Google Inc.

International Business Machines Corporation

Roberto Javier Bayardo

Rakesh Agrawal

Ramakrishnan Srikant

Yiming Ma

Mayur Dhondu Datar

Yifang Liu

Zhenyu Liu

Dimitrios Gunopulos

Allen Pin-Hsiu Huang

Mayank Bawa

Daniel Gruhl

Gerald George Kiernan

Ching-Tien Ho

Ralf Rantzau

Shivakumar Venkataraman

John Christopher Shafer

Weipeng Paul Yan

Adam Isaac Juda

Samuel Ieong

Christos Faloutsos

Uma Mahadevan

Nicholas Charles Fox

Giao Nguyen

Sunita Sarawagi

Monica D Chawathe

Konstantinos Katsiapis

David Chen Chang

Weipeng Yan

Multi-class classification by training a machine learning system based on training inputs each of which includes features and at least one class label. Each training input is assigned a membership value that can indicate if an entity having the features of the training input is a member of the class corresponding to the class label that is also included in the training input. To determine if an entity having test features is a member of several test classes, test inputs can be constructed where each input includes the test features and a class label corresponding to one of the test classes. The test inputs are processed by the trained machine learning system, which produces as outputs test membership values that represent the likelihood that the entity having the features in the test input belong to the test class corresponding to the test class label also included in the test input.

Classification system

A method for enabling a user of a computing device to designate a content item for subsequent review is provided. The computing device is coupled to a memory device and a display. The method is implemented using the computing device. The method includes displaying, by the computing device, a first publication. The method further includes the steps of displaying, by the computing device, a content item at a first time, wherein the content item is displayed in conjunction with the first publication. The method further includes receiving, by the computing device, a first input from the user to designate the content item for subsequent review, generating an indicator that the content item has been designated for subsequent review, causing the memory device to store the indicator, determining a second time to display the content item, and displaying the content item at the second time.

Presentation of non-interrupting content items

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting controlling keywords. In one aspect, a method includes determining that two or more targeting keywords for a same content item source are matched by one or more terms in a content item request, where a first targeting keyword matches the one or more terms with a more specific type of match than the second targeting keyword. If the bid associated with the second targeting keyword does not exceed the bid associated with the first targeting keyword, the first targeting keyword and the second targeting keyword are ranked based on selection scores. The targeting keywords are ranked based at least in part on the quality scores when the bid associated with the second targeting keyword does exceed the bid associated with the first targeting keyword.

Tie breaking rules for content item matching

A system and method to target advertisements to a user search query using a previously entered user search query as well as the current search query. The previously entered search query can accompanying the current search query in a query referrer, or can be determined separately from the current search query. The relevance of the previous search query is determined with respect to the current search query, and if found relevant a combination of the two query terms is used to find a set of advertising keywords from which candidate advertisements to be targeted and served to the user with search results are found. The process can be extended to additional search queries.

Using previous user search query to target advertisements

Pairs of similar objects in a population of objects can be found using a process that includes identifying a comparison vector x in a set of vectors having non-zero features, determining an estimated similarity contribution of a subset of features of the comparison vector x to a similarity between the comparison vector x and each vector in the set of vectors, generating an index that includes features based on a comparison of the similarity contribution with a similarity threshold, and identifying another vector in the set that is similar to the vector x using the index.

Similarity-based searching

A system to provide search query information. The system receives a request for search query information, identifies a set of search queries from a search query log that includes search queries submitted to a search service over a predetermined length of time, and provides the set of search queries. Each of the set of search queries is associated with at least a predetermined number of unique identifiers. Each of the set of search queries is matched to the request for search query information by a combination of exact matches, expanded matches, and broad matches.

Query statistics provider

Pairs of similar vectors in a set of vectors are identified. A comparison vector x is identified, and a set of candidate vectors corresponding to the vector x is identified. For each candidate vector y in the set, a similarity estimate between the comparison vector x and the candidate vector y is determined; if the similarity estimate meets a similarity threshold, a similarity score of the comparison vector x and the candidate vector y is determined; and if the similarity score meets the similarity threshold, the pair of vectors (x, y) is included in a list of similar pairs of vectors.

Pairs of similar vectors (or objects) in a set of vectors (or objects) are identified. A comparison vector x in a set of vectors is identified; a size threshold is determined such that if a similarity between the vector x and a vector y in the set of vectors is equal to or greater than a similarity threshold, then the vector y has a size at least equal to the size threshold, the size of the candidate vector y being determined based on a number of non-zero features in the vector y. A vector having a size less than the size threshold is removed from the set of candidate vectors.

A system and associated method for bulk processing of semi-structured results streams from many different resources ingest bytes, parse as many bytes as practical, and return to process additional bytes. The system processes network packets as they arrive from a computing resource, creating intermediate results. The intermediate results are held in a stack until sufficient information is accumulated. The system then merges the intermediate results to form a single document model. As network packets at one connection are consumed by the system, the system can select another connection at which packets are waiting for processing. The processing of a result at a connection can be interrupted while the system processes the results at another connection. In this manner, the system is able to utilize one thread to process many incoming results in parallel.

System and method for bulk processing of semi-structured result streams from multiple resources

An auditing framework for determining whether a database disclosure of information adhered to its data disclosure policies. Users formulate audit expressions to specify the (sensitive) data subject to disclosure review. An audit component accepts audit expressions and returns all queries (deemed 'suspicious') that accessed the specified data during their execution.

Auditing compliance with a hippocratic database

A privacy-preserving index system addresses the problem of providing a privacy-preserving search over distributed access-controlled content. Indexed documents can be readily reconstructed from inverted indexes used in the search. The privacy-preserving index system builds a centralized privacy-preserving index in conjunction with a distributed access-control enforcing search protocol. The privacy-preserving index utilizes a randomized algorithm for constructing a privacy-preserving index. The privacy-preserving index is strongly resilient to privacy breaches. The privacy-preserving index system allows content providers to maintain complete control in defining access groups and ensuring its compliance, and further allows system implementors to retain tunable knobs to balance privacy and efficiency concerns for their particular domains.

Growing community of inventors

Morgan Hill, CA, United States of America

Roberto Javier Bayardo