Travaux pratiques (TP) - Intelligence artificielle et données

This document explores unsupervised classification methods, focusing primarily on two types: hierarchical and non-hierarchical methods. It details the K-means algorithm, explaining its principles, convergence criteria, and the role of intra-class and inter-class variance. Hierarchical clustering is also discussed, describing dendrograms and aggregation criteria such as single linkage, complete linkage, and variance-based methods. Additionally, the document provides an overview of mixed classification, combining K-means partitioning with hierarchical clustering for a more refined analysis.

This document outlines a course workshop on data mining as part of the Business Analytics & Data Science master's program. It includes a section on principal component analysis (PCA) applied to vehicle data with specific tasks for analysis. The focus is on interpreting correlations and distributions of vehicles using PCA techniques.

The document outlines a practical workshop for Principal Component Analysis (PCA) applied to six continuous variables recorded from data on 20 cars. Using R programming and the FactoMineR and factoextra packages, correlations between variables are analyzed to identify patterns and influences in the factor space. The methodology includes generating a correlation matrix, interpreting factor axes, plotting correlation circles, and examining individual contributions. Results provide insights into the distribution of vehicles across the factor map, supported by graphical visualizations.

This document presents a data mining project focused on classifying road accidents in France for the year 2020. It includes data importation, variable selection, descriptive statistics, correlation analysis, PCA, and K-means clustering. The findings highlight the relationships among various factors contributing to road accidents.

This document discusses the impact of adjusting the 'max_depth' parameter in a machine learning model. It highlights the problem of overfitting and underfitting based on changes to this parameter and identifies the optimal value for achieving the best mean squared error. The findings include specific error values computed for test datasets and emphasize the application of grid search to optimize 'max_depth'.

This document explores the application of regular expressions in R for processing email data. Initially, challenges in parsing email domains are identified, leading to the utilization of specific regular expressions to extract data systematically. Functions like 'grep', 'gsub', 'regexpr', and 'regmatches' are discussed and applied to manipulate, search, and modify email strings effectively. Results include generating a frequency table for email domains from the processed dataset.

This document discusses data reading, the application of regular expressions in R, and how to extract domains from email addresses. It provides coding examples and explanations of the syntax used in regular expressions. The focus is on overcoming challenges associated with unique email formats and counting domain frequencies.

This document outlines a lab project focused on implementing the K-Nearest Neighbors (K-NN) algorithm using the MNIST dataset. It provides objectives, references, and a hands-on approach to loading and visualizing handwritten digits. Students will apply the K-NN classifier to make predictions based on the dataset.

This lab focuses on developing an algorithm to classify images of different shapes (hearts, clubs, diamonds, spades) using various features such as central and Hu moments. It includes steps for segmentation, data selection, feature extraction, and testing the model using a confusion matrix. The goal is to improve the model's accuracy by selecting relevant features.

This document provides hands-on exercises with ensemble learning methods such as bagging, random forests, and boosting using Python's Scikit-learn library. Methods like BaggingClassifier and RandomForestClassifier are explored systematically in the context of reducing variance and improving model accuracy, with practical examples applied on the Digits dataset. Readers are tasked with evaluating model performance through accuracy, variance analysis, and graphical insights. The final sections investigate the role of parameter tuning in improving weak learners and optimizing ensemble models.

This document focuses on regression using decision trees with the Scikit-learn library. The methodology involves training a regression model on sinusoidal data with added noise and exploring model performance by adjusting parameters like max_depth. Practical exercises include modifying the noise level in data and tuning the decision tree parameters for optimized performance. Additionally, it introduces using the Diabetes dataset to evaluate mean squared error and conduct parameter tuning via grid search.