Abstract: For ad campaigns that have multiple advertisements, each associated with an ad creative, which are automatically selected, an online system may bias selection of advertisements away from underestimated advertisements and towards early-selected advertisements with positive user interactions. To increase the likelihood of various advertisements in an ad campaign being evaluated for presentation to users, the online system may: associate a relatively high performance score with each advertisement in an ad campaign, randomly select advertisements from an ad campaign, modify bid amounts associated with advertisements in the ad campaign, or allocate a portion of the ad campaign's budget for allocation across advertisements in the ad campaign. After presenting a threshold number of advertisements in an ad campaign or receiving an instruction from an advertiser, advertisements from the ad campaign may be selected using conventional methods.

Abstract: Methods, systems, and media are provided for a dynamic batch strategy utilized in parallelization of online learning algorithms. The dynamic batch strategy provides a merge function on the basis of a threshold level difference between the original model state and an updated model state, rather than according to a constant or pre-determined batch size. The merging includes reading a batch of incoming streaming data, retrieving any missing model beliefs from partner processors, and training on the batch of incoming streaming data. The steps of reading, retrieving, and training are repeated until the measured difference in states exceeds a set threshold level. The measured differences which exceed the threshold level are merged for each of the plurality of processors according to attributes. The merged differences which exceed the threshold level are combined with the original partial model states to obtain an updated global model state.

Abstract: Event prediction in dynamic environments is described. In an embodiment a prediction engine may use the learnt information to predict events in order to control a system such as for internet advertising, email filtering, fraud detection or other applications. In an example one or more variables exists for pre-specified features describing or associated with events and each variable is considered to have an associated weight and time stamp. For example, belief about each weight is represented using a probability distribution and a dynamics process is used to modify the probability distribution in a manner dependent on the time stamp for that weight. For example, the uncertainty about the associated variable's influence on prediction of future events is increased. Examples of different schedules for applying the dynamics process are given.

Abstract: A computationally efficient method of performing probabilistic linear regression is described. In an embodiment, the method involves adding a white noise term to a weighted linear sum of basis functions and then normalizing the combination. This generates a linear model comprising a set of sparse, normalized basis functions and a modulated noise term. When using the linear model to perform linear regression, the modulated noise term increases the variance associated with output values which are distant from any data points.

Abstract: A scalable clustering system is described. In an embodiment the clustering system is operable for extremely large scale applications where millions of items having tens of millions of features are clustered. In an embodiment the clustering system uses a probabilistic cluster model which models uncertainty in the data set where the data set may be for example, advertisements which are subscribed to keywords, text documents containing text keywords, images having associated features or other items. In an embodiment the clustering system is used to generate additional features for associating with a given item. For example, additional keywords are suggested which an advertiser may like to subscribe to. The additional features that are generated have associated probability values which may be used to rank those features in some embodiments. User feedback about the generated features is received and used to revise the feature generation process in some examples.

Abstract: Methods, systems, and media are provided for a dynamic batch strategy utilized in parallelization of online learning algorithms. The dynamic batch strategy provides a merge function on the basis of a threshold level difference between the original model state and an updated model state, rather than according to a constant or pre-determined batch size. The merging includes reading a batch of incoming streaming data, retrieving any missing model beliefs from partner processors, and training on the batch of incoming streaming data. The steps of reading, retrieving, and training are repeated until the measured difference in states exceeds a set threshold level. The measured differences which exceed the threshold level are merged for each of the plurality of processors according to attributes. The merged differences which exceed the threshold level are combined with the original partial model states to obtain an updated global model state.

Abstract: Event prediction in dynamic environments is described. In an embodiment a prediction engine may use the learnt information to predict events in order to control a system such as for internet advertising, email filtering, fraud detection or other applications. In an example one or more variables exists for pre-specified features describing or associated with events and each variable is considered to have an associated weight and time stamp. For example, belief about each weight is represented using a probability distribution and a dynamics process is used to modify the probability distribution in a manner dependent on the time stamp for that weight. For example, the uncertainty about the associated variable's influence on prediction of future events is increased. Examples of different schedules for applying the dynamics process are given.

Abstract: A scalable clustering system is described. In an embodiment the clustering system is operable for extremely large scale applications where millions of items having tens of millions of features are clustered. In an embodiment the clustering system uses a probabilistic cluster model which models uncertainty in the data set where the data set may be for example, advertisements which are subscribed to keywords, text documents containing text keywords, images having associated features or other items. In an embodiment the clustering system is used to generate additional features for associating with a given item. For example, additional keywords are suggested which an advertiser may like to subscribe to. The additional features that are generated have associated probability values which may be used to rank those features in some embodiments. User feedback about the generated features is received and used to revise the feature generation process in some examples.

Abstract: A computationally efficient method of performing probabilistic linear regression is described. In an embodiment, the method involves adding a white noise term to a weighted linear sum of basis functions and then normalizing the combination. This generates a linear model comprising a set of sparse, normalized basis functions and a modulated noise term. When using the linear model to perform linear regression, the modulated noise term increases the variance associated with output values which are distant from any data points.

Abstract: There are many situations in which it is desired to predict outcomes of events. In an example, an event prediction system is described which receives variables for a proposed event. The system accesses learnt statistics describing belief about weights associated with the variables and uses the weights to determine probability information that the proposed event will have a specified outcome. The process involves combining the accessed statistics and mapping them into a number representing the probability. In another example, a machine learning process using assumed density filtering is used to learn the statistics from data about observed events. The event prediction system may be used as part of any suitable type of system such as an internet advertising system, an email filtering system, or a fraud detection system.