使用 sci-kit 学习的多类多输出回归

Question

I am attempting to use sci-kit learn to develop a Machine Learning program which predicts 9 outputs from 5 inputs but am having trouble.

I have acquired 20,000 instances of the 5 inputs with corresponding 9 outputs for training purposes. The inputs represent the performance measurements of an amplifier. The outputs represent the component sizes which give those performance measurements.

So one row of input variables X may be: [ 8430, 6895, 12735, 208929613, 249]

With the corresponding output variables y: [1000, 400, 1000, 2000, 2500, 1000, 80, 1000, 2000]

After importing all the relevant libraries and assigning the inputs to X and outputs to y, I attempt to create the model as follows:

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
X_train = X_train.values
X_test = X_test.values
y_train = y_train.values
y_test= y_test.values

model = DecisionTreeRegressor()
model.fit(X_train, y_train)
predictions = model.predict(X_test)
accuracy = accuracy_score(y_test, predictions)

After running this code however, I get the following error:

ValueError: multiclass-multioutput is not supported

But reading the SKlearn website says that Decision Trees are inherently multiclass? How should I proceed to fix this error? Or is SKlearn not suitable to this kind of problem? Should I investigate using a neural network instead?

Answer 1

Decision trees are multiclass means that they can deal with a data where the differents samples belong to several classes not that each sample belongs to different classes. you could still implement youre own decision tree and adapt it to your problem by choosing features while building youre tree relatively to the average information gain over the different classes and it would solve the problem. You could also more simply use a basic but different decision tree for each of the different labels while passing to each tree only 1 label per sample. You can also adapt decision trees in many other ways to youre problem or use NN that would be more natural but probably less effective if the data is well structured.