OWAM Framework

Outlier Weighted Autoencoder Modeling Framework : This paper emphasizes the rising importance of traffic modeling for transportation systems optimization. Despite extensive research, previous studies have overlooked outlier impact and real-time analysis. The focus here is on improving vehicle traffic flow prediction, especially in outlier scenarios, both offline and real-time. A new approach, Outlier Weighted Autoencoder Modeling (OWAM), is introduced. Autoencoders identify anomalies for effective abnormal traffic pattern handling. Weighted correlation analysis evaluates the influence of nearby sensors on the target sensor’s flow, enhancing LSTM-based traffic models. OWAM also serves as an outlier-based real-time update strategy, adapting to changing traffic patterns. Testing against various datasets highlights OWAM's accuracy-efficiency balance. Although it doesn't outperform Graph techniques in RMSE, OWAM excels in dynamic outlier-based updating, enriching adaptive traffic prediction models for transportation management.

OWAM:

OWAM with Incremental Updates :

In order to run the code, please follow the below steps:

1. Clone the repository

2. Install the requirements (This code contains two seperate enviorments due to compatibility issues with pytorch)

   A. Install requirements_1.txt for installing the requirements for the OWAM framework (python - 3.10.10)

   pip install -r requirements_1.txt
   

   B. Install requirements_2.txt for installing the requirements for the outlier models -DAADS (Python - 3.9.16)

   pip install -r requirements_2.txt
   

3. Run the Notebooks in the notebooks folder for reproducing the results presented in the thesis report.

   Note: In this project, due to multiple experiments(some variable changes are required to perform eperiments in code), various dependencies and the use of the earlier projects, Notebooks are prefered over standard python files which can be difficult to reproduce results. Run the notebooks in the following order for reproducing the results.

   1. data_preprocessing.ipynb for preprocessing the data and basic data analytics (requirements_1.txt)
   2. autoencoders_daads.ipynb for training the autoencoders and other outlier models using the preprocessed data (requirements_2.txt)
   3. correlation_analysis.ipynb for generating the outlier correlation scores and correlation analysis (requirements_1.txt)
   4. LSTM_OWAM.ipynb for training the LSTM model using the OWAM framework (requirements_1.txt)
   5. LSTM_OWAM_real_time.ipynb for training the LSTM model using the OWAM framework with incremental updates (requirements_1.txt)
   6. results.ipynb for generating the results and plots (requirements_1.txt)

NOTE -

1. All the experiments are already performed and the results are stored in the results folder except results generated from LSTM_OWAM_real_time.ipynb due to the large size of the results. Run this file and then the results can be reproduced by directly running the results.ipynb.

2. The hague dataset can not be shared as it is not public. please mail on xxx to get the information regarding the dataset access. For the other datasets please use the following link and paste them in a new data folder in the repo.

DGCRN (GNN) Model

The data can be downloaded from the below link, paste the data in the data folder in the repo. Hague data can not be downloaded as it is not public. Please mail on xxx to get the information regarding the dataset.

use the following command to run the code -

1. for the Hague dataset:

python train_benchmark.py --model 'DGCRN' --data 'Hague' >> log.txt

2. for the METR-LA dataset:

python train_benchmark.py --model 'DGCRN' --data 'METR-LA' >> log.txt

3. for the PEMS-BAY dataset:

python train_benchmark.py --model 'DGCRN' --data 'PEMS-BAY' >> log.txt

NOTE - Running the model may require high computational power. Please use the --gpu flag to run the model on GPU.