CS 539 Fall 2000 - Assignment 3

Computer Science Department

CS539 Machine Learning
Assignment 3 - Fall 2000

PROF. CAROLINA RUIZ

Due:
First Part: Thursday, September 28, 2000 at 6:00 pm.
Second Part: Thursday, October 05, 2000 at 6:00 pm.

Project Description
Project Assignment
Report Submission and Due Date

PROJECT DESCRIPTION

Construct the most accurate neural network you can for predicting whether the income of a given person is >50K or <= 50K using the census-income dataset from the US Census Bureau which is available at the Univ. of California Irvine Repository.

I have downloaded the dataset into the following directory: /cs/courses/cs539/f00/Projects/Census_Income_Data
You can access the dataset from there.

The census-income dataset contains census information for 48,842 people. It has 14 attributes for each person (age, workclass, fnlwgt, education, education-num, marital-status, occupation, relationship, race, sex, capital-gain, capital-loss, hours-per-week, and native-country) and a boolean attribute class classifying the input of the person as belonging to one of two categories >50K, <=50K.

PROJECT ASSIGNMENT

This project consists of two parts:

Part 1: Due September 28 at 6:00 pm.

Find all mistakes on the C code for the neural network error backpropagation algorithm provided with Chapter 4 of the textbook. The main files you need to use are: backprop.c, backprop.h, and facetrain.c (beware that other files to run the experiment of classifying facial expressions described in the chapter are also provided). To the best of my knowledge, the code has ONE mistake. Submit a written report pointing out this mistake as well as all other mistakes you find in the code.

Part 2: Due October 5th at 6:00 pm.

Construct, train (using error backpropagation), and test the most accurate neural network you can to predict the Salary attribute of the Census-Income data. The following are guidelines to construct and train your neural net:

Code: You must use the C code for the neural network error backpropagation algorithm from Chapter 4 of the textbook. The main files you need to use are: backprop.c and backprop.h
Topology of your Neural Net: Your net should consist of 1 input layer, 1 hidden layer, and 1 output layer. All pairs of nodes in different layers should be connected, and no nodes on the same layer should be connected.
Training Instances: Use the census-income dataset. You can restrict your experiments to a subset of the dataset if your system cannot handle the whole dataset. But remember that the more accurate your system is, the better. Also, note that this dataset has missing values. It is up to you how to fill in appropriate data for those missing values. Also, it is up to you to decide if it's a good idea to discretize continues attributes, and if so, how.
Test Instances: Test data are also available at the UCI. YOU MUST USE AT LEAST THE FIRST 1000 TEST RECORDS FROM THAT TEST DATA IN YOUR EXPERIMENTS.

REPORT AND DUE DATE

Written Report. Please bring your report to my office (FL232) or to class by the due date/time. Your report should contain the following sections that discuss the issues:
1. Code Description: Describe any adaptations of the code that you made.
2. Experiments: For each further experiment you ran describe:
  - Training Data: What data did you use to construct your neural network?
  - Test Data: What data did you use to test your neural network?
  - The topology (number of units in each hidden layer), initial weights, number of iterations of the error backpropagation algorithm, and final weights of each of your neural nets,
  - Any pre or post processing done to improve the accuracy of your net.
  - Accuracy of the resulting neural network.
3. Summary of Results
  - What was the accuracy of the most accurate neural network you obtained?
  - Discuss how this accuracy compares with that of your most accurate decision tree from the previous assignment.
  - Include the most accurate neural network you obtained in your report.
  - Discuss the strengths and the weaknesses of your system.
Oral Report. We will discuss the results from the individual projects during the class on October 5th. Be ready to show your results (prepare transparencies on your results) and to discuss your project solution in class.