Amazon Technologies, Inc.

Microsoft Technology Licensing, LLC

Nina Mishra

Sudipto Guha

Krishnaram Kenthapadi

Shiva Prasad Kasiviswanathan

Yonatan Naamad

Abhimanyu Das

Daniel Blick

Okke Joost Schrijvers

Sreenivas Gollapudi

Alan D Halverson

Alexandros Ntoulas

Umar Syed

Aleksandrs Slivkins

Aleksandra Korolova

Dhivya Eswaran

Lauren Anne Moos

Joshua M Tokle

Morteza Monemizadeh

Tal Wagner

James Alexander Cook

Srikanth Jagabathula

Random cut trees are generated with respective to respective samples of a baseline set of data records of a data set for which outlier detection is to be performed. To construct a particular random cut tree, an iterative splitting technique is used, in which the attribute along which a given set of data records is split is selected based on its value range. With respect to a newly-received data record of the stream, an outlier score is determined based at least partly on a potential insertion location of a node representing the data record in a particular random cut tree, without necessarily modifying the random cut tree.

Outlier detection for streaming data

Computer systems and associated methods are disclosed to implement the non-interactive join of privacy-preserving dataset sketches. In some embodiments, an entity can publish a one-time sketch of their dataset that would enable another entity to join their data without exposing private information. The sketch can map, using a hash function, the identities associated with a first value of the dataset to a data structure, in some embodiments. A same or different entity can join the first sketch with a privacy-preserving second sketch of a second dataset that includes added noise, and can determine an estimate of a number of identities that correspond with specific values of the first and second datasets from the joined dataset. The sketch can be published just one time, and therefore does not require separate new private computations with privacy budgeting for each additional party when a join is desired, in some embodiments.

Privacy-preserving dataset sketches that can be joined non-interactively

Described are systems and methods for determining causal connections between various metrics collected by wearable devices and using those causal connections to provide causal insights to other users. For example, some users may elect to perform one or more self-experiments to explore the impact certain changes in their behavior may have on metrics measured by the user's wearable device. Causal connections determined from those experiments may be used to provide causal insights relating to those metrics to other users who have not performed the same or similar experiments.

Determining causal insights

Techniques for hotspot detection in a dataset are described. A hotspot being a region (or a collection of points) where the value of a function of given any region in the space measures the concentration of points in that region is significantly higher than its other regions of the dataspace. As such, a region that has a denser concentration of points than other regions of the dataspace may be considered a hotspot. In some implementations, hotspot detection includes finding two or more regions to evaluate for high-density in the dataset, a high-density region indicating a potential hotspot and extending a size of the manipulated found two or more regions to determine borders for these regions.

Systems, methods, and apparatus for hotspot detection

Examples of techniques for anomaly detection with feedback are described. An instance includes a technique is receiving a plurality of unlabeled data points from an input stream; performing anomaly detection on a point of the unlabeled data points using an anomaly detection engine; pre-processing the unlabeled data point that was subjected to anomaly detection; classifying the pre-processed unlabeled data point; determining the anomaly detection was not proper based on a comparison of a result of the anomaly detection and a result of the classifying of the pre-processed unlabeled data point; and in response to determining the anomaly detection was not proper, providing feedback to the anomaly detection engine to change at least one emphasis used in anomaly detection.

Anomaly detection with feedback

Techniques for detecting anomalies in streaming graph data are described. For example, an embedding technique of generating a multi-dimensional vector of summations of each weighted edge found in both a random source bounding proper subset and a random destination bounding proper subset associated with a dimension of the epoch graph is detailed. Anomaly detection is performed on the generated multi-dimensional vectors.

Anomaly detection in streaming graphs

At an artificial intelligence system, a random cut tree corresponding to a sample of a multi-dimensional data set is traversed to determine a tree-specific vector indicating respective contributions of individual dimensions to an anomaly score of a particular data point. Level-specific vectors of per-dimension contributions obtained using bounding-box analyses at each level during the traversal are aggregated to obtain the tree-specific vector. An overall anomaly score contribution for at least one dimension is obtained using respective tree-specific vectors generated from one or more random cut trees, and an indication of the overall anomaly score contribution is provided.

Artificial intelligence system providing dimension-level anomaly score attributions for streaming data

Systems, methods, and devices for performing interactive question answering using data source credibility and conversation entropy are disclosed. A speech-controlled device captures audio including a spoken question, and sends audio data corresponding thereto to a server(s). The server(s) performs speech processing on the audio data, and determines various stored data that can be used to determine an answer to the question. The server(s) determines which stored data to use based on the credibility of the source from which the stored data was received. The server(s) may also determine a number of user interactions needed to obtain data in order to fully answer the question and may select a question for a dialog soliciting further data based on the number of user interactions.

Question answering system

Messages are collected and processed to determine topic identifiers that correspond to discussion groups. Queries are received and multiple discussion groups that are relevant to the query are determined based on the messages that are associated with the discussion groups and the topic identifiers associated with the discussion groups. The relevant discussion groups are ranked using a group preference model that simulates the behavior of a hypothetical seeker that considers discussion groups by selecting a message author who is an authority in a particular group, and exploring the discussion groups that are preferred by the selected author. The behavior of the seeker is simulated using a stationary Markov process and is used to generate a probability distribution that is used to rank the relevant discussion groups. The ranked relevant discussion groups are provided in response to the query.

Ranking relevant discussion groups

A set of queries, such as a search log, is divided into commercial queries and non-commercial queries. A first set of query communities is determined from the non-commercial queries and a second set is determined from the commercial queries. The query communities are correlated based on the users who submitted the queries and instances where a query from the first set of query communities was followed by a query from the second set to generate a mapping between the first set of query communities and the second set. Later, a non-commercial query is received from a user, and the mapping is used to predict one or more commercial queries that the user is likely to submit in the future based on the non-commercial query. One or more of the commercial queries are presented to the user according to the mapping with search results responsive to the non-commercial query.

Recommending queries according to mapping of query communities

Described is releasing output data representing a search log, in which the data is suitable for most data mining/analysis applications, but is safe to publish by preserving user privacy. The search log is processed such that a query is only included if a sufficient count of that query is present; noise may be added. User contributions that are considered may be limited to a maximum number of queries. The output may indicate how often (possibly plus noise) that each query appeared. Other output may comprise a query-action graph, a query-inaction graph and/or a query-reformulation graph, with nodes representing queries and nodes representing actions, inactions or reformulations (e.g., clicked URLs, skipped URLs, or selected related queries), and edges between nodes representing action, skip or selection counts (possibly plus noise). The output may correspond to the top results/related queries returned from a search.

Synopsis of a search log that respects user privacy

Techniques and systems are disclosed for returning temporally-aware results from an Internet-based search query. To determine if a query is temporally-based one or more query features are collected and input into a trained classifier, yielding a temporal classification for the query. Further, if a query is classified as temporal, the query results are shifted by determining an alternate set of results for the query, and returning one or more alternate results to one or more users. Based on user interactions with the one or more alternate results, the classifier can be updated, for example, by changing the query to a non-temporal query if the user interactions identify it as such.

Growing community of inventors

Pleasanton, CA, United States of America

Nina Mishra