Advertisement conversion prediction based on unlabeled data

A non-transitory computer-readable storage medium comprising one or more computer-readable instructions that, when executed by at least one processor of a computing device, causes the computing device to:

• receive a notification that a prerequisite event has occurred for an instance of an advertisement rendered on the computing device;
• assign, in response to the prerequisite event, the instance as unlabeled;
• generate a set of feature vectors from the unlabeled instance;
• train, using the set of feature vectors, a machine learning model comprising a logarithmic loss objective function and a first term;
• receive a notification indicating a target event has occurred subsequent to the prerequisite event for the instance;
• re-assign, in response to the target event, the instance as positive;
• update the set of feature vectors associated with the positive instance;
• generate, by the machine learning model, a probability of a target event occurring after a prerequisite event for a new advertisement, and
• determine, using the probability of the target event, a cost for the new advertisement.

A computing system, the system comprising:

• a memory for storing parameters of a machine learning model to determine a probability of a target event occurring subsequent to a prerequisite event for an instance of an advertisement;
• a labeling module, stored in the memory, configured to label the instance as unlabeled, in response to the prerequisite event when the instance is rendered on a user device or on a computing device;
• the labeling module, stored in the memory, further configured to re-assign the unlabeled instance as positive in response to the target event;
• a feature extraction module, stored in the memory, configured to generate feature vectors of the unlabeled instance or of the positive instance;
• a training module, stored in memory, configured to train parameters of the machine learning model, by minimizing a logarithmic loss objective function comprising a first and a second term, the first term optimized for feature vectors derived from unlabeled instances with no target event and the second term is optimized for feature vectors derived from positive instances;
• a prediction module, stored in the memory, configured to provide a probability on whether a target event will occur for a new advertisement, by providing feature vectors of the new advertisement to the machine learning model;
• a pricing module, stored in the memory, configured to provide a cost for distributing the new advertisement based on the probability of a target event;
• and
• at least one physical processor that executes the labelling module, the feature extraction module, the training module, the prediction module, and the pricing module.