The purpose of this assignment is to give you experience developing one of the main algorithms for training feedforward neural networks—backpropagation. In this project, you will evaluate the performance of a feedforward network trained with backprop. For extra credit, you will compare the feedforward network to your previous implementation of the radial basis function network. If you did not implement the RBF network previous, you can do so now. You will also use the same five datasets that you used from Project 1 and Project 4 from the UCI Machine Learning Repository, namely:
1. Breast Cancer — https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Original% 29
This breast cancer databases was obtained from the University of Wisconsin Hospitals, Madison from Dr. William H. Wolberg.
- Glass — https://archive.ics.uci.edu/ml/datasets/Glass+Identification
The study of classification of types of glass was motivated by criminological investigation.
3. Iris — https://archive.ics.uci.edu/ml/datasets/Iris
The data set contains 3 classes of 50 instances each, where each class refers to a type of iris plant.
- Soybean (small) — https://archive.ics.uci.edu/ml/datasets/Soybean+%28Small%29 A small subset of the original soybean database.
5. Vote — https://archive.ics.uci.edu/ml/datasets/Congressional+Voting+Records
This data set includes votes for each of the U.S. House of Representatives Congressmen on the 16 key votes identified by the Congressional Quarterly Almanac.
When using these data sets, be careful of some issues.
- Some of the data sets have missing attribute values. When this occurs in low numbers, you may simplyedit the corresponding values out of the data sets. For more occurrences, you should do some kind of “data imputation” where, basically, you generate a value of some kind. This can be purely random, or it can be sampled according to the conditional probability of the values occurring, given the underlying class for that example. The choice is yours, but be sure to document your choice.
- Most of attributes in the various data sets are either multi-value discrete (categorical) or real-valued.You will need to deal with this in some way. You should apply what is called “one-hot coding” where you create a separate Boolean attribute for each value. For the continuous attributes, you may use one-hot-coding if you wish, but there is actually a better way. Specifically, it is recommended that you normalize them first to be in the range −1 to +1. (If you want to normalize to be in the range 0 to 1, that’s fine. Just be consistent.)
For this project, the following steps are required:
- Download the five (5) data sets from the UCI Machine Learning repository. You can find this repository at http://archive.ics.uci.edu/ml/. All of the specific URLs are also provided above.
- Pre-process each data set as necessary to handle missing data and non-Boolean data (both classes and attributes).
- Implement backpropagation for training feedforward neural networks. You may choose whether or not you wish to use the logistic activation function or the hyperbolic tangent activation function. Remember that this choice affects the update rules because of having different derivatives. It is also your choice whether or not you use momentum.
- Run your algorithms on each of the data sets. These runs should be done with 5-fold cross-validation so you can compare your results statistically. You can use classification error or mean squared error (as appropriate) for your loss function.
- Run your feedforward network trained with backpropagation on each of the data sets. These runs should output the classifications on all of the test examples. If you are doing cross-validation, just output classifications for one fold each. Run your experiments with no hidden layers, one hidden layer, and two hidden layers. It is up to you to tune the number of hidden nodes per layer, and be sure to explain how you did the tuning.
- Extra Credit (for up to 20 additional points): Run your previous implementation of the RBF network (implement here if you did not do it previously) and compare the results to the feedforward network.