Abstract: The present teaching relates to method, system, medium, and implementations for joint prediction. Training data is obtained with information about a plurality of users collected from different sources and ground truth demographics/interests associated with each of the plurality users. Based on the training data, a joint prediction model is trained for simultaneously predicting multiple pieces of demographic/interest information. When information about a user from different sources is received, a joint feature vector is derived therefrom, which is then used by the trained joint prediction model to predict multiple pieces of demographic/interest information about the user.

Abstract: The techniques described herein relate to constructing and using seed audiences. In an embodiment, a method includes loading, by a processing device, a user event sequence, the user event sequence including a plurality of user events and a plurality of corresponding conversions; generating, by the processing device, a plurality of conversion neighborhoods based on the user event sequence, a given conversion neighborhood in the plurality of conversion neighborhood including at least one conversion rule and a set of user events from the plurality of user events; annotating, by the processing device, each conversion neighborhood in the plurality of conversion neighborhoods with categorical labels; and generating, by the processing device, seed audiences for each conversion neighborhood, a given seed audience including a ranked list of user events for each conversion rule associated with the conversion neighborhood.

Abstract: Technologies for template generation using directed acyclic word graphs (DAWGs). The technologies can include receiving a first plurality of titles from a first plurality of title feeds, and sorting the first plurality of titles into a plurality of category sets. And, for each category set of the plurality of category sets, the technologies can include transforming the respective titles belonging to the category set into a trie data structure by separating words in the respective titles into nodes of the trie data structure. For each category set, the technologies can also include transforming the trie data structure into a directed acyclic word graph (DAWG) data structure. Also, for each category set, the technologies can also include generating one or more unique templates based on the DAWG data structure.

Abstract: The techniques described herein relate to constructing and using seed audiences. In an embodiment, a method includes loading, by a processing device, a user event sequence, the user event sequence including a plurality of user events and a plurality of corresponding conversions; generating, by the processing device, a plurality of conversion neighborhoods based on the user event sequence, a given conversion neighborhood in the plurality of conversion neighborhood including at least one conversion rule and a set of user events from the plurality of user events; annotating, by the processing device, each conversion neighborhood in the plurality of conversion neighborhoods with categorical labels; and generating, by the processing device, seed audiences for each conversion neighborhood, a given seed audience including a ranked list of user events for each conversion rule associated with the conversion neighborhood.

Abstract: This teaching relates to predictive targeting. Training data are obtained with pairs of data. Each pair includes an ad opportunity context corresponding to an ad served to a plurality of audiences and a label vector having a plurality of labels, each of which indicates a reaction, with respect to the ad served, of a corresponding one of the audiences in the ad opportunity context. Based on the training data, model parameters of a joint predictive model are learned via machine learning based on an initialized model with initial model parameters by minimizing a loss in an iterative process. The learned joint predictive model is to be used to map an input context of an ad opportunity to an output label vector having a plurality of probabilities, each of which predicts a likelihood of a reaction of a corresponding one of the audiences to the input context of the ad opportunity.

Abstract: Technologies for template generation using directed acyclic word graphs (DAWGs). The technologies can include receiving a first plurality of titles from a first plurality of title feeds, and sorting the first plurality of titles into a plurality of category sets. And, for each category set of the plurality of category sets, the technologies can include transforming the respective titles belonging to the category set into a trie data structure by separating words in the respective titles into nodes of the trie data structure. For each category set, the technologies can also include transforming the trie data structure into a directed acyclic word graph (DAWG) data structure. Also, for each category set, the technologies can also include generating one or more unique templates based on the DAWG data structure.

Abstract: Technologies for template generation using directed acyclic word graphs (DAWGs). The technologies can include receiving a first plurality of titles from a first plurality of title feeds, and sorting the first plurality of titles into a plurality of category sets. And, for each category set of the plurality of category sets, the technologies can include transforming the respective titles belonging to the category set into a trie data structure by separating words in the respective titles into nodes of the trie data structure. For each category set, the technologies can also include transforming the trie data structure into a directed acyclic word graph (DAWG) data structure. Also, for each category set, the technologies can also include generating one or more unique templates based on the DAWG data structure.

Abstract: Technologies for template generation using directed acyclic word graphs (DAWGs). The technologies can include receiving a first plurality of titles from a first plurality of title feeds, and sorting the first plurality of titles into a plurality of category sets. And, for each category set of the plurality of category sets, the technologies can include transforming the respective titles belonging to the category set into a trie data structure by separating words in the respective titles into nodes of the trie data structure. For each category set, the technologies can also include transforming the trie data structure into a directed acyclic word graph (DAWG) data structure. Also, for each category set, the technologies can also include generating one or more unique templates based on the DAWG data structure.

Abstract: The present teaching relates to joint representation of information. In one example, first and second pieces of information are received. Each of the first and second pieces of information relates to one word in a plurality of documents, one of the documents, or one of user to which the documents are given. A model for estimating feature vectors is obtained. The model includes a first neural network model based on a first order of words within one of the documents and a second neural network model based on a second order in which at least some of the documents are given. Based on the model, a first feature vector of the first piece of information and a second feature vector of the second piece of information are estimated. A similarity between the first and second pieces of information is determined based on a distance between the first and second feature vectors.

Abstract: Systems and methods for rewriting query terms are disclosed. The system collects queries and query session data and separates the queries into sequences of queries having common sessions. The sequences of queries are then input into a deep learning network to build a multidimensional word vector in which related terms are nearer one another than unrelated terms. An input query is then received and the system matches the input query in the multidimensional word vector and rewrites the query using the nearest neighbors to the term of the input query.

Abstract: A system stored in a non-transitory medium executable by processor circuitry is provided for generating retargeting keywords based on distributed query word representations. The system includes one or more system databases storing historical web search data. Search retargeting circuitry receives requests to generate sets of retargeting keywords related to one or more categories of an advertisement campaign and pre-processing circuitry retrieves a set of historical web search data related to the one or more categories of the advertisement campaign. Modeling circuitry further applies one or more computational linguistic models to the retrieved set of historical web search data and generates distributed query word representations from the retrieved set of historical web search data. Keyword generator circuitry generates a list of retargeting keywords related to the one or more categories of the advertisement campaign using the generated distributed query word representations.

Abstract: Methods and apparatus are described by which “superphrases” of “seed phrases” representing basic concepts may be identified without having to compare all possible pairs of seed and candidate phrases. According to one class of embodiments, a data structure similar to an inverted index is used for indexing phrases. The elimination of seed and candidate phrase pairs is enabled by building and traversing the index in a particular manner.