Anomaly detection is one of the most practical applications of machine learning and statistics. It is relevant in finance, manufacturing, software operations, healthcare, industrial operations and many more.
During the workshop you will learn the principles and basics techniques of anomaly detection in a hands on fashion. You will code, benchmark and tests from scratch well known algorithms like Isolation Forests, ECOD, mahalanobis distance and PCA.
Contents:
-
Anomaly taxonomy and categorization: how anomalies are usually classified and why is important to think about this in practical applications
-
Evaluation metrics, synthetic data and algorithm benchmarking:: How do we know algorithm A y better than B? How do we know they are even working? In this section we will look into this question and code from scratch the ROC-AUC and precision/recall evaluation metrics
-
Z-score: Yo will implement the simplest possible yet effective anomaly detection algorithm, if you know your data.
-
ECOD: An up and coming anomaly detection workhorse for general use. You will code this algorithm and see how simplicity can be effective.
-
Isolation Forest: One of the most performant and battle tested algorithms for anomaly detection. It involves recursive tree traversing,
...
afficher plus
Anomaly detection is one of the most practical applications of machine learning and statistics. It is relevant in finance, manufacturing, software operations, healthcare, industrial operations and many more.
During the workshop you will learn the principles and basics techniques of anomaly detection in a hands on fashion. You will code, benchmark and tests from scratch well known algorithms like Isolation Forests, ECOD, mahalanobis distance and PCA.
Contents:
-
Anomaly taxonomy and categorization: how anomalies are usually classified and why is important to think about this in practical applications
-
Evaluation metrics, synthetic data and algorithm benchmarking:: How do we know algorithm A y better than B? How do we know they are even working? In this section we will look into this question and code from scratch the ROC-AUC and precision/recall evaluation metrics
-
Z-score: Yo will implement the simplest possible yet effective anomaly detection algorithm, if you know your data.
-
ECOD: An up and coming anomaly detection workhorse for general use. You will code this algorithm and see how simplicity can be effective.
-
Isolation Forest: One of the most performant and battle tested algorithms for anomaly detection. It involves recursive tree traversing, random numbers, averages and approximations but don't be discourage you will build it one step at a time
-
Mahalanobis / PCA: Time permitting you'll code this algorithm whose value is twofold effective in various circumstances and a gateway idea to more sophisticated algorithms such as Autoencoders
Requirements
The workshop is aimed to curios people that are comfortable with coding.
Mathematical ideas will be introduced as necessary but familiarity with notions such as probability, average, mean, standard deviation, etc is advantageous.
The examples and supporting code will be in Python but the workshop can be followed in any language.
Sergio Solorzano Rocha
Sergio Solórzano Rocha earned a PhD in physics from ETH Zürich focusing on computational physics. He published multiple papers on numerical algorithms for physical simulations and analysis. He is currently a senior researcher and developer at Exeon Analytics, where he works on building systems for anomaly detection in cybersecurity.
afficher moins
