Our goal was to build and compare different models from different software tools. We compare a Neural Network from Keras [1] and both a Support Vector Machine and a K-nn regressors from the Scikit-learn [2] framework. We perform model selection and validate via a combination of screening phases and cross-validation, also testing on an internal test set extracted via hold-out.
The aims of project B of the Machine Learning (ML) course held by Professor Alessio Micheli at Department of Computer Science of University of Pisa is to bulid a ML model simulator to solve a multi-target regression task on CUP Dataset provided by the professor and Make a comparison among models . The project contains two different datasets. The Monk and Cup dataset. The Monk dataset is the bench mark dataset used to compare different models.
We study the influence of the various hyper-parameters upon the implementation of a neural network model. This is achieved through intensive use of screening phases which also provide the added benefit of alleviating the computational burden of the many grid-searches. The result of said effort, in the form of a neural network implementing a mini-batch Stochastic Gradient Descent, is then put to the test through a comparison with two main other well known regression models. The former, a Regressor-chain SVR, presents a Radial Basis Function kernel (rbf) with a degree of polynomial kernel function and regularization parameters while the latter, a K-Neighbors Regressor, weights the influence of each neighbor with respect to their euclidean distance from the record fed to it.
- All the detalis about the project can be found on the full report here
[2] https://scikit-learn.org/stable/
[3] https://pandas.pydata.org/
[6] https://seaborn.pydata.org/
- Dawit Mezemir Anelay :[email protected]
- Marco Petix :[email protected]