An audience proposal creator determines a target cost per thousand (CPM) baseline and a demographics CPM baseline for a deal offering audience spots, determines deal constraints based on a target CPM reduction goal, a demographics CPM cap, and the established parameters, and generates rates by selling title for each selling title-weeks for a duration of pending deal, and for each network of a plurality of networks based on the constraints. Target and demo audience rating estimates are acquired based on the target and demo applicable to the advertiser for the plurality of networks, and a distribution of the audience spots generated across the plurality of selling title-weeks, and networks based on the target audience rating estimates, a budget for the pending deal, the generated rates, and available inventory per selling title-weeks, and a proposal generated based on the distribution. An audience processor schedules audience spots across one or more networks for selling title-weeks based on the distribution.

A method (100) of controlling playout of advertisement content during live video streaming at an end-user terminal comprising steps of: receiving (110) advertisement content from an advertisement server; receiving (112) live streamed video content from a content delivery network and playing out the video content; obtaining (114) at least one of image features and audio features of the video content during playout; calculating (116) a content importance rating of video content to be played out during a prediction time window based on said features; and postponing (118) playout of advertisement content scheduled to be played out during the prediction time window if the calculated content importance rating for the prediction time window is above a threshold value.

A utility application for a mobile device inspects data packets from other mobile applications running on the device to gather and record usage data about those applications. Since users of the utility application may not reflect the true population for which the usage data is desired, a system de-biases the data reported from the utility applications using a machine learning model to predict demographics of the users of the utility application. To determine a training data set for the model, the system requests a user to provide a desired user attribute by way of an in-app questionnaire. This enables labeling utility usage data with the demographics, which can be weighted and extrapolated to determine usage across the population as a whole.

In one aspect, a request for web content is received from a user device communicatively coupled to the processing device via the network. In response to receiving the request, user information associated with the user is determined. Predicted responses of the user to each variation of a plurality of variations of the web content are determined using prediction models and the user information. The prediction models include one or more decision trees generated using a splitting criterion requiring a minimum number of positive responses to a variation and a minimum number of negative responses to the variation as a condition of considering the possible split. The variation determined to have a threshold likelihood of yielding a predicted positive response of the predicted responses is selected based on the user information. The variation is transmitted to the user device via the network.

Techniques are provided for k-best option identification of options subject to a supplied tolerance. One technique includes: sampling the options for a first period on a plurality of computers; computing an average and a sample count for each option based on the sampling; splitting the options into a highest group and a lowest group based on the computed averages; selecting a weakest one of the highest group (option A) and a strongest one of the lowest group (option B); and deciding whether or not to terminate based on the supplied tolerance and the selecting of options A and B. In some cases, the technique further includes outputting the highest group and terminating in response to a termination decision; otherwise continue with sampling options A and B for a next period; and updating the computed average and the sample count for options A and B based on corresponding next period sampling.

A plurality of different hosted services each includes enabling logic that enables a set of actions. Usage data for a plurality of different tenants are accessed and actions are grouped into features based upon underlying enabling logic. A correlation score between features is identified based on tenant usage data for those features. A tenant under analysis is selected and usage data for the tenant under analysis is used to identify related features that the tenant under analysis is not using, based upon the correlation scores for the features. An output system is controlled to surface the related features for the tenant under analysis.

Techniques for using machine learning to configure social media campaigns are disclosed. A social relationship management (SRM) service performs supervised machine learning to generate a learned model, at least by: generating feature vectors based on training data including campaign configuration data and one or more campaign success metrics; and performing pattern recognition on the feature vectors to determine one or more preferred campaign configurations. The SRM service publishes messages to one or more social media platforms and receives user interaction data associated with users' interactions with the messages. The SRM service performs unsupervised machine learning to update the learned model based at least in part on the user interaction data. The SRM service receives a request to configure a social media campaign, applies data associated with the request to the learned model to determine a preferred campaign configuration, and configures the social media campaign based on the preferred campaign configuration.

An online system selects for display content items having an option to allow a user to converse with a content provider over an electronic communication system in a way that optimizes for the occurrence of that communication. Content items provided by the online system include links to an online communication system from which the online system can directly observe whether the user engaged in communications with third parties. The online system can thus obtain labeled training data describing communications between users and users' previous interactions with content items and pages of the online system. The trained model is applied to candidate content items to predict a probability that a user will engage in a communication with the content provider if the content is provided to the user, and the model optimizes the selection of content items for which the call to action is to engage in the communication.

A method and apparatus for selecting treatment for visitors to online enterprise channels are disclosed. The method includes receiving information related to a visitor and a current activity of the visitor on an online enterprise channel. The information is transformed to generate transformed data and a plurality of features is extracted from the transformed data. Using the plurality of features, it is determined whether a treatment when rendered to the visitor is capable of increasing a likelihood of the visitor performing a desired action during a current visit to the online enterprise channel. The treatment is selected and rendered if it is determined that the treatment is capable of increasing the likelihood of the visitor performing the desired action. No treatment is rendered if it is determined that no treatment from among the plurality of treatments is capable of increasing the likelihood of the visitor performing the desired action.

Recommendation systems and techniques are described that use linear stochastic bandits and confidence interval generation to generate recommendations for digital content. These techniques overcome the limitations of conventional recommendations systems that are limited to a fixed parameter to estimate noise and thus do not fully exploit available data and are overly conservative, at a significant cost in operational performance of a computing device. To do so, a linear model, noise estimate, and confidence interval are refined by a recommendation system based on user interaction data that describes a result of user interaction with items of digital content. This is performed by comparing a result of the recommendation on user interaction with digital content with an estimate of a result of the recommendation.