Customized Time Series Forecasting via Query Vector Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time series forecasting methods struggle with data deficiency and long-range temporal pattern capture, especially when dealing with diverse temporal patterns across different time series.
Innovation Solution
The proposed framework, known as Customized Time Series Forecasting (CTSF), enhances the expressiveness of encoded temporal patterns by querying and aggregating relevant historic information, and customizes model parameters for each time series using a bidirectional recurrent neural network (RNN) and a historical temporal pattern (HTP) graph.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If global modeling is used to capture temporal patterns across different time series, then pattern generalization is improved, but the ability to capture diverse temporal patterns of individual time series deteriorates
Solution Approach 1:
The patent segments the modeling process into two distinct components: a global model that captures general temporal patterns across all time series, and individual models for each time series that capture specific temporal patterns. This segmentation allows the system to simultaneously achieve pattern generalization through the global model and precise temporal pattern capture through individual models, resolving the contradiction between these two requirements.
2Measurement precision
If a forecasting model is trained for each target time series to achieve customized treatment, then temporal pattern specialization is improved, but data deficiency problems worsen due to lack of sufficient training data
Solution Approach 1:
The patent merges the training process by having individual time series models trained on their respective data while simultaneously being guided by the global model trained on aggregated data from all time series. This combining approach allows each individual model to benefit from both its own time series data and the collective information from all time series, effectively increasing the available training data and resolving the data deficiency problem while maintaining temporal pattern specialization.
3Reliability
If massive training data is used to train neural network models, then model performance is improved, but the requirement for data availability worsens when data is not accessible in real-world scenarios
Solution Approach 1:
The patent creates a universal modeling framework where a global model serves multiple time series simultaneously. This global model is trained on aggregated data from all time series and can be applied to forecast any individual time series, even those with limited data. This multi-functionality allows the system to achieve high model performance through the global model while being adaptable to scenarios with varying data availability, resolving the contradiction between reliability and data availability flexibility.
4Ease of manufacture
If short-term training data is available for a time series, then model training is easier, but the ability to capture long-range historical temporal patterns deteriorates
Solution Approach 1:
The patent introduces the global model as an intermediary that bridges the gap between short-term individual time series data and long-range temporal patterns. The global model is trained on aggregated data from all time series, thereby capturing long-range historical temporal patterns that individual short time series cannot capture alone. This intermediary global model then provides guidance and information to individual models, enabling them to capture long-range patterns even when trained on limited short-term data, thus resolving the contradiction between training ease and information loss.
Data Source
AI summary
The present teaching relates to method, system, medium, and implementations for machine learning. Upon receiving input data associated with a time series, hidden representations associated with the time series in a feature space are obtained and used to generate a query vector in a query space. Such generated query vector is then used to query relevant historic information related to the time series. The query vector and the relevant historic information are aggregated to generate at least one queried vector, which is aggregated with the hidden representations to generate enriched hidden representations that enhance the expressiveness of the hidden representations.