Either on the IRIS dataset or by generating data synthetically, compare the perceptron, SVM (linear kernel) and logistic regression on linearly separable data. When you do this, compare the margins and logit loss obtained by each of the three algorithms in the linearly separable case.

Run the SVM (linear kernel) and logistic regression on the Kaggle Parkinsons/leukemia dataset. Use a binary quantization of your target, and run the SVM and LR on the same quantized targets. To improve logistic regression, you have to preselect features (using LASSO, for example). Test both on a validation set using an appropriate metric that you pick.