# Problem 2: Newsgroup Dataset Optimization

Using any approach, optimize performance of logistic regression on the test set in **news.zip** and compare the performance of your approach to standard SGD. This dataset is the full-dimensional newsgroup dataset (as opposed to the compressed version you worked with previously). The $X$ matrices are stored in sparse matrix format and can be read using scipy.io.mmread. As the dataset is large and high-dimensional, you will have to decide on how best to allocate your computational resources. Try to utilize the sparsity of the data (i.e., don't just convert it to a dense matrix and spend all your time multiplying zeros). You may use any of the techniques covered in class or ideas from outside class (e.g., momentum, variance reduction, minibatches, adaptive learning rates, preprocessing). Describe your methodology and comment on what you found improved performance and why. Plot the performance (negative log likelihood) of your method against standard SGD in terms of the number of gradient evaluations.

Problem 2: Newsgroup Dataset Optimization Using any approach, optimize performance of logistic regression on the test set in news.zip and compare the performance of your approach to standard SGD. This dataset is the full-dimensional newsgroup dataset (as opposed to the compressed version you worked with previously). The X matrices are stored in sparse matrix format and can be read using scipy.io.mmread. As the dataset is large and high-dimensional, you will have to decide on how best to allocate your computational resources. Try to utilize the sparsity of the data (i.e., don't just convert it to a dense matrix and spend all your time multiplying zeros). You may use any of the techniques covered in class or ideas from outside class (e.g., momentum, variance reduction, minibatches, adaptive learning rates, preprocessing). Describe your methodology and comment on what you found improved performance and why. Plot the performance (negative log likelihood) of your method against standard SGD in terms of the number of gradient evaluations