Encoder-Decoder RNN Forecasting for Missing Time-Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for multi-step time series forecasting with missing data suffer from inaccuracies due to imputation techniques that destroy useful information and are not scalable or efficient, especially when data missingness occurs in medium/large windows with high total variation.
Innovation Solution
Employing a system and method using encoder decoder-based RNN learning, where input data is segregated into available and missing data blocks, processed by separate multi-layer RNN encoders, and a combined context vector is generated to forecast target variables without imputation, utilizing backpropagation through time (BPTT) for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If imputation techniques are used to fill missing values, then the data completeness is improved, but the useful information contained in the missingness patterns is destroyed
Solution Approach 1:
The patent extracts and separately processes missing data blocks from the input sequence, feeding them to a dedicated RNN encoder branch. This extraction approach allows the model to preserve the missingness patterns as explicit input features rather than filling them, thereby maintaining the information about why and how data is missing while still achieving complete data processing.
Solution Approach 2:
The patent segments the input data into available data blocks and missing data blocks, processing each through separate RNN encoder branches. This segmentation allows differential handling of different data types, where missing blocks are processed to extract contextual information about the missingness pattern itself, rather than being imputed with potentially misleading values.
2Quantity of substance
If imputation is performed before prediction, then the input data completeness is improved, but the forecasting accuracy deteriorates in scenarios with medium/large windows of consecutive missing time points
Solution Approach 1:
The patent performs preliminary encoding of missing data blocks through a dedicated RNN encoder branch before the main prediction process. This preliminary action extracts contextual information about the missingness patterns and integrates it into the combined context vector, enabling accurate forecasting without requiring imputation of the actual missing values.
Solution Approach 2:
The patent introduces an intermediary processing mechanism where missing data blocks are fed to a separate RNN encoder that acts as a mediator. This intermediary encoder transforms the missing data into contextual representations that inform the prediction process, rather than directly imputing values that could propagate errors through the forecasting model.
3Quantity of substance
If mean imputation is used for high missing data rate, then the data completeness is improved, but the accuracy of results deteriorates
Solution Approach 1:
The patent applies local quality by treating available data and missing data differently through separate RNN encoder branches. Instead of applying a uniform imputation strategy like mean imputation globally, the model locally processes missing data blocks to extract their specific contextual patterns, allowing the handling strategy to adapt to the local characteristics of each missing data segment.
Data Source
AI summary
This disclosure relates generally to a method and system for encoder decoder based RNN learning for time series forecasting in presence of missing data. The present disclosure employs atleast one multi-layer RNN encoder and a decoder for time series forecasting. The method receives a plurality of input data comprising a sequential data transformed into a plurality of windows to obtain a plurality of features. Further, the plurality of features is segregated into an available data blocks and a missing data blocks. The first multi-layer RNN encoder fetches the available data blocks and the second multi-layer RNN encoder fetches the missing data blocks to forecast the target variable from the multi-step time series data. The decoder input is generated by appending the combined context vector with an exogeneous variable at each time step.