Data Science for Business | 6 Real-world Case Studies

Leverage data science to reduce employee turnover by predicting quit risk with variables like job involvement, education, and work-life balance, using logistic regression and random forest models with scikit-learn.

Predict employee turnover by analyzing job involvement, education, satisfaction, performance, and work-life balance; compare logistic regression, neural networks, and random forest classifiers to optimize retention and reduce turnover costs.

Import pandas, NumPy, Seaborn, and Matplotlib; mount the drive and load the human_resources.csv dataset to predict employee turnover and inspect data basics.

Transform and visualize employee data by encoding yes/no features to 0/1, check for missing data with seaborn heatmap, plot histograms, and prune fixed columns for improved feature engineering.

Split the employee data into left and stayed groups, plot counts and percentages, and compare age, salary, and satisfaction. Visualize correlations with a heatmap.

Visualize employee attrition by age and job factors using count plots and seaborn, revealing turnover patterns by age, marital status, job role, involvement, and level.

Visualize attrition with kernel density plots comparing distance from home for left versus stayed employees, and examine years with current manager and total working years by gender and job role.

Prepare training and testing data by cleaning and preprocessing: separate categorical features, apply one-hot encoding, merge with numerical data, and scale with a minmax scaler for modeling.

Explore the intuition behind artificial neural networks, logistic regression, and random forest classifiers, including how dense feedforward networks with weights and activation functions map inputs to outputs.

Assess classification models with confusion matrices, and measure precision, recall, and F1 score to balance accuracy in unbalanced data.

Train and evaluate a logistic regression classifier using train-test split and visualize results with a confusion matrix. Assess accuracy, precision, recall, and the classification report to inform future model choices.

Train and evaluate a random forest classifier, achieving about 90% overall accuracy on the test set and analyzing precision, recall, and the confusion matrix to reveal class-specific performance.

Train and evaluate a neural network classifier with TensorFlow and Keras, using a sequential model of dense layers and a sigmoid output for classification; analyze training history, accuracy, and attrition.

Explore how data science enables market segmentation and targeted ads, using k-means, elbow method, PCA, and EDA with histograms and KDE, plus Autoencoders and Keras on a bank dataset.

Define the marketing problem and business case by segmenting bank customers with unsupervised learning, k-means, elbow method, PCA, and autoencoders for targeted campaigns.

Apply unsupervised learning to segment customers using k-means and PCA on a 9,000-sample credit card dataset, exploring features like balance, purchases, and cash advances.

Visualize and explore a credit card dataset using heatmaps and kernel density estimate plots, identify and fill missing values with the mean, drop duplicates, and examine feature correlations.

understand the theory of k-means clustering, an unsupervised method that groups data via nearest centroids using Euclidean distance, with iterative reassignment and elbow method to pick k.

Apply the elbow method to determine the optimal number of clusters by scaling data with a standard scaler and analyzing k-means inertia across k.

Extend K-means clustering to eight clusters, assign each credit card record to a cluster, and visualize cluster-specific histograms to reveal spending and cash-advance patterns for market segmentation.

Apply principal component analysis to reduce dimensions and visualize clusters, converting three features into two components (PCA one and PCA two) and plotting a PCA-based scatter plot colored by clusters.

Apply autoencoders for dimensionality reduction by building a 7-neuron bottleneck encoder and a symmetric decoder using keras, with 17-input features trained over 25 epochs on credit card df scaled data.

apply autoencoders to reduce 17 features to ten via the encoder, then use k-means and pca to cluster 9000 customers into four groups and visualize with two principal components.

Explore how data science and time series forecasting predict future sales from 1100 stores using features like promotions, holidays, and distance to competition, with the Facebook Prophet tool.

Forecast future sales from historical data using a Facebook profit framework, accounting for seasonality, holidays, promotions, and competition. Explore data import, visualization, and model training across thousands of stores.

Learn to import key libraries and datasets, mount Google Drive, and load CSV files: sales training data and store information, for a case study predicting future sales from historical data.

import and analyze store information data to explore two data frames, visualize the first five rows, and compute the average competition distance in meters and kilometers using pandas describe.

Perform data exploration with seaborn and matplotlib to visualize missing data, distributions, and filtering of open stores, preparing for merging sales and store data in task 3.3.

Explore store data quality by identifying missing values with a heat map, then clean data by filling nan with zeros or the distance mean, and inspect distributions with a histogram.

Merge store info with sales data on the store id, then visualize correlations with sales using a heatmap. Extract year, month, and day from the date to enable time-based analysis.

Explore data visualization of monthly sales and customers, identify peak periods, and compare store types A-D and promo effects using bar and violin plots.

Understand the intuition of Facebook Prophet for time series forecasting with an additive model, capturing seasonality and holidays to predict future sales from historical data.

Train a facebook prophet model on store-specific sales data, rename columns, fit the model, and forecast future daily sales with plots that reveal trend and weekly seasonality.

Data science for business: task six trains a model for store six, integrating public and school holidays to forecast 90 days with Facebook Prophet.

Explore how Toronto's case study uses deep learning to automate chest disease detection. Train Resnet models on x-ray data and classify into four classes: healthy, Covid-19, bacterial pneumonia, viral pneumonia.

Leverage deep learning to detect and classify chest diseases from x-ray images. Clarify the business case and problem statement to apply transfer learning across four classes.

Import and explore a custom chest x-ray dataset, use a pre-trained ResNet-50 model via Keras, and create training and validation generators to classify healthy, Covid-19, bacterial pneumonia, and viral pneumonia.

Visualize 36 train images in a 6x6 grid with labels to illustrate end-to-end deep learning classification and detection on dataset visuals, using matplotlib and label names.

Explore the theory and intuition of convolutional neural networks by examining convolutional layers, kernels, relu, pooling, flattening, and dense classifiers, with a history from LeNet-5 to ResNet and transfer learning.

Explore transfer learning in deep learning by reusing a pre-trained base network, freezing early layers, and training a new classification head for a new task.

Import a pre-trained ResNet-50 model trained on ImageNet, freeze the base layers except the last ten, and prepare to add a classification head in Task 7.

Build and train a resnet-50-based model with a classification head and softmax output for four-class medical image classification (normal, covid-19, viral pneumonia, bacterial pneumonia), using training, validation, and early stopping.

Evaluate the trained deep learning model’s performance by analyzing training and validation accuracy and loss. Validate on unseen test data with transfer learning, reporting 85% accuracy and preparing for deployment.

Explore natural language processing for sentiment analysis of social media posts to predict customer happiness. Use tokenization, CountVectorizer features, and naive Bayes or logistic regression classifiers in scikit-learn.

Import key libraries and load the Amazon Alexa reviews dataset, then inspect the dataframe to understand ratings, variation, and the 0/1 sentiment feedback, preparing for visualization.

Explore a reviews dataset by visualizing missing data with a heatmap and analyzing review length distributions through histograms, including mean, min, max, and longest messages.

Explore the dataset through exploratory data analysis by separating positive and negative reviews, visualize ratings and variations, and generate word clouds from verified reviews to uncover customer themes.

Perform data cleaning by dropping date and rating columns, encoding variations with one-hot get dummies, and concatenating features for model training with feedback as the target.

Remove punctuations from text to convert reviews into numeric format, enabling a machine learning classifier; then remove stop words, perform tokenization, and build a pipeline applying all steps.

Learn to remove stop words and punctuation with nltk's stopwords to reduce noise and reveal meaningful signal, building a clean text preprocessing pipeline for analytics.

Learn count vectorization and tokenization to convert text into a numeric feature matrix for a machine learning classifier, using scikit-learn's count vectorizer and a hands-on mini challenge.

Apply a text cleaning pipeline to the Amazon reviews data, removing punctuation and stop words, then perform count vectorization, and build X and y for subsequent Naive Bayes training.

Explain naive Bayes classifiers through a bank retirement example, detailing prior probability, likelihood, and posterior, and show how to predict retirement eligibility from age and savings.

Train a naive Bayes classifier with scikit-learn’s multinomial NB on encoded text features, using a 3150 by 5200 matrix, train/test split, and fit on X_train and y_train.

Evaluate a trained naive Bayes classifier using the confusion matrix, classification report, and train/test predictions to assess precision and recall for the target class in sentiment analysis.

Train and evaluate a logistic regression classifier using sklearn, assess accuracy, confusion matrix, and classification report, and explore data cleaning, tokenization, and vectorization in text data.

Apply deep learning to detect and localize defects in steel manufacturing images, using a two-stage pipeline with defect classification and ResNet-based segmentation to localize defects.

Explore ai and ml for defect detection and localization in steel manufacturing. Build a two-network pipeline with a Resnet classifier and a u-net segmentation model using 12,000 images with masks.

Import libraries and datasets, review run-length encoding and mask representations, and apply transfer learning with resnet50 pre-trained on image net data to perform pixel-level image segmentation.

Visualize and explore the dataset to understand defect distribution and run-length encoded masks, then assess class balance to set up segmentation with ResNet based transfer learning.

Explore convolutional and residual neural networks with skip connections and transfer learning for defect classification. Use a pre-trained ResNet, freeze early layers, and train only the new dense head.

Develop and train a ResNet-based binary defect detection classifier using transfer learning, with train/validation/test splits, image data generators, and early stopping to prevent overfitting.

Assess a trained ResNet model by loading its json architecture and weights, predicting on the test data, and evaluating accuracy, confusion matrix, precision, recall, and F1 with a tuned threshold.

Understand the intuition of ResNet-based segmentation by combining a U-net style encoder decoder with residual blocks, concatenating contraction and expansion path features to produce a pixel perfect output.

Build and train a ResUnet segmentation model using images and masks with an 80-20 train-test split, residual blocks, upsampling, and tversky loss, for 40 epochs with early stopping and checkpointing.

Assess the trained resnet segmentation model by loading weights and running predictions on test images. Generate masks from encoded pixels and compare to ground-truth masks to evaluate classifier-then-segmentation performance.