Microsoft Technology Licensing, LLC

Emre Mehmet Kiciman

Chun-Kai Wang

Benjamin Livshits

Yi-Min Wang

Kuansan Wang

Albert Gordon Greenberg

Alastair Wolman

John David Dunagan

Jiahe Helen Wang

David A Maltz

Paul Nathan Bennett

Xiaoxin Yin

Heather E Warncke

Romain Thibaux

James Robert Hamilton

John Carlton Platt

Robert L Rounthwaite

Susan Theresa Dumais

Francislav Petrov Penov

Li-wei He

Anders Bertil Vinberg

Jack Wilson Stokes, Iii

George M Moore

Benjamin Zorn

Jaime Brooks Teevan

Stuart Hillary Schaefer

Silviu-Petru Cucerzan

Shuo Chen

Peter Bailey

David G Campbell

Wissam S Kazan

Bo-June Hsu

Nikhil Dandekar

Felipe Luis Naranjo

Aaron C Hoff

Chen Ding

Tuomas Aura

Aleksey Sinyagin

Gilad Lotan

Galen Andrew

Yuzhe Jin

George Danezis

Jeffrey Todd Cohen

Manoj Ajith Prasad

Wenzhao Tan

Dennis B Gannon

Sreeharsha Kamireddy

David Simmons

Andrew Thomas Marshall

Abulimiti Aji

Huizhong Duan

Donald F Ferguson

Danah Boyd

Paul Oka

Yi-Chin Tu

Wissam Kazan

The techniques disclosed herein enable systems to train machine learning models using benign augmentation to enabled resistance various data poisoning attacks. This is achieved by first training a machine learning model using an initial dataset that is trustworthy and originates from a known source. The initial dataset is then modified to include known attack triggers such as syntactic paraphrasing to generate an augmented dataset. The augmented dataset is then used to train a robust machine learning model based using the initially trained machine learning model. The resultant robust machine learning model is then enabled to detect and resist attacks captured by the augmented dataset. The robust machine learning model can be retrained using an untrusted dataset that includes various compromised inputs in conjunction with the augmented dataset. Retraining results in an updated robust machine learning model that can learn and resist various data poisoning attacks on the fly.

Adversarial training to minimize data poisoning attacks

Technologies described herein relate to receiving data from a social networking application, wherein the data is generated by a first individual, and executing an entity extraction algorithm to extract at least one entity from the received data. A hyperlink is displayed on a computing device utilized by a second individual, wherein the hyperlink is based at least in part upon the at least one entity, and the hyperlink is configured to direct the first individual to a search engine and is further configured to cause the search engine to execute a search that pertains to the at least one entity.

Generating activities based upon social data

A method described herein includes acts of extracting at least one entity from an arbitrary web page being viewed by an individual on a computing device and comparing the at least one entity with social network data of the individual, wherein the social network data comprises a plurality of messages generated by members of a social network of the individual. The method further includes the acts of identifying at least one message in the plurality of messages based at least in part upon the comparing of the at least one entity with the social network data of the individual and causing the at least one message to be displayed on the web page in conjunction with the at least one entity.

Social browsing

Technologies pertaining to the creation of personalized data for an individual, including personalized data based on a social networking participant's social networking data, entity extraction, and automatic collage generation are described herein. In one implementation, the personalized data is used to provide a personalized, social-based replacement for a default UI element. In one implementation, a disambiguated named entity is extracted from user generated content in a social networking application, a socially interesting photomontage is automatically generated based on photographs from a social networking application, and the photomontage is rendered with hotspots containing socially relevant annotations.

Social home page

Techniques that may be used for detecting a primary content (e.g., a web page) that the user is viewing and presenting one or more pieces of supplemental content (e.g., social media data) together with the primary content. The supplemental content presented to the user together with the primary content may be content that is matched to the primary content and therefore detected to be relevant to the user. Detection of primary content and matching to supplemental content may be carried out based on a comparison of entities related to the primary and supplemental content. In some embodiments, an analysis of the primary content for entities may include ordering entities according to significance in the primary content and selecting top entities for comparison. Also, in some embodiments, multiple pieces of supplemental content may be displayed to a user categorized based on entity.

Presenting supplemental content in context

Systems, methods, and computer storage media are provided for analyzing a large amount of social media data from a large population of social media users and constructing correlational data models between one or more events that occur within each user's timeline. Social media posts directed to personal experiences of a large number of social media users are extracted. Event timelines are generated for each of the social media users, based on their personal experiences. The event timelines are analyzed with a particular event of interest to measure correlations between events occurring within the timelines and the particular event of interest. Using the measured correlations, a correlational data model is thereby constructed. The correlational data model may be used for application to decision-making calculations by one or more systems in an active or passive manner.

Modeling actions, consequences and goal achievement from social media and other digital traces

Various technologies described herein pertain to disambiguation of a mention of an ambiguous entity in a document. A set of candidate entities can be retrieved from an entity knowledge base based upon the mention of the ambiguous entity, where each of the candidate entities has a respective entity feature representation. Moreover, a document feature representation can be generated based upon features of the document and the respective entity feature representations of the candidate entities. A processor can be caused to select a subset of features from the document feature representation based upon a measure of how discriminative the features from the document feature representation are for disambiguating the mention of the ambiguous entity. A disambiguated result for the mention of the ambiguous entity can be determined based upon the subset of the features. The disambiguated result can be an unknown entity or one of the candidate entities.

Posterior probability pursuit for entity disambiguation

A crowdsourcing data structuring system and method for capturing unstructured data from the Web and adding structure by placing the data in a document that is accessible by others in a cloud computing environment. Using crowdsourcing, the unstructured data is annotated, amended, and verified to add structure to the unstructured data. An anchor and update module convert the data to a pointer that links the document to the data at an information source and stores the pointer in the document rather than the data itself. The data displayed in the document is updated whenever the information source is updated. A contribution module allows users to add data to the document, a validation module allows users to determine the validity of the data linked to in the document, and an expert ranking module allows users to rank the expert or contributor of the data in the document.

Structuring unstructured web data using crowdsourcing

A 'code splitting tool' provides various techniques for automatically analyzing and rewriting existing browser-based applications to introduce dynamic code loading into those applications thereby improving perceived application responsiveness. Structural elements of application code (including functions, classes, etc.) are broken into small 'stubs' with corresponding bodies. Rewritten applications then initially transfer only the portion of the code (including some combination of stubs and bodies) to the client as necessary for initial application execution. Additional stubs and/or bodies are then transferred either on-demand at runtime or in the background. Automated code rewriting proceeds without requiring any application-specific knowledge or changes to existing code prior to code rewriting. Further, the code splitting tool can tailor code rewriting to specific computing devices (computers, PDA's, cell phones, etc.), specific network conditions, and/or specific users, through an automated training process that creates clusters that control code downloads to optimize perceived application responsiveness.

Automated code splitting and pre-fetching for improving responsiveness of browser-based applications

An edge computing platform that provides on-demand delivery of Rich Internet Applications and other applications is disclosed. One embodiment includes an optional manager node and content distribution network (CDN) that include one or more compute nodes. The CDN collects information pertaining to execution of a software application. The CDN aggregates the information and transfers the aggregated information to the manager node. The manager node analyzes the information from the CDN and transfers results of the analysis to the CDN. The CDN receives a software application that is designed to be dynamically updated when executed at the clients. The CDN modifies the software application based on the information from the manager node. The CDN receives a request that pertains to the software application from a client device. The CDN transfers at least a portion of the modified software application to the client.

Edge computing platform for delivery of rich internet applications

A classifier that disambiguates among entities based on a dictionary, such as corpus of documents about those entities, is built by incorporating probabilities that an entity exists that is not in the dictionary. Given a document it is associated by the classifier with an entity. By incorporating out of collection probabilities into the classifier, a higher level of confidence in the match between an entity and a document is achieved.

Entity recognition using probabilities for out-of-collection data

A method described herein includes acts of receiving data from a social networking application, wherein the data is generated by a first individual, and executing an entity extraction algorithm to extract at least one entity from the received data. The method also includes automatically causing a hyperlink to be displayed on a computing device utilized by a second individual, wherein the hyperlink is based at least in part upon the at least one entity, wherein the hyperlink is configured to direct the first individual to a search engine and is further configured to cause the search engine to execute a search that pertains to the at least one entity.

Various technologies described herein pertain to using social activity data to personalize ranking of results returned by a computing operation for a user. For each of the results returned by the computing operation, a respective first affinity of the user to a corresponding result and a respective second affinity of the user to the corresponding result can be calculated and used for ranking the results. The respective first affinity of the user to the corresponding result can be calculated based on correlations between social activity data of the user and social activity data of a first group of historical users that clicked the corresponding result. Moreover, the respective second affinity of the user to the corresponding result can be calculated based on correlations between the social activity data of the user and social activity data of a second group of historical users that skipped the corresponding results.

Ranking based on social activity data

Deployment and execution of a service in a multiple datacenter environment may be facilitated using datacenter execution templates. Developers, business managers, and other interested parties may select and/or modify a declarative execution template embodying multiple factors. The execution template may then be used to generate an execution plan, which is then executed by datacenters which run the service. The execution template may be optimized after deployment of the service, allowing for ongoing adjustment in response to changing business, regulatory, and operational factors.

Datacenter execution templates

Sharing of user preferences is described. In an embodiment a user preference associated with a user is shared with a group of users in order to improve the relevance of results they receive. A database is used to store information detailing a number of groups of users extracted from a social network graph, where the social network graph describes connections between users. On receipt of a user preference associated with a user, a group of users containing the user is selected and the user preference is then shared with everyone in the selected group. In a further embodiment, the groups of users in the database may comprise cohesive groups of users and an extended group associated with each cohesive group. When selecting a group to share preference data with, a cohesive group containing the user is first selected and then the preference data is shared with the corresponding extended group.

Sharing of user preferences

Growing community of inventors

Seattle, WA, United States of America

Emre Mehmet Kiciman