Yahoo! Inc.

Excalibur Ip, LLC

Olivier Chapelle

Sathiya Keerthi Selvaraj

John C Langford

Miroslav Dudik

Alekh Agarwal

Anne Ya Zhang

Method, system, and programs for distributed machine learning on a cluster including a plurality of nodes are disclosed. A machine learning process is performed in each of the plurality of nodes based on a respective subset of training data to calculate a local parameter. The training data is partitioned over the plurality of nodes. A plurality of operation nodes are determined from the plurality of nodes based on a status of the machine learning process performed in each of the plurality of nodes. The plurality of operation nodes are connected to form a network topology. An aggregated parameter is generated by merging local parameters calculated in each of the plurality of operation nodes in accordance with the network topology.

Method and system for distributed machine learning

Approaches and techniques are discussed for ranking the documents indicated in search results for a query based on click-through information collected for the query in previous query sessions. According to an embodiment of the invention, when calculating a relevance score for a particular document, one may overcome positional bias by utilizing click-through information about other documents previously returned in the same search results as the particular document. According to an embodiment, one may utilize Dynamic Bayesian Network, based on said click-through information, to model relevance. According to an embodiment of the invention, one may utilize click-through information to generate targets for learning a ranking function.

Click model for search rankings

In one embodiment, training a ranking model comprises: accessing the ranking model and an objective function of the ranking model; accessing one or more preference pairs of objects, wherein for each of the preference pairs of objects comprising a first object and a second object, there is a preference between the first object and the second object with respect to the particular reference, and the first object and the second object each has a feature vector comprising one or more feature values; and training the ranking model by minimizing the objective function using the preference pairs of objects, wherein for each of the preference pairs of objects, a difference between the first feature vector of the first object and the second feature vector of the second object is not calculated.

Efficient algorithm for pairwise preference learning

Methods, systems, and apparatuses for generating relevance functions for ranking documents obtained in searches are provided. One or more features to be used as predictor variables in the construction of a relevance function are determined. The relevance function is parameterized by one or more coefficients. An ideal query error is defined that measures, for a given query, a difference between a ranking generated by the relevance function and a ranking based on a training set. According to a structured output learning framework, values for the coefficients of the relevance function are determined to substantially minimize an objective function that depends on a continuous upper bound of the defined ideal query error. The query error is determined using a structured output learning technique. The query error is defined as a maximum over a set of permutations.

Determining a relevance function based on a query error derived using a structured output learning technique

Methods, systems, and apparatuses for generating relevance functions for ranking documents obtained in searches are provided. One or more features to be used as predictor variables in the construction of a relevance function are determined. The relevance function is parameterized by one or more coefficients. A query error is defined that measures a difference between a relevance ranking generated by the relevance function and a training set relevance ranking based on a query and a set of scored documents associated with the query. The query error is a continuous function of the coefficients and aims at approximating errors measures commonly used in Information Retrieval. Values for the coefficients of the relevance function are determined that substantially minimize an objective function that depends on the defined query error.

Gradient based optimization of a ranking measure

An improved system and method is provided for training a multi-class support vector machine to select a common subset of features for classifying objects. A multi-class support vector machine generator may be provided for learning classification functions to classify sets of objects into classes and may include a sparse support vector machine modeling engine for training a multi-class support vector machine using scaling factors by simultaneously selecting a common subset of features iteratively for all classes from sets of features representing each of the classes. An objective function using scaling factors to ensure sparsity of features may be iteratively minimized, and features may be retained and added until a small set of features stabilizes. Alternatively, a common subset of features may be found by iteratively removing at least one feature simultaneously for all classes from an active set of features initialized to represent the entire set of training features.

Growing community of inventors

Mountain View, CA, United States of America

Olivier Chapelle