibility has declined, air quality has deteriorated, and many cities are often accompanied by smog weather, which has a relatively la...

AI Technical Summary

Problems solved by technology

[0004] At present, the existing multivariate time series causality analysis mainly focuses on the qualitative Granger causality test between variables, and some studies based on Granger causality only use linear regression to fit data and give results. This method cannot achieve high-accuracy predictions for complex nonlinear data in real life, and has great limitations.

Images