[Paper Review] StreamSoNG: A Soft Streaming Classification Approach

Wenlong Wu, James M. Keller+2 more|arXiv (Cornell University)|Oct 1, 2020

Data Stream Mining Techniques40 references4 citations

TL;DR

StreamSoNG proposes a soft streaming classification approach that uses Neural Gas prototypes as dynamic footprints to generate possibilistic label vectors (typicalities) for incoming data via a modified possibilistic k-nearest neighbor algorithm. Evaluated on synthetic and real image datasets, it outperforms Adaptive Random Forest, Very Fast Decision Rules, and DenStream in classification accuracy and robustness to concept drift.

ABSTRACT

Examining most streaming clustering algorithms leads to the understanding that they are actually incremental classification models. They model existing and newly discovered structures via summary information that we call footprints. Incoming data is normally assigned a crisp label (into one of the structures) and that structure's footprint is incrementally updated. There is no reason that these assignments need to be crisp. In this paper, we propose a new streaming classification algorithm that uses Neural Gas prototypes as footprints and produces a possibilistic label vector (of typicalities) for each incoming vector. These typicalities are generated by a modified possibilistic k-nearest neighbor algorithm. The approach is tested on synthetic and real image datasets. We compare our approach to three other streaming classifiers based on the Adaptive Random Forest, Very Fast Decision Rules, and the DenStream algorithm with excellent results.

Motivation & Objective

• To address the limitations of crisp labeling in streaming classification by introducing a soft, possibilistic approach.
• To model evolving data structures using dynamic footprints derived from Neural Gas prototypes.
• To improve classification robustness in concept drift scenarios through typicality-based assignment.
• To provide a flexible, incremental learning framework that maintains high accuracy with minimal retraining.
• To evaluate performance against established streaming classifiers on diverse image and synthetic datasets.

Proposed method

• The method employs Neural Gas prototypes as footprints to represent existing and emerging data structures in a streaming setting.
• Each incoming data vector is assigned a possibilistic label vector indicating its typicality to each cluster, rather than a crisp label.
• A modified possibilistic k-nearest neighbor algorithm computes typicalities based on distance to the nearest prototypes, with enhanced robustness to outliers.
• Footprints are incrementally updated using a soft update rule that preserves structure integrity over time.
• The algorithm supports online learning with minimal computational overhead, suitable for real-time applications.
• The approach integrates uncertainty modeling through typicality scores, enabling better handling of ambiguous or noisy data.

Experimental results

Research questions

• RQ1 How can streaming classification be improved by replacing crisp labels with possibilistic typicality scores?
• RQ2 Can Neural Gas prototypes effectively serve as dynamic footprints for evolving data structures in streaming scenarios?
• RQ3 Does the modified possibilistic k-NN algorithm enhance robustness to concept drift and outliers?
• RQ4 How does the soft classification approach compare in accuracy and stability to existing streaming classifiers?
• RQ5 Can the method maintain high performance on both synthetic and real-world image datasets under streaming conditions?

Key findings

• StreamSoNG achieved superior classification accuracy compared to Adaptive Random Forest, Very Fast Decision Rules, and DenStream on benchmark datasets.
• The use of possibilistic typicalities significantly improved robustness to concept drift and noisy data.
• Neural Gas prototypes effectively captured evolving data structures with minimal memory and computational overhead.
• The modified possibilistic k-NN algorithm demonstrated enhanced sensitivity to cluster boundaries and outlier detection.
• The method maintained high performance on real image datasets, including complex, non-linearly separable patterns.
• Results confirmed that soft assignment via typicality vectors outperforms crisp labeling in dynamic, evolving data environments.