Iterative online learning to improve targeted advertising

A computer-implemented method, comprising:

• selecting a first probability distribution function characterized by a first mean vector and a first covariance matrix, the first probability distribution function representing a probability of selecting values, from a first range of values, for first model parameters associated with a model configured to receive a vector representing a website as an input and output a predicted probability of conversion;
• selecting a second probability distribution function characterized by a second mean vector and a second covariance matrix, the second probability distribution function representing a probability of selecting values, from a second range of values, for second model parameters associated with the model;
• sampling a plurality of sets of model parameter values based on the first probability distribution and the second probability distribution function, each set of model parameter values from the plurality of sets of model parameter values defining an instance of the model from a plurality of instances of the model;
• calculating, using a first instance of the model from the plurality of instances of the model, a first plurality of conversion probabilities for each website from a plurality of websites;
• calculating using a second instance of the model from the plurality of instances of the model, a second plurality of conversion probabilities for each website from the plurality of websites;
• selecting a subset of websites from the plurality of websites for which a conversion probability from at least one of the first plurality of conversion probabilities or the second plurality of conversion probabilities is greater than a predetermined probability threshold;
• obtaining an indication of a plurality of actual conversion rates, each actual conversion rate from the plurality of conversion rates associated with a website from the subset of websites;
• updating the first probability distribution function based on the plurality of actual conversion rates;
• and
• updating the second probability distribution function based on a difference between the first plurality of conversion probabilities and the plurality of actual conversion rates such that an updated second probability distribution function is characterized by at least one of a shifted second mean vector or a different second covariance matrix.

A computer-implemented method, comprising:

• selecting a mean vector and a covariance matrix that at least partially define a probability distribution function representing a probability of selecting values, from a range of values, for each model parameter from a plurality of model parameters associated with a model, the model configured to receive a vector representing a website as an input and output a predicted probability of conversion when the website is visited;
• sampling a plurality of sets of model parameter values based on the probability distribution function, each set of model parameters values from the plurality of sets of model parameter values defining an instance of the model from a plurality of instances of the model;
• for each instance of the model, calculating the predicted probability of conversion for a first plurality of websites;
• selecting a second plurality of websites, each website from the second plurality of websites being a website from the first plurality of websites associated with the predicted probability of conversion greater than a predetermined probability threshold as calculated by at least one instance of the model;
• for each website from the second plurality of websites, selecting a bid factor for determining a cost of displaying an impression, the bid factor based on the predicted probability of conversion;
• for each visitation event from a plurality of visitation events of a website from the second plurality of websites, receiving an indication that an impression has been displayed to a user;
• obtaining a plurality of actual conversion rates, from the plurality of actual conversion rates being determined for each website from the second plurality of websites based on a number of conversion events associated with the plurality of visitation events;
• and
• updating the probability distribution based on the actual conversion rates such that probabilities of possible values for at least one model parameter from the plurality of model parameters has at least one of a shifted mean or a different covariance relative to probabilities of possible values for that model parameter prior to updating the probability distribution function.