Modification on a multivariate time series anomaly detection model

This is a simple project where you are required to modify a Graph neural network/Graph deviation network model that is based on (https://github.com/d-ailin/GDN)
The github model is an anomaly detection model. The steps include:
1. Sensor Embedding: uses embedding vectors to capture the unique characteristics of each variable
2. Graph Structure Learning: learns a graph structure representing dependence relationships between variable
3. Graph Attention-Based Forecasting: forecasts future values of each variable based on a graph attention function over its neighbors
4. Graph Deviation Scoring: identifies deviations between actual and predicted values using threshold calculation

Modification on the model:
Apply one of the below to set thresholds and select the best threshold to identify extreme values.
1. nonparametric dynamic error thresholding (NDT) from (https://github.com/khundman/telemanom)
OR
2. POT (peaks over threshold) model of EVT (extreme value theory) from (https://github.com/NetManAIOps/OmniAnomaly)

Implement the model on these datasets:
1. SMD dataset
https://github.com/NetManAIOps/OmniAnomaly/tree/master/ServerMachineDataset

2. MIT-BIH Supraventricular Arrhythmia Database (MBA)
https://physionet.org/content/svdb/1.0.0/

3. Soil Moisture Active Passive (SMAP) dataset

4. Mars Science Laboratory (MSL) dataset
https://www.kaggle.com/datasets/imkrkannan/mars-weather-data

This project should incude:
-Exploratory Data Analysis
-Data preprocessing
-Model building and training
-Provide visualization and list of the labelled anomalies in time series

Evaluation metrics: Precision, Recall, AUC, F1