Microsoft Technology Licensing, LLC

Rovi Technologies Corporation

Daniel Sitton

Nir Nice

Dror Kremer

Roy Varshavsky

Michael Feldman

Eran Yariv

Noam Koenigstein

Ulrich Paquet

Shahar Keren

Limor Lahiani

Benny Schlesinger

Hen Fitoussi

Yoram Yaacovi

Kfir Karmon

Ron Jack Karidi

Oded Elyada

Keren Master

Shay Ben-Elazar

Shai Roitman

Shimon Shlevich

Ori Bar-Ilan

Yossef Hai Ben David

Amit Perelstein

Ori Folger

Noga Amit

Meitar Ronen

Amnon Catav

Shay Ben-elazar

Yair Zohav Zagdanski

Non-limiting examples of the present disclosure relate to application of artificial intelligence (AI) processing to generate classifications of user activity for a group of users. For example, a classification prediction is generated indicating whether students in an educational class are predicted, over a predetermined time period, to have a high or low activity level based on contextual analysis of multiple types of user-driven events. As user activity data is typically quite robust, the present disclosure applies dimensionality reduction processing to efficiently manage user activity data and further improve accuracy in generating downstream binary classifications. A dimensionality reduction transformation of user activity data results in a low-dimensional representation of input feature data that is contextually relevant for generating a binary classification. Derived classifications are then utilized to generate data insights pertaining to user activity levels of one or more users. Data insights are can then be rendered for presentation.

Automated generation of predictive insights classifying user activity

The present disclosure describes automated generation of early warning predictive insights derived from contextual analysis of user activity data of a distributed software platform. Predictive insights are automatically generated from analysis of user activity through implementation of trained artificial intelligence (AI) modeling. User activity data is accessed pertaining to user interactions by a plurality of users a software data platform. The trained AI modeling generates a plurality of mobility determinations that identify changes in patterns of user behavior over a current temporal filter associated with the user activity data. The plurality of mobility determinations is curated using business logic rules that evaluate a relevance of the mobility determinations. One or more predictive insights may be generated and presented via a graphical user interface notification. Exemplary notifications help provides insights into how user behavior has changed and why that is, thereby fostering understanding of predictions that can lead to actionable results.

Automated generation of early warning predictive insights about users

A data set may be distributed over many data stores, and a query may be distributively evaluated by several data stores with the results combined to form a query result (e.g., utilizing a MapReduce framework). However, such architectures may violate security principles by performing sophisticated processing, including the execution of arbitrary code, on the same machines that store the data. Instead of processing queries, a data store may be configured only to receive requests specifying one or more filtering criteria, and to provide the data items satisfying the filtering criteria. A compute node may apply a query by generating a request including one or more filter criteria, providing the request to a data node, and applying the remainder of the query (including sophisticated processing, and potentially the execution of arbitrary code) to the data items provided by the data node, thereby improving the security and efficiency of query processing.

Filtering queried data on data stores

Example apparatus and methods perform matrix factorization (MF) on a collaborative filter based usage matrix to create a multi-dimensional latent space that embeds users, items, and features. A full distance matrix is extracted from the latent space. The full distance matrix may be extracted from the latent space by defining a distance metric between item pairs based on the multi-dimensional representation in the latent space. The full distance matrix may be populated with values computed for item pairs using the distance metric. A plurality of vectors associated with a multi-dimensional Euclidean space are produced from the full distance matrix. The plurality of vectors produce a navigable data set. The plurality of vectors may be produced in a manner that minimizes strain on the distances vectors. A representation of the navigable data set may be presented as, for example, a virtually traversable landscape that supports an interactive user experience.

Recommendation system with multi-dimensional discovery experience

A recommendation system for optimizing content recommendation lists is disclosed. The system dynamically tracks a list interaction history of a user, which details that user's interactions with a plurality of different lists presenting different recommended items to that user. The system automatically correlates one or more list preferences with that user based on the list interaction history, and builds a recommendation list with a plurality of candidate items having different recommendation confidences. The recommendation list is built such that each candidate item with a higher recommendation confidence is prioritized over each candidate item with a lower recommendation confidence according to the one or more list preferences correlated to that user.

Adaptive recommendation system

A technique for efficiently factoring a matrix in a recommendation system. Usage data for a large set of users relative to a set of items is provided in a usage matrix R. To reduce computational requirements, the usage matrix is sampled to provide a reduced matrix R'. R&#x2032; is factored into a user matrix U&#x2032; and an item matrix V. User vectors in U&#x2032; and V are initialized and then iteratively updated to arrive at an optimal solution. The reduced matrix can be factored using the computational resources of a single computing device, for instance. Subsequently, the full user matrix U is obtained by fixing V and analytically minimizing an error in UV=R+error. The computations of this analytic solution can be divided among a set of computing devices, such as by using a map and reduce technique. Each computing device solves the equation for different respective subset of users.

Efficient modeling system for user recommendation using matrix factorization

A proximity matching system may use broadcast wireless identifiers transmitted by users' devices to match users with other nearby users. The identifiers may be collected by a plurality of agents, then the identifiers may be matched with pre-defined profiles to generate physically proximate users by a remote service. The group of proximate users may be provided to various applications and consumed with summarized properties or individual properties, depending on the approved privacy settings as selected by the users. In some embodiments, the broadcast wireless identifiers may be personal area network identifiers, local area network identifiers, cellular network identifiers, or other broadcast identifier. In some embodiments, the agents may not establish a peer to peer or other connection with the broadcasting device. The agents may be fixed or mobile agents, and the proximity of users may be generated through links between nearby agents in a meshed fashion.

Wireless identifiers for proximity applications

Messages to a focal user are organized by relevance of the message originators. A visual representation of the messages includes a focal user representation (textual or graphic) and multiple contact representations (textual or graphic). The contact representations are displayed at respective relevance distances from the focal user representation. Text regions present the contents of messages from the source contacts, e.g., using graphic novel-style word balloons. The contact representations can be positioned on screen in maps, radar format, or other configurations. Users can filter contacts according to relevance, and can filter messages by pertinence.

Relevance oriented graphical representation of discussion messages

Large-scale event processing systems are often designed to perform data mining operations by storing a large set of events in a massive database, applying complex queries to the records of the events, and generating reports and notifications. However, because such queries are performed on very large data sets, the processing of the queries often introduces a significant delay between the occurrence of the events and the reporting or notification thereof. Instead, a large-scale event processing system may be devised as a large state machine organized according to an evaluation plan, comprising a graph of event processors that, in realtime, evaluate each event in an event stream to update an internal state of the event processor, and to perform responses when response conditions are met. The continuous monitoring and evaluation of the stream of events may therefore enable the event processing system to provide realtime responses and notifications of complex queries.

Large-scale event evaluation using realtime processors

Information is automatically located which is relevant to source content that a user is viewing on a user interface without requiring the user to perform an additional search or navigate links of the source content. The source content can be, e.g., a web page or a document from a word processing or email application. The relevant information can include images, videos, web pages, maps or other location-based information, people-based information and special services which aggregate different types of information. Related content is located by analyzing textual content, user behavior and connectivity relative to the source. The related content is scored for similarity to the source. Content which is sufficiently similar but not too similar is selected. Similar related content is grouped to select representative results. The selected content is filtering in multiple stages based on attribute priorities to avoid unnecessary processing of content which is filtered out an early stage.

Online relevance engine

Growing community of inventors

Tel Aviv, Israel

Daniel Sitton