Anomaly Detection: Machine Learning, Deep Learning, AutoML

Identify anomalies across time-based, non-time-based, and image data by exploring supervised and unsupervised approaches; distinguish time series from non-time data.

Explore time-based anomaly detection across bank transactions, vendor purchases, and call center data to uncover anomalies. Apply supervised or unsupervised methods and image analysis to detect fraud and predictive maintenance.

Explore distinction between anomalies and outliers, showing every outlier is an anomaly but not vice versa, using box plots with IQR and control charts with mean and std dev.

Detect outliers with the z-score method by comparing data points to the mean using mu and sigma, flagging z-scores beyond 3 (or 2) in Python with NumPy and pandas.

Explore the interquartile range method to detect outliers using Q1 and Q3, compute IQR, and visualize with a box plot and data examples in Python.

Explore distance-based outlier detection using mean-based thresholds, such as mean plus or minus two standard deviations, with Euclidean, Manhattan, and Mahalanobis distances, plus clustering and nearest-neighbor insights in high-dimensional data.

Dbscan groups closely packed data points and flags outliers in low-density regions, avoiding a predefined cluster count and using epsilon and minimum points to define core, border, and noise.

Explore density-based, centroid-based, probabilistic, and hierarchical clustering methods, including dbscan, k-means, gmm, and lof, to identify anomalies in employee attrition data, with preprocessing and pca visualization.

Compare cluster and non cluster anomaly detection: cluster methods form data groups and flag points far from centroids, while non cluster methods assess each point against local density or norms.

Learn how histogram based outlier score (HBOS) identifies anomalies by constructing feature histograms, computing probability densities, and multiplying densities to flag unusual points with a 95th percentile threshold.

Identify anomalies efficiently with the isolation forest, a non clustering algorithm that evaluates each point by distance through random partitioning, recursive splits, and path-lengths in sparse regions.

Apply isolation forest to fraud detection in motor claims, and use explainable ai to reveal drivers of decisions, with label encoding and scaling at 1% contamination rate to identify anomalies.

Adopt a hybrid approach that clusters with GMM and detects local anomalies with LOF within clusters, improving precision and scalability in anomaly detection.

Explainable ai for anomaly detection using isolation forest and shape to reveal drivers of fraud in motor insurance, guiding proactive controls.

Explore anomaly detection in a supervised learning setup with labeled data, using get_dummies encoding and standard scaler preprocessing, and evaluate with accuracy, confusion matrix, TPR, TNR, and MCC.

Examine an imbalanced dataset case study in manufacturing using predictive maintenance, and compare methods like smote, random forest, and class weight adjustments to improve minority class detection.

learn how image anomaly detection converts images to vectors and uses isolation forest to flag anomalies, using OpenCV and image search library to detect cracks and counterfeit items.

Learn anomaly detection for images using isolation forest, with HSV color histograms as features, training on normal images to distinguish anomalies, and saving/loading the model for predictions.

Detect anomalies in time series by removing seasonal and trend components to obtain residuals, then use moving averages and interquartile range, with profit for change points and forecasting.

Use the profit library for time series anomaly detection, forecasting with ds and y labels and date formatting, loading Google Drive data in Colab, fitting the model, and plotting anomalies.

Explore generative adversarial networks, where a generator creates fake data and a discriminator distinguishes it in an alternating minimax game, enabling unsupervised learning and synthetic data for anomaly detection.

Explore Boltzmann machines, a generative, stochastic neural network with visible and hidden layers that learn dependencies and generate synthetic data, including restricted Boltzmann machines and deep belief networks.

Explore autoencoders, an unsupervised neural network that encodes input data into a latent space and decodes it back, using mean squared error to detect anomalies via reconstruction error.

Build unsupervised anomaly detection for insurance fraud using a 37-feature dataset with auto encoder, gan, and deep belief network; three anomalies detected by auto encoder and gan, one by dbn.