Microsoft Technology Licensing, LLC

Qi Guo

Xianren Wu

Shan Zhou

Bo Hu

Erik Eugene Buchanan

Lei Ni

Daniel Sairom Krishnan Hewlett

Dan Liu

Ryen William White

Krishnaram Kenthapadi

Gungor Polatkan

Fernando David Diaz

Rohan Ramanath

Anish Ramdas Nair

Sahin Cem Geyik

Cagri Ozcaglar

Lester Gilbert Cottle, Iii

Yi Wu

Wei Guang Lu

Baolei Li

Xiaoyi Zhang

Xuhong Zhang

Xuebin Yan

Mingzhou Zhou

Anthony Hsu

Wenxiang Chen

Wensheng Sun

Yu Gong

Haitong Tian

Siyao Sun

Pei-Lun Liao

Keqiu Hu

Chenya Zhang

Patrick Cheung

A device includes a memory that stores a prefetching model. A control module receives a content page including one or more links each associated with selectable content and collects data associated with the content page. The collected data includes at least one of first data indicative of respective relationships between each of the links and a viewport of the device and second data indicative of characteristics of the viewport. The control module further assigns, using the prefetching model, respective scores to each of the links based on the collected data, and selectively generates, based on the assigned scores, a request to prefetch the selectable content associated with at least one of the links.

Prefetching using dynamic user model to reduce latency

An indication of a plurality of different entities in a social networking service is received, including at least two entities having a different entity type. A plurality of user profiles in the social networking service is accessed. A first machine-learned model is used to learn embeddings for the plurality of different entities in a d-dimensional space. A second machine-learned model is used to learn an embedding for each of one or more query terms that are not contained in the indication of the plurality of different entities in the social networking service, using the embeddings for the plurality of different entities learned using the first machine-learned model, the second-machine learned model being a deep structured semantic model (DSSM). A similarity score between a query term and an entity is calculated by computing distance between the embedding for the query term and the embedding for the entity in the d-dimensional space.

Joint representation learning of standardized entities and queries

In an example embodiment, a platform is provided that utilizes information available to a computer system to feed a neural network. The neural network is trained to determine both the probability that a searcher would select a given potential search result if it was presented to him or her and the probability that a subject of the potential search result would respond to a communication from the searcher. These probabilities are combined to produce a single score that can be used to determine whether to present the searcher with the potential search result and, if so, how high to rank the potential search result among other search results. During the training process, a rescaling transformation for each input feature is learned and applied to the values for the input features.

Rescaling layer in neural network

In an example embodiment, a platform is provided that utilizes information available to a computer system to feed a neural network. The neural network is trained to determine both the probability that a searcher would select a given potential search result if it was presented to him or her and the probability that a subject of the potential search result would respond to a communication from the searcher. These probabilities are essentially combined to produce a single score that can be used to determine whether to present the searcher with the potential search result and, if so, how high to rank the potential search result among other search results. In a further example embodiment, embeddings used for the input features are modified during training to maximize an objective.

Embedding layer in neural network for ranking candidates

In an example embodiment, position bias and other types of bias may be compensated for by using two-phase training of a machine-learned model. In a first phase, the machine-learned model is trained using non-randomized training data. Since certain types of machine-learned models, such as those involving deep learning (e.g., neural networks) require a lot of training data, this allows the bulk of the training to be devoted to training using non-randomized training data. However, since this non-randomized training data may be biased, a second training phase is then used to revise the machine-learned model based on randomized training data to remove the bias from the machine-learned model. Since this randomized training data may be less plentiful, this allows the deep learning machine-learned model to be trained to operate in an unbiased manner without the need to generate additional randomized training data.

Two-stage training with non-randomized and randomized data

In an example embodiment, position bias is addressed by introducing an inverse propensity weight into a loss function used to train a machine-learned model. This inverse propensity weight essentially increases the weight of candidates in the training data that were presented lower in a list of candidates. This achieves the benefit of counteracting the position bias and increases the effectiveness of the machine-learned model in generating scores for future candidates. In a further example embodiment, a function is generated for the inverse propensity weight based on responses to contact requests from recruiters. In other words, while the machine learned-model may factor in both the likelihood that a recruiter will want to contact a candidate and the likelihood that a candidate will respond to such a contact, the function generated for the inverse propensity weight will be based only on training data where the candidate actually responded to a contact.

Position debiasing using inverse propensity weight in machine-learned model

Techniques for implementing a learning semantic representations of sparse entities using unsupervised embeddings are disclosed herein. In some embodiments, a computer system accesses corresponding profile data of users indicating at least one entity of a first facet type associated with the user, and generating a graph data structure comprising nodes and edges based on the accessed profile data, with each node corresponding to a different entity indicated by the accessed profile data, and each edge directly connecting a different pair of nodes and indicating a number of users whose profile data indicates both entities of the pair of nodes. The computer system generating a corresponding embedding vector for the entities based on the graph data structure using an unsupervised machine learning algorithm.

Unsupervised learning of entity representations using graphs

In an example, an indication of a plurality of different entities in a social networking service is received, including at least two entities having a different entity type. Then a plurality of user profiles in the social networking service are accessed. A machine-learned model is then used to calculate, based on co-occurrence counts reflecting a number of user profiles in the plurality of user profiles in which corresponding nodes co-occurred, a similarity score between a first node and second node by computing distance between the first node and the second node in a d-dimensional space on which a plurality of entities are mapped, the similarity score generated using a generalized linear mixed model having a global coefficient vector applied to global function pertaining to the co-occurrence counts and a first random effects coefficient vector applied to a random effects per-country function.

Smart suggestions personalization with GLMix

In an example embodiment, two machine learned models are trained. One is trained to output a probability that a searcher having a member profile in a social networking service will select a potential search result. The other is trained to output a probability that a member corresponding to a potential search result will respond to a communication from a searcher. Features may be extracted from a query, information about the searcher, and information about the member corresponding to the potential search result and fed to the machine learned models. The outputs of the machine learned models can be combined and used to rank search results for returning to the searcher.

Factored model for search results and communications based on search results

Techniques for processing search queries are described. Consistent with some embodiments, a computer system generates a profile vector representation for each of several user profiles based on the user profile data of the user profiles, and then stores the vector representations. A subsequent query is processed to generate a query vector representation for the query. A neural network is used to generate a similarity score for each pairing of the query vector representation and a profile vector representation. Finally, some number of user profiles having the highest similarity scores are provided as search results.

Techniques for querying user profiles using neural networks

In an example, a plurality of user profiles in a social networking service are accessed. A heterogeneous graph structure having a plurality of nodes connected by edges is generated, each node corresponding to a different entity in the social networking service, each edge representing a co-occurrence of entities represented by nodes on each side of the edge in at least one of the user profiles. Weights are calculated for each edge of the heterogeneous graph structure, the weights being based on co-occurrence counts reflecting a number of user profiles in the plurality of user profiles in which corresponding nodes co-occurred. The heterogeneous graph structure is embedded into a d-dimensional space. A machine-learned model is then used to calculate a similarity score between a first node and second node by computing distance between the first node and the second node in the d&#xac;-dimensional space.

Standardized entity representation learning for smart suggestions

In an example, a deep learning network is used to calculate a similarity score between a first query in a social networking service and each of one or more suggestable entities in the social networking service. The suggestable entities are determined via a first machine learned model. The deep learning network takes as input the suggestable entities as well as a history of interactions with a graphical user interface of a social networking service by a first member of the social networking service, a history of queries performed via the graphical user interface by the first member, and the first query itself.

Personalized deep models for smart suggestions ranking

A system and method includes receiving a search query and obtaining, from a database, member data of a member. For each of a plurality of nonlinear models, the nonlinear model is traversed based on a comparison of characteristics to conditions to obtain a score, wherein, among the nonlinear models, at least one characteristic is an inferred characteristic based on at least one of: activities by the member in an online networking system; and connections by the member in the online networking system. The score obtained from each of the nonlinear models is combined to obtain a combined score and a user interface to displays information related to the member based, at least in part on the combined score.

Growing community of inventors

Sunnyvale, CA, United States of America

Qi Guo