Course summary

Pagina 1

Algorithm Assumptions Requirements Varie Anomaly detection Univariate anomaly detection Simple univariate thresholding No outliers in data Adaptive thresholding No outliers in data Exponential Smoothing No outliers in data Median absolute deviation (MAD) Strong to outliers Grubbs’ outlier test Data normally distributed Multivariate anomaly detection Elliptic envelope fitting Multivariate normal distribution of data % of outliers in the data (estimated) One-class SVM No anomalies in the training data Parametron nu = %outliers Kernel → gamma Data with a lot of features Isolation forest Contamination ratio (% of critical paths) High dimensional and poorly scaled dataset KNN K = number of points K-means K= cluster centroids Local Outliers Factor (LOC) Forecasting-based Methods Moving averages Differencing time series with a linear trend Autocorrelation cyclical or seasonal time series ARIMA p=autoregression d=distance order q=moving averages Recurrent Neural Networks (RNN) Long short-term memory (LSTM) Time series data Algorithm Assumptions Requirements Time complexity Clustering Distance measures Euclidean Jaccard Percentage of identical attributes Cosine Angle between vectors of the 2 points Edit Smallest number of insertion and deletion to transform the string x in y Hamming Number of components in which 2 binary vectors differ Hierarchical clustering → distance between points O(n2) Top-down process Bottom-up process Elbow analysis Representative-based clustering K-means Euclidean space Globular clusters Normalize data before K needed in advance initialization O(knt) k-medoids Non Euclidean space K needed in advance O(n2) BFR algorithm Cluster normally distributed around centroid Large quantity of data Expectation maximization Cluster with a probability distribution…