The Wise son:
This is a very interesting discussion. There are many differences between the two, but in my opinion, the most important one regards the assumptions that each method relies on. SM is known for its many underlying assumptions, while ML in general almost doesn’t rely on any assumption. I assume you all agree with me.
The Simple son:
Well, to my mind the real difference is the inference method. ML relies on dividing the data to training and testing sets while SM relies on distributions and Neeman-Pearson likelihood testing. Therefore SM results in a P-Value, while ML in an ad hoc assessment such as confusion matrix. KIS - keep it simple.
The Wicked son:
The main difference is of course the prediction. ML tries to use the accumulated data to predict the future, while SM predicts the past. Statisticians always warn against using their models for predictions, while ML engineers seem like they are reinventing the wheel. How about some practice to answer the question: “The next person you meet in the street, is he/she more likely to use SM or ML?”
SM may use this R code:
your_SM_model <- R::glm(SM_or_ML ~ Current_year + age + gender + coffee_intake + income, data=world_data, family=’binomial’)
new_data <- data.frame(Year=2022, age=32, gender=female, coffee_intake= 5, income = 18,000)
predict(your_SM_model, new_data, type="response")
And ML can use this R code:
library(neuralnet)
nn <- neuralnet(SM_or_ML ~ Current_year + age + gender + coffee_intake + income,data=your_train_set, hidden=c(2,1), linear.output=TRUE, threshold=0.01)
nn.results <- compute(nn, your_test_set)
prediction = nn.results$net.result
He who couldn’t ask:
Guys, I am shaking, both SM and ML are frightening… the formulas and codes, gosh… I would rather talk about other things.. I’ve made a delicious peach-Pi if anybody wants, it goes well with a sunset beach cocktail and with all of these sophisticated models.