University of Thessaly (UTH)
Η εφαρμογή της τεχνητής νοημοσύνης στoν τραπεζικό κλάδο: πολυσταδιακή βαθιά μάθηση για τον εντοπισμό απάτης και την ερμηνευτικότητα μοντέλων
Abstract
dc:descriptionThe banking sector stands at the threshold of a profound transformation, and the driving force behind this revolution is Artificial Intelligence (AI). Innovative AI applications have already been put forth to address challenges across various domains, including credit scoring, risk assessment, enhancing client experiences, and portfolio management. Among the most pressing challenges in the banking sector is the detection of fraudulent activities within streams of transactions. Recently, deep learning models have emerged to tackle this specific issue, focusing on the identification and prediction of potential fraudulent events. The primary objective is twofold: first, to estimate the unknown distribution of normal and fraudulent transactions, and subsequently, to pinpoint anomalies that may signal potential fraud. Within the pages of this dissertation, we delve into a groundbreaking multistage deep learning model designed to effectively handle incoming transaction streams and detect fraudulent activities. Our approach advocates the utilization of autoencoders for feature selection and the acquisition of latent data space representations based on nonlinear optimization models. On the carefully curated set of significant features, we then apply a deep Convolutional Neural Network (CNN) to detect instances of fraud. This amalgamation of techniques involves two distinct processing blocks and is tailored to identify fraud based on the exposed latent data representation rather than the initial data. Another facet of this dissertation aims to introduce a novel enhancement to the LIME algorithm, a tool used for interpreting machine learning models. Beyond model prediction, the growing emphasis in contemporary research lies in the ability to decipher what a model has learned. The interconnectedness of the myriad interpretation methods proposed in the field and the identification of common evaluation concepts remain unclear. Our proposed technique employs an autoencoder to generate new instances around the data point under scrutiny, in contrast to the conventional method of simple data creation using a Gaussian distribution centered around the point of interest.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zioviris, Georgios
- Ζιοβίρης, Γεώργιος
Subjects
dc:subject × 22- Τεχνητή νοημοσύνη
- Ερμηνεύσιμη τεχνητή νοημοσύνη
- Αποτελεσματικότητα, Τραπεζικό σύστημα
- Ερμηνεύσιμη μηχανική μάθηση
- Αυτοκωδικοποιητές
- Συνελικτικό νευρωνικό δίκτυο
- Ανίχνευση απάτης
- Μείωση διαστάσεων
- Μοντέλο μακράς βραχυχρόνιας μνήμης
- Artificial intelligence
- Interpretable artificial intelligence
- Interpretable machine learning
- Banking efficiency
- Convolutional neural networks
- Autoencoders
- Fraud detection
- Dimensionality reduction
- LSTM
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Computer and Information Sciences
- Natural Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/56313