Bootstrapped Time-Window Event Prediction for Software Applications
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for predicting events in software applications are limited by biased training data and inefficient use of time-domain data, leading to inaccurate predictions and increased resource consumption.
Innovation Solution
A time-domain bootstrapped event prediction model is trained using randomly selected bootstrap division points to divide training data into non-overlapping time windows, enabling the use of both 'baked' and 'unbaked' data, and ensembling multiple models to improve accuracy and reduce resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained using all available training data at once, then the model can learn from comprehensive data, but the processing time and memory requirements increase significantly
Solution Approach 1:
The training data is divided into multiple non-overlapping time windows, and the model is trained separately on each time window. This segmentation allows the model to process data in smaller batches, reducing memory requirements and processing time while maintaining prediction accuracy through ensembling of multiple models trained on different time segments.
2Measurement precision
If traditional machine learning models are trained using all available training data at once, then the model can learn from comprehensive data, but the memory requirements increase significantly
Solution Approach 1:
The training data is divided into multiple non-overlapping time windows, and the model is trained separately on each time window. This segmentation allows the model to process data in smaller batches, reducing memory requirements while maintaining prediction accuracy through ensembling of multiple models trained on different time segments.
3Measurement precision
If machine learning models are trained on rare one-time events with limited data, then the model can predict these specific events, but the generalizability of the model is greatly limited
Solution Approach 1:
The training data is divided into multiple non-overlapping time windows, creating multiple training datasets from the same comprehensive data. This allows the model to learn patterns across different time periods, improving generalizability while maintaining accuracy for rare events through the ensembling of multiple models trained on different time segments.
Data Source
AI summary
Certain aspects of the present disclosure provide techniques for training and using time-domain bootstrapped event prediction models to predict the occurrence of an event within a software application. An example method generally includes receiving a data set of user activity within a software application. A request to predict a likelihood of an event occurring with respect to the software application based on the user activity is received. A likelihood of the event occurring is predicted using an event prediction model. The event prediction model is generally configured to predict the likelihood of the event occurring based on a likelihood over each of a plurality of non-overlapping time windows. A likelihood of the event occurring within a first time window is conditioned on a likelihood of the event occurring within a second time window. One or more actions are taken within the software application based on the predicted likelihood.