Semi-Supervised Heterogeneous Transfer Learning for Tabular Data — A Study on Solutions and Applications in Mobile Networks
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Science |
Master's thesis
Authors
Date
Department
Major/Subject
Mcode
SCI3115
Degree programme
Language
en
Pages
62
Series
Abstract
This thesis addresses the under-explored area of applying semi-supervised heterogeneous domain adaptation (HetDA) methods to tabular data, specifically focusing on scenarios encountered in real-world applications such as mobile networks. The motivation arises from the lack of comprehensive solutions tailored for tabular data, despite its prevalence in various domains. The work centers around the challenge of effectively applying domain adaptation (DA) methods to tabular data, as the existing literature mainly focuses on image classification and sentiment analysis tasks. Furthermore, the study aims to bridge the gap between theoretical advancements and practical applications in mobile networks, particularly in solving the Secondary Carrier Prediction (SCP) problem. The methodological approach involves a detailed exploration of existing DA techniques, with a specific emphasis on HetDA methods. Through a systematic review, key insights are drawn from recent literature, guiding the selection of methods for evaluation and implementation. Results from the study demonstrate the effectiveness of selected methods, including Conditional Weighting Adversarial Network (CWAN), Joint Mean Embedding Alignment (JMEA), and Kernel Heterogeneous Domain Alignment (KHDA), when applied to tabular data in multiclass classification tasks with few labeled data available. These methods showcase promising performance in terms of data efficiency. Evaluation demonstrates that HetDA techniques can achieve comparable or even superior performance to traditional methods using just 10% of the labeled data, making them a valuable solution to reduce data collection and retention. Key contributions include not only providing practical solutions for DA in tabular data but also shedding light on the importance of adapting DA techniques to domain-specific challenges, such as those encountered in mobile networks. Additionally, the study identifies potential areas for improvement and future research directions, aiming to foster advancements in DA methodologies tailored for tabular data.Description
Supervisor
Hämäläinen, WilhelmiinaThesis advisor
Boström, HenrikAlkhatib, Amr