How to evaluate te testing data from the trained data?

Question

I am new at ML, and still trying to understand some concepts, so I figured I could ask here and maybe finally understand.

How does it work the whole splitting data from a data set?

Before answering, here is what I "know" and "understand" so far:

I have a data set, I'll call it "ds". ds has 10 features, f1,f2...f10, would be represented by each column of the ds. Now, I also have an array of labels that goes with each row of ds, which represents what each sample of data outputs.

Samples	F1	F2	F3	F4	F5	F6	F7	F8	F9	F10	Labels
1	ele1	ele2	ele3	ele4	ele5	ele6	ele7	ele8	ele9	ele10	1
2	ele11	ele12	ele13	ele14	ele15	ele16	ele17	ele18	ele19	ele20	0
...	...	...	...	...	...	...	...	...	...	...	...
n	n	n+1	n+2	n+3	n+4	n+5	n+6	n+7	n+8	n+9	1

Next, when testing ML algorithms, I'm suppose to split the ds into training data (training_data) and testing data (testing_data), and same thing with the labels, training label (training_labels), and testing labels (testing_labels). So now I have 4 pieces of information.

In order to check the algorithm's performance, I have to run the algorithm to each set of training data (rows...), and it would produce one result for each row, to which I would like to call "trained data". Once the algorithm has been run on the training data, I will also have an array of "trained_labels" ->[trained_label1, trained_label2, ... trained_label_n ] which it's size will match with the number of rows of the training_data.

Finally, I'm supposed to use this newly created "trained_labels" and compare it to the testing_labels and based on the accuracy (if all the trained_labels match all the testing_labels then it would be 100% accurate), I will be able to tell how good was the algorithm

Assuming a proportion 80% training and 20% testing, here is the big question that I still I am not able to wrap my head around, the size of the trained_labels will be different to the size of the testing labels, so given that fact, how can I campare the accuracy?

Like coding that with python syntax would be:

accuracy = (testing_labels == trained_labels).sum() / len(testing_labels) *100,

So if I could get some guidance it would be very appreciated.

Kind regards

P.S: I don't think slicing the training_labels size to match testing_labels size would be the answer.

Answer 1

I think there is a slight misconception here. When you are training your data, the algorithm is not producing a "trained label", it is using the labels to try to find the best fit of the model on your data to maximize its performance (getting it as close as possible as trained labels that we know). So your models literally "knows" the value of the training sample and labels.
Now what we want to do is evaluate whether this model that we fitted is performing decently on data it has not learned directly on. This is why we have the test data, which is data not seen yet by the algo. We will then compare the labels given by our model for the data it has not seen yet with the true results (test labels) that we know.
Regarding the 80/20 split or different cross validation techniques, we want to optimize the amount of data we train our model on. Basically you want your model to be trained on as much of "production like" data as possible, depending on your use case you might change the split arbitrarily or use cross validation to optimize the amount of data you train/test on.