Cooperative Coevolution versus Monolithic Evolutionary Search for Semi-Supervised Tabular Classification

Jamal Toutouh

cs.NEApr 1, 2026

Cooperative Coevolution versus Monolithic Evolutionary Search for Semi-Supervised Tabular Classification

Authors: Jamal Toutouh

Organizations: ITIS, Universidad de Málaga Spain

Abstract

This paper studies semi-supervised tabular classification in the extreme low-label regime using lightweight base learners. The paper proposes a cooperative coevolutionary method (CC-SSL) that evolves (i) two feature-subset views and (ii) a pseudo-labeling policy, and compares it to a matched monolithic evolutionary baseline (EA-SSL) and three lightweight SSL baselines. Experiments on 25 OpenML datasets with labeled fractions {1%,5%,10%} evaluate test MacroF1 and accuracy, together with evolutionary and pseudo-label diagnostics. CC-SSL and EA-SSL achieve higher median test MacroF1 than the lightweight baselines, with the largest separations at 1% labeled data. Most CC-SSL vs. EA-SSL comparisons are statistical draws on final test performance. EA-SSL shows higher best-so-far fitness and higher diversity during search, while time-to-target is comparable and generations-to-target favors EA-SSL in several multiclass settings. Pseudo-label volume, ProbeDrop, and validation optimism show no significant differences between CC-SSL and EA-SSL under the shared protocol.

Figures & tables

Figure 1. Parameter tuning results summary

$lf$	Method	Min	Median	IQR	Max	Wins
0.01	CC-SSL	0.540	0.675	0.214	0.948	7
0.01	EA-SSL	0.579	0.694	0.235	0.953	7
0.01	ST	0.357	0.484	0.167	0.936	-
0.01	LS	0.343	0.421	0.090	0.760	-
0.01	HCo	0.343	0.496	0.185	0.912	-
0.01	Best Class.	0.522	0.632	0.146	0.948	-

Table 1. Descriptive statistics for MacroF1 results for 11 binary datasets

$lf$	Method	Min	Median	IQR	Max	Wins
0.01	CC-SSL	0.368	0.615	0.255	0.872	11
0.01	EA-SSL	0.387	0.628	0.275	0.913	11
0.01	ST	0.313	0.399	0.242	0.709	-
0.01	LS	0.022	0.171	0.181	0.383	-
0.01	HCo	0.295	0.463	0.146	0.661	-
0.01	Best Class.	0.320	0.558	0.205	0.777	-

Table 2. Descriptive statistics for MacroF1 results for 11 multiclass datasets

Panel of Figure 2 — Figure 2. Median best-so-far fitness per generation

Panel of Figure 3 — Figure 3. Median population diversity per generation, measured as mean pairwise distance

Panel of Figure 4 — Figure 4. Result distributions of TTT.

Panel of Figure 5 — Figure 5. Result distributions of GTT.

Panel of Figure 6 — Figure 6. Distributions of pseudo-labels added

Panel of Figure 7 — Figure 7. Distributions of optimism metric

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

ID	Name	Rows	Feat.	Classes
3	kr-vs-kp	3196	37	2
15	breast-w	699	10	2
29	credit-approval	690	16	2
31	credit-g	1000	21	2
37	diabetes	768	9	2
38	sick	3772	30	2

Appendix

Table 3. Benchmark datasets (OpenML-CC18) and basic characteristics. Rows = number of observations

n

, Feat. = number of features

p

, Classes = number of classes

cl

Stage 1: Population sizes and generations
Configuration parameter(s)	Tentative values
$(\|popA\|,\|popB\|,G)$	$\{(6,6,50),\ (8,8,38),\ (10,10,30)\}$
Stage 2: Evolutionary operator probabilities
$p^{(A)}_{cx}$	$\{0.70,\ 0.85,\ 0.95\}$
$p^{(A)}_{mut}$	$\{0.25,\ 0.45,\ 0.65\}$
$p^{(B)}_{cx}$	$\{0.70,\ 0.85,\ 0.95\}$

Appendix

Table 4. Parameter configurations explored in the sensitivity analysis

$lf$	Method	Min	Median	IQR	Max	Wins
0.01	CC-SSL	0.632	0.836	0.179	0.963	5
0.01	EA-SSL	0.639	0.836	0.189	0.966	5
0.01	ST	0.537	0.667	0.234	0.943	-
0.01	LS	0.522	0.654	0.233	0.939	-
0.01	HCo	0.522	0.690	0.212	0.939	-
0.01	Best Regular	0.653	0.714	0.172	0.952	-

Appendix

Table 5. Descriptive statistics for accuracy results for 11 binary datasets

$lf$	Method	Min	Median	IQR	Max	Wins
0.01	CC-SSL	0.374	0.676	0.241	0.873	11
0.01	EA-SSL	0.387	0.670	0.256	0.913	11
0.01	ST	0.323	0.452	0.258	0.750	-
0.01	LS	0.102	0.257	0.232	0.519	-
0.01	HCo	0.316	0.473	0.173	0.678	-
0.01	Best Regular	0.327	0.548	0.194	0.782	-

Appendix

Table 6. Descriptive statistics for accuracy results for 11 multiclass datasets

Method	3	15	29	31	37	38	44	50	1049	1067	1480
$lf$ =0.01
CC-SSL	0.850(0.026)	0.948(0.020)	0.835(0.069)	0.540(0.063)	0.675(0.049)	0.842(0.041)	0.893(0.016)	0.561(0.073)	0.650(0.052)	0.652(0.034)	0.612(0.044)
EA-SSL	0.876(0.030)	0.953(0.010)	0.835(0.042)	0.579(0.046)	0.694(0.069)	0.851(0.026)	0.897(0.028)	0.579(0.068)	0.657(0.078)	0.649(0.032)	0.608(0.046)
ST	0.450(0.292)	0.936(0.066)	0.357(0.070)	0.412(0.091)	0.445(0.191)	0.484(0.048)	0.668(0.243)	0.530(0.068)	0.587(0.089)	0.608(0.148)	0.417(0.135)
LS	0.343(0.020)	0.760(0.145)	0.357(0.070)	0.412(0.000)	0.426(0.155)	0.484(0.000)	0.377(0.002)	0.395(0.000)	0.468(0.014)	0.503(0.083)	0.421(0.214)
HCo	0.343(0.314)	0.912(0.098)	0.375(0.250)	0.412(0.101)	0.407(0.081)	0.496(0.048)	0.592(0.345)	0.516(0.075)	0.597(0.117)	0.639(0.064)	0.484(0.165)

Appendix

Table 7. MacroF1 results comparison for 11 binary datasets, median(IQR). Shaded: CC-SSL/EA-SSL significantly better than all baselines, Bonferroni-corrected Wilcoxon, p < 0.01

Method	11	12	14	16	22	23	28	46	54	307	1468
$lf$ =0.01
CC-SSL	0.538(0.031)	0.872(0.065)	0.563(0.054)	0.728(0.095)	0.615(0.048)	0.418(0.069)	0.823(0.090)	0.855(0.235)	0.503(0.060)	0.368(0.040)	0.688(0.062)
EA-SSL	0.539(0.039)	0.913(0.018)	0.593(0.070)	0.745(0.125)	0.628(0.037)	0.409(0.084)	0.881(0.036)	0.864(0.047)	0.521(0.075)	0.387(0.072)	0.687(0.066)
ST	0.406(0.088)	0.709(0.061)	0.329(0.188)	0.621(0.130)	0.399(0.101)	0.353(0.132)	0.532(0.159)	0.599(0.389)	0.313(0.078)	0.317(0.061)	0.318(0.149)
LS	0.383(0.307)	0.346(0.166)	0.118(0.093)	0.077(0.036)	0.186(0.056)	0.171(0.065)	0.022(0.005)	0.228(0.099)	0.259(0.090)	0.047(0.061)	0.022(0.000)
HCo	0.406(0.072)	0.661(0.100)	0.463(0.077)	0.619(0.093)	0.407(0.105)	0.353(0.095)	0.519(0.141)	0.565(0.358)	0.386(0.193)	0.295(0.063)	0.487(0.161)

Appendix

Table 8. MacroF1 results comparison for 11 multiclass datasets, median(IQR). Shaded: CC-SSL/EA-SSL significantly better than all baselines, Bonferroni-corrected Wilcoxon, p < 0.01

Method	3	15	29	31	37	38	44	50	1049	1067	1480
$lf$ =0.01
CC-SSL	0.852(0.025)	0.952(0.019)	0.836(0.063)	0.643(0.068)	0.719(0.063)	0.963(0.007)	0.898(0.016)	0.632(0.071)	0.854(0.039)	0.820(0.035)	0.674(0.029)
EA-SSL	0.877(0.029)	0.957(0.010)	0.836(0.041)	0.667(0.041)	0.732(0.056)	0.966(0.007)	0.902(0.026)	0.639(0.057)	0.877(0.041)	0.825(0.021)	0.669(0.051)
ST	0.537(0.161)	0.943(0.051)	0.556(0.120)	0.667(0.343)	0.602(0.223)	0.937(0.005)	0.677(0.196)	0.597(0.055)	0.829(0.070)	0.837(0.037)	0.663(0.106)
LS	0.522(0.045)	0.819(0.087)	0.556(0.120)	0.700(0.000)	0.654(0.043)	0.939(0.000)	0.606(0.001)	0.653(0.000)	0.879(0.000)	0.847(0.003)	0.594(0.366)
HCo	0.522(0.165)	0.924(0.076)	0.522(0.193)	0.690(0.150)	0.649(0.163)	0.939(0.004)	0.695(0.202)	0.615(0.053)	0.815(0.086)	0.836(0.036)	0.611(0.160)

Appendix

Table 9. Accuracy results comparison for 11 binary datasets, median(IQR). Shaded: CC-SSL/EA-SSL significantly better than all baselines, Bonferroni-corrected Wilcoxon, p < 0.01

Method	11	12	14	16	22	23	28	46	54	307	1468
$lf$ =0.01
CC-SSL	0.676(0.043)	0.873(0.067)	0.567(0.053)	0.728(0.095)	0.620(0.049)	0.432(0.031)	0.825(0.080)	0.864(0.183)	0.504(0.045)	0.374(0.044)	0.682(0.048)
EA-SSL	0.670(0.066)	0.913(0.018)	0.593(0.062)	0.753(0.131)	0.630(0.030)	0.421(0.036)	0.881(0.037)	0.873(0.044)	0.520(0.071)	0.387(0.067)	0.688(0.074)
ST	0.457(0.109)	0.750(0.067)	0.423(0.110)	0.630(0.123)	0.452(0.070)	0.369(0.064)	0.606(0.141)	0.635(0.154)	0.350(0.080)	0.323(0.065)	0.340(0.191)
LS	0.489(0.420)	0.408(0.190)	0.197(0.101)	0.177(0.064)	0.257(0.082)	0.346(0.111)	0.102(0.003)	0.519(0.279)	0.343(0.054)	0.114(0.055)	0.111(0.000)
HCo	0.452(0.142)	0.678(0.086)	0.473(0.062)	0.618(0.093)	0.432(0.082)	0.373(0.074)	0.557(0.115)	0.630(0.133)	0.398(0.121)	0.316(0.072)	0.512(0.181)

Appendix

Table 10. Accuracy results comparison for 11 multiclass datasets, median(IQR). Shaded: CC-SSL/EA-SSL significantly better than all baselines, Bonferroni-corrected Wilcoxon, p < 0.01

Gene	Symbol	Type	Domain / constraint
View builder $a$ (Population $\mathcal{A}$ )
Mask (view 1)	$m^{(1)}$	binary vector	$\{0,1\}^{d}$ , $\lVert m^{(1)}\rVert_{0}\geq k_{\min}$
Mask (view 2)	$m^{(2)}$	binary vector	$\{0,1\}^{d}$ , $\lVert m^{(2)}\rVert_{0}\geq k_{\min}$
Projection flag (view 1)	$p^{(1)}$	binary	$\{0,1\}$
Projection flag (view 2)	$p^{(2)}$	binary	$\{0,1\}$
Projection dim (view 1)	$k^{(1)}$	integer	$\{2,\dots,k_{\max}\}$

Appendix

Table 11. Genotype specification shared by CC-SSL and EA-SSL. CC-SSL evolves

a

and

b

in separate populations; EA-SSL concatenates the same genes into a single genotype

Semi-Supervised Learning

Semi-supervised learning uses both labeled and unlabeled data to improve prediction when reliable annotations are scarce.

See all 164 papers

Multi-Objective Evolutionary Semi-Supervised Learning for Resource-Aware Smart City Classification
Oct 20, 2026Francisco José Sedeño, Jamal ToutouhSemi-Supervised Learning Tabular Classification
ITIS Software, University of Malaga, Malaga, Spain
When Does Self-Supervised Pretraining Help Tabular Models? A Study of Label Scarcity and Missing Data
Aug 25, 2026Sahand MazroueiSelf-Supervised Pre-Training Tabular ML
Faculty of Mathematics and Computer Science, Kharazmi University
SemiScope: Disentangling Classifier Tuning and Joint Optimization in Semi-Supervised Security Classification
Jun 30, 2026Rui Shu, Tianpei Xia, Jingzhu HeSelf-Training Cybersecurity
North Carolina State University · ShanghaiTech University

Oct 20, 2026J/K move · Enter open · S save

Multi-Objective Evolutionary Semi-Supervised Learning for Resource-Aware Smart City Classification

Francisco José Sedeño, Jamal Toutouh

ITIS Software, University of Malaga, Malaga, Spain

Semi-Supervised Learning Tabular Classification Feature Selection Multi-Objective Evolutionary Optimization

Open paper AskPDF arXiv

Smart-city applications often combine scarce labeled data with abundant unlabeled observations. Semi-supervised learning (SSL) exploits both labeled and unlabeled data to improve classification under limited supervision. This article presents MOEA-SSL, a multi-objective evolutionary approach for lightweight semi-supervised tabular classification that jointly optimizes predictive quality and feature usage. Experimental evaluation on three Smart City classification datasets, considering 1%, 5%, and 10% labeled data, shows that MOEA-SSL achieves competitive predictive performance compared with established lightweight SSL methods while significantly reducing feature usage relative to EA-SSL in all evaluated conditions. The computed Pareto fronts also provide lower-feature-use alternatives with limited degradation in classification performance.

Abstract

Semi-Supervised Learning

Explore similar work

Multi-Objective Evolutionary Semi-Supervised Learning for Resource-Aware Smart City Classification

When Does Self-Supervised Pretraining Help Tabular Models? A Study of Label Scarcity and Missing Data

SemiScope: Disentangling Classifier Tuning and Joint Optimization in Semi-Supervised Security Classification