Herein, we propose a machine learning method based on pedestrian trajectory data to classify public space usage states and discriminate unknown usage states. Aggregated feature values for each small cell were regarded as feature vectors representing the usage state. They were classified into usage state “types” via principal component analysis and x-means clustering. During validation using actual data, 16 types appearing at specific times and days were identified, and 1.1% of the test data were determined to be “new usage states” not found in the training data. This method helps understand long-term and complex variations in public space utilization patterns.