This is a university project and part of the 2024 exam of a course regarding Deep Learning at the University of Pavia. The project consists of constructing an MLP that is able to classify two distinct classes of a dataset.
Moreover, the project has the requirement of finding some ways of analysing the final model in order to retrieve and understand how the decision process works, at least in the first few layers of the model.
The dataset is composed of 13 features and 1 target variable. The features can be numerical, categorical or binary. The target variable is binary.
After preprocessing the dataset, different models are trained and evaulated. The
best model is selected and analysed. It is a MLP with 3 hidden layers, with
respectively 12, 34 and 34 neurons, ReLU activation functions, and a final ouput
layer with a sigmoid activation function and a single neuron. See the file
DL-exam/Project_DL.ipynb for a more detailed description of the model, the training
process and the selection of the best model.
The model is able to classify the two classes with an accuracy of 0.8125. The confusion matrix is shown below:
The analysis of the model is performed by visualizing the weights of the model
and then recursively checking which features are the most important for the
determination of the output. A more detailed description of the analysis can be
seen again in file DL-exam/Project_DL.ipynb. Here
I will just report the final chart showing what part of the decision is taken by
each of the 13 features of the dataset (the chart shows on the left 27 features
since in preprocessing the number of features was increased by the one-hot
encoding of the categorical features).
As we can see there isn't any feature that is particularly dominant. There are though some who have slightly more or less influence. For instance, Feature 1 (count the right part of the bar) has an appearent influence on the decision: indeed it seems as though when the outcome of the model is 1 it's more likely that the value is either 1 or 2, while when the outcome is 0 it's more likely that the value is either 0 or 3