Abstract: A neural network system can select content based on user and item content embeddings in an approach that can be updated in real time on the user device without server support. Requests for content sent to the server can include an anonymous user embedding that includes data describing the user's inputs. The content that is nearest to the user embedding in a joint embedding space can be returned as suggested content.

Abstract: A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.

Abstract: Technologies for skill taxonomy management are described. Embodiments include extracting an input text from an online system and applying an unsupervised generative text machine learning model to the input text. The text generator generates a set of sentences based on a job title included in the input text. One or more skills are extracted from the set of sentences. The extracted one or more skills correspond to one or more skills in a skill taxonomy. A frequency distribution is generated over the extracted one or more skills. The one or more skills are ranked based on the frequency distribution. Based on the ranking, a subset of the extracted one or more skills is generated. The subset of the extracted one or more skills is provided to a downstream operation, process, or service of the online system.

Abstract: Systems and methods are provided for generating a user click history table and a random bucket training table, generating training data for training a user-type-affinity machine learning model by combining the user click history table and the random bucket training table, and training the user-type-affinity machine learning model with the generated training data. The systems and methods further provide for generating a user click prediction table and generating user-type-affinity prediction values for each of the plurality of users by inputting the user click prediction table into the user-type-affinity machine learning model.

Abstract: A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.

Abstract: 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.

Abstract: Techniques for using deep reinforcement learning for training a recommendation model for an online service are disclosed herein. In some embodiments, a computer-implemented method comprises training a recommendation model using deep reinforcement learning and a Markov decision process, where the Markov decision process has a state space including state embeddings of a plurality of reference users, an action space including action embeddings of the plurality of reference users, and a reward function. The reward function may be configured to issue a first reward based on current impression interaction data and a second reward based on a measurement of engagement of the reference user with the online service.

Abstract: Methods, systems, and computer programs are provided for presenting career information based on career transitions of members. One method comprises generating, using a machine-learning (ML) model, an embedding for a current job position of a member of an online service. The model is obtained by training a ML program with training data for job transitions of members. Further, the method includes generating, by the ML model, embeddings for career transitions, of members of the online service, that occurred within a predetermined time period. For each career transition, a similarity value is calculated between the embedding of the career transition and the embedding for the current job position. Further, the method includes operations for ranking the career transitions based on the similarity values, generating a career insight for the member based on the ranked career transitions, and causing presentation of the career insight on a user interface.

Abstract: A neural network system can select content based on user and item content embeddings in an approach that can be updated in real time on the user device without server support. Requests for content sent to the server can include an anonymous user embedding that includes data describing the user's inputs. The content that is nearest to the user embedding in a joint embedding space can be returned as suggested content.

Abstract: 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.

Abstract: Techniques for predicting specialty data for a knowledge base using a machine learning model are disclosed herein. In some embodiments, a computer-implemented method comprises: for each skill in a plurality of skills, computing a skill-to-specialty distribution for specialties using a first machine learning model; for each skill in the plurality of skills, computing a user-to-skill distribution for the plurality of skills based on feature data of a first user of an online service using a second machine learning model; computing a user-to-specialty distribution for the plurality of specialties based on the skill-to-specialty distribution and the user-to-skill distribution, the user-to-specialty distribution comprising a corresponding user-to-specialty probability value for each specialty in the plurality of specialties given the first user; and using the user-to-specialty distribution in an application of the online service.

Abstract: A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.

Abstract: A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.

Abstract: 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.

Abstract: A neural network system can select content based on user and item content embeddings in an approach that can be updated in real time on the user device without server support. Requests for content sent to the server can include an anonymous user embedding that includes data describing the user's inputs. The content that is nearest to the user embedding in a joint embedding space can be returned as suggested content.

Abstract: 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.

Abstract: An on-line social network system is configured to generate a news feed for a member by processing updates originating from different sources using different first pass ranker models. The first pass ranker models generate respective sets of raw scores, which are calibrated based on a consistent scale of feed engagement metrics of interest, such as a click through rate. The calibrated scores are then used as training data to train a second pass ranker and/or as input into the second pass ranker at the time when the second pass ranker is to generate respective ranks for items in an inventory of updates identified as potentially of interest to a focus member and to select a subset of items from the inventory based on the generated respective ranks.

Abstract: Multimodal data sets of a given entity (e.g., a user) can be processed using a plurality of different machine learning schemes, such as a recurrent neural network and a fully connected neural network. Representations generated by the networks can be combined in an additive layer and further in a multiplicative layer that emphasizes informative modalities and tolerates less informative modalities.

Abstract: An on-line social network system is configured to generate a news feed for a member by processing updates originating from different sources using different first pass ranker models. The first pass ranker models generate respective sets of raw scores, which are calibrated based on a consistent scale of feed engagement metrics of interest, such as a click through rate. The calibrated scores are then used as training data to train a second pass ranker and/or as input into the second pass ranker at the time when the second pass ranker is to generate respective ranks for items in an inventory of updates identified as potentially of interest to a focus member and to select a subset of items from the inventory based on the generated respective ranks.

Abstract: Systems and methods are disclosed that recommend one or more electronic presentations to a user based on one or more factors. These factors may include contextual information, behavioral information, profile information, or combinations of the foregoing. Contextual information may include content and/or features extracted from a given electronic presentation. Behavioral information may include user behavioral data, such as the number of times a user has viewed a presentation, the amount of the presentation viewed by the user, presentations previously viewed by the user, and other such behavioral data. Profile information may include user professional profile information, such as skills the user has identified as possessing, employment history information, and other such user professional profile information.