Aristotle University Of Thessaloniki (AUTH)
Local interpretation of machine learning models
Abstract
dc:descriptionWhile Artificial Intelligence paves the way for new advancements and opens up new possibilities, its use in certain domains may be prohibited. Indeed, when human lives are at stake, AI-driven systems' decisions must be reliable and trustworthy. Explainability is one way to meet these criteria. However, in order to achieve user-level explainability, interpretability concerns must be addressed first. This is especially noticeable in AI systems comprised of machine learning pipelines. Although some models are intrinsically interpretable, the top performing ones are uninterpretable. Random forest models, for example, are complex models that excel at tasks like fraud detection and disease probability prediction. Even though their subcomponents, the decision trees, can be interpreted, random forests cannot inherently provide reasons for their decisions. Furthermore, neural networks that achieve superhuman abilities in analysing both structured and unstructured data are almost always uninterpretable by design because of their large number of parameters. These models perform amazingly well at many tasks, including hate speech recognition and predictive maintenance. This thesis investigates both models, seeking to address their interpretability issues. Starting with neural networks, to address their interpretability, LioNets, a technique that actively considers the neural network's architecture and latent representations, was developed. LioNets, which have been tested on textual and time-series data, produce neighbours in the latent space to ensure adjacency between the generated neighbours and an instance whose prediction attempts to interpret. LionForests was designed to explain random forests models, providing rules as explanations. To accomplish so, it makes use of the model's knowledge, extracting every decision path from all the trees regarding a decision and effectively combining them to build a rule of low complexity. Originally developed for binary classification tasks, this technique has been extended to cope with multi-class and multi-label classification, as well as regression tasks. By introducing such techniques, to enable complex models' interpretability across several supervised tasks, specific properties were taken into account in the process, while new ones like conclusiveness have been introduced. The latter, which is strictly related to rule-based explanations, ensures that the explanation does not contain misleading or incorrect elements. To evaluate the performance of these new techniques, several metrics from the literature have been employed, as well as a few novel ones proposed. The truthfulness metric, in particular, was introduced as a variation of faithfulness, with the goal of simulating curious users who will evaluate on their own whether an explanation is correct or not. In the context of this thesis, a meta-explanation technique was also investigated with the goal of producing more accurate explanations by aggregating various interpretations for a specific instance using truthfulness. A common characteristic of all the contributions mentioned previously is the scope of the produced interpretations. While there are interpretability techniques that provide explanations that apply to the entire structure of the model, called global interpretations, this thesis focuses on techniques that provide interpretations that are specific to individual instances. This concept is also known as local interpretability.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Mollas, Ioannis
- Μολλάς, Ιωάννης
Subjects
dc:subject × 14- Μηχανική μάθηση
- Επεξηγήσιμη τεχνητή νοημοσύνη
- Ερμηνεύσιμη μηχανική μάθηση
- Τοπική ερμηνευσιμότητα
- Machine learning
- Explainable artificial intelligence
- Interpretable machine learning
- Local interpretability
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54400