Abstract: A training system is described herein for generating a prediction model that relies on a feature space with reduced dimensionality. The training system performs this task by producing partitions, each of which corresponds to a subset of aspect values (where each aspect value, in turn, may correspond to one or more attribute values). The training system then produces instances of statistical information associated with the partitions. Each instance of statistical information therefore corresponds to feature information that applies to a plurality of aspect values, rather than a single aspect value. The training system then trains the prediction model based on the feature information. Also described herein is a prediction module that uses the prediction model to make predictions in various online contexts.

Abstract: In one embodiment, an evolving advertising system automatically optimizes internet advertising. A data storage unit 250 may store an evolving advertisement unit 320 with an advertisement characteristic according to an initial configuration parameter. A communication interface 280 may transmit the evolving advertisement unit 320 as part of a primary website 310. A processor 220 may alter the evolving advertisement unit 320 automatically upon a trigger event by changing the advertisement characteristic to follow an automatically generated configuration parameter to optimize an advertisement performance metric.

Abstract: Various technologies described herein pertain to estimating predictive accuracy gain of a potential feature added to a set of features, wherein an existing predictor is trained on the set of features. Outputs of the existing predictor for instances in a dataset can be retrieved from a data store. Moreover, a predictive accuracy gain estimate of a potential feature added to the set of features can be measured as a function of the outputs of the existing predictor for the instances in the dataset. The predictive accuracy gain estimate can be measured without training an updated predictor on the set of features augmented by the potential feature.

Abstract: Embodiments described herein are directed to methods and systems for performing neural network computations on encrypted data. Encrypted data is received from a user. The encrypted data is encrypted with an encryption scheme that allows for computations on the ciphertext to generate encrypted results data. Neural network computations are performed on the encrypted data, using approximations of neural network functions to generate encrypted neural network results data from encrypted data. The approximations of neural network functions can approximate activation functions, where the activation functions are approximated using polynomial expressions. The encrypted neural network results data are communicated to the user associated with the encrypted data such that the user decrypts the encrypted data based on the encryption scheme. The functionality of the neural network system can be provided using a cloud computing platform that supports restricted access to particular neural networks.

Abstract: Embodiments described herein are directed to methods and systems for performing neural network computations on encrypted data. Encrypted data is received from a user. The encrypted data is encrypted with an encryption scheme that allows for computations on the ciphertext to generate encrypted results data. Neural network computations are performed on the encrypted data, using approximations of neural network functions to generate encrypted neural network results data from encrypted data. The approximations of neural network functions can approximate activation functions, where the activation functions are approximated using polynomial expressions. The encrypted neural network results data are communicated to the user associated with the encrypted data such that the user decrypts the encrypted data based on the encryption scheme. The functionality of the neural network system can be provided using a cloud computing platform that supports restricted access to particular neural networks.

Abstract: Technologies pertaining to tuning a hyper-parameter configuration of a learning algorithm are described. The learning algorithm learns parameters of a predictive model based upon the hyper-parameter configuration. Candidate hyper-parameter configurations are identified, and statistical hypothesis tests are undertaken over respective pairs of candidate hyper-parameter configurations to identify, for each pair of candidate hyper-parameter configurations, which of the two configurations is associated with better predictive performance. The technologies described herein take into consideration the stochastic nature of training data, validation data, and evaluation functions.

Abstract: Technologies pertaining to computing a tiering policy that defines how digital items are desirable stored across a plurality of different storage tiers are described herein. A data repository that comprises data that is indicative of historic user interaction with a search engine is accessed. Subsequently, a tiering policy for digital items that are retrievable by way of the search engine is computed based at least in part upon the data that is indicative of the historic user interaction with the search engine. Retrieval times for digital items in the data storage tiers differ across the data storage tiers.

Abstract: A service that performs automatic selection and recommendation of featurization(s) for a provided dataset and machine learning application is described. The service can be a cloud service. Selection/recommendation can cover multiple featurizations that are available for most common raw data formats (e.g., images and text data). Provided a dataset and a task, the service can evaluate different possible featurizations, selecting one or more based on performance, similarity of dataset and task to known datasets with featurizations known to have high predictive accuracy on similar tasks low predictive error, training via learning algorithms to take multiple inputs, etc. The service may include a request-response aspect that provides access to the best featurization selected for the given dataset and task.

Abstract: Technologies pertaining to computing a tiering policy that defines how digital items are desirable stored across a plurality of different storage tiers are described herein. A data repository that comprises data that is indicative of historic user interaction with a search engine is accessed. Subsequently, a tiering policy for digital items that are retrievable by way of the search engine is computed based at least in part upon the data that is indicative of the historic user interaction with the search engine. Retrieval times for digital items in the data storage tiers differ across the data storage tiers.

Abstract: Various technologies described herein pertain to estimating predictive accuracy gain of a potential feature added to a set of features, wherein an existing predictor is trained on the set of features. Outputs of the existing predictor for instances in a dataset can be retrieved from a data store. Moreover, a predictive accuracy gain estimate of a potential feature added to the set of features can be measured as a function of the outputs of the existing predictor for the instances in the dataset. The predictive accuracy gain estimate can be measured without training an updated predictor on the set of features augmented by the potential feature.

Abstract: Various technologies described herein pertain to estimating predictive accuracy gain of a potential feature added to a set of features, wherein an existing predictor is trained on the set of features. Outputs of the existing predictor for instances in a dataset can be retrieved from a data store. Moreover, a predictive accuracy gain estimate of a potential feature added to the set of features can be measured as a function of the outputs of the existing predictor for the instances in the dataset. The predictive accuracy gain estimate can be measured without training an updated predictor on the set of features augmented by the potential feature.

Abstract: Technologies pertaining to tuning a hyper-parameter configuration of a learning algorithm are described. The learning algorithm learns parameters of a predictive model based upon the hyper-parameter configuration. Candidate hyper-parameter configurations are identified, and statistical hypothesis tests are undertaken over respective pairs of candidate hyper-parameter configurations to identify, for each pair of candidate hyper-parameter configurations, which of the two configurations is associated with better predictive performance. The technologies described herein take into consideration the stochastic nature of training data, validation data, and evaluation functions.

Abstract: A training system is described herein for generating a prediction model that relies on a feature space with reduced dimensionality. The training system performs this task by producing partitions, each of which corresponds to a subset of aspect values (where each aspect value, in turn, may correspond to one or more attribute values). The training system then produces instances of statistical information associated with the partitions. Each instance of statistical information therefore corresponds to feature information that applies to a plurality of aspect values, rather than a single aspect value. The training system then trains the prediction model based on the feature information. Also described herein is a prediction module that uses the prediction model to make predictions in various online contexts.

Abstract: Various technologies described herein pertain to estimating predictive accuracy gain of a potential feature added to a set of features, wherein an existing predictor is trained on the set of features. Outputs of the existing predictor for instances in a dataset can be retrieved from a data store. Moreover, a predictive accuracy gain estimate of a potential feature added to the set of features can be measured as a function of the outputs of the existing predictor for the instances in the dataset. The predictive accuracy gain estimate can be measured without training an updated predictor on the set of features augmented by the potential feature.

Abstract: The claimed subject matter provides a method for visualizing machine learning accuracy. The method includes receiving a plurality of training instances for the machine learning system. The method also includes receiving a plurality of results for the machine learning system. The plurality of results corresponds to the plurality of training instances. The method further includes providing an interactive representation of the training instances and the results. The interactive representation supports identifying inaccuracies of the machine learning system attributable to the training instances, the features used to obtain a featurized form of the training instance, and/or a model implemented by the machine learning system.

Abstract: Described is processing the search results obtained from a search engine to determine advertisements that match properties of those search results. For example, the URL-related information or domain-related information in the search results may be used to select an advertisement. Also described are various conditions that may need to be met before an advertisement is selected. Further, an advertisement may be modified to include information in the search results, such as to insert a competing company's name that appears in a search result into the advertisement. Also described is a system for returning keywords in response to a URL/domain.

Abstract: In one embodiment, an evolving advertising system automatically optimizes internet advertising. A data storage unit 250 may store an evolving advertisement unit 320 with an advertisement characteristic according to an initial configuration parameter. A communication interface 280 may transmit the evolving advertisement unit 320 as part of a primary website 310. A processor 220 may alter the evolving advertisement unit 320 automatically upon a trigger event by changing the advertisement characteristic to follow an automatically generated configuration parameter to optimize an advertisement performance metric.

Abstract: Technologies pertaining to computing a tiering policy that defines how digital items are desirable stored across a plurality of different storage tiers are described herein. A data repository that comprises data that is indicative of historic user interaction with a search engine is accessed. Subsequently, a tiering policy for digital items that are retrievable by way of the search engine is computed based at least in part upon the data that is indicative of the historic user interaction with the search engine. Retrieval times for digital items in the data storage tiers differ across the data storage tiers.

Abstract: Described is using per-user profile data (e.g., maintained in a browser cookie) as a factor in selecting advertisements to be presented to a user for a current context such as containing query keywords. For example, an advertiser may be willing to bid more if the current context's keywords match the user profile data that indicates a particular area of interest to the user and advertiser. Also described is updating the per-user profile data with the current context if doing so increases the expected utility of the per-user profile data, e.g., increases the predicted amount of revenue from advertisement clicking. Also described is other advertisement personalization based upon the per-user profile data, e.g., the ranking and/or appearance of the advertisements.

Abstract: Methods and computer-readable media are provided for receiving keyword expansions from expansion providers and selecting a set of keyword expansions that are used for advertisement selection. Keyword expansions that correspond to a particular search query or text from a browsed web page are received from an expansion provider. Feature data is extracted from each keyword expansion, and may include properties of the keyword expansion or the expansion provider. A score is assigned to each keyword expansion, and based on the score, a set of keyword expansions is selected from the keyword expansions received from the expansion provider. The set of keyword expansions is used to select relevant advertisements for presentation to the user.