FSF Architecture for IoT Traffic Forecasting in SDNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current IoT traffic forecasting methods in Software-Defined Networks (SDNs) lack effective feature selection techniques, leading to suboptimal Quality of Service (QoS) routing, particularly when dealing with limited data sets and nonlinear relationships, and existing methods either require extensive data for Deep Learning or rely on fixed feature selection thresholds.
Innovation Solution
The development of an end-to-end trainable feature selection-forecasting architecture (FSF) that dynamically selects features using trainable importance scores and gamma-gated feature selection units, allowing for automatic and adaptive feature selection, thereby improving forecasting accuracy and reducing training time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Deep Learning techniques (LSTM, 1D CNN) are used for forecasting, then forecasting performance is improved, but data set size requirement increases
Solution Approach 1:
The patent extracts and selects only the most relevant features from the input data using feature selection techniques (mRMR, mutual information, correlation analysis) before feeding them to the forecasting model. This reduces the dimensionality of the data from the full dataset to a curated subset of informative features, allowing the model to achieve high performance with less data.
Solution Approach 2:
The patent applies different processing strategies to different features based on their characteristics. Important features are selected and retained while less important ones are discarded. The forecasting model processes selected features with higher computational resources while handling remaining features more efficiently, optimizing the trade-off between performance and data requirements.
2Device complexity
If feature selection is not used, then model complexity is reduced, but forecasting error increases
Solution Approach 1:
The patent performs feature selection as a preliminary step before the main forecasting process. By pre-selecting informative features using statistical measures (mutual information, correlation coefficients, mRMR scoring), the system prepares optimized input data that reduces forecasting error without requiring complex model architectures.
Solution Approach 2:
The patent introduces feature selection metrics and scoring mechanisms as intermediary components between the raw data and the forecasting model. These intermediaries (mRMR scorer, correlation analyzer, feature importance calculators) act as mediators that transform raw features into a refined set of informative inputs, improving forecasting accuracy without increasing model complexity.
3Device complexity
If fixed feature selection threshold is used, then computational complexity is reduced, but adaptability to different data conditions decreases
Solution Approach 1:
The patent implements dynamic feature selection where the selection process adapts to different data conditions and contexts. Instead of using fixed thresholds, the system calculates feature importance scores (mutual information, correlation coefficients) that are specific to each forecasting task and data set, allowing the model to adaptively select features based on the actual data characteristics.
Solution Approach 2:
The patent changes the parameters of feature selection based on the specific forecasting problem and data characteristics. Different scoring functions (mRMR, mutual information, correlation) are applied depending on the data conditions, and the selection threshold is adjusted dynamically rather than being fixed, enabling the system to adapt to varying data scenarios.
4Measurement precision
If more features are selected, then forecasting accuracy is improved, but training time increases
Solution Approach 1:
The patent applies partial feature selection by selecting only the most informative subset of features rather than using all available features. By applying multiple selection criteria (mRMR, mutual information, correlation analysis) and choosing a limited number of top-ranked features, the system achieves sufficient forecasting accuracy with reduced training time compared to using all features.
Solution Approach 2:
The patent replaces brute-force approaches (using all features or exhaustive search) with efficient statistical measures for feature evaluation. Instead of computationally expensive methods, the system uses analytical solutions (correlation coefficients, mutual information calculations) that provide good feature rankings with lower computational cost, reducing training time while maintaining accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The subject of invention is related to a method which develops a novel feature selection-forecasting architecture, called FSF, that is able to perform automatic, dynamic selection of features in order to minimize the forecasting error in IoT traffic forecasting for predictive Quality of Service (QoS) routing in Software-Defined Networks (SDNs). By minimizing the forecasting error, the FSF architecture allows higher quality IoT traffic predictions that improve the QoS in routing on SDNs.