CS6140 Machine Learning

HW4A Ensembles: AdaBoost, Active Learning, ECOC

Make sure you check the syllabus for the due date. Please use the notations adopted in class, even if the problem is stated in the book using a different notation.

PROBLEM 1 Adaboost code [50 points]

Implement the boosting algorithm as described in class. Note that the specification of boosting provides a clean interface between boosting (the meta-learner) and the underlying weak learning algorithm: in each round, boosting provides a weighted data set to the weak learner, and the weak learner provides a predictor in return. You may choose to keep this clean interface (which would allow you to run boosting over most any weak learner) or you may choose to more simply incorporate the weak learning algorithm inside your boosting code.

Decision Stumps as simple classifiers

Each predictor will correspond to a decision stump, which is just a feature-threshold pair (f,t); in other words a single-split decision tree. Note that for each feature, you may have many possible thresholds which we shall denote . Given an instance, a decision stump predict +1 if the input instance has a feature value exceeding the threshold otherwise, it predicts -1. To create the various thresholds for each feature you should

sort the training examples by their f_i values
remove duplicate values, and
construct thresholds that are midway between successive feature values.

Boosting with Decision Stumps

Run your Adaboost code on the Spambase dataset

"Optimal" Decision Stumps: Run your implementation of boosting with "optimal" decision stumps on the training data. After each round, you should compute (1) the local "round" error for the decision stump returned, (2) the current training error for the weighted linear combination predictor at this round, (3) the current testing error for the weighted linear combination predictor at this round, and (4) the current test AUC for the weighted linear combination predictor at this round.
- Create three plots: One for the local "round" error (which should go up as rounds increase), one for the training and test error (which should both go down as rounds increase), and one for the test AUC (which should go up as the rounds increase). You should boost until you see "convergence" in test error or AUC.
- For the final weighted linear combination that is produced, create an ROC curve on the test data and compare your results to those you obtained in previous assignments.
You should think carefully about how you can efficiently generate the required results above. For example, I would suggest keeping a running weighted linear prediction value (before thresholding at zero) for each training and testing instance: when each new round predictor is created, you can simply update your running weighted linear prediction value and then easily compute training and testing error rates (by thresholding these values at zero), as well as testing AUCs (by ranking the instances by these values).
"Randomly Chosen" Decision Stumps: Repeat the procedure above for "randomly chosen" decision stumps. Note that you will almost certaily have to boost for more rounds to "converge".

PROBLEM 2 [50 points] Adaboost on UCI datasets

UCI datasets: AGR, BAL, BAND, CAR, CMC, CRX, MONK, NUR, TIC, VOTE. (These are archives which I downloaded a while ago. For more details and more datasets visit http://archive.ics.uci.edu/ml/). The relevant files in each folder are only two:
   * .config : # of datapoints, number of discrete attributes, # of continuous (numeric) attributes. For the discrete ones, the possible values are provided, in order, one line for each attribute. The next line in the config file is the number of classes and their labels.
    * .data: following the .config convention the datapoints are listed, last column are the class labels.
You should write a parser that given the .config file, reads the data from the .data file.

A. Run the Adaboost code on the UCI data and report the results. The datasets CRX, VOTE are required, rest are optional

B. Run the algorithm for each of the required datasets using c% of the datapoints chosen randomly for training, for several c values: 5, 10, 15, 20, 30, 50, 80. Test on a fixed fold (not used for training). For statistical significance, you can repeat the experiment with different randomly selected data or you can use cross-validation.

C: Active Learning Run your code from PB1 on Spambase, CRX, VOTE dataset to perform Active Learning. Specifically:

- start with a training set of about 5% of the data (selected randomly)

- iterate: train the Adaboost for T rounds; from the datapoints not in the training set; select the 2.5% ones that are closest to the separation surface (boosting score F(x) closest to 0) and add these to the training set (with labels). Keep training the ensemble, every T boosting rounds add data to training set until the size of the training set reaches 60% of the data.

How is the performance improving with the training set increase? Compare the performance of the Adaboost algorithm on the c% randomly selected training set with c% actively-built training set for several values of c : 5, 10, 15, 20, 30, 50. Perhaps you can obtain results like these

PROBLEM 3 [50 points] Error Correcting Output Codes

Run Boosting with ECOC functions on the 20Newsgroup dataset with extracted features. The zip file is called 8newsgroup.zip because the 20 labels have been grouped into 8 classes to make the problem easier. The features are unigram counts, preselected by us to keep only the relevant ones.

There are no missing values here! The dataset is written in a SPARSE FORMAT: "label featureId:featureValue featureId:featureValue featureId:featureValue ...". The features not listed are not missing values, they have zero values which were not written down to save space. In a full-matrix format, these values would be 0.

ECOC are a better muticlass approach than one-vs-the-rest. Each ECOC function partition the multiclass dataset into two labels; then Boosting runs binary. Having K ECOC functions means having K binary boosting models. On prediction, each of the K models predicts 0/1 and so the prediction is a "codeword" of length K 11000110101... from which the actual class have to be identified.

You can use the following setup for 20newsgroup data set.

- Use the exhaustive codes with 127 ECOC functions as described in the ECOC paper, or randomly select 20 functions.

- Use all the given features

- For each ECOC function, train an AdaBoost with decision stumps for 200 or more iterations

The above procedure takes a few minutes (Cheng's optimized code, running on a Haswell i5 laptop) and gives us at least 70% accuracy on the test set.