Interactive semantic data exploration for error discovery

A method for interactive exploration of data and classification predictions made by a concept classifier trained for a target concept, the method comprising:

• displaying, in a user interface, visual representations of a plurality of data items in a star coordinate space spanned by a plurality of anchor concepts each mapping the data items to respective finite real-valued scores, wherein positions of the visual representations of the data items are based on the scores for the plurality of anchor concepts and wherein attributes of the visual representations of the data items discriminate between labeled and unlabeled data items and reflect data item classifications as labeled or as predicted by the concept classifier, respectively;
• by a user, manipulating the anchor concepts in the user interface by at least one of moving, modifying a definition of, adding, or removing an anchor concept;
• responsive to user manipulation of the anchor concepts in the user interface, updating the positions of the visual representations of the data items in the star coordinate space;
• by the user, inspecting the positions and attributes of the visual representations of the data items to discover outliers, inspecting contents of data items identified as outliers or associated with discrepancies to discover a misclassified unlabeled data item, and, upon such discovery, labeling the misclassified data item for inclusion in updated training data;
• and
• responsive to user labeling of the data item, by a hardware processor executing machine-readable instructions, retraining the concept classifier based on the updated training data and updating predictions by the retrained concept classifier, and causing the attributes of the visual representations of the data items to be updated in the user interface based on the updated predictions.

A system for interactively visualizing predictions made by a concept classifier trained for a target concept, the system comprising:

• one or more hardware processors;
• and
• one or more machine-readable media storing instructions for execution by the one or more hardware processors, the instructions, when executed by the one or more hardware processors, causing the one or more hardware processors to perform operations comprising:
• causing display, in a user interface, of visual representations of a plurality of data items in a star coordinate space spanned by a plurality of anchor concepts, each anchor concept mapping the data items onto a finite real-valued score, wherein positions of the visual representations of the data items in the star coordinate space are based on the scores for the plurality of anchor concepts, and wherein attributes of the visual representations of the data items discriminate between labeled and unlabeled data items and reflect data item classifications as labeled or as predicted by the concept classifier, respectively;
• responsive to user manipulation of the anchor concepts in the user interface, updating the positions of the visual representations of the data items;
• and
• responsive to user selection and labeling of one or more of the data items, retraining the concept classifier and updating the attributes of the visual representations of the data items based on updated predictions by the retrained concept classifier.