Abstract
This paper has applied the matrix Gaussian distribution of the likelihood function of the complete data set to reduce time complexity of multi-output relevance vector regression from O(VM^3) to O(V^3 +M^3), where V and M are the number of output dimensions and basis functions respectively and V < M. Our experimental results demonstrate that the proposed method is more competitive and faster than the existing methods like Thayananthan et al. (2008). Its computational efficiency and accuracy can be attributed to the different model specifications of the likelihood of the data, as the existing method expresses the likelihood of the training data as the product of Gaussian distributions whereas the proposed method expresses it as the matrix Gaussian distribution.
Original language | English |
---|---|
Pages (from-to) | 217-230 |
Number of pages | 14 |
Journal | Economic Modelling |
Volume | 81 |
Early online date | 20 Apr 2019 |
DOIs | |
Publication status | Published - 1 Sept 2019 |
Keywords
- relevance vector regression
- machine learning
- sparse Bayesian learning