The article introduces COBWEB, an incremental conceptual clustering system designed to maximize inference ability. Conceptual clustering is a machine learning task that organizes data into a classification scheme without requiring a pre-classified dataset. COBWEB is inspired by environmental and performance concerns, focusing on incremental processing and computational efficiency. The system uses a heuristic measure called category utility to guide the search for optimal classification trees. These trees are represented probabilistically, with each node summarizing the objects classified under it. COBWEB employs hill-climbing strategies and bidirectional operators (merging and splitting) to navigate through the space of possible hierarchies. The article evaluates COBWEB's effectiveness in various domains, demonstrating its ability to form classes that are useful for predicting unknown object properties. Experiments in the domains of congressional voting records and soybean disease cases show that COBWEB can achieve high accuracy in inferring diagnostic conditions and attribute values, highlighting the system's utility for real-world applications.The article introduces COBWEB, an incremental conceptual clustering system designed to maximize inference ability. Conceptual clustering is a machine learning task that organizes data into a classification scheme without requiring a pre-classified dataset. COBWEB is inspired by environmental and performance concerns, focusing on incremental processing and computational efficiency. The system uses a heuristic measure called category utility to guide the search for optimal classification trees. These trees are represented probabilistically, with each node summarizing the objects classified under it. COBWEB employs hill-climbing strategies and bidirectional operators (merging and splitting) to navigate through the space of possible hierarchies. The article evaluates COBWEB's effectiveness in various domains, demonstrating its ability to form classes that are useful for predicting unknown object properties. Experiments in the domains of congressional voting records and soybean disease cases show that COBWEB can achieve high accuracy in inferring diagnostic conditions and attribute values, highlighting the system's utility for real-world applications.