RENAL Goodness-of-Fit Test

Code for the REcurrent NeurAL (RENAL) Goodness-of-Fit test proposed in the paper Recurrent Neural Goodness-of-Fit Test for Time Series by Aoran Zhang, Wenbin Zhou, Liyan Xie, and Shixiang Zhu.

See below for information about the framework and implementations.

Architecture of RENAL Framework

Real-world observations are compared to model-generated sequences, with darker blue indicating better fits. We first use a recurrent neural network $\phi$ to extract conditionally independent history embeddings. We then construct their transition probability matrices using these embeddings and evaluate the fit with a chi-square discrepancy test. A tutorial on using RENAL can be found in example_usage.ipynb.

File Usage

• nn_model.py includes neural network models used in the framework, such as RNNs and LSTMs.

• tpp.py and ts_data.py contain codes for generating and visualizing temporal point processes and time series data.

• embedding_binning.py implements the embedding binning process to transform continuous state spaces into discrete state spaces and estimate transition probability matrices.

• example_usage.ipynb demonstrates how to use the framework with example data.

• ksd.py/mmd.py/linear_time.py/kernelgof.py/util.py contain baseline implementations for kernel-based goodness-of-fit tests.

Reference

Zhang A, Zhou W, Xie L, Zhu S. Recurrent Neural Goodness-of-Fit Test for Time Series.