Isolation-Inspired Machine Learning
Description
This Open access book belongs to What if one of the most challenging problems in machine learning-clustering in high-dimensional, complex data-could be solved not with deep learning, but through simple space partitioning in linear time? This book introduces a groundbreaking family of isolation-based algorithms, starting from the widely adopted Isolation Forest to the more recent Isolation Kernel and Isolation Distributional Kernel (IDK). Together, these methods offer an effective and efficient approach to anomaly detection, clustering, classification, and similarity search across both vector databases and complex data types such as time series, trajectories, and graphs.
Designed for machine learning researchers, data scientists, and professionals working with large or structured datasets, the book demonstrates how isolation partitions-created by isolating each point from the rest-can outperform many conventional techniques, including deep learning, in both speed and task-specific accuracy. It presents a compelling case where traditional NP-hard problems, like discovering clusters of arbitrary shapes and densities, are solved in linear time through isolation-inspired thinking. In the domain of spatial transcriptomics, IDK-based clustering has been shown to surpass even GPU-accelerated deep learning methods, running faster on CPUs while maintaining superior clustering performance.
Beyond algorithmic innovation, the book emphasizes intuitive insights and lessons learned over years of research that led to these breakthroughs. It challenges readers to rethink the role of learning in problem-solving and shows how understanding the problem is the necessary step in creating simpler and better solutions faster. With minimal mathematical prerequisites-only a basic understanding of core data mining tasks-this book invites a broad range of readers to explore how isolation-inspired methods can redefine efficiency and effectiveness in machine learning, dispelling the common assumption that 'deep learning is the answer'.
More details
Person
Kai Ming Ting is a Full Professor at the School of Artificial Intelligence, Nanjing University. He is best known for developing Isolation Forest, a widely adopted anomaly detection algorithm, which has more than 10,000 citations in either scholar.google.com or patents.google.com, and for introducing the Isolation Kernel and Isolation Distributional Kernel (IDK). These isolation-based methods have transformed data mining and machine learning by enabling fast, effective solutions to tasks such as anomaly detection, clustering, and retrieval in vector databases as well as complex data types including time series, trajectories, graphs, and spatial transcriptomics. Notably, IDK-based methods often outperform deep learning models while running efficiently on CPUs. Professor Ting has served on program committees for leading AI and data mining conferences such as AAAI, ACM SIGKDD, IEEE ICDM, and ICML, and has received research grants from the NSFC, US Air Force, Australian Research Council, and Toyota InfoTechnology Center. His work has earned accolades including the ACM SIGKDD ANDEA Test of Time Award and IEEE ICDM best paper award.
Content
.- Introduction.
.- Isolation-based Anomaly Detectors
.- Isolation Kernel for Similarity Measurement.
.- Isolation Distributional Kernel.
.- IDK-based Clustering: Metamorphosing an NP-hard problem into a Linear-time problem.
.- Similarity Measures for Complex Data Objects.
.- Using Isolation Partitions for Hashing and Retrieval.
.- Explainable Artificial Intelligence via Isolation Partitions.
.- Other Uses of Isolation Partitions and Key Questions Raised.
.- Mass Estimation and Mass-based Similarity.
.- When Deep Learning is not the Answer.
.- Key Takeaways and Future Directions.