---
kind: concept
slug: support-vector-machine
title: Support vector machine
summary: >-
  A classifier in several roles: an RBF SVM on fixed features (97.35% on CK+), a linear SVC on TF-IDF spam (0.991 accuracy), a linear SVM on EHR features (AUC up to 89.8%), an SVM on terrain sensor statistics (95.87%), an SVR baseline, and an active-learning query model. Theory: gradient descent on linear networks reaches the hard-margin SVM. Numbers are author-reported.
tags: [classification, kernel-methods, supervised-learning, semi-supervised-learning, implicit-bias, text-classification, class-imbalance, clinical-decision-support, tactile-sensing, regression]
papers: ["1305.1707", "1604.03225", "1608.01413", "1806.00468", "1904.08067", "2103.07678", "2204.06343", "2104.07784", "2112.15538", "2405.11619", "2406.13547", "2411.04700", "2412.17991", "2506.08274", "2604.26818", "2605.04666", "2609.07408"]
related: [active-learning, adaboost, ensemble-learning, feature-selection, feature-scaling, constrained-conditional-models, implicit-bias-of-gradient-descent, interpretable-machine-learning, one-class-classification, text-classification, text-embedding, random-forest, intrusion-detection, lasso-regression, semi-supervised-learning, multi-layer-perceptron, long-short-term-memory, temporal-convolutional-network]
status: draft
updated_at: 2026-11-01
---

## Overview

Ten of the sixteen cards train an SVM and use its output inside a larger
system, or as the classifier itself. It gives a class label for a fixed-length
vector [arxiv:1604.03225 §5.3], a spam label for TF-IDF vectors
[arxiv:2405.11619 p. 8], or a weight per firewall rule [arxiv:2406.13547 §3].
It gives a decision label for an EHR order from ranked features
[arxiv:2605.04666 p. 3], and a label for a spiral drawing from image statistics
[arxiv:2609.07408 p. 11]. In active learning it gives the decision values that
pick the next pixels to label [arxiv:2104.07784 §4.1] [alg. 1]. It gives the
final label for a graph-based
semi-supervised classifier [arxiv:2604.26818 §3]. It gives a grasp label from
sensor statistics [arxiv:2204.06343 §4.2.2] and a terrain label from IMU and
force-torque statistics [arxiv:2411.04700 §3.1]. It also gives confidence scores
that a later search combines [arxiv:1608.01413 §4.1] [§4.3] [§4.4].

Two cards use the SVM as a point of comparison. The joint-angle card uses a
support vector regressor as its only non-sequential baseline
[arxiv:2412.17991 §2.1.6]. A benchmark treats the SVM as one of 14 models and
asks how sensitive it is to feature scaling [arxiv:2506.08274 §5.1]. The other
five cards do not run an SVM. A theory card treats it as a limit, not a
component [arxiv:1806.00468 §3]. A survey lists it among text classifiers
[arxiv:1904.08067 §4.6] but runs no experiments [arxiv:1904.08067 §1] [§5]. A
review of class imbalance names SVM-based remedies and reports no SVM numbers
[arxiv:1305.1707 §2.2] [§4]. A review of disease diagnosis lists SVM among the
most frequent algorithms in its 102 studies, and gives no SVM results of its own
[arxiv:2112.15538 §6] [fig. 10]. A review of remote sensing for mineral
exploration describes SVMs for lithology, alteration and mineral mapping, and
runs no experiments of its own [arxiv:2103.07678 §5.2.2] [§1].

The shared part of the trained cases is the training step: the features are
fixed before the SVM sees them, and the SVM is fitted on labelled examples
[arxiv:1604.03225 §5.3] [arxiv:1608.01413 §4] [arxiv:2605.04666 p. 3]
[arxiv:2204.06343 §4.2.2] [arxiv:2411.04700 §3.1]. The active-learning card
refits the SVM on the growing labelled set at each round [arxiv:2104.07784
alg. 1]. The graph-cut card is the
exception on the label side: the SVM is fitted on labels that a graph assigns
to unlabelled points, and only on those it is confident about
[arxiv:2604.26818 §3]. What differs is how the SVM is set up. The facial
expression card uses an RBF kernel from libsvm and tunes C and gamma by grid
search [arxiv:1604.03225 §5.3]. The grasp card searches kernel, C and gamma over
a grid, separately for each gripper, action and object set, and standardises
the feature vector [arxiv:2204.06343 §4.1] [§4.2.2]. The terrain card searches C,
gamma and the kernel, and reports the best value per input set
[arxiv:2411.04700 §3.1] [tab. 2] [tab. 3] [tab. 4]. The arithmetic card uses the
Illinois-SL package with its default settings and does not state the kernel, C
or the regulariser [arxiv:1608.01413 §4.4]. The spam card uses a linear kernel
and does not state C [arxiv:2405.11619 p. 9]. The firewall card fixes C at 0.5,
chosen from five values, and compares an L1 with an L2 penalty
[arxiv:2406.13547 §4.1]. The EHR card uses a linear SVM and does not state C,
the library or the feature scaling [arxiv:2605.04666 p. 3]. The image card runs
the SVM with library default settings and gives no parameter values
[arxiv:2609.07408 p. 11]. The joint-angle card gives one fixed setting for its
regressor, gamma = 0.01 and C = 1, and describes no search
[arxiv:2412.17991 §2.1.6].

The theory card is different in kind. It is a theory paper with no experiments.
It does not use an SVM as a component. It describes which margin solution
gradient descent reaches on a linear network, and for fully connected networks
that solution is the hard-margin SVM direction [arxiv:1806.00468 §3]. The
result rests on assumptions the paper states but does not prove
[arxiv:1806.00468 §3].

Seventeen cards are compiled so far, so this page covers the uses they describe
and nothing more.

## Variants

### RBF SVM on boosted features

Card: [Geometric Feature-Based Facial Expression Recognition in Image Sequences Using Multi-Class AdaBoost and Support Vector Machines](/papers/1604.03225).
AdaBoost first selects features. For each selected feature the paper keeps the
maximum displacement over the sequence (type one) or the maximum change in angle
and distance (type two), so each video gives one vector with two values per
feature [arxiv:1604.03225 §5.3]. An RBF SVM is trained on these vectors. With
400 selected features (800 dimensions) it reports 97.35% average accuracy under
5-fold cross-validation on CK+ [arxiv:1604.03225 tab. 6]. With 200 features it
reports 95.50% and with 100 features 93.20%; the AdaBoost-DTW classifier itself
reports 95.17% [arxiv:1604.03225 tab. 3] [tab. 4] [tab. 5]. So the SVM beats
that classifier only at 200 and 400 features.

### Binary and multiclass SVM scorers in constrained inference

Card: [Solving General Arithmetic Word Problems](/papers/1608.01413).
A binary SVM scores whether each quantity is needed for the solution. A
multiclass SVM scores which operation sits at the lowest common ancestor of
each quantity pair [arxiv:1608.01413 §4.3] [§4.4]. A beam search of width 200
combines these scores under a positive-answer and, for "how many" questions, an
integer-answer constraint [arxiv:1608.01413 §4.1]. The relevance SVM labels
94.7% of AI2 quantities correctly in the relaxed sense, and 89.1% of problems
fully correctly [arxiv:1608.01413 tab. 5]. The authors report 72.0% on AI2 with
all constraints and 78.0% with the positive constraint only [arxiv:1608.01413
tab. 6], so the constraint setting changes the headline number.

### Linear SVC on TF-IDF text

Card: [Novel Interpretable and Robust Web-based AI Platform for Phishing Email Detection](/papers/2405.11619).
The paper merges six public spam collections into 82,486 emails and turns each
email into TF-IDF weights [arxiv:2405.11619 p. 7] [eq. 1]. A linear-kernel SVC
is trained on an 80/20 split, with random state 42 [arxiv:2405.11619 p. 8] [p. 9].
Its chosen model reports accuracy 0.991 and F1 0.99 on the 16,498-email test
split [arxiv:2405.11619 tab. 2]. The same table gives the SVC on Word2Vec
features an F1 of 0.71 and 0.82 [arxiv:2405.11619 tab. 2], so the choice of
features moved the score more than the choice of classifier did.

Two cautions come from the card. The 0.991 is the best of eleven rows, all
scored on the same test split, and the paper does not say whether any choice was
made on a validation split [arxiv:2405.11619 tab. 2] [p. 8]. The labels are spam
and ham, not phishing, so the score does not show phishing detection
[arxiv:2405.11619 p. 11].

### Linear SVM on rule-firing features

Card: [ModSec-Learn: Boosting ModSecurity with Machine Learning](/papers/2406.13547).
Each SQL-injection rule of the OWASP Core Rule Set becomes one binary feature,
which is 1 when the rule fires on a query [arxiv:2406.13547 §3]. A linear SVM
with an L1 or an L2 penalty learns one weight per rule, and the learned weights
replace the fixed severity points of ModSecurity [arxiv:2406.13547 §3].
The training set has 40,000 samples and the test set 10,000, balanced between
classes [arxiv:2406.13547 §4.1]. Author-reported TPR at 1% FPR at paranoia level
4 is 99.02 for the SVM and 68.55 for vanilla ModSecurity [arxiv:2406.13547 tab. 1].

The L1 and L2 SVM rows in that table are identical, so the penalty shows up only
as sparsity in the weights. The 18 zero-weight rules are reported for logistic
regression with L1, not for the SVM [arxiv:2406.13547 §4.4]. The paper states
that the other hyperparameters keep their defaults [arxiv:2406.13547 §4.1].

### Linear SVM on top-ranked features

Card: [Feature importance analysis for patient management decisions](/papers/2605.04666).
Each candidate temporal feature gets a single-feature AUC, computed on 2,900
patients. The top 1, 3 or 30 features by that AUC train a linear SVM, which is
tested on the remaining patients [arxiv:2605.04666 p. 3] [p. 5]. The feature set
has 9,223 candidates, but the SVM tables cover only 10 lab orders and 10
medications [arxiv:2605.04666 p. 3] [tab. 1] [tab. 2]. On those 20 decisions the
held-out AUC is 69.6–87.7% with the top feature and 80.5–89.8% with the top 30
[arxiv:2605.04666 tab. 1] [tab. 2]. The paper reports no external baseline
[arxiv:2605.04666 tab. 1] [tab. 2]. Its ranking is greedy, and the authors say
this may not be the best choice [arxiv:2605.04666 p. 5].

### SVM with default settings on image statistics

Card: [Impact of canny edge detection preprocessing on performance of machine learning models for Parkinson's disease classification](/papers/2609.07408).
Each spiral image is reduced to eleven statistics. Recursive feature elimination
keeps, for each model, the features ranked 2 or lower [arxiv:2609.07408 p. 9]
[p. 10]. The SVM is one of nine classifiers run with library default settings
[arxiv:2609.07408 p. 11]. The feature selection runs on the whole training set
before cross-validation, so the reported CV scores may be optimistic
[arxiv:2609.07408 alg. 3] [p. 22]. Its AUC is 0.85 on the original training images and
0.83 after Canny preprocessing [arxiv:2609.07408 fig. 12] [fig. 13]. On the
augmented set DS2 it is 0.94 [arxiv:2609.07408 fig. 13]. Its model size grows
from 7 KB on DS0 to 113–304 KB on DS2 [arxiv:2609.07408 fig. 40], and its
prediction time on DS2 is 9,000–30,000 ms, the highest of the nine models there
[arxiv:2609.07408 fig. 41]. The paper does not define how size or time were
measured [arxiv:2609.07408 p. 19] [p. 20].

### Grid-searched SVM on grasp-sensor statistics

Card: [Single-grasp Deformable Object Discrimination: the Effect of Gripper Morphology, Sensing Modalities, and Action Parameters](/papers/2204.06343).
Each squeeze gives a time series from the gripper sensors. The paper computes
per-channel statistics, among them skewness, maximum, kurtosis, median, standard
deviation and sum, and standardises the resulting vector [arxiv:2204.06343 §4.1].
The SVM is tuned over kernel, C and gamma, separately for each gripper, action
and object set [arxiv:2204.06343 §4.2.2]. The stated C range is 1e-5 to 1e-3
[arxiv:2204.06343 §4.2.2].

Author-reported accuracy depends on the gripper and the set. On the objects set
with the qb SoftHand, the best SVM reaches about 86% at the slow speed
[arxiv:2204.06343 §5.1]. On 20 polyurethane foams with the Barrett Hand it
reaches 94–99% across actions, with the SVM and the LSTM [arxiv:2204.06343 §5.2]
[fig. 5]. On the same foams with the Robotiq 2F-85 it reaches 68–77% across
speeds [arxiv:2204.06343 §5.2]. Transfer across finger configurations fails for
the qb SoftHand and the Barrett Hand [arxiv:2204.06343 §5.4] [fig. 6]. The
paper calls the SVM much less demanding than the LSTM, but gives no timings or
sizes [arxiv:2204.06343 §7].

### Terrain classifier on force-torque and IMU statistics

Card: [Field Assessment of Force Torque Sensors for Planetary Rover Navigation](/papers/2411.04700).
Each force-torque sensor (FTS) contributes 35 inputs, from six channels and a
lever estimate, each with five statistics. Each IMU contributes 15 accelerations
with the same statistics, so the full input has 225 values [arxiv:2411.04700 §3].
The SVM is a support vector classifier, tuned by grid search over C, gamma and
the kernel [arxiv:2411.04700 §3.1]. The grid covers C in {0.1, 1, 10, 100},
gamma in {1, 0.1, 0.01, 0.001}, and the linear, RBF, polynomial and sigmoid
kernels [arxiv:2411.04700 §3.1]. The data are split 75/25 into training and test
sets [arxiv:2411.04700 §3.1] [tab. 2].

The best test accuracies are 85.84% with IMU only (polynomial kernel), 95.58%
with FTS only (linear kernel) and 95.87% with both (linear kernel)
[arxiv:2411.04700 tab. 2] [tab. 3] [tab. 4]. The FTS-only and combined SVMs
reach 100% training accuracy, so the test figures are the ones to read
[arxiv:2411.04700 tab. 3] [tab. 4]. The paper does not state whether the SVM
inputs were scaled, and the code scales only when a flag is passed
[arxiv:2411.04700 §3.1] [code: ml/svm.py#L93-L97]. Timing on one laptop CPU
gives 24.3 s for the full grid search, 0.13 s for one fit with the best
parameters, and 7.6 ms to predict the 339-sample test set
[arxiv:2411.04700 tab. 8]. The field test uses one rover, a half-scale
model of the ExoMars rover, on four terrain classes [arxiv:2411.04700 §1] [§3.1].

### Support vector regressor as a baseline for joint angles

Card: [Online Adaptation for Myographic Control of Natural Dexterous Hand and Finger Movements](/papers/2412.17991).
The paper regresses the wrist and finger angles from forearm sensor signals. Its
baseline is a support vector regressor (SVR) with a Gaussian radial basis
kernel, gamma = 0.01 and C = 1, on a single 200 ms feature window
[arxiv:2412.17991 §2.1.6] [tab. 1]. It is the only non-sequential baseline
[arxiv:2412.17991 §2.1.6]. On the standard session with nine subjects it reaches
11.3° angular RMSE and r² 0.72. The temporal convolutional and LSTM models reach
8.4° and 8.0° with r² 0.85 [arxiv:2412.17991 §3.1]. On the free-form offline
session with eight subjects the SVR reaches 15.4°, against 10.4° and 9.9° for
the two sequence models [arxiv:2412.17991 §3.2]. The card gives the SVR a
response delay of 120 ms on average, and under 25 ms for the sequence models
[arxiv:2412.17991 §3.2] [fig. 6].

### RBF SVM as the query model in active learning

Card: [A survey of active learning algorithms for supervised remote sensing image classification](/papers/2104.07784).
The loop trains a classifier on a small labelled set of pixels, scores the
unlabelled pool, and sends the q best pixels to a user for labelling. Each
round then repeats [arxiv:2104.07784 alg. 1]. With an RBF-kernel SVM, the paper
tunes the kernel and penalty by 5-fold cross-validation, but the grid and the
chosen values are not stated [arxiv:2104.07784 §6.2]. Margin sampling takes, for
a one-against-all SVM, the smallest absolute decision value over classes, which
is the pixel closest to a boundary [arxiv:2104.07784 §4.1]. The paper reports
margin sampling as generally the best of the SVM heuristics it tested, against
random sampling, on three datasets [arxiv:2104.07784 fig. 4] [§7]. Adding a
diversity term to MCLU seems to help slightly, except on AVIRIS Indian Pines
[arxiv:2104.07784 fig. 6] [§7]. The results appear only as curves, so no
numbers are given in the text [arxiv:2104.07784 §7]. The paper also says that
KL-max needs one retrained model per candidate and is hard to use with SVMs
[arxiv:2104.07784 §5.1].

## When to use which

| Situation | Use | Why |
|---|---|---|
| Fixed-length vector per example, features selected beforehand, several hundred dimensions | RBF SVM with grid-searched C and gamma | This is the setup behind the best reported result in [arxiv:1604.03225 tab. 6]. The grid ranges and feature scaling are not stated [arxiv:1604.03225 §5.3]. |
| Sparse, high-dimensional text such as TF-IDF over a corpus | Linear-kernel SVC | It has the best accuracy in [arxiv:2405.11619 tab. 2] on one split. C is not stated [arxiv:2405.11619 p. 9], so set it and report it. Pick the features first: Word2Vec rows in the same table score much lower. |
| One auditable weight per binary feature, with some features dropped | Linear SVM or logistic regression with an L1 penalty | The learned weights replace hand-set points [arxiv:2406.13547 §3]. Zero weights remove rules, which the authors report for logistic regression at PL4 [arxiv:2406.13547 §4.4]. The results are on one author-built SQLi test set. |
| Many decisions, few features each, and a per-decision AUC target | Linear SVM on the top-k features by single-feature AUC | Top-30 has the higher range in both tables: 80.5–89.8% against 69.6–87.7% for top-1 [arxiv:2605.04666 tab. 1] [tab. 2]. The ranking is greedy [arxiv:2605.04666 p. 5]. Rank on training patients only, because the paper does not state how the SVM was trained [arxiv:2605.04666 p. 3]. |
| Latency or model size matters, and the features are image statistics on augmented data | Compare cheaper models first; use the SVM only if its accuracy justifies the cost | In [arxiv:2609.07408 fig. 41] the SVM is the slowest of nine models on DS2. Random forest had the higher AUC on DS2 (0.97 against 0.94) and a fixed size of 61 KB [arxiv:2609.07408 fig. 13] [fig. 40]. |
| Per-trial sensor statistics, one label per trial, and a gripper that the model will see in training | SVM on the statistics with a grid search, or an LSTM on the raw series | On the Barrett Hand foams the SVM and LSTM both reach 94–99% [arxiv:2204.06343 §5.2] [fig. 5]. The paper reports no timings for the SVM [arxiv:2204.06343 §7]. Transfer across finger configurations fails for two of the four devices [arxiv:2204.06343 §5.4]. |
| Terrain or contact class from a fixed window of sensor statistics, where inference must be cheap on a CPU | Linear or RBF SVC on FTS statistics, with the scaling stated and reported | With FTS the SVM reaches 95.58% against 96.17% for a one-hidden-layer network [arxiv:2411.04700 tab. 3] [tab. 6]. One fit takes 0.13 s and 339 predictions take 7.6 ms on one laptop [arxiv:2411.04700 tab. 8]. Scaling is not stated [arxiv:2411.04700 §3.1]. |
| Several decisions that must agree, such as a valid expression tree | SVM scorers feeding a constrained search | The constraints are enforced at inference over the SVM scores [arxiv:1608.01413 §4.1]. Wrong quantity-schema extraction causes most relevance errors on IL [arxiv:1608.01413 §5.5]. |
| Rare class, where the majority class dominates the fit | An SVM as one remedy among sampling, cost-sensitive and boosting methods, checked on your data | The review gives no numbers for SVM remedies. It lists the boosting SVM as ignoring the class distribution [arxiv:1305.1707 tab. 2]. Compare it with the other families it lists [arxiv:1305.1707 §2]. |
| Real-valued features on very different scales, any kernel | Scale the features, and check the effect on your own data | In the benchmark, SVM accuracy changed significantly with the scaler (Wilcoxon p = 0.0001) [arxiv:2506.08274 app. A]. On Dry Bean, the best scaler raised SVM accuracy from 0.5803 to 0.9273 [arxiv:2506.08274 app. A]. The benchmark's kernel is disputed (see Contradictions). |
| Continuous targets from a sliding window, with a non-sequential baseline needed for comparison | SVR with a fixed window as the floor a sequence model must beat | On the standard session the SVR is 2.9–3.3° worse in angular RMSE than the sequence models [arxiv:2412.17991 §3.1]. The card reports the SVR only on this dataset [arxiv:2412.17991 §3.1] [§3.2]. |
| Labels are expensive, a large unlabelled pool exists, and a human labels one batch per round | Margin sampling with an RBF SVM, with a diversity term if batches repeat | Margin sampling was generally best on the SVM curves, against random sampling [arxiv:2104.07784 fig. 4] [§7]. The curves give no numbers in the text, and kernel and C are not stated [§6.2]. Re-run the comparison on your data before relying on it. KL-max is hard to use with SVMs [§5.1]. |
| Remote-sensing mapping of lithology, alteration or minerals from bands or hyperspectral data, with labelled samples | SVM as one of several classifiers, compared with random forest and CNNs on your data | The review lists SVM, random forest and CNNs as methods applied to this task, with no comparative numbers [arxiv:2103.07678 §5.2] [tab. 2]. It says no single approach solves every case, so the choice must be tested per dataset [arxiv:2103.07678 §5, §6]. |
| You need a reproducible classifier | Fix kernel, C, penalty and scaling in writing | No single card gives all of these. The firewall card gives C = 0.5 [arxiv:2406.13547 §4.1]. The arithmetic card points to library defaults [arxiv:1608.01413 §4.4]. The benchmark's text and its code name different kernels [arxiv:2506.08274 §3.2]. The terrain card leaves scaling unstated [arxiv:2411.04700 §3.1]. |
| Few labels, many unlabelled points, and a neighbourhood graph that suits the features | Graph-inferred labels, then an SVM on the confident points | In the authors' tables this beats the supervised SVM with the same kernel in all 36 rows and manifold regularization in 29 [arxiv:2604.26818 fig. 4]. The threshold ε and the graph weights set the result, and the paper does not state the cubic kernel or how γ maps to LIBSVM's C [arxiv:2604.26818 §6.2] [§6.1]. Reproduce one row before relying on it. |
| You want to predict which margin solution gradient descent reaches on a deep linear model | Theory only: fully connected gives the hard-margin SVM direction | The claim holds under the paper's assumptions, which it takes as hypotheses [arxiv:1806.00468 §3]. The convolutional case gives a different penalty [arxiv:1806.00468 §3, Thm. 2a]. No experiments back either claim [arxiv:1806.00468 §1–§6]. |

## Implementation sketch

Three pipelines are sketched. The first is the RBF pipeline after feature
selection. The grid and scaling are open choices the paper leaves unstated. The
second is the linear pipeline on sparse text. The third is one round of margin
sampling with an RBF SVM. The arithmetic pipeline needs its scorers and beam
search, and the theory card has no algorithm to implement.

```pseudocode
# RBF pipeline. Per video: maximum signed value over frames, per selected feature  [§5.3]
x = max_over_frames(selected_feature_values)  # length 2 * M for M features
# Grid search over C and gamma; ranges not stated  [§5.3]
best = grid_search(C_range, gamma_range, cross_validate)
# Scaling before the kernel: not stated  [§5.3]
# Train an RBF-kernel SVM with the chosen settings  [§5.3]
model = train_svm(X_train, y_train, kernel="rbf", C=best.C, gamma=best.gamma)
# One label per video  [§5.3]
label = model.predict(x)
```

```pseudocode
# Linear pipeline on TF-IDF text  [arxiv:2405.11619 p. 7-9]
# Fit the vocabulary and IDF on training text only; the paper does not say it did
vectoriser = fit_tfidf(train_texts)
X_train = vectoriser.transform(train_texts)
X_test = vectoriser.transform(test_texts)
# Linear kernel; C is not stated, so choose and report it  [arxiv:2405.11619 p. 9]
model = fit_linear_svc(X_train, y_train, C=C, random_state=42)
# Label 1 is spam, 0 is ham  [arxiv:2405.11619 p. 11]
label = model.predict(X_test)
```

```pseudocode
# One round of margin sampling with an RBF SVM  [arxiv:2104.07784 §4.1] [alg. 1]
def margin_sampling_round(X_lab, y_lab, X_pool, q):
    # Retrain on the labelled pixels; C and gamma not stated  [§6.2]
    model = train_svm(X_lab, y_lab, kernel="rbf")
    # One-against-all decision values per pool pixel  [§4.1]
    scores = model.decision_function(X_pool)  # pool x classes
    # Smallest absolute value over classes: closest to a boundary  [§4.1]
    margin = abs(scores).min(axis=1)  # pool
    # Send the q smallest margins to the user for labelling  [alg. 1]
    return argsort(margin)[:q]
```

## Contradictions

**Kernel of the benchmark's SVM.** The benchmark's text says the SVM uses an RBF
kernel [arxiv:2506.08274 §3.2]. Its card records that the released classification
code uses a linear kernel and that the RBF kernel is used only for regression
([Impact of Feature Scaling](/papers/2506.08274), Paper vs code). Which kernel
produced the reported SVM accuracy is therefore unresolved, and this page does
not settle it.

**SVM on augmented images.** The image card's conclusion says the SVM is among
the models that "excelled on augmented datasets" [arxiv:2609.07408 p. 35]. Its
median DS2 accuracy for the SVM is 0.50 [arxiv:2609.07408 fig. 42], while the
pooled accuracy from the DS2 confusion matrix is 0.86 [arxiv:2609.07408 fig. 16].
The card computed the pooled value itself, and the paper does not explain the gap.
The DS2 AUC of 0.94 [arxiv:2609.07408 fig. 13] is a different measure and does
not resolve it.

**Incompatible C ranges.** The grasp card searches C from 1e-5 to 1e-3
[arxiv:2204.06343 §4.2.2]. The terrain card
selects C = 0.1 from {0.1, 1, 10, 100} [arxiv:2411.04700 tab. 2]. The two grids
do not overlap, and the papers use different sensors, features and scaling. So
the C values are not transferable between these tasks, and this page does not
say which range is right for a given problem.

Apart from these, the cards use SVMs for different purposes, and their numbers
are not comparable. The classification cards report accuracy, AUC or F1 on
different tasks and datasets. The firewall card reports TPR at a fixed FPR. The
joint-angle card reports angular RMSE and r² for a regressor. The theory card,
the survey and the three reviews report no numbers for the SVM.

**Kernel SVM and dimension.** The mineral review says the SVM projects data into
a space of smaller dimension [arxiv:2103.07678 §5.2.2]. The card for that paper
notes that kernel SVMs usually map to a higher-dimensional space, and the paper
gives no kernel setting. The page does not settle which description applies to
the SVMs the reviewed studies used.

The survey's pitfalls section reads a comparative remark about SVM efficiency in
its section 6 as a summary of cited work, not as a result of the survey
[arxiv:1904.08067 §6]. This page does not use that remark.

## Papers

- [Geometric Feature-Based Facial Expression Recognition in Image Sequences Using Multi-Class AdaBoost and Support Vector Machines](/papers/1604.03225): RBF SVM on AdaBoost-selected landmark-motion features, 97.35% author-reported on CK+.
- [Solving General Arithmetic Word Problems](/papers/1608.01413): binary and multiclass SVM scorers inside a constrained beam search over expression trees.
- [Text Classification Algorithms: A Survey](/papers/1904.08067): survey that lists SVM among text classifiers with a string-subsequence kernel; no experiments, no SVM numbers.
- [Implicit Bias of Gradient Descent on Linear Convolutional Networks](/papers/1806.00468): theory that gradient descent on fully connected linear networks converges in direction to the hard-margin SVM, under stated assumptions.
- [Single-grasp Deformable Object Discrimination: the Effect of Gripper Morphology, Sensing Modalities, and Action Parameters](/papers/2204.06343): grid-searched SVM on grasp-sensor statistics; author-reported 94–99% on Barrett Hand foams, level with an LSTM.
- [Novel Interpretable and Robust Web-based AI Platform for Phishing Email Detection](/papers/2405.11619): linear SVC on TF-IDF features for spam email, 0.991 accuracy author-reported on one split.
- [ModSec-Learn: Boosting ModSecurity with Machine Learning](/papers/2406.13547): linear SVM and logistic regression over binary rule-firing features, learned weights in place of ModSecurity severity points.
- [Field Assessment of Force Torque Sensors for Planetary Rover Navigation](/papers/2411.04700): SVM on FTS and IMU statistics for terrain classes, 95.87% author-reported test accuracy with both sensors.
- [Online Adaptation for Myographic Control of Natural Dexterous Hand and Finger Movements](/papers/2412.17991): SVR with a Gaussian kernel as the only non-sequential baseline for joint-angle regression; 11.3° against 8.0–8.4° RMSE on the standard session.
- [The Impact of Feature Scaling In Machine Learning: Effects on Regression and Classification Tasks](/papers/2506.08274): benchmark in which the SVM's accuracy is significantly sensitive to the scaler; the SVM kernel is disputed between text and code.
- [Feature importance analysis for patient management decisions](/papers/2605.04666): linear SVM on the top 1, 3 or 30 temporal EHR features ranked by single-feature AUC; 69.6–89.8% author-reported AUC on 20 decisions.
- [Semi-supervised learning with max-margin graph cuts](/papers/2604.26818): graph-inferred labels, then a kernel hinge-loss SVM on confident points; author-reported lower error than supervised SVM and manifold regularization on UCI.
- [Impact of canny edge detection preprocessing on performance of machine learning models for Parkinson's disease classification](/papers/2609.07408): SVM with library defaults on eleven image statistics; AUC 0.94 on augmented images, with the accuracy figures in conflict.
- [Class Imbalance Problem in Data Mining Review](/papers/1305.1707): review of class-imbalance remedies that names SVM-based methods; no experiments and no numbers.
- [A survey of active learning algorithms for supervised remote sensing image classification](/papers/2104.07784): RBF SVM as the query model; margin sampling generally best in author-reported curves, no tabulated numbers.
- [Machine learning based disease diagnosis: A comprehensive review](/papers/2112.15538): review of 102 studies in which SVM is among the most frequent algorithms; no experiments.
- [A review of machine learning in processing remote sensing data for mineral exploration](/papers/2103.07678): review that describes SVM for lithology, alteration and mineral mapping; no experiments, no SVM numbers.