Multi-label classification in ML Python - ML Experiment: Train & Evaluate

Experiment - Multi-label classification

Problem:We want to classify images where each image can have multiple labels at the same time, like tagging a photo with 'beach', 'sunset', and 'people'. Our current model predicts labels but tends to overfit the training data.

Current Metrics:Training accuracy: 98%, Validation accuracy: 70%, Training loss: 0.05, Validation loss: 0.45

Issue:The model is overfitting: it performs very well on training data but poorly on validation data.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 92%.

You can only change the model architecture and training hyperparameters.

Do not change the dataset or data preprocessing steps.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

ML Python

import numpy as np
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, BatchNormalization
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import EarlyStopping

# Assume X_train, y_train, X_val, y_val are preloaded numpy arrays

model = Sequential([
    Dense(128, activation='relu', input_shape=(X_train.shape[1],)),
    BatchNormalization(),
    Dropout(0.4),
    Dense(64, activation='relu'),
    BatchNormalization(),
    Dropout(0.3),
    Dense(y_train.shape[1], activation='sigmoid')  # sigmoid for multi-label
])

model.compile(optimizer=Adam(learning_rate=0.001),
              loss='binary_crossentropy',
              metrics=['accuracy'])

early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

history = model.fit(X_train, y_train,
                    epochs=50,
                    batch_size=32,
                    validation_data=(X_val, y_val),
                    callbacks=[early_stop],
                    verbose=0)

train_acc = history.history['accuracy'][-1] * 100
val_acc = history.history['val_accuracy'][-1] * 100
train_loss = history.history['loss'][-1]
val_loss = history.history['val_loss'][-1]

print(f'Training accuracy: {train_acc:.2f}%')
print(f'Validation accuracy: {val_acc:.2f}%')
print(f'Training loss: {train_loss:.4f}')
print(f'Validation loss: {val_loss:.4f}')

Added Dropout layers after dense layers to reduce overfitting by randomly turning off neurons during training.

Added BatchNormalization layers to stabilize and speed up training.

Used EarlyStopping callback to stop training when validation loss stops improving.

Reduced learning rate to 0.001 for smoother training.

Results Interpretation

Before: Training accuracy was 98%, validation accuracy was 70%. The large gap shows overfitting.

After: Training accuracy dropped to 90%, validation accuracy improved to 87%. Loss values also show better generalization.

Adding dropout and batch normalization, along with early stopping and a lower learning rate, helps reduce overfitting and improves the model's ability to generalize to new data in multi-label classification.

Bonus Experiment

Try using a convolutional neural network (CNN) architecture instead of a simple dense network for multi-label classification on image data.

💡 Hint

CNNs can better capture spatial features in images, which often improves multi-label classification performance.

Practice

1. What is the main difference between multi-label classification and multi-class classification?

easy

A. Multi-label classification uses regression, multi-class uses classification.

B. Multi-label classification assigns only one label, multi-class assigns multiple labels.

C. Multi-label classification is used only for images, multi-class for text.

D. Multi-label classification assigns multiple labels to one example, multi-class assigns only one.

Solution

1. Step 1: Understand multi-label classification

   Multi-label classification means each example can have more than one correct label at the same time.

2. Step 2: Compare with multi-class classification

   Multi-class classification means each example can have only one label from many possible classes.

3. Final Answer:

   Multi-label classification assigns multiple labels to one example, multi-class assigns only one. -> Option D

4. Quick Check:

   Multi-label = multiple labels, multi-class = single label [OK]

Hint: Remember: multi-label means many labels per example [OK]

Common Mistakes:

• Confusing multi-label with multi-class
• Thinking multi-label assigns only one label
• Mixing up classification with regression
• Assuming multi-label is only for images

2. Which of the following is a correct way to represent labels for multi-label classification in Python?

easy

A. labels = [0, 1, 2]

B. labels = [[1, 0, 1], [0, 1, 0]]

C. labels = 'cat,dog,bird'

D. labels = 3

Solution

1. Step 1: Understand label representation for multi-label

   Multi-label classification uses a list or array where each position represents a label, with 1 or 0 indicating presence or absence.

2. Step 2: Check options for correct format

   labels = [[1, 0, 1], [0, 1, 0]] shows a list of lists with 1s and 0s, correctly representing multiple labels per example.

3. Final Answer:

   labels = [[1, 0, 1], [0, 1, 0]] -> Option B

4. Quick Check:

   Multi-label uses binary vectors per example [OK]

Hint: Use binary lists to show multiple labels [OK]

Common Mistakes:

• Using a single integer for labels
• Using a string instead of list
• Using a flat list without nested structure
• Confusing multi-class label format with multi-label

3. Given this Python code snippet for multi-label classification predictions:

import numpy as np
preds = np.array([[0.8, 0.1, 0.6], [0.3, 0.7, 0.2]])
threshold = 0.5
binary_preds = (preds > threshold).astype(int)
print(binary_preds)

What is the output?

medium

A. [[1 1 1]
[0 0 0]]

B. [[0 1 0]
[1 0 1]]

C. [[1 0 1]
[0 1 0]]

D. [[0 0 0]
[1 1 1]]

Solution

1. Step 1: Apply threshold to predictions

   Compare each value in preds with 0.5: values > 0.5 become 1, else 0.

2. Step 2: Convert boolean to int and print

   First row: 0.8>0.5=1, 0.1>0.5=0, 0.6>0.5=1; Second row: 0.3>0.5=0, 0.7>0.5=1, 0.2>0.5=0.

3. Final Answer:

   [[1 0 1] [0 1 0]] -> Option C

4. Quick Check:

   Thresholding preds > 0.5 = binary labels [OK]

Hint: Compare each prediction to threshold for binary output [OK]

Common Mistakes:

• Confusing > with >=
• Not converting boolean to int
• Mixing rows and columns in output
• Using wrong threshold value

4. You trained a multi-label model but it always predicts only one label per example. What is the most likely cause?

medium

A. Using softmax activation instead of sigmoid in the output layer

B. Using sigmoid activation instead of softmax in the output layer

C. Using binary cross-entropy loss

D. Using a threshold of 0.1 for predictions

Solution

1. Step 1: Understand output activations for multi-label

   Multi-label models use sigmoid activation to allow independent probabilities per label.

2. Step 2: Identify problem with softmax

   Softmax forces probabilities to sum to 1, so only one label gets high probability, limiting multi-label predictions.

3. Final Answer:

   Using softmax activation instead of sigmoid in the output layer -> Option A

4. Quick Check:

   Softmax limits to one label, sigmoid allows many [OK]

Hint: Use sigmoid for multi-label, softmax for single-label [OK]

Common Mistakes:

• Confusing softmax and sigmoid activations
• Ignoring loss function compatibility
• Setting threshold too low or high
• Assuming threshold fixes activation issues

5. You have a dataset where each image can have multiple tags like 'beach', 'sunset', and 'people'. You want to build a multi-label classifier. Which metric is best to evaluate your model's performance?

hard

A. Precision, Recall, and F1-score calculated per label and averaged

B. Accuracy (percentage of exact matches of all labels)

C. Mean Squared Error

D. Confusion matrix for single-label classification

Solution

1. Step 1: Understand evaluation needs for multi-label

   Exact match accuracy is too strict because all labels must match perfectly, which is rare.

2. Step 2: Choose suitable metrics

   Precision, Recall, and F1-score per label, then averaged, give a balanced view of performance on each label.

3. Final Answer:

   Precision, Recall, and F1-score calculated per label and averaged -> Option A

4. Quick Check:

   Use per-label metrics averaged for multi-label evaluation [OK]

Hint: Use per-label precision/recall for multi-label metrics [OK]

Common Mistakes:

• Using strict accuracy that ignores partial matches
• Using regression metrics like MSE
• Using single-label confusion matrix
• Ignoring label imbalance in metrics