Mindset: Intent-driven Search

Mindset was an early demonstration of our ongoing machine learning research applied to the problem of web search. We trained an automated text classifier which determines whether any given web page is mostly "commercial" (e.g. it's main purpose is to sell you something) or not. We used new advanced algorithms recently developed at Yahoo! Research to train a classifier which is accurate and yet still fast enough to classify web pages on the fly as they show up in search results.

The demo provides a slider widget for users to explicitly specify their intent. Leaving the slider in the middle means they want to use the original Yahoo! search result order. Moving the slider to the extreme right means they want the top results to be those which the classifier is most confident are "non-commercial".

Putting the slider in between those two positions means a blend -- somewhat faithful to the original ordering but also tending to bring the more obvious non-commercial pages to the top. Similarly, sliding into the left half indicates blending of commercialness confidence vs original ordering. Regardless of the exact user interface mechanism (which was not the focus of our work), Mindset demonstrates how machine learning can enable explicit user intent to be harnessed to improve the search experience.