Microsoft Technology Licensing, LLC

Satya Pradeep Kanduri

Mikhail Parakhin

Marcelo De Barros

Qiang Wu

Manish Mittal

Nabeel Kaushal

Cynthia Yu

Adam Edlavitch

A web search system uses humans to rank the relevance of results returned for various sample search queries. The search results may be divided into groups allowing training and validation with the ranked results. Consistent guidelines for human evaluation allow consistent results across a number of people performing the ranking. After a machine learning categorization tool, such as MART, has been programmed and validated, it may be used to provide an absolute rank of relevance for documents returned, rather than a simple relative ranking, based, for example, on key word matches and click counts. Documents with lower relevance rankings may be excluded from consideration when developing related refinements, such as category and price sorting.

Shopping search engines

Outlier images&#x2014;those images that differ substantially from other images in a set&#x2014;can be automatically identified. One or more penalty values can be assigned to each image that quantifies how different that image is from others in the set. A threshold can be determined based on the set of penalty values. Each image whose penalty values are above the threshold is an outlier image. The penalty values can be the sum of per-pixel penalty values multiplied by the number of pixels with nonzero penalty values. A per-pixel penalty value can be the difference between a color value for that pixel and a predetermined range of color values, based on corresponding pixels in other images. The per-pixel penalty value can be determined for each component color and then optionally summed together. The threshold penalty values can be adjusted to provide for greater, or less, sensitivity to differences among the images.

Automated identification of image outliers

Growing community of inventors

Redmond, WA, United States of America

Satya Pradeep Kanduri