Rule discovery program, rule discovery process, and rule discovery apparatus

A rule discovery process executed by a computer, the process comprising:

• receiving a plurality of sample data items each formed of a multimedia data item and a text data item distinctively associated with that multimedia data item;
• selecting evaluation positions of a data array constituting each of said multimedia data items in said sample data items, wherein the evaluation positions are positions at which evaluation is to be made;
• integrating said selected evaluation positions into a plurality of evaluation regions which are to be evaluated;
• calculating, for each of the plurality of sample data items, numerical values representing a feature of data in the individual evaluation regions;
• generating a prediction formula by analyzing correlation of the text data items included in the plurality of sample data items with the calculated numerical values representing the feature of data in the individual evaluation regions of the received multimedia data items, the prediction formula predicting, from data in the plurality of evaluation regions of a given multimedia data item, a single text data item that represents a feature of the given multimedia data item as a whole;
• evaluating accuracy of said prediction formula so as to obtain an evaluation result;
• and
• outputting said prediction formula as a rule that indicates relationships between the multimedia data items and text data items, when said evaluation result satisfies a predetermined criterion;
• wherein said selecting evaluation positions, said integrating selected evaluation positions, said calculating numerical values, said generating a prediction formula, said evaluating accuracy, and said outputting are repeated with a different combination of evaluation positions of said data array until a predetermined condition for completion is satisfied, and
• wherein said integrating forms the plurality of evaluation regions facing each other across one or more positions of the data array other than said selection evaluation positions of the data array by integrating adjacent ones of said selected evaluation positions of the data array as a candidate region for said evaluating by representing adjacent pixels using a value of a single variable, calculates a degree of dispersion of values of data elements in the integrated adjacent evaluation positions in each of the evaluation regions using the single variable, and cancels integration of the evaluation positions when the degree of dispersion is greater than a predetermined value.

A rule discovery apparatus comprising:

• a processor executing:
• an input reception unit which receives a plurality of sample data items each formed of a multimedia data item and a text data item distinctively associated with that multimedia data item;
• a search unit which selects evaluation positions of a data array constituting each of said multimedia data items;
• an integration unit which integrates said selected evaluation positions into a plurality of evaluation regions which are to be evaluated;
• a prediction formula generation unit which calculates, for each of the plurality of sample data items, numerical values representing a feature of data in the individual evaluation regions, and generates a prediction formula by analyzing correlation of the text data items included in the plurality of sample data items with the calculated numerical values representing the feature of data in the individual evaluation regions of the received multimedia data items, the prediction formula predicting, from data in the plurality of evaluation regions of a given multimedia data item, a single text data item that represents a feature of the given multimedia data item as a whole;
• a prediction formula evaluation unit which evaluates accuracy of said prediction formula so as to obtain an evaluation result;
• an output unit which outputs said prediction formula whose evaluation result satisfies a predetermined criterion, as a rule that indicates relationships between the multimedia data items and text data items;
• and
• a display unit which displays the rule on a display screen;
• wherein the operations of said search unit, said integration unit, said prediction formula generation unit, said prediction formula evaluation unit, and said output unit are repeated with a different combination of evaluation positions of said data array until a predetermined condition for completion is satisfied, and
• wherein said integration unit forms the plurality of evaluation regions facing each other across one or more positions of the data array other than said selection evaluation positions of the data array by integrating adjacent ones of said selected evaluation positions of the data array as a candidate region for said evaluating by representing adjacent pixels using a value of a single variable, calculates a degree of dispersion of values of data elements in the integrated adjacent evaluation positions in each of the evaluation regions using the single variable, and cancels integration of the evaluation positions when the degree of dispersion is greater than a predetermined value.