Mobile Ad Delivery Pacing Using Location Conversion Prediction

The patent segments the location prediction task into multiple components: geo-fence event detection, geo-block event detection, and conversion probability prediction. Each component is handled by separate machine learning models trained on specific datasets, allowing the system to process location data in modular stages rather than requiring a single large-scale model, thus reducing overall processing requirements while maintaining accuracy.

The system performs preliminary actions by pre-processing location data into structured formats (geo-fence events, geo-block events) before prediction. Historical location data is pre-trained to create location prediction models that can be applied to new data without requiring intensive real-time processing, shifting computational burden to offline training phases.

The patent extracts only the most relevant features from comprehensive location data for prediction. The system identifies and extracts key parameters such as geo-fence events, geo-block events, and location patterns that are most predictive of conversion, discarding redundant information. This extraction approach maintains prediction accuracy while significantly reducing the volume of data requiring processing.

Location data is pre-processed and structured into standardized formats (geo-fence events, geo-block events) in advance. This preliminary structuring organizes raw location data into prediction-ready formats with extracted features, eliminating the need for complex real-time data processing during prediction operations.

The patent applies local quality by calibrating location prediction models with campaign-specific parameters and thresholds. Different campaigns can have customized prediction criteria, geo-fence configurations, and conversion definitions tailored to their specific objectives. This allows each campaign to be optimized independently without requiring complete model redesign, managing complexity through localized adjustments rather than global changes.

The system uses universal location prediction models that can serve multiple campaigns simultaneously. The core prediction engine processes location data in a standardized way, and campaign-specific optimization is achieved through configurable parameters and thresholds rather than requiring separate models for each campaign. This multi-functional approach reduces overall system complexity while maintaining campaign-specific optimization capabilities.