Time Series Pattern Prediction Using Bayesian Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time series prediction models face challenges in multi-step ahead prediction, handling outliers, and efficiently processing multiple time series, often requiring extensive calculations and suffering from accuracy issues due to accumulated errors.
Innovation Solution
A time series pattern prediction device and method utilizing a multi-layer Bayesian network to predict patterns across various time scales by segmenting time series data, generating segment pattern sets, and constructing a Bayesian network that trains causal relationships between different time scales.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional time series prediction models (ARIMA) are used, then prediction performance is guaranteed for time series satisfying assumptions, but the models cannot handle outliers and require specific assumptions like normality and independent identity distribution
Solution Approach 1:
The patent changes the fundamental parameters of the prediction model by using Bayesian networks with symbolic patterns instead of traditional statistical assumptions. This allows the model to handle outliers and non-normal data while maintaining prediction reliability through probabilistic reasoning and causal relationship modeling.
Solution Approach 2:
The patent replaces the mechanical statistical assumptions (normality, independence) with a Bayesian network framework that uses symbolic patterns and causal relationships. This substitution allows the system to handle complex, non-stationary time series data with outliers without requiring strict distributional assumptions.
2Productivity
If multi-step ahead prediction is performed using direct prediction method, then multiple future steps are predicted at once, but a large amount of calculation is required
Solution Approach 1:
The patent segments the time series data into multiple segments and creates a multi-layer Bayesian network where each layer corresponds to different time scales. This segmentation allows the model to perform multi-step ahead prediction by querying different layers, significantly reducing the calculation burden compared to direct prediction methods.
Solution Approach 2:
The patent introduces a new dimension by creating a multi-layer Bayesian network structure where layers represent different time scales. This dimensional transformation enables efficient multi-step prediction through the layered architecture, avoiding the computational explosion of direct multi-step prediction methods.
3Productivity
If iterative prediction method is used, then multiple future steps are predicted step by step, but accuracy decreases due to accumulated errors from feeding predicted values back as input
Solution Approach 1:
The patent segments the prediction task across multiple independent layers in the Bayesian network, each layer handling different time scales. This segmentation allows the model to predict multiple future steps without iterative feedback, eliminating accumulated errors while maintaining high accuracy through parallel probabilistic reasoning.
Solution Approach 2:
The patent introduces symbolic patterns as intermediaries between the time series data and prediction outputs. These symbolic patterns serve as mediators that capture causal relationships and temporal dependencies without requiring iterative feedback, thereby maintaining prediction accuracy while enabling multi-step ahead forecasting.
4Measurement precision
If tree-based models (LightGBM, XgBoost) are used for time series prediction, then prediction performance improves, but the models require a lot of calculations and are prone to overfitting
Solution Approach 1:
The patent replaces tree-based mechanical learning algorithms with a Bayesian network framework that uses symbolic patterns and causal reasoning. This substitution maintains high prediction performance while reducing calculation requirements and eliminating overfitting through probabilistic modeling and explicit causal relationship representation.
Solution Approach 2:
The patent changes the fundamental approach from parameter estimation (tree-based) to probabilistic reasoning (Bayesian networks). This parameter change allows the model to achieve high accuracy without extensive calculations or hyper-parameter tuning, while the probabilistic framework naturally prevents overfitting through regularization inherent in Bayesian inference.
5Productivity
If deep learning models (RNN, LSTM) are used for time series prediction, then prediction capability improves, but the models perform worse than tree-based models or require sensitive hyper-parameter tuning
Solution Approach 1:
The patent substitutes deep learning neural networks with a Bayesian network framework that uses symbolic patterns. This substitution eliminates the need for complex hyper-parameter tuning while maintaining or improving prediction capability through explicit causal relationship modeling and probabilistic reasoning that is more interpretable and easier to operate.
Solution Approach 2:
The patent changes from learning parameters (neural network weights) to probabilistic parameters (conditional probabilities in Bayesian networks). This parameter change simplifies the operational complexity by eliminating sensitive hyper-parameter tuning, while the symbolic pattern representation maintains high prediction capability through human-interpretable causal relationships.
6Measurement precision
If a multi-layer Bayesian network is used to predict patterns at various scales, then prediction accuracy and causal relationship identification improve, but the model structure becomes more complex
Solution Approach 1:
The patent segments the complex prediction task into multiple manageable layers, each handling specific time scales. This segmentation makes the model structure more organized and easier to manage, while the layered architecture naturally captures causal relationships across different temporal scales, improving prediction accuracy without proportionally increasing complexity.
Solution Approach 2:
The patent creates a universal multi-layer Bayesian network framework that can handle multiple time scales and predict various patterns simultaneously. This multi-functional design improves prediction accuracy across different horizons while the shared probabilistic framework reduces overall model complexity compared to separate models for each time scale.
Data Source
AI summary
Disclosed is a time series pattern prediction device, which includes a segment pattern set generation unit that divides time series data into a plurality of segments and generates a segment pattern set based on a plurality of unit patterns corresponding to the plurality of segments, a network generation unit that generates a Bayesian network based on the segment pattern set, and a time series pattern prediction unit that generates a prediction pattern based on the Bayesian network.