Univerza na Primorskem Fakulteta za matematiko, naravoslovje in informacijske tehnologije
-->
SI | EN

ponedeljek, 9. september 2024 Vanja MILESKI: Time Series transformations for churn prediction with CNNs

Jutri, v ponedeljek, 9. septembra 2024, bo ob 16.00 uri izvedeno
predavanje v okviru PONEDELJKOVEGA SEMINARJA RAČUNALNIŠTVA IN INFORMATIKE
Oddelkov za Informacijske znanosti in tehnologije UP FAMNIT in UP IAM.

ČAS/PROSTOR: 9. september 2024 ob 16.00 na daljavo prek Zoom-a.

----------------------------------------
PREDAVATELJ: Vanja MILESKI
----------------------------------------

After graduating from the Faculty of Computer and Information Science at the University of Ljubljana in 2015, Vanja Mileski started working at the Jožef Stefan Institute (JSI). He was a Master's student at the International Postgraduate School Jožef Stefan and a student researcher at the JSI. After finishing his Master's studies, he applied his knowledge of data mining in the private sector as a Data Scientist in the retail, telecommunications, banking, stock market and insurance sectors. His current research interests include Time-Series classification, Deep Learning, ResNet and Inception architectures as well as LLMs.

-----------------------------------------------------------------------------------------------
NASLOV: Time Series transformations for churn prediction with CNNs
-----------------------------------------------------------------------------------------------

POVZETEK:

Churn prediction represents a pivotal task in the retail sector, focusing on identifying customers with a high risk of attrition. We analyze a multi-year dataset from a large Slovenian retailer, encompassing detailed customer demographics, purchasing behaviors, and transaction records. To assess model performance, we employ a rolling-window cross-validation approach on temporally ordered data. Convolutional Neural Networks (CNNs) have been utilized for Time Series forecasting in this context. The model's performance is evaluated using rolling-window cross-validation on the chronologically structured data. We hypothesize that applying various Time Series transformations, such as downsampling, smoothing and other manipulations, can enhance the predictive accuracy of CNN models compared to raw time series inputs.

Seminar bo potekal na daljavo, s pričetkom ob 16:00 uri prek Zoom-a na sledeči povezavi:

https://upr-si.zoom.us/j/297328207?pwd=S3Zpdk1VR3pjckNtWkQwKzlvcDR5UT09

Vabljeni.