Keras for Classification (Binary & Multi-Class)

Classification forms the backbone of many deep learning applications:

• Sentiment analysis
• Image recognition
• Spam filtering
• Disease classification
• Quality control
• Product categorisation

We’ll use two standard real-world datasets:

• IMDB Reviews → Binary classification
• Fashion-MNIST → Multi-class classification

Binary Classification - (IMDB Sentiment)

IMDB Dataset Overview

Dataset contains:

• 25,000 movie reviews (train)
• 25,000 movie reviews (test)
• Labels: 0 = negative, 1 = positive
• Reviews are pre-tokenised as integer sequences

Load data:

We limit vocabulary to 10,000 most common words.

Preprocess the IMDB Data

Reviews are variable-length sequences.
We pad them so all sequences have equal length.

Build a Binary Classification Model

We’ll use:

• An Embedding layer to convert integers → dense word vectors
• A GlobalAveragePooling1D to reduce the sequence
• A Dense output with sigmoid activation

Why this architecture?

• Embedding: learns word meanings
• Pooling: simple sequence reduction
• Sigmoid output: perfect for binary classification

Compile the Binary Model

Binary classification uses:

• Loss: binary_crossentropy
• Activation: sigmoid
• Metric: accuracy

Train the Model

Training is fast because the model is lightweight.

Evaluate the Model

Typical accuracy: ~85–88%
More advanced models (LSTM/Conv1D) reach 90–92%, which you’ll learn in a later section.

Predictions

Multi-Class Classification - (Fashion-MNIST)

Fashion-MNIST Overview

This dataset is like MNIST but with clothing items:

• T-shirts
• Trousers
• Pullovers
• Dresses
• Coats
• Sandals
• Shirts
• Sneakers
• Bags
• Ankle boots

Load dataset:

Preprocess the Data

Normalise pixel values:

Flatten for dense layers:

Build a Multi-Class Classifier

Multi-class classifiers end with softmax output and integer labels.

Compile the Multi-Class Model

Key points:

• softmax output
• sparse_categorical_crossentropy for integer labels (0–9)

Train the Model

Expect around 88–92% accuracy without CNNs. (CNNs will significantly improve in a later section.)

Evaluate

Making Predictions

Confusion Matrix

We can use scikit-learn’s confusion matrix:

Key Differences Between Binary and Multi-Class Models