Synthetically Generated CDR

Python was used to create the dataset. Starting from a real dataset of 300 rows, we were able to generate synthetic datathat largely preserves the synthetic properties of the real data. How we created the dataset will be presented in the nextparagraphs. The most important aspect of the CDR is the way the cell id of the telecom tower is chosen. Using the list of all Cell Towers in Romania, after the userchoosed the area of interest, only the cell ids of the towers in this area are stored in the dataset (See the paper listed for the entire procedure). In this way cell_id, longitude and latitude fields are created. Longitude and latitude represent the geographical coordinates of the tower.For example the Romanian tower with cell_id 26.133.003 has longitude 27,60346 and latitude 47,143381.A unique_id is used to differentiate between the different lines in the CDRs used. It is created using uuid4, from theuuid library in python. uuid4 allows the creation of unique ids using random numbers and digits. Example: 49d791e7-a38e-4d33-b947-2e49d3c734b0. The identifying call elements are generated using 15 random numbers from 0 to 9. Forthis, the python random library was used, with the choice method. The caller id is also generated randomly, using the random library from python. Some of the features were generated randomly, and some have been generated synthetically. For generation we usedSynthetic Data Vault (SDV) , which is a library in python for generating synthetic tabular data. The initial CDR waspre-processed so that it could be inserted into the synthetic data generation library, the timestamp column was splitinto several columns in order to generate each detail that could be used in a Machine Learning model.For example, the timestamp of the call for one of the callers is 06.12.2023/13:20:01. Out of this we made multiplefeatures, using the pandas library. After preprocessing, the data is entered into the SDV library. The first step is to generate a metadata for the dataframe used. This step is necessary for calling the method that generates the synthetic data. This is also done via SDV. For the generation we used FAST ML from SDV which is a synthesizer optimized for modeling speed. The generated dataset preserves (to a large extent) the properties of the real dataset, considering that 2000 data entries were generated from a dataset with only 300 entries.The hour of the call, minute of the call, second of the call columns keep the normal ranges for this type of data, 0-24,0-59, respectively 0-59. To see an analysis between real and generated data, please check the paper listed in the description.