Event Log Feature Transformation for ML Predictive Accuracy

The patent transforms binary event indicators into continuous feature values by applying decay functions over time. Instead of simple 0/1 values, each event generates a feature value that decays continuously from an initial value toward zero, with the decay rate controlled by a decay factor. This parameter transformation preserves temporal information while maintaining computational tractability through the use of standardized decay functions.

The patent adds a temporal dimension to binary event data by introducing time-based decay features. Each event is not just recorded as occurred or not, but is transformed into a feature that exists across multiple time points with diminishing magnitude. This dimensional expansion from static binary values to dynamic decaying features captures both occurrence and temporal patterns without requiring storage of complete event histories.

The patent extracts only the essential temporal and frequency information from complete event logs by computing aggregate feature values. Instead of storing individual event timestamps and details, the system computes feature values that summarize event frequency and recency through decay functions. This extraction process retains predictive information while dramatically reducing storage requirements by eliminating redundant raw event data.

The patent uses computationally inexpensive decay functions that can be applied to event data without requiring permanent storage of raw events. The decay features serve as temporary representations that capture essential patterns for prediction while allowing original detailed event logs to be discarded or not stored at all. This approach trades computational processing for storage efficiency.

The patent transforms binary event parameters into continuous decaying feature parameters that encode temporal patterns. This parameter enrichment provides machine learning models with more informative inputs that capture event frequency and recency, enabling better predictive accuracy. The transformation maintains computational efficiency because the decay features can be pre-computed and updated incrementally without requiring complex processing during model training.

The patent performs feature transformation in advance before model training by computing decay-based feature values from event logs. This preliminary processing converts raw binary events into informative continuous features that are ready for immediate use in training. The pre-computed features eliminate the need for complex temporal processing during model training, maintaining training speed while significantly improving input quality and predictive accuracy.