Microsoft Technology Licensing, LLC

Alibaba Damo (Hangzhou) Technology Co., Ltd.

Bolin Ding

Chi Wang

Surajit J Chaudhuri

Danyel Aharon Fisher

Joshua Stanley Allen

Robyn Dominik Moritz

Vivek Ravindranath Narasayya

Sergey M Yekhanin

Janardhan Dattatreya Kulkarni

Harsha Prasad Nori

Silu Huang

Josh D Benaloh

Jacob Rubin Lorch

Srikanth Kandula

Kaushik Chakrabarti

Jingren Zhou

Sudipto Das

Manoj A Syamala

Melissa Erin Chase

Srinath T V Setty

Paul Luo Li

Olga Ohrimenko

Gaoxiang Xu

Robert Grandl

Alexander Norton Meade

Jingjing Wang

Anil Atmanand Shanbhag

Aleksandar Vitorović

Matthaios Olma

Azade Nazi

Zitao Li

An item recommendation method, a device and a learning system for protecting user privacy are provided. The method includes: obtaining a locally collected scoring matrix, the scoring matrix being used to describe a scoring situation of multiple users on multiple items; locally training an item expression matrix and a user expression matrix corresponding to the scoring matrix, and sending a first user expression matrix obtained by local training to a server, so that the server aggregates and processes user expression matrices independently uploaded by the multiple data devices, and sends an aggregated second user expression matrix to the multiple data devices; and predicting degree of preferences of the multiple users for the multiple items according to the item expression matrix obtained by the local training and the second user expression matrix, so that item recommendations are performed for the multiple users according to the degrees of preference.

Item recommendation method and device for protecting user privacy and learning system

In automated machine learning, an approximate best configuration can be selected among multiple candidate machine-learning configurations by progressively sampling training and test datasets for the iterative training and testing of the configurations while progressively pruning the set of candidate configurations based on associated estimated confidence intervals for their respective performance.

Efficient configuration selection for automated machine learning

This document relates to hardware protection of differential privacy techniques. One example obtains multiple instances of encrypted telemetry data within a secure enclave and processes the encrypted telemetry data to obtain multiple instances of unencrypted telemetry data. The example also processes, within the secure enclave, the multiple instances of unencrypted telemetry data to obtain a perturbed aggregate. The example also releases the perturbed aggregate from the secure enclave.

Hardware protection for differential privacy

Methods, systems, and computer programs are presented for obtaining histogram data from a database utilizing an interface with histogram-related options. One method includes an operation for providing, by a server, an application programming interface (API), to access the database, which includes a histogram request, to obtain histogram data from the database, with histogram options. The server receives, from a client device, a first histogram request including histogram options. Additionally, the method includes an operation for identifying bins for the histogram based on the one or more histogram options. For each bin, the server accesses the database to obtain data for each bin. The server returns, to the client device, the histogram data for the histogram as a table containing bin values for all the bins, where the client device is configured to present the histogram to a user based on the histogram data.

Database query for histograms

Methods, systems, apparatuses, and computer-readable storage medium are described herein for remotely analyzing testing results based on LDP-based data obtained from client devices in order to determine an effect of a software application with respect to its features and/or the population in which the application is tested. The analysis is based on a series of statistical computations for conducting hypothesis tests to compare population means, while ensuring LDP for each user. For example, an LDP scheme is used on the client-side that privatizes a measured value corresponding to a usage of a resource of the client. A data collector receives the privatized data from two sets of populations. Each population's clients have a software application that may differ in terms of features or user group. The privatized data received from each population is analyzed to determine an effect of the difference between the software applications of the different populations.

Remote testing analysis for software optimization based on client-side local differential privacy-based data

Generally discussed herein are devices, systems, and methods for database management. A method may include determining a first hyperloglog (HLL) sketch of a first column of data, determining a second HLL sketch of a second column of data, estimating an inclusion coefficient based on the first and second HLL sketches, and performing operations on the first column of data or the second column of data in response to determining the inclusion coefficient is greater than, or equal to, a specified threshold.

Database management using hyperloglog sketches

A facility for reporting on original values on behalf of each of a plurality of users&#x2014;each falling within a domain partitioned into a plurality of segments of uniform size&#x2014;is provided. For each of the plurality of users, the facility (a) randomly selects an upward-rounding window size that is smaller than the segment size; (b) for each source value in a domain, randomly determines a mapping of each segment of the domain to a segment identifier value; (c) determines an original value for the user; (d) adds the upward-rounding window size to the user's original value to obtain a window-augmented original value; (e) identifies a segment containing the window-augmented original value; (f) identifies a segment identifier value mapped-to from the identified segment using the determined mapping; and (g) transmits a reporting communication on the user's behalf reporting a value based on the identified segment identifier value.

Collection of sensitive data&#x2014;such as software usage data or other telemetry data&#x2014;over repeated collection cycles in satisfaction of privacy guarantees

Provided are methods and systems for generating physical database design tuning recommendations. Given a workload, the system analyzes the workload to identify and recommend a set of rowstore and columnstore indexes optimal for the performance of the workload. The system is designed to estimate the size of the columnstore index (at the granularity of each column) without actually building the index, estimate the improvement in query performance that each columnstore index would result in when built, and automatically derive the workload used for the physical design tuning task by analyzing stored query execution history data. This automatic workload derivation is orthogonal to columnstores and can be used even when columnstore indexes are not being used.

Workload-driven recommendations for Columnstore and Rowstore indexes in relational databases

A processing unit can determine a first subset of a data set including data records selected based on measure values thereof. The processing unit can determine an index mapping a predicate to data records associated with that predicate and approximation values of the records. The processing unit can process a query against the first subset to provide a first result and a first accuracy value, determine that the first accuracy value does not satisfy an accuracy criterion, and process the query against the index. In some examples, the processing unit can process the query against a second subset including data records satisfying a predetermined predicate. In some examples, the processing unit can receive data records and determine the first subset. Data records can include respective measure values. Data records with higher measure values can occur in the first subset more frequently than data records with lower measure values.

Aggregate-query database system and processing

Techniques for presenting data visualizations are described. Consistent with some embodiments of the present invention, a data visualization system includes an approximate query engine to derive approximate query results and present approximate data visualizations, in interactive time. The data visualization system also includes a precise query engine, thereby enabling a data analyst to selectively request that a precise query result be generated for a specific query, so that the data analyst can confirm a data observation made when viewing the approximate visualization for that specific query. The precise visualization is presented in a manner that allows the data analyst to view, and thus confirm or disprove, a data observation recorded when viewing the approximate visualization, and also compare the precise visualization with the previously generated and presented approximate visualization.

Fast approximate results and slow precise results

One or more approximations of query output in a data analytics platform are controlled. The one or more approximations are controlled by generating values of error metrics associated with placements of samplers in one or more query execution plans associated with the query, and injecting a plurality of samplers into the query execution plans, using the determined values of the error metrics, in lieu of storing samples of input to the query prior to execution of the query.

Controlling approximations of queries

A visualization system suitable for visualizing differences between data sets is provided. The system determines a first data set and a second data set along with a difference between the first and second data set. A value corresponding to each unit of display of the second data set is plotted on at least one graph. The visualization system determines, for each unit of display of the second data set, whether the first data set has a value for the unit of display. In response to the first data set having the value, the visualization system plots a difference of the value of the first data set on the at least one graph. In response to the first data set not having the value, the visualization system plots a missing structure element on the at least one graph. The at least one graph is displayed to a user.

Growing community of inventors

Redmond, WA, United States of America

Bolin Ding