Hazard Function Estimation for Time-Series Event Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques face limitations in accurately estimating the time of event occurrence from time-series data, particularly when dealing with complex temporal changes and handling time-series inputs, as they fail to grasp temporal variations and require pre-assumed distributions, leading to limited precision in estimating events like traffic accidents.
Innovation Solution
An event occurrence time learning device and method that utilize a hazard function to estimate the likelihood of event occurrence based on time-series data, incorporating a hazard estimation unit and parameter estimation unit to optimize the likelihood function, allowing for precise estimation without pre-assuming distributions and handling temporal changes through Recurrent Neural Networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional survival analysis and deep learning techniques are used to estimate event occurrence time, then a model of non-linear relationship can be constructed, but it requires pre-assuming event occurrence distribution which limits precision when complicated temporal changes are involved
Solution Approach 1:
The patent transforms the estimation approach by changing from assuming a fixed event occurrence distribution to dynamically estimating the hazard function h(t) that represents the instantaneous likelihood of event occurrence at time t. This parameter change allows the model to adapt to complicated temporal changes in different domains (traffic accidents, medical events, equipment failures) without requiring pre-assumed distribution forms.
2Device complexity
If pre-assumed event occurrence distribution is used, then the model construction is simplified, but it becomes impossible to properly understand temporal changes in the inputs
Solution Approach 1:
The patent introduces the hazard function h(t) as an intermediary that bridges the input time-series data and the event occurrence time estimation. This intermediary function dynamically captures temporal changes in the input data without requiring pre-assumed distribution forms, thereby preserving temporal information while maintaining model tractability through the use of Recurrent Neural Networks to approximate the hazard function.
3Productivity
If conventional techniques estimate event occurrence time without handling time-series data, then calculation is simplified, but it is impossible to obtain information about movements and speeds of objects
Solution Approach 1:
The patent applies Recurrent Neural Networks to continuously process time-series data, maintaining a running understanding of the system state that incorporates temporal dynamics. This continuous processing enables the model to capture movement and speed information of objects over time while efficiently calculating the hazard function at each time step, thus preserving useful information without sacrificing calculation efficiency.
Data Source
AI summary
A hazard estimation unit 21 estimates a likelihood of an occurrence of an event according to a hazard function, with respect to each of a plurality of pieces of time-series data that are a series of multiple pieces of data to which an event occurrence time relevant to the data is given in advance and that include time-series data in which the event did not occur and time-series data in which the event occurred. A parameter estimation unit 22 estimates a parameter of the hazard function so as to optimize a likelihood function expressed by including the event occurrence time given with respect to each of the plurality of pieces of time-series data and the likelihood of the occurrence of the event estimated with respect to each of the plurality of pieces of time-series data.