This is a good starting point showing the breakdown of logistic regression on a small dataset to predict SUV purchase data. This is an extension of the SUV data logistic regression analysis listed here. Note this was a rather small dataset so it is difficult to find too much to implement but this process would be useful for larger datasets. What you will see in this rendition is an implementation of a larger confusion matrix, indexing of columns, adding new columns, and an explanation regarding why we might want explore the splitting of the data based on a larger confusion matrix. This also explores the use of the multilabel confusion matrix that is part of the sklearn package. Finally, this separates the purchasing data into Male and Female purchases and demonstrates the methods you would use to explore how other models might fare for the less-accurately predicted male data. If you were to run the logistic regression after scaling the data, this is the basic breakdown you will see of true positives, false positives, false negatives and true negatives (0 labels represent walk-ways, while the 1 labels represent the purchases):
The first thing that needs to be performed differently to accomplish a deeper dive into the data is to enumerate the gender column and in this instance, likely due to alphabetical order, you will see male purchasers identified by a "1", and female by a "0". Here are the first 10 colyumns of the modified set:
| User ID | Gender_int | Age | EstimatedSalary | Purchased |
|---|---|---|---|---|
| 15624510 | 1 | 19 | 19000 | 0 |
| 15810944 | 1 | 35 | 20000 | 0 |
| 15668575 | 0 | 26 | 43000 | 0 |
| 15603246 | 0 | 27 | 57000 | 0 |
| 15804002 | 1 | 19 | 76000 | 0 |
| 15728773 | 1 | 27 | 58000 | 0 |
| 15598044 | 0 | 27 | 84000 | 0 |
| 15694829 | 0 | 32 | 150000 | 1 |
| 15600575 | 1 | 25 | 33000 | 0 |
| 15727311 | 0 | 35 | 65000 | 0 |
Then, with the help of setting up an array of conditions and using the "logical_and" as well as the numpy "select" function we can create the new column categorizing male, female, purchased, and non-purchased data:
| User ID | Gender_int | Age | EstimatedSalary | Purchased | Gender_purchase_even = Purchased |
|---|---|---|---|---|---|
| 15624510 | 1 | 19 | 19000 | 0 | 1 |
| 15810944 | 1 | 35 | 20000 | 0 | 1 |
| 15668575 | 0 | 26 | 43000 | 0 | 3 |
| 15603246 | 0 | 27 | 57000 | 0 | 3 |
| 15804002 | 1 | 19 | 76000 | 0 | 1 |
| 15728773 | 1 | 27 | 58000 | 0 | 1 |
| 15598044 | 0 | 27 | 84000 | 0 | 3 |
| 15694829 | 0 | 32 | 150000 | 1 | 2 |
| 15600575 | 1 | 25 | 33000 | 0 | 1 |
| 15727311 | 0 | 35 | 65000 | 0 | 3 |
The next hurdle is to rerun the testing and training data with regard to the newly added column. In this example this column was assigned a variable (Z) when the original confusion matrix was created. Simply run the LogisticRegression() model with this new data as your target column, rename the columns and rows with the .columns(), and .index() properties of your new dataframe, and apply labels that describe whether it was a man, woman, and whether they purchased or not to those properties. In this example, in order to see how all of the numbers break down percentagewise, use the dataframe.sum() method, and re-writing the values of the dataframe as the ratio of each cross-calculated value over the sum (DF = DF.astype('float')/DF.sum()). Here you can see the confusion matrix:
If you want to see a little breakdown of the numbers for each category, you can create a small bar chart. Simply be careful to label correctly as you will see the values are presorted based on which category is present the most.
From the expanded confusion matrix, it should be pretty clear that the models are better at predicting Female purchasers (just over 1% error for female data, but over 7% of the error is attributed to male data), than Male purchasers, so it is worth exploring if we could improve the Male data only. In case you want to see the counts, you can see the breakdowns here by constructing a title for each category, constructing selected values of that 4 category column, and constructing 2X2 confusion matrices from there. You can see these counts here: 
This next step is to set up a list of classification models with 2 columns; one with the models and the other with the arguments for each model. Another task is to create a function that utilizes the "eval()" function to concatenate the models in your list, with the arguments, then fit and transform those models to come up with prediction data for each model. Have the function return the accuracy percentages. Then, create a new list, and a for loop that will run through the initial list using the function created and capture the accuracy results. For this process, I tried the Random Forest Classifier, Naive Bayes, Support Vector Machines (linear and polynomial), and Linear Regression. Here is a chart of the accuracy results, and there were minor improvements for all models with the exception of Gaussian Naive Bayes:
This process might provide you with slighly different results if you ran different times, but you are unlikely to see much variance. The goal is to try to find a model that reliably and significantly performs better on the male data than logistic regression would. Then you could run logistic regression on the female data, run the improved model on the male results, and then combine the prediction results into one more efficient set. You can easilly try different arguments for the string inputs within the second column of the list of classifications. This is just a general process to explore which modeling methods might work best for a particular set of your data.
Ultimately this methodology will allow you to try different models, and split your accuracy results to see what data might benefit from a different model.