├── data                      
│   ├── raw                   # Folder where we can find the raw files to be taken to create the graph
│   ├── processed             # Folder for the processed dataset to be used for training our model
├── models                    # Folder for stored model
├── pytorch-geometric         # Folder for source files
├── LICENSE
└── README.md

We used a transformers to highlight the deep meaning that is contain in the aggregation of name and categories columns. For that we choose a language agnostic bert based model to transform a concatenation of the name and categories of each point of interest to a vector.

We choose to have all the nearest neighbors search by country for the machine to be able to compute in an acceptable time for metrics to be calculated in a smaller set. We calculated that around only 0.002% of true links are lost and could not be found afterward while performing edges classification.

The embedding associated with each "text" column ("name" + "categories" columns concatenated) is used to compute a 'cosine' distance (1-{cosine similarity score}) which will allow using a the k-nearest neighbors algorithm to create a graph.
We also choose to get a threshold on distance above which we will not select the neighbor given by the k-nearest neighbors algorithm. For that we focused on the known duplicates POI in a random selected part of the half of the training set and took out some statistics :

The latitude and longitude value of each POI is used to compute a haversine distance which will allow using a the k-nearest neighbors algorithm to create a graph.
We also choose to get a threshold on distance above which we will not select the neighbor given by the k-nearest neighbors algorithm. For that we focused on the known duplicates POI in a random selected part of the half of the training set and took out some statistics :

Creation of a classification batched base learning model.