Informatica

Most classification algorithms involve subjective inputs or hyperparameters to be determined prior to performing the classification. When taking different input or hyperparameter values, each classification algorithm will comprise a collection of classifiers. In this work, we propose a data-driven methodology for assessing similarity in consensus agreement within such a collection of classifiers, and between two classification algorithms, conditional on the dataset of interest. The core of our approach lies in considering the variability introduced by different hyperparameter values for each algorithm when performing such comparisons. We address these problems by evaluating the similarity through consensus agreement and by proposing the application of asymmetric similarity indices based on the Jaccard coefficient. We present the proposed methodology on two publicly available datasets.

One of the biggest difficulties in telecommunication industry is to retain the customers and prevent the churn. In this article, we overview the most recent researches related to churn detection for telecommunication companies. The selected machine learning methods are applied to the publicly available datasets, partially reproducing the results of other authors and then it is applied to the private Moremins company dataset. Next, we extend the analysis to cover the exiting research gaps: the differences of churn definitions are analysed, it is shown that the accuracy in other researches is better due to some false assumptions, i.e. labelling rules derived from definition lead to very good classification accuracy, however, it does not imply the usefulness for such churn detection in the context of further customer retention. The main outcome of the research is the detailed analysis of the impact of the differences in churn definitions to a final result, it was shown that the impact of labelling rules derived from definitions can be large. The data in this study consist of call detail records (CDRs) and other user aggregated daily data, 11000 user entries over 275 days of data was analysed. 6 different classification methods were applied, all of them giving similar results, one of the best results was achieved using Gradient Boosting Classifier with accuracy rate 0.832, F-measure 0.646, recall 0.769.