A solution to Grab's AI for SEA competition
- Linux (Prediction in
10 - Final Solution - Booking Demand Prediction.ipynbutilizes parallelism which doesn't work on Windows)
- Python 3.6+
- jupyter
- pygeohash
- simdkalman
- sklearn
- scipy==1.2
- statsmodels
- pystan
- fbprophet
- tensorflow==2.0.0-alpha0
NOTE: You only need to run 10 - Final Solution - Booking Demand Prediction.ipynb. The other notebooks just explain the methodology behind the solution.
pip3 install -r requirements.txt
or:
pip3 install pygeohash simdkalman sklearn scipy==1.2 statsmodels pystan fbprophet tensorflow==2.0.0-alpha0 --upgrade
Note that you can also do it inside the jupyter notebooks by adding ! right before the commands:
!pip3 install pygeohash simdkalman sklearn scipy==1.2 statsmodels pystan fbprophet tensorflow==2.0.0-alpha0 --upgrade
First, open jupyter using:
jupyter notebook
This will open a window in your browser. From there, open 10 - Final Solution - Booking Demand Prediction.
If you haven't installed jupyter yet, just enter the following to your commandline:
pip3 install jupyter
Focus on the desired cell to run and hold shift then enter. It is recommended to run the cells in order as presented.
And viola, done!
In hyperlocal forecasting, we need to fit a model for each timeseries (by location). On the other hand, we only need a single spatiotemporal model for all of them. So why use hyperlocal models? For better accuracy.
The following are the errors for the spatiotemporal models:
| Models | RMSE |
|---|---|
| 1st degree polynomial regressor | 0.048594363494371254 |
| 2nd degree polynomial regressor | 0.04912314854368942 |
| 3rd degree polynomial regressor | 0.0495815584173599 |
| Random Forest | 0.07614693023190335 |
| XGBoost | 0.04855022108270493 |
| Neural Network | 0.049747 |
While the following are the errors for the hyperlocal models:
| Models | RMSE |
|---|---|
| FBProphet | 0.03973981575859182 |
| Theta Method | 0.035792962736291206 |
| Kalman Filter | 0.03900183400744959 |
For more information, please check out notebooks 4 to 8
https://facebook.github.io/prophet/docs/quick_start.html
https://simdkalman.readthedocs.io/en/latest/
Please check out the journal articles in the journal folder. The researcher found the paper The Optimized Theta Model very enlightening.
The period of the cycles are of length 96. This is way too large for SARIMA, making it slow.
The researcher first suspected global temporal anomalies after observing the mean demand over time
An autoencoder was then trained on the density maps for each timestep. Fair enough, anomalies were found:
These anomalies happened on day 17 from 9:30am to 12:45pm. Perhaps there was an outage? A server malfunction? The researcher doesn't know.
To avoid these anomalies from affecting the results, days 1-19 were ignored while doing crossvalidation. Additionally, an additional regressor was added to the FBProphet model to handle these anomalies.
Apparently, the mahalobis distance of the frequencies, demand mean, and time of the day of the timeseries doesn't affect the RMSEs that much. See the following plots (sorry for the misleading labels!):
RMSE vs. Mahalanobis Distance of the Frequencies
RMSE vs. Demand Mean
RMSE vs. Time of the Day
Thus, no external variables were used. For robustness, HuberRegression was used to determine the weights of the predictions.
Simple averaging also works well. In fact, it works 0.36% better than using the weights found using HuberRegression. The weights were used anyway since they are proportional to the models' performance.
| Models | RMSE | Weight |
|---|---|---|
| FBProphet | 0.03973981575859182 | 0.29395973 |
| Theta Method | 0.035792962736291206 | 0.38132017 |
| Kalman Filter | 0.03900183400744959 | 0.309658 |
For more information, please check out 9 - Ensembling.ipynb