This training focuses on distance or density-based machine learning models that primarily take into account the local character of data points. Encompassing Density Estimation, Nearest Neighbors and Clustering, these algorithms are necessary tools in our Machine Learning toolkit.

The Local Models training starts by introducing the topic of Local Models and their applications in supervised as well as in unsupervised machine learning. We then discuss the relevance of different distance metrics and normalization methods, before delving into Density Estimation methods such as Kernel Density Estimation as well as parametric alternatives. Continuing to nearest neighbors algorithms like k-Nearest Neighbors and Approximate Nearest Neighbors, we finally arrive at unsupervised methods for Clustering, such as k-Means, Expectation Maximization and DBSCAN. This theoretical knowledge is applied in practice during a two-part lab session.

The training includes theory, demos, and hands-on exercises. After this training you have gained knowledge about:

• Distance Metrics
• Normalization
• Density Estimation
• k-Nearest Neighbors
• Approximate Nearest Neighbors
• Clustering
• k-Means
• Expectation-Maximization
• Hierarchical Clustering
• DBSCAN

The training starts with the definition of anomaly and some applications of anomaly detection. Then we talk about anomaly detection in the context of time series where we mainly focus on how to detect anomalies across many time series. After that the two main types of anomaly detection are discussed: supervised and unsupervised. Then we dive into the unsupervised anomaly detection algorithms and explain some of them in more detail. Finally we discuss data preprocessing steps and performance evaluation metrics.

In the lab exercise we apply the acquired knowledge to find faulty bearings, based on their vibration measurements. We do this by extracting features from the vibration time series and using unsupervised anomaly detection methods. The lab is available both in R and Python languages.

The training includes theory and hands-on exercises. After this training you will have gained knowledge about:

• Anomaly detection within and across time series
• Extract features from time series
• Types of anomaly detection: supervised, unsupervised
• Some methods for supervised detection
• Unsupervised anomaly detection algorithms
• Clustering based
• Kernel density estimation and Gaussian mixture model
• One-class SVM
• KNN-based methods, Local Outlier Factor
• Isolation forest
• Preprocessing steps, like normalization and PCA
• Evaluation metrics: AUC, precision, recall