Anna LeontjevaKunsing, Sven2019-10-152019-10-152018http://hdl.handle.net/10062/66197Masinõppe meetodeid on tihti kasutatud klientide lahkumise ennustamiseks teenindussektoris. Käesolev töö annab ülevaate metoodikatest ja keskendub lähemalt juhumetsade ja närvivõrgu mudelile ja nende rakendamisele reaalsetele kliendiandmetele. Lisaks tavapärastele profiiliandmetele püütakse klientidevahelise tehinguinfo põhjal leida, kas selliste andmete lisamine parandab esialgse mudeli karakteristikuid. Parimaks mudeliks osutub profiiliandmetele tuginev juhumetsade mudel saagisega 59%, täpsusega 62% ja esitustäpsusega 4% mida võib pidada ebapiisavaks, et seda reaalses äritegevuses rakendada. Võrgustikuandmed, mille kaasamiseks mudelitesse kasutatakse node2vec algoritmi ja kliendi egograafis aset leidnud varasemaid klientide lahkumisi ei aidanud kaasa esialgse mudeli parandamisele.Machine learning methods have often been used to predict client churn in the service sec-tor. This paper gives an overview of some methods used and focuses on the random forest and neural network models and their application on real-life client data. In addition to us-ing the ordinary profile data, it seeks to find out whether the amendment of the data set with inter-client transaction data helps to improve the characteristics of the original model. The best model proposed is one based on random forests with recall of 59%, accuracy of 62% and precision of 4%, which can be regarded as insufficient to deserve a real-life business implementation. Node2vec algorithm and previous churn cases in each client’s ego graph were used to process network data to add it to the usual profile data for model training. Unfortunately, the addition of such data did not help to improve the original model.etThesis