Microsoft Technology Licensing, LLC

Dan Shacham

Uri Merhav

Peide Zhong

Qi He

Jaewon Yang

How Jing

Angela Jiang

Craig Martell

Fan Yang

Fei Chen

Feng Guo

Qin Iris Wang

Ji Yan

Mohsen Jamali

Siyuan Zhang

Hong H Tam

Tianhao Lu

Steven Curtis McClung

Junzhe Miao

Yunpeng Xu

Zhanpeng Fang

Tao Xiong

Sebastian Alexander Csar

Shuai Wang

Described herein is a technique for mapping the raw text of a job title of an online job posting to an entity embedding, associated with an entity or entry of a title taxonomy. The raw text of the job title is first encoded to generate a multilingual word embedding in a multilingual word embedding space. Then, the vector representation of the job title, as represented in the multilingual word embedding space is translated, using a neural network, to a vector representation of the job title in the entity embedding space. Finally, a nearest neighbor search is performed to identify an entity embedding associated with an entity or entry in the title taxonomy that has a vector representation that is closest in distance to the vector output by the neural network.

Techniques for improving standardized data accuracy

In an example embodiment, a system is provided whereby a machine learning model is trained to predict a standardization for a given raw title. A neural network may be trained whose input is a raw title (such as a query string) and a list of candidate titles (either title identifications in a taxonomy, or English strings), which produces a probability that the raw title and each candidate belong to the same title. The model is able to standardize titles in any language included in the training data without first having to perform language identification or normalization of the title. Additionally, the model is able to benefit from the existence of 'loan words' (words adopted from a foreign language with little or no modification) and relations between languages.

Language agnostic machine learning model for title standardization

The disclosed embodiments provide a system that identifies duplicate entities. During operation, the system selects training data for a first machine learning model based on confidence scores representing likelihoods that pairs of entities in an online system are duplicates. Next, the system updates parameters of the first machine learning model based on features and labels in the training data. The system then identifies a first subset of additional pairs of the entities as duplicate entities based on scores generated by the first machine learning model from values of the features for the additional pairs and a first threshold associated with the scores. The system also determines a canonical entity in each of the duplicate entities based on additional features. Finally, the system updates content outputted in a user interface of the online system based on the identified first subset of the additional pairs.

Identifying duplicate entities

A system and method for inferring appropriate courses for recommendation based on member characteristics is disclosed. A social networking system receives a request for recommended courses, wherein the request is associated with a member of the social networking system. The social networking system identifies a group of members who are similar to the first member. The social networking system creates a list of recently learned skills by members of the group of members similar to the member. For a particular skill in the list of skills, the social networking system determines whether the member possesses the particular skill. In accordance with a determination that the member does not possess the particular skill, the social networking system identifies at least one course that teaches the particular skill from a list of courses. The social networking system transmits the identified course to the client device for display as a recommended course.

Inferring appropriate courses for recommendation based on member characteristics

In an example embodiment, a first DCNN is trained to output a value for a first metric by inputting a plurality of sample documents to the first DCNN, with each of the sample documents having been labeled with a value for the first metric. Then a plurality of possible transformations of a first input document are fed to the first DCNN, obtaining a value for the first metric for each of the plurality of possible transformations. A first transformation is selected from the plurality of possible transformations based on the values for the first metric for each of the plurality of possible transformations. Then a second DCNN is trained to output a transformation for a document by inputting the selected first transformation to the second DCNN. The second input document is fed to the second DCNN, obtaining a second transformation of the second input document.

Training a neural network using another neural network

In an example embodiment, a machine learning algorithm is used to train an objective prediction model to output a prediction value for an input member of a social networking service and a potential objective, based on member attribute information and action information. At prediction time, member attribute information and action information for a first user may be fed to the objective prediction model to obtain prediction values for a plurality of different potential objectives, one of which can be selected based on the prediction values. The selected objective can then be used to optimize coordinates, in a latent representation space, mapped to a plurality of different entities in a social network structure.

Modification of base distance representation using dynamic objective

Hybrid classification system and method are described. The method commences when an input detector detects a raw input string that represents a value of a category in a member profile maintained by the on-line social network. The machine learning classifier derives a standardized value corresponding to the raw input string. The trigger module provides the raw input string to the correcting filter. The correcting filter determines a corrected standardized value corresponding to the raw input string based on the raw input string and a corrective rule. The label module then identifies the member profile as associate with the corrected standardized value.

Hybrid classification system

In an example embodiment, a solution that automatically predicts an industry for a candidate company is provided. An existing industry classifier is trained using a first machine learning algorithm, the first machine learning algorithm taking as input first training data and existing industries listed in an industry taxonomy. A new industry classifier is trained using a second machine learning algorithm, the second machine learning algorithm taking as input second training data and new industries listed in an industry taxonomy. Then the candidate company is fed into the existing industry classifier, producing one or more predicted existing industries corresponding to the candidate company. The candidate company is also fed into the new industry classifier, producing one or more predicted new industries corresponding to the candidate company. One or more final predicted industries are selected from among the one or more predicted existing industries and the one or more predicted new industries.

Parent company industry classifier

In an example embodiment, for each of a plurality of different titles in a social network structure, the title is mapped into a first vector having n coordinates, while kills are mapped into a second vector having n coordinates. The first and second vectors are stored in a deep representation data structure. One or more objective functions are applied to at least one combination of two or more of the vectors in the deep representation data structure. Then, an optimization test on each of the at least one combination is performed using a corresponding objective function output for each of the at least one combination of two or more of the vectors, and, for any combination that did not pass the optimization test, one or more coordinates for the vectors in the combination are altered so that the vectors in the combination become closer together within an n-dimensional space.

Skill-based title prediction model

In an example embodiment, for each of a plurality of different entities in a social network structure, the entity is mapped into a vector having n coordinates. The vector for each of the plurality of different entities is stored in a deep representation data structure. One or more objective functions are applied to at least one combination of two or more of the vectors in the deep representation data structure. Then, an optimization test on each of the at least one combination of two or more of the vectors is performed using a corresponding objective function output for each of the at least one combination of two or more of the vectors, and, for any combination that did not pass the optimization test, one or more coordinates for the vectors in the combination are altered so that the vectors in the combination become closer together within an n-dimensional space.

Embedded deep representation of social network taxonomy

In an example embodiment, for each of one or more input documents: a first value is determined for the first metric for a first transformation of the input document by passing the first transformation to s first Deep Convolutional Neural Network (DCNN), a second transformation of the input document is determined by passing the input document to a second DCNN, the second transformation of the input document is passed to the first DCNN, obtaining a second value for the first metric for the second transformation of the input document, the first and second transformations being of the first transformation type, and a difference between the first value and the second value for the input document is determined. Then it is determined whether to change the system over from the first DCNN to the second DCNN based on the difference between the first value and the second value.

Changeover from one neural network to another neural network

In an example embodiment, a first plurality of images stored on a computing device is identified, each image having an indication that it depicts a first member of a social networking service. The first plurality of images is used as training data to a first machine learning algorithm to train a first machine learning algorithm model corresponding to the first member, the first machine learning algorithm model corresponding to the first member designed to calculate a member likelihood score for a candidate image. Then a second plurality of images stored on the computing device is obtained. Each image of the second plurality of images is fed to the first machine learning algorithm model corresponding to the first member, obtaining a member likelihood score for each of the second plurality of images. Then, based on the member likelihood scores for the second plurality of images, one or more member images are selected.

Cold start machine learning algorithm

An online social networking system extracts terms from an unstructured job title record. The system searches a job role taxonomy database with the extracted terms to identify job roles. For each job role identified, the system extracts a plurality of additional terms appearing in the unstructured job title record. For each additional term, the system maps the additional term to a standardized modifier, thereby identifying a job seniority modifier, a job specialty modifier, a job accreditation modifier, and a job status modifier for each additional term. The system creates a multi-dimensional standardized job title for the member profile or job posting by writing the job role, the job seniority modifier, the job specialty modifier, the job accreditation modifier, and the job status modifier to a standardization record in a standardization database.

Multi-dimensional job title logical models for social network members

An online social networking system receives an unstructured job title record from a profile of a member or a job posting. The system extracts a raw job title from the unstructured job title record, and extracts a first seniority level from the raw job title. The first seniority level is a seniority modifier associated with the raw job title. The system determines a second seniority level. The second seniority level is a company seniority within the company associated with the unstructured job title record. The system determines a third seniority level. The third seniority level is a seniority score for the member or the job posting. The system compares the seniority score with a second seniority score, and communicates with the member, or transmits the job posting to the member, based on the comparison of the seniority score and the second seniority score.

Deriving multi-level seniority of social network members

In an example embodiment, an optimal cropping of a digital image is determined. A machine learning algorithm is used to generate a professionalism score for the digital image, the utilizing a machine learning algorithm comprising a training mode where a plurality of sample images with labeled professionalism scores are used to train a classification function in a model that produces as professionalism score as output; and an analysis mode where the model is used to generate a professionalism score for the digital image. Then, the professionalism score is used as an input to a discrete variable optimization algorithm to determine an optimum cropped version of the digital image from a plurality of possible cropped versions of the digital image using the classification function.

Growing community of inventors

Sunnyvale, CA, United States of America

Dan Shacham