Walmart Apollo, LLC

Verizon Media Inc.

Verizon Patent and Licensing Inc.

Mridul Jain

Saigopal Thota

Gajendra Alias Nishad Kamat

Xun Luan

Antriksh Akshesh Shah

Albin Kuruvilla

Andrew Feng

Peter Cnudde

Giridhar Addepalli

Navinder Pal Singh Brar

Pruthvi Raj Eranti

Jun Shi

Saurabh Agrawal

Pawan Kumar Gupta

Deepak Goyal

Santos Kumar Das

Francis Hsu

Saumya Agarwal

Mitesh Sinha

Saurabh Tewari

Sai Kiran Reddy Malikireddy

Sébastien Jean-Maurice Olivier Péhu

Sebastien Jean-Maurice Olivier Pehu

Rewati Mahendra Ovalekar

Neil Mohan Reddy Palleti

Rijul Magu

Ashraful Arefeen

Pragya Jain

The present teaching relates to distributed deep machine learning on a cluster. In one example, a request is received for estimating one or more parameters associated with a machine learning model on a cluster including a plurality of nodes. A set of data is obtained to be used for estimating the one or more parameters. The set of data is divided into a plurality of sub-sets of data, each of which corresponds to one of the plurality of nodes. Each sub-set of data is allocated to a corresponding node for estimating values of the one or more parameters based on the sub-set of data. Estimated values of the one or more parameters obtained based on a corresponding sub-set of data allocated to the node, are received from each of the plurality of nodes. The one or more parameters of the machine learning model are estimated based on the estimated values of the one or more parameters generated by at least some of the plurality of nodes.

Method and system for distributed deep machine learning

A distributed system including multiple processors associated with non-transitory computer-readable media storing computing instructions. The computing instructions, when collectively executed on the multiple processors, cause the multiple processors collectively to perform certain acts. The acts can include executing multiple iterations until a stopping condition is satisfied, by, for each of the multiple iterations: (i) processing a set of input nodes at the multiple processors using a set of criteria to generate first data at the multiple processors, wherein the set of input nodes is different at each of the multiple iterations; (ii) determining a list of output nodes using adjacency rows of an adjacency list at different ones of the multiple processors, such that each output node of the list of output nodes is one hop from a respective input node of the set of the input nodes; and (iii) updating the set of the input nodes for a subsequent iteration of the multiple iterations based on the list of output nodes when the stopping condition is not satisfied. The acts also can include outputting second data based at least in part on the first data. Other embodiments are disclosed.

Traversing an adjacency list on distributed processors

A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform ingesting streaming events for processing by multiple models; mapping each of the streaming events to a model of the multiple models; storing each of the streaming events in a respective queue in a respective sequence store, such that a respective one of the multiple models retrieves (i) a respective one of the streaming events in the respective sequence store associated with the respective one of the multiple models and (ii) a respective key corresponding to the respective one of the streaming events from a leaf store, to asynchronously perform the respective machine-learning inferencing, wherein the multiple models run independently and in parallel on multi-tenant threads. Other embodiments are disclosed.

Asynchronous multi-tenant model inferencing on streaming databases

This application relates to apparatus and methods for automatically associating customer data to a corresponding customer. A computing device may receive linking data identifying a plurality of links, where each like associates at least two nodes that each represent customer data. The computing device may partition the linking data into multiple partitions, and cause a union find algorithm to be executed for each partition in parallel to associate each node with a parent ID. The computing device may iteratively execute a global shuffle algorithm to place all same nodes in a same partition, and may assign a same parent ID to the same nodes. The computing device may iteratively execute a path compression algorithm across all partitions to generate a graph output that associates all child nodes of a same parent node with the same parent ID.

Method and apparatus for determining associated data linkages

A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform: generating linkage scores between nodes at least based on a machine learning model; creating links between the nodes to form connected components based on the linkage scores exceeding a predetermined threshold; generating an actual matching linkage set of the nodes linked in the connected components by using a relaxed blocking criteria; and generating a quality score for the connected components. Other embodiments are disclosed.

Automatically measuring quality scores for connected components using machine learning models

A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform: bundling multiple registered queries of a dataset using a scheduling technique, wherein the dataset is homogenous in schema; running a single table scan of the dataset to process the multiple registered queries of the dataset in parallel; and generating a respective output responsive to each of the multiple registered queries. Other embodiments are disclosed.

Optimizing scans using query planning on batch data

This application relates to apparatus and methods for determining confidence levels in associated data using machine learning algorithms. In some examples, a computing device may generate training graph data where each training graph connects at least two nodes by an edge, and each node represents data. The computing device may train a machine learning algorithm based on the generated training data. The computing device may then receive linked data, which associates at least two nodes, each representing data, with each other. The computing device may generate graph data based on the linking data, to provide to the machine learning algorithm as input. The computing device may then execute the machine learning algorithm on the generated graph data to generate values for each of its edges. The values may identify, for each edge, a confidence level in the connection between the two nodes for that edge.

Method and apparatus for determining data linkage confidence levels

A distributed system including multiple processing nodes. The distributed system can perform certain acts. The acts can include receiving a set of input nodes and a set of criteria. The acts can include obtaining an adjacency list representing a large connected component. The large connected component can include nodes, edges, and edge metadata. A quantity of the nodes of the large connected component can exceed 1 billion. The adjacency list can be distributed across the multiple processing nodes. The nodes of the large connected component can include the input nodes. The acts also can include performing one or more iterations of traversing the large connected component until a stopping condition is satisfied. Each iteration can include processing a set of input nodes at the multiple processing nodes using the set of criteria to generate first data at the multiple processing nodes, determining a set of output nodes such that each output node of the set of output nodes is one hop from a respective input node of the set of input nodes, consolidating the first data from the multiple processing nodes to a first processing node of the multiple processing nodes, processing the first data at the first processing node; and assigning the set of input nodes for a subsequent iteration of the one or more iterations based on the set of output nodes when the stopping condition is not satisfied. The acts further can include outputting second data based on the first data received and processed at the first processing node during the one or more iterations. Other embodiments are disclosed.

Traversing a large connected component on a distributed file-based data structure

Systems and methods of real-time modeling pipeline inferencing are disclosed. At least one model configured to calculate at least one metric from one or more features is deployed. A model inferencing pipeline configured to extract the one or more features from a customer-specific data pipeline is implemented for the at least one mode. The model inferencing pipeline is generated using a training data set extracted from a cross-customer data pipeline. The at least one metric is calculated using the one or more features extracted from the customer-specific data pipeline.

System and method for real-time modeling inference pipeline

A distributed system including multiple processing nodes. The distributed system can perform certain acts. The acts can include receiving a first conflation event identifying a first node and a second node. The first node can be part of a first set. The first set can include a sole parent node stored at a first processing node of the multiple processing nodes. The second node can be part of a second set. The second set can include a sole parent node stored at a second processing node of the multiple processing nodes. The first and second sets can be disjoint sets. The first conflation event can be received at an event-driven stream application at one of the multiple processing nodes. The acts also can include conflating the first set and the second set into a conflated set. The conflated set can include the first and second nodes. The conflated set can include a sole parent node. Data for each node of the conflated set can be stored at the sole parent node of the conflated set at a processing node of the multiple processing nodes. Other embodiments are disclosed.

Event-driven identity graph conflation

The present teaching relates to a method and system for validating labels of training data. A first group of data records associated with the training data are received, wherein each of the first group of data records includes a vector having at least one feature and a first label. For each of the first group of data records, a second label is determined based on the at least one feature in accordance with a first model. Thereafter, a loss based on the first label associated with the data record and the second label is obtained, and the data record having an incorrect first label is classified when the loss meets a pre-determined criterion. Upon classifying the data records, a sub-group of the first group of data records is generated, wherein each of the data records included in the sub-group has the incorrect first label.

Growing community of inventors

Cupertino, CA, United States of America

Mridul Jain