Interpretabilidade em Modelos de Sistemas de Recomendação
In the past few years, we have seen increased adoption of Deep Learning models in several Machine Learning tasks. This is mainly due to their ability to achieve unprecedented predictive performance in several applications, with special relevance in computer vision, natural language processing, and,...
Saved in:
| Main Author: | |
|---|---|
| Format: | Dissertation |
| Language: | Portuguese |
| Published: |
ProQuest Dissertations & Theses
01-01-2021
|
| Subjects: | |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the past few years, we have seen increased adoption of Deep Learning models in several Machine Learning tasks. This is mainly due to their ability to achieve unprecedented predictive performance in several applications, with special relevance in computer vision, natural language processing, and, more recently, recommender systems. However, in certain domains, model interpretability is a requirement. In such cases, Deep Learning models – typically more complex and therefore known as “black-box” models – are generally replaced by simpler models, easily interpretable by users, yet with lower predictive ability. There is a major difficulty in optimizing, in parallel, predictive performance and model interpretability. The adoption of black-box models triggers a crucial need for plausibility and verifiability of decision making.Led by the awareness of this problem in the adoption of complex models, the Artificial Intelligence scientific community has dedicated an increasing effort on eXplainable Artificial Intelligence (XAI). This increased research on XAI has resulted in the proliferation of novel methods, however, there is still not enough convergence on concept definitions and there is a lack of consensus on how to evaluate the quality and veracity of explanations. While most of the literature focuses on post-hoc interpretability, i.e. surrogate methods that try to explain black-box models from outside, there is an ongoing debate about the limitations of such methods, in opposition to the usage and research on intrinsically interpretable models.Having the above in consideration, this dissertation proposes an intrinsically interpretable Deep Learning model. We develop a new multi-layer neural network architecture that allows the definition of interpretable variables in a dataset that, according to a priori knowledge, has a monotonic relation with the response variable. This technique is based on the imposition of monotonicity constraints in the model, using two alternative approaches: (1) by imposing constraints on the model’s parameters in specific layers, or (2) by changing the training process, applying an additional component to the loss function that penalizes non-monotonic gradients. The monotonicity approaches implemented with the architecture were evaluated, in the first stage, using a dataset provided by the host company, with the purpose of improving personalized offers to customers. To enable the reproducibility of this work, similar experiments were additionally conducted using a public dataset, for which the objective is to predict the prices of used cars. In both cases, results show that it is possible to improve the interpretability of black-box models without compromising their predictive performance. As a result, obtained models maintain predictive ability while simultaneously conforming to domain knowledge, this way allowing reliable knowledge discovery from data, and improving trustworthiness in decision making. |
|---|---|
| ISBN: | 9798383330579 |