spdal

Single-Pass Discard-After-Learn — hyperellipsoid classifiers for online streaming data.

Each training sample is processed once and then discarded. No full dataset is ever stored. All classifiers implement scikit-learn's partial_fit / predict interface.

---

Installation

pip install spdal

Development mode:

git clone https://github.com/PeemapatW/single-pass-discard-after-learn
cd single-pass-discard-after-learn
pip install -e ".[dev]"

---

Quick Start

from sklearn.datasets import make_classification
from spdal import LRHE

X, y = make_classification(n_samples=500, random_state=42)

clf = LRHE()
clf.fit(X[:400], y[:400])
print(clf.predict(X[400:]))          # array of class labels
print(len(clf.neuron_list))          # number of learned prototypes

Incremental learning

from spdal import TRACED
import numpy as np

X, y = make_classification(n_samples=500, random_state=42)
classes = np.unique(y)

clf = TRACED()
for i in range(0, 400, 50):
    clf.partial_fit(X[i:i+50], y[i:i+50], classes=classes)

print(clf.predict(X[400:]))

---

Classifiers

Class	Full name	Year
VEBF	Versatile Elliptic Basis Function	2010
SCIL	Streaming Chunk Incremental Learning	2019
LRHE	Learning with Recoil in Hyperellipsoidal Structure	2020
SHEF	Scalable Hyper-Ellipsoidal Function	2020
D4	Diversion of Data Distribution Direction	2026
TRACED	Trend-Adaptive Classification with Ellipsoidal Disambiguation	2026

Comparison

Feature	VEBF	SCIL	LRHE	SHEF	D4	TRACED
Learning mode	Single	Chunk	Single	Single	Chunk	Chunk
Neurons per class	Multiple	Multiple	Multiple	Multiple	One	Multiple
Width formula	Init: avg pairwise dist; Update: width + center shift; Merge: sqrt(2π|λ|)	Update: width + center shift; Expand: sqrt(1+η·max_psi)·w; Merge: 1.96 sqrt(|λ|/n)	Init: avg pairwise dist; Update: width + center shift; Merge: sqrt(2π|λ|)	Update: recursive covariance; Width = r·sqrt(λ)	Init: avg pairwise dist; Update: α·sqrt(2πλ) + (1-α)·(width + center shift)	Init: mean NN dist; Update: γ·r·sqrt(λ) + (1-γ)·(width + center shift); Merge: r·sqrt(λ)
Creation threshold	Hyperellipsoidal test	Hyperellipsoidal test	Hyperellipsoidal test	Euclidean distance (dynamic threshold)	N/A (always merges)	Euclidean distance (dynamic threshold)
Ambiguity resolution	—	—	Shift-and-shrink	Discriminant projection	Coincident: principal-axis subspace	Coincident: principal-axis subspace; Exterior: EMA displacement + expansion

---

Parameters

VEBF

from spdal import VEBF
clf = VEBF(theta=0, delta=1)

Parameter	Default	Description
theta	0	Overlap threshold for neuron merging
delta	1	Width scaling from pairwise distances for initial width [$\delta > 0$]

SCIL

from spdal import SCIL
clf = SCIL(N0=3, eta=2, delta=1, theta=0)

Parameter	Default	Description
N0	3	Min samples for an active neuron in prediction
eta	2	Width expansion scaling factor
delta	1	Width scaling from pairwise distances for initial width [$\delta > 0$]
theta	0	Overlap threshold for neuron merging

LRHE

from spdal import LRHE
clf = LRHE(alpha=0.99, theta=0, delta=1)

Parameter	Default	Description
alpha	0.99	Shrink multiplier during recoil [$\alpha \in [0,1]$]; paper-recommended (gradual shrinking)
theta	0	Overlap threshold for neuron merging
delta	1	Width scaling from pairwise distances for initial width [$\delta > 0$]

SHEF

from spdal import SHEF
clf = SHEF(M=3, r=1.5)

Parameter	Default	Description
M	3	Min samples before adaptive threshold triggers
r	1.5	Ellipsoid radius scaling constant [$r > 0$]

D4

from spdal import D4
clf = D4(width_parameter=1, reduce_dims=0, delta=1, norm=2, threshold=15)

Parameter	Default	Description
width_parameter	1	Blend: 1 = pure statistical width, 0 = pure expansion-based [$\text{width_parameter} \in [0,1]$]
reduce_dims	0	Axes to drop in coincident-region disambiguation
delta	1	Width scaling from pairwise distances for initial width [$\delta > 0$]
norm	2	Lp norm for distance calculation
r	1.5	Statistical ellipsoid radius width scaling factor [$r > 0$]
threshold	15	Angle threshold (degrees) for axis pairing

D4 maintains one neuron per class. When two nearest neurons belong to different classes, it select their axes using parallel and compactness criteria and assigns the class with the smaller projected distance in that subspace.

Note: Theorem 2 of the D4 paper contains a sign error in the proof. This does not affect the algorithm or experimental results. See technical note for details.

TRACED

from spdal import TRACED
clf = TRACED(
    alpha=0.5, beta=0.01, delta=2, width_parameter=1,
    reduce_dims=1, N0=3, r=2.507, norm=2,
    method='overlap-outside', distance_metric='boundary',threshold=15
)

Parameter	Default	Description
alpha	0.5	EMA weight for displacement smoothing (0 = disabled) [$\alpha \in [0,1]$]
beta	0.01	EMA weight for expansion-rate smoothing (0 = disabled) [$\beta \in [0,1]$]
delta	2	Initial dynamic threshold scaling (mean NN distance × delta) [$\delta > 0$]
width_parameter	1	Blend: 1 = pure statistical width, 0 = pure expansion-based [$\text{width_parameter} \in [0,1]$]
reduce_dims	1	Axes to drop in coincident-region disambiguation
N0	3	Min samples for an active neuron in prediction and adaptive threshold triggers
r	sqrt(2π)	Statistical ellipsoid radius width scaling factor [$r > 0$]
norm	2	Lp norm for distance calculation
method	'overlap-outside'	Corrections to apply: 'overlap', 'outside', or both
distance_metric	'boundary'	'boundary' or 'center'
threshold	15	Angle threshold (degrees) for axis pairing

TRACED resolves two ambiguous regions:

• Coincident (x inside multiple classes) — principal-axis subspace projection (like D4)
• Exterior (x outside all neurons) — predicts using EMA-smoothed displacement and expansion as a trend model

---

sklearn Interface

All classifiers are sklearn.base.BaseEstimator subclasses and support:

clf.fit(X, y)                              # full batch training
clf.partial_fit(X, y, classes=classes)     # incremental update
clf.predict(X)                             # returns array of class labels
clf.classes_                               # array of known class labels
clf.neuron_list                            # list of neuron dicts

Compatible with scikit-learn pipelines and cross-validation tools that support partial_fit.

---

Neuron Schema

Learned prototypes are stored in clf.neuron_list as a list of dicts:

{
    'y':             class_label,
    'center':        np.ndarray,     # prototype position
    'cov':           np.ndarray,     # covariance matrix
    'eig_component': np.ndarray,     # PCA eigenvectors
    'width':         np.ndarray,     # semi-axis lengths
    'n':             int,            # sample count
    # SCIL, D4, TRACED only:
    'variance':      np.ndarray,     # eigenvalues
    # TRACED only:
    'displacement':  np.ndarray,     # EMA displacement vector
    'expansion':     np.ndarray,     # EMA per-axis expansion rates
}

---

Development

# Run tests
pytest tests/ -v

# Run a single test class
pytest tests/test_classifiers.py::TestTRACED -v

# Build for PyPI
pip install build && python -m build

---

References

1. VEBF — Jaiyen, S., Lursinsap, C., & Phimoltares, S. (2010). A Very Fast Neural Learning for Classification Using Only New Incoming Datum. IEEE Transactions on Neural Networks, 21(3), 381–392. [paper]
2. SCIL — Junsawang, P., Phimoltares, S., & Lursinsap, C. (2019). Streaming chunk incremental learning for class-wise data stream classification with fast learning speed and low structural complexity. PLOS ONE, 14(9), e0220624. [paper]
3. LRHE — Jindadoungrut, K., Phimoltares, S., & Lursinsap, C. (2020). Neural Learning With Recoil Behavior in Hyperellipsoidal Structure. IEEE Access, 8, 114643–114655. [paper]
4. SHEF — Rungcharassang, P., & Lursinsap, C. (2020). Scalable Hyper-Ellipsoidal Function with Projection Ratio for Local Distributed Streaming Data Classification. IEEE Access, 8, 105460–105474. [paper]
5. D4 — Wongsriphisant, P., Plaimas, K., & Lursinsap, C. (2026). Markov-based continuous learning with diversion of data distribution direction for streaming data in limited memory. Expert Systems With Applications, 298, 129818. [paper]
   • Technical note (Theorem 2): docs/D4_theorem2_note.md
6. TRACED — Wongsriphisant, P., Plaimas, K., & Lursinsap, C. (2026). TRACED: Trend-Adaptive Classification with Ellipsoidal Disambiguation for Resolving Exterior and Coincident Regions in Data Streams. Information Sciences, 743, 123338. [paper]

---

Citation

If you use this library in your research, please cite the relevant paper(s):

@article{WONGSRIPHISANT2026D4,
    title = {Markov-based continuous learning with diversion of data distribution direction for streaming data in limited memory},
    journal = {Expert Systems with Applications},
    volume = {298},
    pages = {129818},
    year = {2026},
    issn = {0957-4174},
    doi = {https://doi.org/10.1016/j.eswa.2025.129818},
    url = {https://www.sciencedirect.com/science/article/pii/S0957417425034335},
    author = {Peemapat Wongsriphisant and Kitiporn Plaimas and Chidchanok Lursinsap},
}

@article{WONGSRIPHISANT2026TRACED,
    title = {TRACED: Trend-adaptive classification with ellipsoidal disambiguation for resolving exterior and coincident regions in data streams},
    journal = {Information Sciences},
    volume = {743},
    pages = {123338},
    year = {2026},
    issn = {0020-0255},
    doi = {https://doi.org/10.1016/j.ins.2026.123338},
    url = {https://www.sciencedirect.com/science/article/pii/S0020025526002690},
    author = {Peemapat Wongsriphisant and Kitiporn Plaimas and Chidchanok Lursinsap},
}

---

Notes

• The original monolithic implementation is preserved at deprecated/spdal.py for reference.
• Refactoring into the modular src/spdal/ package structure, docstrings, and parameter naming were performed by Claude (Anthropic) and reviewed by the project owner.

Changelog

See CHANGELOG.md. Latest — 0.4.0: VEBF, LRHE, SHEF and TRACED are brought closer to the algorithms as described in their papers, so their results may differ from 0.3.0; LRHE and SCIL gain small numerical safeguards; and D4 no longer takes an r argument. Details and paper references are in the changelog.