Code to fully reproduce benchmark results (and to extend for your own purposes) from the paper: LIMESegment: Meaningful, Realistic Time Series Explanations

The goal of this package is to provide a modular way of adapting LIME to Time Series data. It provides methods for:

• Segmenting a time series
• Perturbing a time series
• Measuring similarity between time series
• Putting these all together to find explanations in the form 'This segment was most import for the overall classification'

• Clone repository
• Install dependencies 'pip3 install -r requirements.txt'

Content

• Utils:
  • data.py : Contains code generating all synthetic datasets used for experiments & loading UCR datasets used for expleriments
  • models.py : Contains code generating and training all classification models used to test explanations:
    • KNN: K Nearest Neighbour Algorithm implemented with Scikit learn
    • CNN:1D Convolutional Neural Network implemented with keras
    • LSTMFCN: State of the Art Hybrid LSTM and and LSTMFCN implemented by Karim et al. : https://github.com/titu1994/LSTM-FCN
  • explanations.py: Contains code generating explanations:
    • LIMEsegment: Our proposed adaptation of LIME to time series data
    • Neves: The proposed adapatation of LIME to time series data of Neves et al.: https://boa.unimib.it/retrieve/handle/10281/324847/492202/Manuscript.pdf
    • Leftist: The proposed adaptation of LIME to time series data of Guilleme et al.: https://ieeexplore.ieee.org/abstract/document/8995349
  • Contains code for NNSegement, our proposed segmentation algorithm and is a building block of LIMESegment
  • perturbations.py: Contains code for the perturbation strategies evaluated:
    • RBP: Realistic Background Peturbation, a frequency based perturbation strategy proposed in the paper and is a building block of LIMESegment
    • Zero, random noise and Gaussian blur perturbations
  • metrics.py: Contains code for the Robustness and Faithfulness measures used to evaluate each explanation module
  • constants.py: Contains details for loading and processing the UCR Time Series datasets used in experiments
• Experiments: Jupyter notebooks containing reproducable implementations of our paper experiments:
  • Segmentation: Evaluate NNSegment against state-of-the-art segmentation approaches
  • Background Perturb: Evaluate RBP against other peturbation strategies
  • Locality: Evaluate the use of DTW in LIMESegment against the Euclidean distance measure
• RobustnessFaithfulness.py: Evalue LIMESegment against Neves and Leftist on overall explanations
• Data: Contains UCR datasets used in experiments. Note that the apnea dataset used in the Segmentation Tests is not included in this repo but can be downloaded from: https://www.physionet.org/content/apnea-ecg/1.0.0/)

Running LIMESegment

Run:

  from Utils.explanations import LIMESegment
  explanations = LIMESegment(ts, model, model_type, distance, window_size, cp, f)

Returns segment importance vector as returned by LIMESegment LIMESegment takes as input:

• ts: TS array of shape T x 1 where T is length of time series
• model: Trained model on dataset array of shape n x T x 1
• model_type: String indicating if classificaton model produces binary output "class" or probability output "proba", default "class"
• distance: Distance metric to be used by LIMESegment default is 'dtw'
• window_size: Window size to be used by NNSegment default is T/5
• cp: Number of change points to be determinded by NNSegment default is 3
• f: Frequency parameter for RBP default is T/10