Please add the dependency versions R & Python in the dependencies section.

Hi Ian,

I took another look at the report. I was wondering if you could add that we optimized for the max_depth and min_sample_split and also, that we balanced the classifier to the analysis? (based on the code below)

parameters ={'max_depth': np.arange(1, 30, 5), 
             'min_samples_split': np.arange(2, 400, 20)}

tree = DecisionTreeClassifier(class_weight='balanced')
clf = GridSearchCV(tree, parameters, cv = 10)
clf.fit(X_train, y_train)

I missed it last time, but I think it's important to specify in our analysis.

Thank you,
Betty

These are the suggestions provided by peers on our project:

1. Convert the feature importance table to a bar graph to provide better visualization and summary of results.

2. Include details about analysis such as the hyperparameters used for building the classifier (i.e. k for cross-validation, min_sample_split value, max_depth used).

3. Add to future direction about how our decision tree classifier would compare to a Naive Bayes classifier.

I think its always helpful to be extra clear in the usage instructions, so I would suggest that you indicate all the commands you suggest are expected to be run from the root of this project.

Mechanics

• The link in your proposal repo only redirect to the release page of the project repo. You can also create a direct link to v1.0 of the repo with https://github.com/UBC-MDS/DSCI_522_SF_crime/tree/v0.1.

• Would be nice to show in README that the data importing script does run and maybe print the first few rows of the data.

Reasoning

• Maybe this wasn’t covered at the time of writing your proposal, you might want to mention how you are going to train and tune your decision tree in terms of tuning and transformation of the predictions, in addition to what packages you will use. Since a good portion of the predictors are categorical, are you going to use all classes as is? Maybe some classes for these predictors will be quite rare as well?

Other comments

• Might be interesting to plot the incidents on a map of SF as part of an initial data exploration. With 2 million data points the map might look very busy though, might need to subsample or make a heat map instead.

Hi @bettybhzhou, how do you think we can divide the work for the pipeline?

Your python & R scripts would be more readable and modular if they all had a main function.

Not clear in overview what non-processed means? And maybe provide some examples of what encompasses processing as well to help the reader understand what problem you are trying to solve.

Processed:

• arrest, booked
• arrest, cited
• prosecuted by outside agency
• juvenile cited
• complainant refused to prosecute
• district attorney refused to prosecute
• not prosecuted
• juvenile booked
• psychopathic case
• located
• prosecuted for lesser offense

Not processed

• none
• cleared-contact juvenile for more info
• juvenile diverted
• juvenile admonished
• exceptional clearance
• unfounded

Mechanics

• Might want to add version number for the packages in the dependency section.
• Might want to add a link to the final report in README.
• Might want to explain the usage of the scripts a little more. Say, 01_clean_data.py has three arguments.

Quality

• In the documentation at the top of a python or R script, you can explain what the arguments are in addition to showing an usage example.

Report

• Would be nice to link the description of variables in Section 1.0 to the variable names in subsequent sections. Say, there are x and y in Table 2, are they longitude and latitude?
• Also, I am not 100% sure what is stolen, buliding, vehicle neither when I just look at the Table. Are they among the 50 words extracted? Maybe list these 50 words somewhere?
• How did you choose the depth of your decision tree?
• How did you compute the testing accuracy in section 4.0?
• Style-wise, would be nice to replace the 'snake_case' with the usual case throughout.

Given that docs/san_francisco_report.md depends on all the other things you are making, the all target probably only needs to depend on docs/san_francisco_report.md.

Figures should be as small as possible (yet still clearly readable and understandable). So that you can easily read the figure captions and surrounding text while looking at the figure. In the case of the figures in your report, the text size in the figures is great, but the plot areas and the bars/boxes can be much smaller. This is specifically only really a problem for figure 1 in your report.

It is recommended to list all files (not just the folder they are in) for the targets and dependencies in a Makefile to keep the dependency tree working well.