Salesforce Com, Inc.

Salesforce, Inc.

Other

Arun Kumar Jagota

Matthew Daniel Fuchs

Dmytro Kudriavtsev

Pawan Nachnani

Stanislav Georgiev

Parth Vijay Vaishnav

Dai Duong Doan

Danil Dvinov

Chenghung Ker

Ajitesh Jain

Kaushal Bansal

Fang Wong

Lei Ming

Rakesh Ganapathi Karanth

Cem Gurkok

Amrita Dasgupta

Piranavan Selvanandan

Govardana Sachithanandam Ramachandran

Sancho Savio Pinto

Saurin G Shah

Rahul Mathias Madan

Shravani Madhavaram

Gregory Haardt

Susan S Levine

David Hacker

Navin Kumar Ramineni

George Vitchev

Hawro Mustafa

Marc Joseph Delurgio

Venkata Murali Tejomurtula

Kevin Han

Amit Vijayant

System determines, for first value at first time in time series, first estimate based on value and velocity for time series, first lag, and first time. System determines, for first value, second estimate based on value and velocity for time series, second lag, and first time. System determines first weight based on difference between second value, at second time in time series, and first estimate and second weight based on difference between second value and second estimate. System determines, for second value, first forecast based on value and velocity for time series, first lag, and second time. System determines, for second value, second forecast based on value and estimated velocity for time series, second lag, and second time. System determines, for second value, combined forecast based on first forecast weighed by first weight and second forecast weighed by second weight. If combined forecast satisfies threshold, system outputs alert.

Multi-scale exponential-smoothing forecaster for time series data

An online system performs predictions for real-time tasks and near real-time tasks based on available network bandwidth. A client device receives a regression based machine learning model. Responsive to receiving a task, the client device determines an available network bandwidth for the client device. If the available network bandwidth is below a threshold, the client device uses the regression based machine learning model to perform the task. If the client device determines that the network bandwidth is above the threshold, the client device extracts features of the task, serializes the extracted features, and transmits the serialized features to an online system, causing the online system to use a different machine learning model to perform the task based on the serialized features.

Real-time predictions based on machine learning models

A system tokenizes raw values and corresponding standardized values into raw token sequences and corresponding standardized token sequences. A machine-learning model learns standardization from token insertions and token substitutions that modify the raw token sequences to match the corresponding standardized token sequences. The system tokenizes an input value into an input token sequence. The machine-learning model determines a probability of inserting an insertion token after an insertion markable token in the input token sequence. If the probability of inserting the insertion token satisfies a threshold, the system inserts the insertion token after the insertion markable token in the input token sequence. The machine-learning model determines a probability of substituting a substitution token for a substitutable token in the input token sequence. If the probability of substituting the substitution token satisfies another threshold, the system substitutes the substitution token for the substitutable token in the input token sequence.

Machine-learnt field-specific standardization

An online system performs predictions for real-time tasks and near real-time tasks that need to be performed by a deadline. A client device receives a real-time machine learning based model associated with a measure of accuracy. If the client device determines that a task can be performed using predictions having less than the specified measure of accuracy, the client device uses the real-time machine learning based model. If the client device determines that a higher level of accuracy of results is required, the client device sends a request to an online system. The online system provides a prediction along with a string representing a rationale for the prediction.

Determining rationale for a prediction of a machine learning based model

Adaptive field-level matching is described. A system identifies first elements in a field of a prospective record for a database, and second elements in the field of a candidate record, in the database, for matching the prospective record. The system identifies features corresponding to any of the first elements that are identical to any of the second elements, any of the first elements that are absent from the second elements, and any of the second elements that are absent from the first elements. A machine-learning model uses the features to determine a field match score for the candidate record's field. Another machine-learning model weighs the field match score and weighs another field match score for another field of the candidate record to determine a record match score for the candidate record. If the record match score satisfies a threshold, the system identifies the candidate record as matching the prospective record.

Adaptive field-level matching

A system receives a record which includes a string and separates the string into a number of tokens, including a token and another token. The system identifies a pattern that includes an entity, another entity, and a number of entities that equals the number of tokens, and another pattern that includes the same number of entities as the number of tokens. The system determines a combined probability that combines a probability based on the number of entries in the entity's dictionary which stores the token, and another probability based on a number of character types in the other entity that match characters in the other token. If the combined probability associated with the pattern is greater than another combined probability associated with the other pattern, the system matches the record to a system record based on recognizing the token as the entity and the other token as the other entity.

Adaptive recognition of entities

System determines first and second scores based on applying function to features of first and second values in fields in first and second records, respectively. System determines first priority based on first score and second priority based on second score for displaying first and second values in fields in first profile. System revises, based on feedback associated with first value and second value, parameter associated with function and determines third score based on applying function, associated with revised parameter, to feature of third value in field in third record. System determines fourth score based on applying function, associated with revised parameter, to feature of fourth value in field in fourth record and determines third priority, based on third score, for displaying third value in field in second profile and fourth priority, based on fourth score, for displaying fourth value in field in second profile.

Machine learning from data steward feedback for merging records

A model is trained to create a probability distribution of counts based on counts of distinct values stored by person profiles in a field. The model is trained to create another probability distribution of counts based on other counts of other distinct values stored by the person profiles in another field. The count of distinct values stored by a person profile in the field is identified. Another count of distinct values stored by the person profile in the other field is identified. A score is determined based on a cumulative distribution function of the count under the probability distribution of counts. Another score is determined based on the cumulative distribution function of the other count under the other probability distribution of counts. If the score and the other score combine in an overall score that satisfies a threshold, a message is output about the person profile being suspected of corruption.

Discovering suspicious person profiles

A system creates graph of nodes connected by edges. Each node represents corresponding value of corresponding attribute and is associated with count of corresponding value. Each edge is associated with count of instances that values represented by corresponding connected nodes are associated with each other. The system identifies each node associated with first count as first set of keys, and deletes each node associated with first count. The system identifies each edge associated with second count as second set of keys, and deletes each edge associated with second count. The system identifies each node associated with third count as third set of keys, and deletes each node associated with third count. The system identifies each edge associated with fourth count as fourth set of keys, and deletes each edge associated with fourth count. The system uses each set of keys to search and match records.

Generating adaptive match keys

Determine first count of first records storing first value in first field, second count of second records storing second value in second field, third count of third records storing third value in third field. Determine count threshold using first, second and third counts, dispersion measure based on dispersion of values stored in second field by first records and other dispersion measure based on other dispersion of values stored in third field by first records. Train machine-learning model to determine dispersion measure threshold based on dispersion and other dispersion measures. If first count is greater than count threshold, and dispersion measure is greater than dispersion measure threshold, create match index based on first and second fields. Receive prospective record storing first value in first field, second value in second field. Use match index to identify record storing first value in first field, second value in second field as matching prospective record.

Adaptive match indexes

A method and system of matching field values of a field type are described. Blurring operations are applied on a first and second values to obtain blurred values. A first maximum score is determined from first scores for blurred values, where each one of the first scores is indicative of a confidence that a match of the first and the second values occurs with knowledge of a first blurred value. A second maximum score is determined from second scores for the blurred values, where each one of the second scores is indicative of a confidence that a non-match of the first and the second values occurs with knowledge of the first blurred value. Responsive to determining that the first maximum score is greater than the second maximum score, an indication that the first value matches the second value is output.

Method and a system for fuzzy matching of entities in a database system based on machine learning

A system creates a graph of nodes connected by edges, the nodes including: i) a first node associated with a first value and a count of the first value, and ii) a second node associated with a second value and a count of the second value, the edges including an edge that connects the first and second nodes and is associated with a count of instances of the first value being stored with the second value. The system includes each node and each associated with clique count less than clique threshold in keys sets and deletes each node and each edge associated with clique count less than clique threshold. The system identifies triplet nodes connected by triplet edges. If estimated clique count for triplet values represented by triplet nodes is less than clique threshold, the system includes triplet values in keys set and identify triplet of nodes as analyzed.

Generating adaptive match keys based on estimating counts

Search query result set count estimation is described. A system parses data set query that includes first query attribute and second query attribute. The system identifies first hierarchy of connected nodes including a first node representing a first query attribute, and a second hierarchy of other connected nodes including a second node representing a second query attribute. The system identifies a directed arc connecting first correlated node in first hierarchy to second correlated node in second hierarchy. The system identifies cross-hierarchy probabilities of correlations between values of a first attribute represented by the first correlated node and values of a second attribute represented by the second correlated node. The system outputs query result set estimated count generated from cross-hierarchy probabilities, probabilities that values of first attribute are associated with values corresponding to first node, and probabilities that values of second attribute are associated with values corresponding to second node.

Search query result set count estimation

A system determines a name probability based on a first name dataset frequency of a first name value stored by a first name field in a personal record and a last name dataset frequency of a last name value stored by a last name field in a personal record. The system determines at least one other probability based on another dataset frequency of another value stored by another field in the personal record and an additional dataset frequency of an additional value stored by an additional field in the personal record. The system determines a combined probability based on the name probability and the at least one other probability. The system increments a count of identifiable personal records for each personal record that has a corresponding combined probability that satisfies an identifiability threshold. The system outputs a message based on the count of identifiable personal records.

Growing community of inventors

Sunnyvale, CA, United States of America

Arun Kumar Jagota