Enhancing product search engine results using user click history

摘要

In an example embodiment, previous search queries and clicked-on results are retrieved. This results in one or more pairs, each pair containing a query from the search term database and a first set of search engine results from the click database. Then a score is calculated for each feature within the one or more pairs, and a second set of search queries is boosted using the scores for the features.

主权项

1. A method for enhancing search engine results, comprising:
retrieving user history data including one or more pairs, each pair containing a query searched for by a user in a search engine and a first set of search engine results for the query, the first set of search engine results including only results that the user has clicked on; grouping the pairs into groupings, each grouping having pairs sharing a single identical query; for each grouping:
tokenizing one or more fields of each search result in the first set of search engine results into a first set of features, wherein each feature is a characteristic of a product that is a subject of a result in the first set of search engine results;simulating a search using the single identical query contained in the grouping against a raw search engine, the raw search engine not boosting the single identical query and producing a second set of search engine results different than the first set of search engine results by virtue of at least one of the second set of search engine results being new and unseen by other users;gathering features from the second set of search engine results into a second set of features;for each feature in the first or second set of features:
computing a first probability of a search result in the first set of search engine results having the feature;computing a second probability of a search result in a result set comprising the first set of search engine results and the second set of search engine results having the feature;computing a score for the feature based on the first probability and the second probability, by computinglog2⁢p⁡(t|x)p⁡(t),  where p(t|x) is the first probability and p(t) is the second probability; and causing search terms in future queries to be boosted by adding one or more of the features based upon the computed scores for the features.