danthe96 / CIoTS

Causal Inference on Time Series

CIoTS

This repository implements an approach for Causal Inference on Time Series. It is heavily based on a paper by Chen, "A time series causal model".
It also contains the base implementation of Chen's algorithm and a simple time series data generator.

Installation

pip install -U -r requirements.txt

How to use

To start, first generate a causally dependent time series.
Params:

• dimensions (required) - the number of time series to generate
• max_p (required) - the max time lag of any causal effects, i.e. for a point in time t there is at most a causal relationship X<sub>t-p</sub> --> Y<sub>t</sub>
• data_length - the number of data points to generate
• incoming_edges - the number of causal relationships per time series

from CIoTS import CausalTSGenerator

generator = CausalTSGenerator(dimensions=3, max_p=4, data_length=10000, incoming_edges=2)
data = generator.generate()

To visualize based on what model the dataset has been generated, we can visualize the partial time graph.

plt.title('Original graph')
generator.draw_graph()

To estimate the graph, we run Chen's algorithm and render the result as graph.

from CIoTS import pc_chen, partial_corr_test, draw_graph

predicted_graph = pc_chen(partial_corr_test, data, p=4, alpha=0.05)

plt.title('Estimated graph')
draw_graph(predicted_graph, dimensions=3, max_p=4)

We can also print metrics of the algorithm's accuracy.

from CIoTS import evaluate_edges

pd.DataFrame(evaluate_edges(generator.graph, predicted_graph), index=[0])

Our research

The above tutorial runs Chen's base algorithm. Our research will focus on estimating max_p automatically and estimating the graph iteratively. This is still work in progress.