CETM26: Machine Learning and Data Mining Report

Assignment Task

Assignment 1: Artificial Neural Networks

Aims

This coursework aims to provide a practical context to concepts covered in early weeks surrounding artificial neural networks. You will be required to conduct practical work with Python, and to write a brief report outlining your work.

Task - Feedforward Neural Networks

You will implement a feedforward artificial neural network (ANN) within Python (using Keras) for the purposes of solving a binary classification task as a mini-research project. For this network, you will be provided a dataset from the finance sector. A description of the dataset is provided in the Data Fields section of this document below.

You are expected to appropriately read in the training data, construct an ANN, train the network, and then evaluate it on the testing data.

For this you should consider aspects of the network architecture, such as how many hidden layers and nodes are required, for an ‘optimal’ solution. When constructing your network you should only consider changing this parameter, leaving others stationary, and the solver as ‘SGD’. Largely the data has already been transformed ready for the task; however, you should consider how many inputs you wish to provide to the network, both feature-wise, and number of examples.

Alongside the Python code, you will write a small report ( 1500 words maximum ) outlining your solution, the architecture chosen, any processing of the dataset, as well as evaluative results. For the purposes of this report, you should carefully consider experimental design, showing comparisons between various different architectures you’ve tried, using evaluative metrics to demonstrate an overall good solution to the task.

Dataset

You are provided a single CSV file for the purposes of this assignment. These data are heavily modified from the Bank Marketing Dataset available at the UCI Machine Learning Repository (Moro et al., 2014)

You can view the Data Description at the UCI Repository page. Where data originally had 3 or more attribute types for a categorical variable, these have been changed to numeric for you.