TrafficAnalysis

This is the term project of COMP4434, which aims to use different models to classify daily inflows/outflows of crowds into groups with different dates (weekends/weekdays).

Get started

1. Install JetBrains PyCharm Professional and Anaconda3.
2. Open PyCharm, clone this project and open it.
3. [IMPORTANT] Create a folder called data and move the dataset files to this folder.
4. Enter File > Settings > Project: TrafficAnalysis > Project Interpreter.
5. Click the "gear" button on the top-right corner of the dialog, and then click Add....
6. Select Conda Environment on the left, and click OK button.
7. Click OK button again to close the Settings window.
8. Wait until PyCharm is ready.
9. If a message pops up on the bottom-right corner of the screen which asks you if you want to turn on Scientific Mode, do not hesitate to turn it on.
10. PyCharm may suggest you to install missing packages. Allow it to do so.
11. You are ready to go.

Data preprocessing

Currently, we do not use weather data to train data models.

date

Originally, date is a simple 1d array of strings like "2015110101". After the preprocessing, date becomes a 2d array with three columns.

Column 1: Since the dates in this dataset starts from 1 Nov 2015, the strings of date are converted to the number of days since 1 Nov 2015. For example, "2015111524" will be converted to 14 since 15 - 01 = 14.

Column 2: The last 2 characters of the string "2015110101" means the number of time slot in this day. "01" means the first time slot which is 00:00 - 00:30. It is converted to integers which start from zero. So, "01" will be converted to 0.

Column 3: If this day is Saturday or Sunday, this value will be 1. Otherwise, it will be 0. This is the label data for future training and testing.

data

The inflow/outflow data is normalized using preprocessing.scale() function from scikit-learn.

Data cleansing

In the given data, we found that some time slots are missing in some of the dates. So we iterate the date array to figure out those "dirty" dates and remove these data entries.

Data modelling mechanism

We plan to train and test the data model using the whole inflow/outflow data of a single day (with 48 time slots) as a single input entry. Therefore, there are 138 valid days (as input entries), and each input entry has 48 * 2 * 32 * 32 float number elements. It is stored in the data array.

There are only 2 possible values of output for each input. 1 means that this day is Saturday or Sunday. 0 means that this day is a weekday. It is stored in the label array.

Various classification models would be applied.

Data model training and evaluation

We use the following classifiers to train our model:

• SVM classfier
• SGD classfier
• k-NN classfier
• Bagging classifier

According to our experiment result, 4-Nearest-Neighbor classifier performs the best among the others with the accuracy of 96%.

Contributor

• @EririSawamura, @EvilCharles, @113741090a: data model training and evaluation.
• @CrabAss, @liao-victor: data preprocessing.
• @CrabAss: project coordinator.

Acknowledgement

The dataset used in this experiment is provided in the following paper:

Junbo Zhang, Yu Zheng, Dekang Qi. Deep Spatio-Temporal Residual Networks for Citywide Crowd Flows Prediction. In AAAI 2017.

This paper is available on arXiv. The dataset is available in lucktroy/DeepST.