Loading...

Replacing Deep Neural Networks with New Methods to Increase their Interpretability

Rajaei, Pedram | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57255 (19)
  4. University: Sharif University of Technology
  5. Department: Mathematical Sciences
  6. Advisor(s): Tefagh, Mojtaba; Motahari, Abolfazl
  7. Abstract:
  8. Recently, many bioinformatics prediction models have been produced using deep neural networks. However, due to the nature of deep neural network models, the impact of features on the final answer is not fully apparent. Moreover, due to the large number of features in bioinformatics problems, both significant computational resources and a considerable amount of time are required for model learning. Nevertheless, the selection and use of optimal features in bioinformatics is often overlooked. This is because microarray methods often generate a vast number of features, and selecting the optimal features from this large number can be very challenging and time-consuming. Additionally, the features selected by feature selection methods may not be interpretable in deep neural networks. In this thesis, we aim to introduce a novel method for solving the problem of classifying normal and cancer cells using a combination of feature selection methods and neural additive models, which achieves a relatively high accuracy (over 70%) and is highly interpretable. This model has been applied to 10,000 tabular thyroid cancer and normal datasets. Using this model, the features that are most correlated with the target variable are easily selected. Then, classification is performed using these features. Not only has the accuracy of the proposed model on the dataset been more than 70%, but due to the high interpretability of neural additive models, the impact of each feature on the problem's solution is evident. Also, due to the reduction in the dimensionality of the problem, both the required computational resources are reduced, and the model's speed is increased.
  9. Keywords:
  10. Genome Analysis ; Feature Selection ; Interpretability ; Bioinformatics ; Deep Additive Networks ; Deep Neural Networks

 Digital Object List

 Bookmark

...see more