Car destination prediction model

The goal of this project is to create a predictive model to predict the position destination of a car based on a date and a starting position.

System requirements

The has been done using Python and Spark as the main technologies. To be able to run the notebooks you need to have installed the following:

• Spark 2.2
• Python 3.6. Also, these packages are required:
  • notebook
  • findspark
  • numpy
  • pandas
  • scikit-learn
  • python-geohash
  • matplotlib
  • gmaps

If you have Anaconda installed in your computer, you can easily get your Python environment ready by loading python-environment.yml, which contains all the dependencies. You can do it by simply running:

conda env create -f python-environment.yml

Although is not necessary to perform the data processing and running the model, you will need a Google Maps Javascript API Key to visualize maps with gmaps in some notebook. After you activate it in the Google Developers Console, you must add it to your environment by:

export GOOGLE_API_KEY=[Your fantastic API KEY goes here]

Project structure

The project contains the following type of files:

• Jupyter notebooks. They contain the code for the project implementation. You will better understand the project by following in this order:
  • data-cleansing.ipynb: Contains the code for read and explore the raw dataset, make some data cleanup transformations and visualization (maps). It produces as a result the file processed-dataset.csv
  • features-preparation: Normalize the data, expands dimensionality, and in general compute new features which could be useful depending on the model that choose later. It produces featured-dataset.csv
  • random-forest-model.ipynb: Implements Random Forest Prediction Model.
  • k-nearest-model.ipynb: Implements K-Nearest Neighbor Prediction Model.
• Python script.
  • predict-destination.py: This script runs the models generated in the notebooks to predict the heading of a vehicle based on its starting position and time.
• Models. The trained models are stored in the following files:
  • random_forest_model.pkl
  • k_nearest_model.pkl
• Analysis Documentation. There is a PDF file which details all the analysis, decision making, and discuss the code of the implementation: predictive-analytics-connected-car.pdf

Running the models

To ease the evaluation of the model, I've created a simple script in Python so that you can play with different values and see the prediction.

To run the script, Spark is not needed, and only numpy, scikit-learn and geohash Python packages are required. However, if you loaded the environment which I provided with the project, you'll have everything you need to go.

From a command line if you type:

./predict-destination.py -h

You will get help on how to use it:

usage: predict-destination.py [-h] {forest,knn} time latitude longitude

positional arguments:
  {forest,knn}  Predictive model to use, can be either forest or knn
  time          Start trip time, with the format "yyyy-MM-dd HH:mm:ss". It
                must be between quotation marks. For instance, you coud use:
                "2017-05-24 12:26:37"
  latitude      Latitude of the trip start position. For instance, you could
                use: 47.409291
  longitude     Longitude of the trip end position. For instance, you could
                use: 8.546942

optional arguments:
  -h, --help    show this help message and exit

For example, if you wanted to make a prediction using the K-Nearest Neighbor Model:

./predict-destination.py knn "2017-05-29 18:23:27" 32.989318 -97.263840

---

And that's all. Enjoy the code! Feedback is welcome ;-)