Machine Learning Practical: 6 Real-World Applications

Explore how machine learning classifies breast tumor images as malignant or benign using features like radius, texture, and area from fine needle aspiration.

Visualize the breast cancer dataset in a Jupyter notebook using Seaborn. Create pair plots, count plots, scatter plots, and a heatmap of feature correlations to differentiate malignant from benign tumors.

Evaluate the model on unseen data to test generalization and avoid overfitting. Use a confusion matrix to compare y_test and y_predict and identify type I and II errors.

Explore convolutional neural networks, from basic neural networks to building a multi-layer convolutional network with feature detectors, pooling, and flattening, trained on a 70,000-image dataset to classify 28x28 images.

Explore how convolutions use a kernel to scan images, detect features with a feature detector, and produce feature maps, with blur and sharpen effects, stride, pooling concepts, and flattening.

Learn how convolutional networks apply relu activation after feature detection to create sparse feature maps, use maxpooling to reduce dimensionality, then flatten and feed a fully connected classifier.

Train a convolutional neural network on 28x28 grayscale images using Keras, preparing train/test/validation sets, normalizing by 255, reshaping to 28x28x1, and training with Adam and categorical crossentropy for 50 epochs.

Explore the correlation matrix as a pre-modeling heat map of inter-feature relationships to ensure features are independent and avoid linear dependencies before building a machine learning model.

identify patterns in users' financial habits to minimize churn in subscription products and predict why subscribers unsubscribe to reengage them, using product-related data from a fintech finance tracking product.

Build and evaluate logistic regression classifier for machine learning practical, using x_train and y_train. Fit on training data, predict x_test, and assess with confusion matrix, accuracy, precision, recall, and f1_score.

Run k-fold cross validation with scikit-learn's cross_val_score on a logistic regression classifier across 10 folds to report ~63–64% accuracy, then analyze feature coefficients to identify key predictors.

Import libraries and the financial data csv, then run an initial exploratory data analysis to inspect columns and describe numerical features, confirming no missing values and preparing for histograms.

Explore model building by applying support vector machines with linear and rbf kernels and a random forest with 100 trees, compare results and validate with 10-fold cross-validation.

Normalize the amount to a minus one to one range with a scikit-learn standard scaler, drop unused columns, and split the data into X and Y for fraud prediction.

Explore key deep learning concepts such as activation functions (sigmoid and ReLU), feed forward, backpropagation, supervised learning, and dataset splits with dropout to manage underfitting and overfitting.

Use decision trees to classify fraudulent transactions, train with X_train and Y_train, evaluate with score and confusion matrix, achieving 47 mislabelings out of about 500.

Apply under sampling to balance fraud and non-fraud cases by matching non-fraudulent samples to fraudulent counts, train-test split, and evaluate with a confusion matrix to compare performance.