Microsoft Technology Licensing, LLC

Jaewon Yang

Qi He

Xiao Yan

Baoxu Shi

Feng Guo

Fei Chen

Lin Zhu

How Jing

Dan Shacham

Uri Merhav

Yanen Li

Jacob Bollinger

Mikhail Obukhov

Shan Li

Yiming Wang

Wenjia Ma

Abhishek Gupta

Bee-Chung Chen

Kevin Chang

Xianren Wu

Anish Ramdas Nair

Cagri Ozcaglar

Boyi Chen

Mustafa Emre Kazdagli

Peide Zhong

Liang Tang

Liwei Wu

Joey Bai

Zhuliu Li

Sufeng Niu

Minji Yoon

Feishe Chen

Baohua Huang

Weizhi Meng

Wei Kang

Angie Zhu

Farzard Eskafi

Shiqi Wu

Techniques for implementing a performance-adaptive sampling strategy towards fast and accurate graph neural networks are provided. In one technique, a graph that comprises multiple nodes and edges connecting the nodes is stored. An embedding for each node is initialized, as well as a sampling policy for sampling neighbors of nodes. One or more machine learning techniques are used to train a graph neural network and learn embeddings for the nodes. Using the one or more machine learning techniques comprises, for each node: (1) selecting, based on the sampling policy, a set of neighbors of the node; (2) based on the graph neural network and embeddings for the node and the set of neighbors, computing a performance loss; and (3) based on a gradient of the performance loss, modifying the sampling policy.

Performance-adaptive sampling strategy towards fast and accurate graph neural networks

Described herein are techniques for using a graph neural network to encode online job postings as embeddings. First, an input graph is defined by processing one or more rules to discover edges that connect nodes in an input graph, where the nodes of the input graph represent job postings or standardized job attributes, and the edges are determined based on analyzing a log of user activity directed to online job postings. Next, a graph neural network (GNN) is trained based on an edge prediction task. Finally, once trained, the GNN is used to derive node embeddings for the nodes (e.g., job postings) of the input graph, and in some instances, new online job postings not represented in the original input graph.

Encoding a job posting as an embedding using a graph neural network

Techniques for mining training data for use in training a dependency model are disclosed herein. In some embodiments, a computer-implemented method comprises: obtaining training data comprising a plurality of reference skill pairs, each reference skill pair comprising a corresponding first reference skill and a corresponding second reference skill, the plurality of reference skill pairs being included in the training data based on a co-occurrence of the corresponding first and second reference skills for each reference skill pair in the plurality of reference skill pairs, the co-occurrence comprising the corresponding first and second reference skills co-occurring for a same entity; and training a dependency model with a machine learning algorithm using the training data, the dependency model comprising a logistic regression model or a data gradient boosted decision tree (GBDT) model. The dependency model may then be used to identify corresponding dependency relations for a plurality of target skill pairs.

Mining training data for training dependency model

Embodiments of the disclosed technologies incorporate taxonomy information into a cross-lingual entity graph and input the taxonomy-informed cross-lingual entity graph into a graph neural network. The graph neural network computes semantic alignment scores for node pairs. The semantic alignment scores are used to determine whether a node pair represents a valid machine translation.

Machine translation of entities

In an example embodiment, a screening question-based online screening mechanism is provided to assess job applicants automatically. More specifically, job-specific questions are automatically generated and asked to applicants to assess the applicants using the answers they provide. Answers to these questions are more recent than facts contained in a user profile and thus are more reliable measures of an appropriateness of an applicant's skills for a particular job.

Machine learned model framework for screening question generation

Techniques for scoring data items using a machine-learned model are provided. In one technique, multiple skills are identifying based on a job posting. Multiple attribute values of the job posting are identified. For each identified skill, multiple probabilities are identified, each probability corresponding to a different attribute value of the identified attribute values. The probabilities are input into a machine-learned model to generate multiple scores. Multiple skills of a candidate user are identified. An affinity score between the job posting and the candidate user is generated based the scores and the skills of the candidate user.

Generating a machine-learned model for scoring skills based on feedback from job posters

Apparatuses, computer readable medium, and methods are disclosed for verifying skills of members of an online connection network. The apparatus, computer readable medium, and methods may include a method including responding to a first member of the online connection network indicating a skill possessed by the first member by selecting a skill verification user interface (UI) to present to a second member of the online connection network where the first member and the second member are connected via the online connection network. The method may further include presenting the skill verification UI to the second member, where the skill verification UI presents an indication of the first member, an indication of the skill, and a query regarding a competence level of the skill possessed by the first member. The method may further include receiving a response to the query and determining a skill validation value of the skill for the first member based on the response and a machine learning model.

Skill validation

In an example embodiment, a framework to infer a user's value for a particular attribute based upon a multi-task machine learning process with uncertainty weighting that incorporates signals from multiple contexts is provided. In an example embodiment, the framework aims to measure a level of a user attribute under a certain context. Rather than attempting to devise a universal, one-size-fits-all value for the attribute, the framework acknowledges that the user's value for that attribute can vary depending on context and factors in the context under which the user's attribute levels are measured. Multiple contexts are defined depending on different situations where users and entities such as companies and organizations need to evaluate user attribute levels. Signals for attribute levels are then collected for each context. Machine learning models are utilized to estimate attribute values for different contexts. Multi-task deep learning is used to level attributes from different contexts.

Multi-task learning framework for multi-context machine learning

Techniques are provided for using machine learning techniques to analyze textual content. In one technique, a potential item is identified within a document. An analysis of the potential item is performed at multiple levels of granularity that includes two or more of a sentence level, a segment level, or a document level. The analysis produces multiple outputs, one for each level of granularity in the multiple levels of granularity. The outputs are input into a machine-learned model to generate a score for the potential item. Based on the score, the potential item is presented on a computing device. In response to user selection of the potential item, an association between the potential item and the document is created. The association may be used later to identify a set of users to which the document (or data thereof) is to be presented.

Machine learning techniques for analyzing textual content

During operation, the system obtains a first embedding produced by an embedding model from an input string representing an entity and a hierarchy of clusters of embeddings generated by the embedding model from a set of standardized entities. Next, the system searches the hierarchy of clusters for a subset of the embeddings that are within a threshold proximity to the first embedding in a vector space. The system then calculates embedding match scores between the input string and a first subset of the standardized entities represented by the subset of the embeddings based on distances between the subset of the embeddings and the first embedding in the vector space. Finally, the system modifies, based on the embedding match scores, content outputted in response to the input string within a user interface of an online system.

Semantic matching and retrieval of standardized entities

Techniques for ranking skills using an ensemble machine learning approach are described. The outputs of two heterogenous, machine-learned models are combined to rank a set of skills that may be possessed by an end-user of an online service. Some subset of the highest-ranking skills is then presented to the end-user with a recommendation that the skills be added to the end-user's profile. The ensemble learning technique involves a concept referred to as 'boosting', in which a weaker performing model is enhanced (e.g., 'boosted') by a stronger performing model, when ranking the set of skills. Accordingly, by using a combination of models, better results are achieved than might be with either one of the individual models alone. Furthermore, the approach is scalable in ways that cannot be achieved with heuristic-based approaches.

Techniques for suggesting skills

Techniques for selecting candidates using a personalized model are disclosed herein. In some embodiments, a computer system, for each candidate of a plurality of candidates, generating a corresponding confidence score for a combination of the candidate, a particular viewer, and a particular attribute based on a scoring model, with the corresponding confidence score being configured to indicate a likelihood that the particular viewer will select the corresponding candidate as a preference with respect to the particular attribute. The computer system then selects a subset of the plurality of candidates based on the corresponding confidence scores of the candidates in the subset, and causes the subset of candidates to be displayed on a computing device of the viewer along with a prompting for the viewer to select one of the selected subset of candidates as the preference with respect to the particular attribute.

Candidate selection using personalized relevance modeling system

A news feed system provided with an on-line social network system determines that a news feed is to be constructed for a viewer. The news feed system accesses the viewer's profile and other information associated with the viewer, accesses an inventory of activities that have been identified as potentially of interest to the viewer, and calculates relevance score for each item inventory of activities using the combined predictions methodology. The activities are then arranged for presentation to the viewer via a news feed web page, using respective calculated relevance scores.

Combined predictions methodology

Techniques for improving search using generalized linear mixed models are disclosed herein. In some embodiments, a computer-implemented method comprises: receiving a search query comprising at least one search term and being associated with a user; extracting features from corresponding profiles of a plurality of candidates; for each one of the candidates, generating a corresponding score based on a generalized linear mixed model comprising a generalized linear query-based model and a random effects user-based model; selecting a subset of candidates from the plurality of candidates based on the corresponding scores; and causing the selected subset of candidates to be displayed to the user in a search results page for the search query.

Generalized linear mixed models for improving search

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.

Growing community of inventors

Sunnyvale, CA, United States of America

Jaewon Yang