Welcome to LGM-Geocoding’s documentation!¶
LGM-Geocoding¶
A python library for accurate classification of best geocoding sources per coordinate pair.
LGM-Geocoding is a python library which implements a full Machine Learning workflow for training classification algorithms on annotated datasets that contain mappings between coordinate pairs and the ideal geocoding source for them and producing models that provide accurate predictions about ideal geocoding sources for previously unseen geocoding pairs. Geocoding implements a series of training features, regarding the different coordinate pairs that are available for each geocoder, the interaction between them and neighboring geospacial data (namely road data). Further, it encapsulates grid-search and cross-validation functionality, based on the scikit-learn toolkit, assessing as series of classification models and parameterizations, in order to find the most fitting model for the data at hand. Indicatively, we succeed a 66.04% accuracy with the Random Forest classifier compared to the baseline approach that achieves accuracy of 55.05% (see References).
The source code was tested using Python 3.6 and Scikit-Learn 0.22 on a Linux server.
Description¶
The module consists of the following steps:
- Features extraction
This step takes into account the features configuration given in config.py and creates X_train and X_test feature pairs grouped by folds and ready to be utilized by machine learning algorithms in the next steps.
- Algorithm selection
A list of given classifiers given in config.py are evaluated in a nested cross-validation scheme in order to find which performs the best on the features sets created in the previous step.
- Model selection
Given a selected classifier, this step tries to find its best configuration.
- Model training
Utilizing the knowledge from the previous step, a model is trained on the whole available data using the optimal configuration. This model is then saved to disk for later usage.
- Model deployment
This step loads the optimal model from disk and uses it in order to classify a set of unseen, unlabeled, test data. Classification results come in a form of (predicted-label, score) pairs, suggesting the model’s confidence about each prediction.
Setup procedure¶
Download the latest version from the GitHub repository, change to the main directory and run:
pip install -r pip_requirements.txt
It should install all the required libraries automatically (scikit-learn, numpy, pandas etc.).
Usage¶
The execution of the project starts with the Features extraction step initializing the pipeline’s root folder which the following steps will refer to in order to output their results. Each step can be executed as follows:
- Features extraction
python -m geocoding.features_extraction -fpath <fpath>
where <fpath> is the path to the csv file containing train data.
- Algorithm selection
python -m geocoding.algorithm_selection -experiment_path <exp_path>
where <exp_path> is the path to the folder created from the first step.
- Model selection
python -m geocoding.model_selection -classifier <clf_name> -experiment_path <exp_path>
where <clf_name> is the classifier’s name to be optimized in order to build the model and <exp_path> same as before.
- Model training
python -m geocoding.model_training -experiment_path <exp_path>
where <exp_path> same as before.
- Model deployment
python -m geocoding.model_deployment -experiment_path <exp_path> -fpath <fpath>
where <exp_path> same as before and <fpath> is the path to the csv file containing the test data.
Documentation¶
Source code documentation is available from linkgeoml.github.io.
References¶
Alexis et al. Improving geocoding quality via learning to integrate multiple geocoders. SSDBM ’20.
