Abstract: Technologies for graph-based labeling of digital content items include, in some embodiments, for digital content items received from user systems by an application system, generating and storing a content graph. The content graph can include labeled nodes that correspond to digital content items that have labels, unlabeled nodes that correspond to digital content items that do not have labels, and edges that indicate relationships between content items. Edge data for an edge between an unlabeled node and an adjacent node can be retrieved from the content graph. Responsive to a set of inputs that includes the retrieved edge data and embedding data associated with the unlabeled node, a machine learning model trained on labeled nodes and edges of the content graph can assign a label to the unlabeled node.

Abstract: In an example embodiment an approximate nearest neighbor framework is provided to query user activity data to find users who are similar to users who have been “matched” to a particular piece of content but who otherwise would not have been matched on their own. The users who have been matched may be called a seed set of users, which are known in real-time, or near-real-time. Use of the approximate nearest neighbor framework allows the system to expand instantly the initial seed set of users to other similar users to rapidly distribute relevant pieces of content to active users, increasing liquidity of the system. Additionally, the target set of specific users to which a notification is sent about the pieces of content can also be expanded, increasing the recall rate.

Abstract: In an example embodiment, user interactions with a graphical user interface are modeled to derive an efficient representation that is highly available through a framework. This representation enables downstream analysis as to the relevancy of the user interactions through libraries leveraging standardized activity representations. With these components, it becomes possible to derive user intent in a modular fashion, domain by domain, while decoupling many system aspects, and also providing high capacity and precise intent information to leverage for personalization.

Abstract: In an example, a piece of content is obtained. The piece of content is segmented into a plurality of segments. Each of the plurality of segments is divided into a plurality of units. Then, for each of the plurality of units for each segment, a quality model is used to pass the unit through a long short-term memory (LSTM) corresponding to the unit, causing an embedding of the unit and generating one or more parameters. For each of the plurality of segments for the piece of content, the quality model is used to pass the segment, and one or more parameters obtained from LSTMs corresponding to units within the segment, through an LSTM corresponding to the segment, causing an embedding of the segment. The piece of content is then classified as spam based on the embeddings of the plurality of segments for the piece of content, using the quality model.

Abstract: In an example embodiment, user interactions with a graphical user interface are modeled to derive an efficient representation that is highly available through a framework. This representation enables downstream analysis as to the relevancy of the user interactions through libraries leveraging standardized activity representations. With these components, it becomes possible to derive user intent in a modular fashion, domain by domain, while decoupling many system aspects, and also providing high capacity and precise intent information to leverage for personalization.

Abstract: In an example embodiment an approximate nearest neighbor framework is provided to query user activity data to find users who are similar to users who have been “matched” to a particular piece of content but who otherwise would not have been matched on their own. The users who have been matched may be called a seed set of users, which are known in real-time, or near-real-time. Use of the approximate nearest neighbor framework allows the system to expand instantly the initial seed set of users to other similar users to rapidly distribute relevant pieces of content to active users, increasing liquidity of the system. Additionally, the target set of specific users to which a notification is sent about the pieces of content can also be expanded, increasing the recall rate.

Abstract: In some embodiments, a computer system detects user-entered text that has been entered in a search field of a search engine via a user interface of a computing device of a user, determines a context representation for the user-entered text based on one or more search queries submitted by the user within a particular amount of time before the user-entered text was entered, generates a corresponding score for each one of a plurality of auto-completion candidates based on the auto-completion candidate and the context representation, and causes at least a portion of the plurality of auto-completion candidates to be displayed in an auto-complete user interface element of the search field based on the corresponding scores of the at least a portion of the plurality of auto-completion candidates prior to the user-entered text being submitted by the user as part of a search query.

Abstract: A system and method for ranking query-advertisement combinations is disclosed. Embodiments use an online component to enhance and rank query ad combinations. The query ad combination is then reranked with a trained factorization machine. The subsequent list of ranked query-ad combinations is then output. The output may be to an auction for determine ad-query combinations having the greatest expected revenue.

Abstract: Disclosed are systems, methods, and non-transitory computer-readable media for guided query recommendations. A search system generates search query recommendations for a user based on activity data associated with the user. In one technique, the search system generates a search query recommendation based on a search query sequence identified from the activity data of the user. For example, the search query sequence is used as input into a machine learning model, such as a sequence to sequence model trained on historical search query sequences that resulted in a targeted action. In another technique, the search system generates a search query recommendation based on multi-session query data of the user. For example, the search system generates a multi-session embedding vector representing the multiple activity sessions of the user. The multi-session embedding vector is used as input in a classification model that assigns probability values to candidate search terms.

Abstract: The disclosed embodiments provide a system for processing data. During operation, the system obtains a first sentence representing a first sequence of actions between a user and a set of jobs. Next, the system applies a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first set of output embeddings. The system then combines the first set of output embeddings into a first session embedding that encodes the first sequence of actions. Finally, the system outputs the first session embedding for use in characterizing job-seeking activity of the user.

Abstract: The disclosed embodiments provide a system for processing data. During operation, the system obtains a first sentence representing a first sequence of actions between a user and a set of jobs. Next, the system applies a language model to token embeddings of a first set of tokens in the first sentence and position embeddings of token positions in the first sentence to produce a first set of output embeddings. The system then combines the first set of output embeddings into a first session embedding that encodes the first sequence of actions. Finally, the system outputs the first session embedding for use in characterizing job-seeking activity of the user.

Abstract: Methods, systems, and computer programs are presented for improving search mechanisms for guest users or unidentified members of an online service. One method includes operations for detecting a first search query for a guest user and for initializing a session for the guest user in response to the first search query. Further, the method includes operations for logging activities of the guest user during the session and detecting a second search query on the online service for the guest user while the session is active. Results are obtained in response to the second search query, and the results are prioritized based on the second search query and the activities logged of the guest user. Further, the plurality of results is presented on a computing device of the guest user.

Abstract: In some embodiments, a computer system detects user-entered text that has been entered in a search field of a search engine via a user interface of a computing device of a user, determines a context representation for the user-entered text based on one or more search queries submitted by the user within a particular amount of time before the user-entered text was entered, generates a corresponding score for each one of a plurality of auto-completion candidates based on the auto-completion candidate and the context representation, and causes at least a portion of the plurality of auto-completion candidates to be displayed in an auto-complete user interface element of the search field based on the corresponding scores of the at least a portion of the plurality of auto-completion candidates prior to the user-entered text being submitted by the user as part of a search query.

Abstract: Techniques for generating individualized learning paths are provided. A skill dependency graph is generated that indicates, for each pair of connecting nodes in the graph, a first skill in the pair as a prerequisite of a second skill in the pair. A set of destination skills is determined that a user is to obtain to achieve a possible career goal. Based on the skill dependency graph and the set of destination skills, one or more prerequisite skills that the user should obtain prior to obtaining the set of destination skills are identified. Based on the set of destination skills, the one or more prerequisite skills, and information about the user, an individualized learning path is generated that comprises a sequence of learning resources that allows the particular user to obtain a set of skills. The individualized learning path is presented on a screen of a computing device of the user.

Abstract: In an example, a piece of content is obtained. The piece of content is segmented into a plurality of segments. Each of the plurality of segments is divided into a plurality of units. Then, for each of the plurality of units for each segment, a quality model is used to pass the unit through a long short-term memory (LSTM) corresponding to the unit, causing an embedding of the unit and generating one or more parameters. For each of the plurality of segments for the piece of content, the quality model is used to pass the segment, and one or more parameters obtained from LSTMs corresponding to units within the segment, through an LSTM corresponding to the segment, causing an embedding of the segment. The piece of content is then classified as spam based on the embeddings of the plurality of segments for the piece of content, using the quality model.

Abstract: Systems and methods for rewriting queries based on data features are disclosed. A data source contains a history of queries and results associated with the queries. A source query is received and a subset of the data source is extracted that is related to the query. Costs are then associated with the associations of the queries and results, and an absorbed cost is determined for each query. The queries having the lowest absorbed cost are recommended for rewrites.

Abstract: A system and method for ranking query-advertisement combinations is disclosed. Embodiments use an online component to enhance and rank query ad combinations. The query ad combination is then reranked with a trained factorization machine. The subsequent list of ranked query-ad combinations is then output. The output may be to an auction for determine ad-query combinations having the greatest expected revenue.

Abstract: Systems and methods for building an index for matching queries and bidded terms are disclosed. The expansion of the bid-terms includes a collective expansion using graph mining techniques on a subgraph extracted from user actions on a <query, URL> bi-partite graph. The extracted subgraph is specific to an AdGroup (obtained using all the available context) and is personalized. The extracted subgraph can be tuned to various objectives such as head/torso/tail, specific vs generic etc. Multiple relevance measures of candidate query nodes are calculated with reference to multiple source nodes or context, and the final rankings are biased to commercial relevance using personalized random-walks.

Abstract: Systems and methods for building a search index for query recommendation and ad matching are disclosed. The system accesses a query-URL graph and extracts a subgraph related to an ad campaign. The subgraph is annotated according to desired criteria. The sub graph is reversed and the reversed annotated subgraph is ranked to find nodes of importance. The nodes of importance are then used to build a preference vector which is used to find a stationary distribution of the sub graph. A plurality of random walks of the sub graph is performed to build a corpus of words. The corpus of words are input into a language model to learn associations, from which the top query terms associated with an ad campaign are found and indexed. The index is then inverted for recommending ads for received query terms.

Abstract: Systems and methods for rewriting queries based on data features are disclosed. A data source contains a history of queries and results associated with the queries. A source query is received and a subset of the data source is extracted that is related to the query. Costs are then associated with the associations of the queries and results, and an absorbed cost is determined for each query. The queries having the lowest absorbed cost are recommended for rewrites.