The final report was very comprehensive and covered the entire process of working with and learning the data well. The visualizations of the data were relevant and helped provide intuition about the structure of the data, as well as the labels that were being predicted. It was also apparent that a lot of thought had gone into structuring the testing framework. The 5-fold cross validation seems like it was a good way to train and test the data, especially because the data was time-series. Another strong point was the use of the F-Score to demonstrate the performance of each model. The discussion explaining this was helpful, and it was useful for comparing the models.
In terms of areas for improvement, I agree with your analysis that the Naive Bayes classifier was suspicious in predicting the same value for all of the prediction set. An examination of the parameters learned by this classifier may have revealed why, and also may have provided another insight into the structure of the data. I also would have liked to see more discussion about why the AdaBoost classifier trained on the second set performed the best, possibly explained in terms of the similarities between the data from different years. Lastly, it does seem suboptimal to not use all of the data provided. I would be interested in seeing your results in applying a GLRM to learn some of these excluded data entries, like LOCATION. Thank you for the interesting project and writeup!
The final report was very comprehensive and covered the entire process of working with and learning the data well. The visualizations of the data were relevant and helped provide intuition about the structure of the data, as well as the labels that were being predicted. It was also apparent that a lot of thought had gone into structuring the testing framework. The 5-fold cross validation seems like it was a good way to train and test the data, especially because the data was time-series. Another strong point was the use of the F-Score to demonstrate the performance of each model. The discussion explaining this was helpful, and it was useful for comparing the models.
In terms of areas for improvement, I agree with your analysis that the Naive Bayes classifier was suspicious in predicting the same value for all of the prediction set. An examination of the parameters learned by this classifier may have revealed why, and also may have provided another insight into the structure of the data. I also would have liked to see more discussion about why the AdaBoost classifier trained on the second set performed the best, possibly explained in terms of the similarities between the data from different years. Lastly, it does seem suboptimal to not use all of the data provided. I would be interested in seeing your results in applying a GLRM to learn some of these excluded data entries, like LOCATION. Thank you for the interesting project and writeup!