即使我尝试使用训练数据进行预测，sklearn Logistic Regression 的准确性也太低

Question

I am currently trying to use Logistic Regression on some vectors and I use the sklearn library.

Here is my code. I first the files that contain the data and the assign the values to arrays.

# load files
xvectors_train = kaldiio.load_scp('train/xvector.scp')

# create empty arrays where to store the data
x_train = np.empty(shape=(len(xvectors_train.keys()), len(xvectors_train[list(xvectors_train.keys())[0]])))
y_train = np.empty(len(xvectors_train.keys()), dtype=object)

# assign values to the empty arrays
for file_id in xvectors_train:
  x_train[i] = xvectors_train[file_id]
  label = file_id.split('_')
  y_train[i] = label[0]
  i+=1

# create a model and train it
model = LogisticRegression( max_iter = 200, solver = 'liblinear')
model.fit(x_train, y_train) 

# predict 
model.predict(x_train)

#score
score = model.score(x_train, y_train)

For some reason even if I use the x_train data for my predictions the score is about 0.32. Shouldn't it be 1.0, because the model already knows the answers for those? If I use my test data the score is still like 0.32.

Does anyone know what the problem could be?

Answer 1

There isn't any obvious problem, and the result looks normal: your test score is very similar to your training score.

Most models try to learn the rules/params that generalize to new data, but NOT memorizing your existing training data, which means "Shouldn't it be 1.0, because the model already knows the answers for those?" is not true...

If you are actually seeing that your test set score is significantly lower than your training score (eg, 0.32 vs 1.0), then it means your model is badly overfitting and needs to be fixed.