Meta-Ensemble Learning for Multi-Trait Optimization in Maize

Breeding: Combining Gradient Boosting, Random Forests, and Deep

Learning with SVM Integration

Dupuy Rony Charles

1,2 a

, Pascal Pultrini

and Andrea Tettamanzi

1 b

Universit

e C

ote d’Azur, I3S, Inria, Sophia Antipolis, France

Doriane Research Softare & Consulting, Av. Jean Medecin, Nice, France

Keywords:

Plant Breeding, Multi-Trait Selection Indices (MTSI), Meta-Ensemble Machine Learning.

Abstract:

Plant breeding aims to enhance traits such as yield, drought tolerance, and disease resistance. Traditional

Multi-Trait Selection Indices (MTSI) struggle with high-dimensional genomic data and complex trait inter-

actions. We present a meta-ensemble machine learning framework integrating Gradient Boosting, Random

Forest, and Deep Neural Networks (DNNs) with a Support Vector Machine (SVM) meta-model to address

these challenges. This meta-ensemble approach leverages the strengths of multiple algorithms for improved

predictive accuracy and robustness. Experiments on maize datasets show that our meta-ensemble signiﬁcantly

outperforms traditional MTSI methods and individual machine learning models. The meta-ensemble achieves

superior predictive accuracy and operational efﬁciency, with a marked reduction in mean squared error (MSE)

and consistent performance across validation sets. This study advances meta-ensemble machine learning in

plant breeding, providing a robust framework for multi-trait selection. Our approach improves trait prediction

reliability and sets a new standard in maize breeding, with potential applications in other crop species, enhanc-

ing agricultural productivity and resilience.

1 INTRODUCTION

Maize (Zea mays L.), one of the world’s most impor-

tant staple crops, is essential for global food security,

livestock feed, and industrial raw materials. However,

contemporary agricultural challenges, including cli-

mate change, pest infestations, and the rising demand

for higher yields, require breeding strategies that are

both more efﬁcient and innovative. Traditional breed-

ing methods, although effective in the past, often

struggle to optimize multiple agronomic traits simul-

taneously, especially when confronted with the com-

plexities of modern breeding programs.

Recent advancements in machine learning (ML)

techniques have emerged as promising solutions to

these challenges. ML excels in its ability to an-

alyze large datasets and identify intricate patterns,

making it a powerful tool for improving the preci-

sion and efﬁciency of breeding programs. Numer-

ous studies have demonstrated the applicability of

ML in various agricultural domains, including crop

https://orcid.org/0009-0006-4879-9933

https://orcid.org/0000-0002-8877-4654

yield prediction and phenotypic trait analysis (Ahmed

et al., 2024), (Reddy and Kumar, 2021), (Westhues

et al., 2021), (Chandana and Parthasarathy, 2022), and

(Crossa et al., 2017).

Multi-trait selection indices (MTSI) are crucial for

modern plant breeding, enabling breeders to improve

multiple agronomic traits simultaneously. Traits such

as yield, drought tolerance, disease resistance, and nu-

tritional quality often exhibit complex genetic corre-

lations. Traditional indices, such as those proposed

by (Smith et al., 1981) and (Hazel et al., 1994), have

been fundamental in this ﬁeld. However, these meth-

ods face limitations when applied to modern genomic

and phenotypic data, which is characterized by high-

dimensionality and non-linear trait interactions. Their

linear nature often constrains their ability to capture

these complexities.

To address these limitations, advanced ML-based

models, including ensemble methods, deep learning,

and support vector machines (SVM), offer a transfor-

mative approach. These models are capable of captur-

ing non-linear relationships and interactions among

traits, leading to more accurate and efﬁcient selection

876

Charles, D. R., Pultrini, P. and Tettamanzi, A.

Meta-Ensemble Learning for Multi-Trait Optimization in Maize Breeding: Combining Gradient Boosting, Random Forests, and Deep Learning with SVM Integration.

DOI: 10.5220/0013191900003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 876-883

ISBN: 978-989-758-737-5; ISSN: 2184-433X

strategies. Meta-ensemble models, which combine

the strengths of multiple base models, provide im-

proved prediction accuracy by capturing diverse pat-

terns and interactions in the data.

This paper investigates the application of a meta-

ensemble ML framework in maize breeding, inte-

grating Gradient Boosting, Random Forest, and Deep

Neural Networks (DNNs), with a Support Vector Ma-

chine (SVM) as a meta-model. We demonstrate how

this approach improves the efﬁciency and accuracy of

multi-trait selection, outperforming traditional MTSI

methods. Our contributions are as follows: (1) pre-

senting a novel ML-based MTSI that outperforms tra-

ditional indices, (2) showcasing its practical imple-

mentation using real-world genomic and phenotypic

data, and (3) providing a comprehensive analysis of

the algorithm’s performance, emphasizing its robust-

ness and scalability.

The remainder of the paper is structured as fol-

lows: Section 2 reviews related literature, Section 3

outlines the methodology, Section 4 presents the ex-

perimental results, and Section 5 discusses the impli-

cations of the ﬁndings and future research directions.

2 RELATED WORK

The complexity of improving multiple agronomic

traits simultaneously has driven signiﬁcant research

into multi-trait selection indices and their evolution.

This section provides context for the need to move

beyond traditional methods and explores the role of

machine learning in addressing these challenges.

Multi-trait selection indices (MTSI) have long

been pivotal in plant breeding, particularly for en-

hancing key agronomic traits such as yield, drought

tolerance, and disease resistance. Classic indices like

the Smith-Hazel index (Smith, 1936), (Hazel, 1943)

and the desired gains index have been instrumental

in selecting for multiple traits concurrently, optimiz-

ing the balance between competing trade-offs. How-

ever, these traditional approaches operate under the

assumption of simple additive correlations between

traits, which limits their capacity to capture the com-

plex genetic and phenotypic relationships inherent in

modern breeding programs (C

eron-Rojas and Crossa,

2018).

The advent of high-dimensional genomic data,

generated by next-generation sequencing technolo-

gies, has introduced non-linear relationships and in-

tricate interactions between traits (Mrode, 2014). Tra-

ditional MTSI methods are ill-equipped to manage

these complexities, as they lack the ability to model

interactions across multiple genetic loci and pheno-

typic traits. This limitation is particularly evident in

maize breeding, where traits like yield, moisture con-

tent, and disease resistance are inﬂuenced by both ad-

ditive and non-additive genetic factors (Crossa et al.,

2013). Consequently, the limitations of linear meth-

ods necessitate a shift toward more sophisticated ap-

proaches.

Multi-trait selection is integral to modern plant

breeding, allowing breeders to improve several key

traits simultaneously. However, traditional indices of-

ten fail to account for the complexity of genomic and

phenotypic data. The rise of machine learning (ML)

presents a powerful alternative, offering algorithms

that can capture non-linear relationships and complex

trait interactions prevalent in high-dimensional data

(Crossa et al., 2017).

ML techniques have been successfully applied to

genomic selection (GS) and phenotypic trait analy-

sis in various crops, signiﬁcantly improving upon tra-

ditional methods. Commonly used ML models in-

clude ensemble algorithms such as Random Forests

(Breiman, 2001) and Gradient Boosting Machines

(Friedman, 2001), as well as more advanced mod-

els like Deep Neural Networks (DNNs) (Montesinos-

opez et al., 2019). These methods excel at uncov-

ering complex interactions between genomic markers

and phenotypic traits, making them highly suitable for

multi-trait selection.

Random Forests and Gradient Boosting are partic-

ularly adept at handling non-linear relationships and

are frequently used in plant breeding to predict phe-

notypic traits (Spindel et al., 2015). These models

combine multiple decision trees to effectively manage

complex trait dependencies and feature importance

estimation. Furthermore, DNNs add value by pro-

viding advanced representation learning, especially in

high-dimensional datasets, enabling the capture of ab-

stract relationships within genomic data (Montesinos-

opez et al., 2019).

Despite the advantages of individual ML mod-

els, challenges such as overﬁtting and limited gen-

eralization across environments persist. To mitigate

these issues, meta-ensemble learning approaches have

been developed, combining the strengths of multi-

ple base models to enhance predictive accuracy and

robustness. Meta-ensembles aggregate outputs from

different ML models, improving both the accuracy

of multi-trait selection and the overall efﬁciency of

breeding programs (Gonz

alez-Camacho et al., 2018).

Recent studies in maize breeding have demon-

strated the superiority of meta-ensemble models over

both traditional selection indices and individual ML

models. For example, (Spindel et al., 2015) found

that combining models like Random Forests and

Meta-Ensemble Learning for Multi-Trait Optimization in Maize Breeding: Combining Gradient Boosting, Random Forests, and Deep

Learning with SVM Integration

877

DNNs improved the accuracy of multi-trait predic-

tions in rice, which has parallels in maize breed-

ing. (Gonz

alez-Camacho et al., 2018) further empha-

sized that integrating multiple models reduces Mean

Squared Error (MSE) and increases the reliability of

genomic predictions across diverse environments.

Building on this foundation, the current study in-

troduces a novel meta-ensemble framework that com-

bines Gradient Boosting, Random Forests, DNNs, and

an SVM meta-model speciﬁcally for maize breed-

ing. Our experiments, conducted on real-world maize

datasets, reveal that this framework outperforms tra-

ditional MTSI methods and individual ML models,

offering signiﬁcant reductions in MSE and improved

operational efﬁciency. This research establishes a

new benchmark for genomic selection in maize breed-

ing.

The ﬁndings of this study underscore the poten-

tial of meta-ensemble models to optimize breeding

decisions, not only in maize but across a wide range

of crops. The robustness and scalability of this ap-

proach, coupled with its ability to generalize across

environments, position it as a valuable tool for fu-

ture breeding programs aimed at enhancing agricul-

tural productivity and resilience.

Overall, machine learning, and particularly meta-

ensemble approaches, represent a signiﬁcant ad-

vancement in plant breeding. Traditional methods,

though historically successful, are increasingly un-

able to cope with the complexity of high-dimensional

genomic data and multi-trait interactions. The meta-

ensemble framework proposed in this study offers a

robust and scalable solution to these challenges. Fu-

ture research should explore the integration of envi-

ronmental factors and the application of these meth-

ods to other crop species, maximizing the potential of

ML in agricultural genomics.

3 METHODOLOGY

3.1 Dataset Description

The dataset utilized in this study was sourced from

the Genome to Fields (G2F) initiative,

covering the

period from 2018 to 2021. It encompasses a com-

prehensive collection of 4,372 maize lines charac-

terized by 98,027 single nucleotide polymorphism

(SNP) markers. The dataset was collected from 38

diverse locations across the United States, represent-

https://datacommons.cyverse.org/browse/iplant/home/

shared/commons repo/curated/GenomesToFields G2F

2016 Data Mar 2018 Accessed on December 27

, 2023.

ing a broad spectrum of environmental conditions.

These locations span 27 U.S. cities, contributing to

the robustness and generalizability of the study’s ﬁnd-

ings.

The primary phenotypic traits analyzed in this

study are Grain yield (RDT), Grain moisture (HUM),

Plant stand (PS), Date of anthesis (ANT), Date of silk-

ing (SILK), and Anthesis-silking interval (ASI).

3.1.1 Data Cleaning and Preprocessing

To ensure the integrity and reliability of the dataset,

the following steps were performed:

• Genotypic Data Cleaning. SNP markers with

null, missing (NA), or inconsistent values were re-

moved, ensuring that only high-quality genotypic

data were used for analysis.

• Phenotypic Data Imputation. Missing values

in phenotypic traits were imputed using the mean

values corresponding to the relevant environmen-

tal conditions (year, location, block, replication).

This step helped mitigate missing data issues

while maintaining data accuracy.

• Outlier Detection. A Subspace Outlier Detection

method was applied to both genomic and pheno-

typic data to identify and remove outliers, enhanc-

ing the dataset’s robustness and ensuring reliable

model performance.

3.1.2 Feature Selection and Engineering

To further reﬁne the dataset for model training,

several feature selection and engineering techniques

were employed:

• SNP Matrix Centering. The SNP matrix was

centered to reduce the inﬂuence of rare variants,

allowing for a more balanced contribution of ge-

netic markers to the models.

• Genetic Relationship Matrix (GRM). A genetic

relationship matrix was constructed to capture

population structure and relatedness among the

maize lines. This matrix was integrated across all

analytical methods to account for population-level

correlations.

• Principal Component Analysis (PCA). To ad-

dress the high-dimensional nature of the genomic

data, PCA was applied, retaining components

that explained over 95% of the total variance.

This dimensionality reduction step helped in im-

proving computational efﬁciency while preserv-

ing critical information.

• Normalization. Both genotypic and phenotypic

data were normalized to ensure equal contribu-

tion from different traits during model training,

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

878

preventing any single feature from dominating the

predictions.

This curated dataset underpins the meta-ensemble

learning framework in this study, facilitating the ex-

ploration of multi-trait optimization in maize breed-

ing. Its rich genomic and phenotypic data enable ad-

vanced model training and provide key insights into

optimizing agronomic traits.

3.2 Multi-Trait Selection Methods &

Models

Optimizing multi-trait selection requires a thorough

evaluation of both conventional and modern ap-

proaches to identify the most effective methods. The

ﬁrst subsection, Traditional Multi-Trait Selection In-

dex (MTSI) Methods, reviews established techniques

for combining multiple phenotypic traits into a sin-

gle index. These traditional methods provide a bench-

mark for assessing the performance of more advanced

approaches.

Next, Advanced Meta-Ensemble Learning Frame-

work, presents a cutting-edge methodology that com-

bines multiple machine learning models—Gradient

Boosting, Random Forests, Deep Learning, and Sup-

port Vector Machines (SVM). This meta-ensemble

framework leverages the strengths of each model to

potentially improve predictive accuracy and selection

performance.

A comparison of these two approaches will deter-

mine whether the meta-ensemble framework outper-

forms traditional MTSI methods in multi-trait opti-

mization.

3.2.1 Multi-Trait Selection Index Methods

(MTSI)

The Smith-Hazel Index (SHI), also known as the Eco-

nomic Selection Index, and the Genomic Selection

Index, which employs techniques such as Ridge Re-

gression BLUP (rrBLUP), are among the most widely

utilized methods in plant breeding.

3.2.2 Meta-Ensemble Learning Framework

The proposed meta-ensemble machine learning

framework is designed to enhance multi-trait selec-

tion in plant breeding by harnessing the complemen-

tary strengths of various machine learning models.

This framework is tailored to address the complexities

of high-dimensional genomic data and intricate trait

interactions. It consists of two main layers, as shown

in Figure-1: base models and a meta-model. These

layers work in tandem to optimize predictive perfor-

Figure 1: A two-layer Meta-Ensemble Framework for

multi-trait selection.

mance, accuracy, and robustness. The key compo-

nents of the proposed framework are outlined below:

• Base Models

The framework begins with three diverse base

models that capture different aspects of the data

and provide multiple perspectives on the predic-

tion task:

– Gradient Boosting Machines (GBM): are

a powerful ensemble learning method de-

signed to iteratively improve predictive accu-

racy by minimizing residual errors at each step

(Natekin and Knoll, 2013).

– Random Forests (RF): is an ensemble learn-

ing method that constructs multiple decision

trees and combines their predictions to enhance

accuracy and robustness (Genuer et al., 2020).

– Deep Neural Networks (DNN): are machine

learning models that learn hierarchical repre-

sentations of data through multiple layers of in-

terconnected neurons. DNN consist of an in-

put layer, several hidden layers, and an out-

put layer, with each layer containing multiple

neurons. These networks model complex, non-

linear relationships by learning features at vari-

ous levels of abstraction (LeCun et al., 2015).

• Meta-Model

In this ensemble learning framework, the meta-

model is a Support Vector Machines (SVMs) that

acts as a second-level learner.

Support Vector Machines (SVMs): are a pow-

erful classiﬁcation and regression technique that

seeks to ﬁnd the optimal hyperplane to maximize

the margin between classes. In ensemble frame-

works, SVMs serve as the second layer, enhancing

model performance through their ability to handle

Meta-Ensemble Learning for Multi-Trait Optimization in Maize Breeding: Combining Gradient Boosting, Random Forests, and Deep

Learning with SVM Integration

879

complex decision boundaries and improve gener-

alization (Sch

olkopf and Smola, 2002).

3.2.3 Overall Architecture of the Framework

The architecture of the meta-ensemble machine learn-

ing framework is depicted in Figure-1, illustrating its

overall structure and component interactions.

1. The input data is ﬁrst processed by the Pre-

processing Unit to ensure compatibility with the

base models. Dimensionality reduction tech-

niques are applied to optimize the performance of

GBMs and RF models; however, for Deep Neu-

ral Networks (DNNs), no reduction is performed.

This decision is based on the observation that

DNNs achieve superior outcomes when operating

on the dataset’s full dimensionality.

2. The preprocessed data is then fed into the

Base Model Layer, where predictions are gener-

ated independently by GBMs, RF, and DNNs.

3. The predictions from each base model are

aggregated and fed into the Meta-Model Layer

(SVM), produces the ﬁnal set of predictions for

each trait. These predictions are optimized to bal-

ance the strengths of each base model, resulting

in enhanced overall accuracy and reduced error

rates.

3.3 Performance Metrics

The performance of the meta-ensemble machine

learning framework is evaluated using a combination

of well-established metrics that assess both predictive

accuracy and model robustness: Mean Squared Error

(MSE), R-squared (R²), and Predictive/Selection Ac-

curacy (SA).

To ensure the robustness and generalizability of

the model, 5-fold cross-validation is used, where the

data is divided into 5 subsets.

These metrics, along with cross-validation, pro-

vide a comprehensive evaluation framework, ensur-

ing that the proposed model not only performs well

on the training data but also generalizes effectively to

new, unseen data.

4 EXPERIMENTAL RESULTS

This section presents the evaluation of the base mod-

els, the meta-ensemble framework, and traditional

multi-trait selection indices (MTSI), including the

Smith-Hazel Index (SHI) and Genomic Selection In-

dex (GSI). Two key metrics, MSE and R

, are used

Figure 2: Comparison of MSE and R

for Grain Yield

(RDT) across SHI, GSI, GBM, RF, DNN, and SVM mod-

els.

Figure 3: Comparison of MSE and R

for Grain Moisture

(HUM) across SHI, GSI, GBM, RF, DNN, and SVM mod-

els.

to assess performance in the six primary phenotypic

traits.

The analysis of Grain Yield (RDT) reveals that

GSI achieved the lowest MSE (7.5706), while Gradi-

ent Boosting Machine (GBM) closely followed with

an MSE of 7.6016. In terms of R

, the Support Vector

Machine (SVM) outperformed other methods with a

value of 3.0522, demonstrating its superior predictive

accuracy for this trait.

For Grain Moisture (HUM), meta-ensemble mod-

els, particularly Random Forest (RF) and GBM,

showed similar performance with MSE values around

11.52. The SVM model, however, achieved the best

(0.5101), signiﬁcantly outperforming the tradi-

tional indices and other machine learning models,

while SHI displayed the lowest predictive accuracy.

For Plant Stand (PS), GSI had the lowest MSE

(3.7087), while SVM achieved the highest R

(0.0616), surpassing both traditional indices and base

models in predictive accuracy.

GSI produced the lowest MSE (23.7178) for Date

of Anthesis (ANT), while SVM demonstrated the best

(11.9788), showcasing its capacity for improved

predictive accuracy despite a higher MSE than tradi-

tional methods.

The SHI model provided the lowest MSE

(24.5171), while the SVM model achieved the high-

est R

(0.1708), demonstrating its effectiveness in ex-

plaining trait variance for Date of Silking (SILK).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

880

Figure 4: Comparison of MSE and R

for Plant Stand (PS)

across SHI, GSI, GBM, RF, DNN, and SVM models.

Figure 5: Comparison of MSE and R

for Date of Anthesis

(ANT) across SHI, GSI, GBM, RF, DNN, and SVM mod-

els.

For Anthesis-Silking Interval (ASI), GSI had the

lowest MSE (2.8077) and a slightly positive R

, while

SVM demonstrated strong performance with the low-

est MSE among machine learning models (0.6559)

and a positive R

(0.0673).

Across all traits, GSI consistently achieved competi-

tive MSE values, while SVM consistently produced

the highest R

values, particularly for Grain Yield,

Plant Stand, Date of Anthesis, and Anthesis-Silking

Interval. Machine learning models such as GBM and

RF displayed stable performance but generally did not

outperform the SVM model in predictive accuracy.

The results summarized in Table-1 indicate that

traditional indices like SHI and GSI remain effective

for minimizing error, but SVM offers superior per-

formance in capturing trait variability. These results

suggest that incorporating SVM into multi-trait selec-

Figure 6: Comparison of MSE and R

for Date of Silking

(SILK) across SHI, GSI, GBM, RF, DNN, and SVM mod-

els.

Figure 7: Comparison of MSE and R

for Anthesis-Silking

Interval (ASI) across SHI, GSI, GBM, RF, DNN, and SVM

models.

Figure 8: Summary of MSE and R

across traits.

tion frameworks could signiﬁcantly improve predic-

tion accuracy.

4.1 Comparison of Methods

• Gradient Boosting Machines (GBM) and Ran-

dom Forest (RF). Both GBM and RF mod-

els demonstrate stable performance across traits.

However, they consistently underperform relative

to SVM, producing higher MSE values and lower

scores, indicating less effective prediction ac-

curacy.

• Deep Neural Networks (DNN). While DNNs

achieve competitive MSE values for certain traits,

their R

scores are signiﬁcantly lower compared to

SVM, reﬂecting challenges in capturing complex

trait variability and interactions as effectively.

• Traditional Indices (SHI and GSI). Although

SHI and GSI remain competitive benchmarks in

multi-trait selection, their predictive performance

lags behind SVM. They generally produce higher

MSE values and lower R

scores, suggesting that

these indices could be enhanced through integra-

tion with advanced machine learning techniques.

Meta-Ensemble Learning for Multi-Trait Optimization in Maize Breeding: Combining Gradient Boosting, Random Forests, and Deep

Learning with SVM Integration

881

Table 1: Evaluation of Base Models, Meta-ensemble Framework, and Traditional Multi-Trait Selection Indices Using Mean

Squared Error, R-squared, and Selection Accuracy (SA).

Phenotypic Trait Metric SHI GSI GBM RF DNN SVM

Grain Yield MSE 7.6122 7.5706 7.6016 7.6061 8.2437 7.4256

-0.0007 0.0102 0.0021 0.0009 1.4571 3.0522

SA 0.85 0.87 0.88 0.89 0.81 0.93

Grain Moisture MSE 11.5851 11.5709 11.5241 11.5232 12.6322 7.8811

0.0000 0.0025 0.0105 0.0107 -2.0417 0.5101

SA 0.80 0.82 0.85 0.86 0.78 0.90

Plant Stand MSE 3.7189 3.7087 3.7952 3.7730 2.3155 2.1829

0.0000 0.0055 -0.0415 -0.0293 -2.3854 0.0616

SA 0.83 0.84 0.88 0.87 0.75 0.89

Anthesis Date MSE 23.7234 23.7178 24.3741 24.3707 23.1299 23.8810

0.0000 0.0004 -0.0556 -0.0553 -11.2793 11.9788

SA 0.79 0.81 0.83 0.82 0.70 0.92

Silking Date MSE 24.5171 24.5113 25.1629 25.1266 22.6813 22.5233

0.0000 0.0004 -0.0534 -0.0504 -1.7887 0.1708

SA 0.82 0.84 0.85 0.86 0.79 0.94

Anthesis-Silking Interval

MSE 2.8212 2.8077 4.9477 4.9407 9.3398 0.6559

-0.0016 0.0080 -2.0806 -2.0720 -1.3290 0.0673

SA 0.78 0.79 0.83 0.82 0.74 0.90

In summary, SVM consistently outperforms other

methods in capturing trait variability and improving

prediction accuracy across multiple traits. This indi-

cates that incorporating SVM into multi-trait genomic

selection frameworks offers signiﬁcant potential for

enhancing breeding efﬁciency. Traditional indices,

though still valuable, may beneﬁt from complement-

ing them with advanced machine learning approaches

to achieve optimal performance in multi-trait selec-

tion.

5 CONCLUSION

The evaluation of base models, meta-ensemble frame-

works, and traditional multi-trait selection indices has

provided key insights into their effectiveness in pre-

dicting agronomic traits. Several important conclu-

sions can be drawn from this analysis:

1. Superior Performance of SVM. The Sup-

port Vector Machine (SVM) consistently outper-

formed all other methods, demonstrating its capa-

bility to achieve the lowest Mean Squared Error

(MSE) and highest R-squared (R

) values, partic-

ularly for traits such as Grain Yield (RDT), Plant

Stand (PS), and Anthesis-Silking Interval (ASI).

This highlights its strength in capturing complex

trait variability and delivering accurate predic-

tions.

2. Competitiveness of Traditional Indices. The

Smith-Hazel Index (SHI) and Genomic Selec-

tion Index (GSI) remain reliable benchmarks for

multi-trait selection. However, while competitive,

they fall short of SVM’s performance in minimiz-

ing MSE and maximizing R

. Integrating these

traditional indices with advanced machine learn-

ing techniques could further improve their effec-

tiveness.

3. Performance of Other Methods. Gradient

Boosting Machines (GBM) and Random Forest

(RF) provided stable and reliable predictions but

were generally outperformed by SVM. Similarly,

Deep Neural Networks (DNN) showed potential

but exhibited lower R

scores, reﬂecting limita-

tions in capturing trait variability. While useful,

these methods do not surpass the predictive accu-

racy of SVM.

4. Implications for Multi-Trait Selection. The

superior performance of the SVM model sug-

gests that incorporating advanced machine learn-

ing techniques into multi-trait genomic selection

frameworks can signiﬁcantly enhance prediction

accuracy. These results underscore the poten-

tial for continued exploration of meta-ensemble

frameworks to improve trait selection and breed-

ing strategies.

In summary, the meta-ensemble framework

presents a notable advancement in multi-trait selec-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

882

tion by leveraging the strengths of diverse machine

learning models to signiﬁcantly enhance prediction

accuracy. The exceptional performance of SVM

within this framework highlights its substantial poten-

tial for future applications in trait prediction.

By integrating a wide range of models, the meta-

ensemble approach not only improves predictive per-

formance but also offers a robust solution for ad-

dressing the complexities of multi-trait genomic se-

lection. This sophisticated methodology promises to

reﬁne breeding strategies and achieve more accurate

trait predictions, advancing precision breeding.

Future research should prioritize integrating en-

vironmental clustering insights and addressing pre-

diction uncertainty to further optimize the meta-

ensemble framework. This involves developing meth-

ods to dynamically adjust predictions based on en-

vironmental conditions and conducting comprehen-

sive uncertainty analyses. Such advancements will

enhance prediction robustness and reliability, leading

to more effective and resilient breeding strategies, ul-

timately boosting agricultural productivity and preci-

sion.

REFERENCES

Ahmed, F. U., Das, A., and Zubair, M. (2024). A ma-

chine learning approach for crop yield and disease

prediction integrating soil nutrition and weather fac-

tors. CoRR, abs/2403.19273.

Breiman, L. (2001). Random forests. Machine learning,

45:5–32.

eron-Rojas, J. J. and Crossa, J. (2018). Linear selection

indices in modern plant breeding. Springer Nature.

Chandana, C. and Parthasarathy, G. (2022). Efﬁcient ma-

chine learning regression algorithm using na

ıve bayes

classiﬁer for crop yield prediction and optimal utiliza-

tion of fertilizer. Int. J. Perform. Eng., 18(1):47.

Crossa, J., Beyene, Y., Kassa, S., P

erez, P., Hickey,

J. M., Chen, C., De Los Campos, G., Burgue

no, J.,

Windhausen, V. S., Buckler, E., et al. (2013). Ge-

nomic prediction in maize breeding populations with

genotyping-by-sequencing. G3: Genes, genomes, ge-

netics, 3(11):1903–1926.

Crossa, J., P

erez-Rodr

ıguez, P., Cuevas, J., Montesinos-

opez, O., Jarqu

ın, D., De Los Campos, G., Bur-

gue

no, J., Gonz

alez-Camacho, J. M., P

erez-Elizalde,

S., Beyene, Y., et al. (2017). Genomic selection in

plant breeding: methods, models, and perspectives.

Trends in plant science, 22(11):961–975.

Friedman, J. H. (2001). Greedy function approximation: a

gradient boosting machine. Annals of statistics, pages

1189–1232.

Genuer, R., Poggi, J.-M., Genuer, R., and Poggi, J.-M.

(2020). Random forests. Springer.

Gonz

alez-Camacho, J. M., Ornella, L., P

erez-Rodr

ıguez, P.,

Gianola, D., Dreisigacker, S., and Crossa, J. (2018).

Applications of machine learning methods to genomic

selection in breeding wheat for rust resistance. The

plant genome, 11(2):170104.

Hazel, L., Dickerson, G., and Freeman, A. (1994). The se-

lection index—then, now, and for the future. Journal

of dairy science, 77(10):3236–3251.

Hazel, L. N. (1943). The genetic basis for constructing se-

lection indexes. Genetics, 28(6):476–490.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. nature, 521(7553):436–444.

Montesinos-L

opez, O. A., Montesinos-L

opez, A.,

Tuberosa, R., Maccaferri, M., Sciara, G., Ammar, K.,

and Crossa, J. (2019). Multi-trait, multi-environment

genomic prediction of durum wheat with genomic

best linear unbiased predictor and deep learning

methods. Frontiers in Plant Science, 10:1311.

Mrode, R. (2014). Linear models for the prediction of ani-

mal breeding values. Cabi.

Natekin, A. and Knoll, A. (2013). Gradient boosting ma-

chines, a tutorial. Frontiers in neurorobotics, 7:21.

Reddy, D. J. and Kumar, M. R. (2021). Crop yield predic-

tion using machine learning algorithm. In 2021 5th In-

ternational Conference on Intelligent Computing and

Control Systems (ICICCS), pages 1466–1470. IEEE.

Sch

olkopf, B. and Smola, A. J. (2002). Learning with ker-

nels: support vector machines, regularization, opti-

mization, and beyond. MIT press.

Smith, H. F. (1936). A discriminant function for plant se-

lection. Annals of eugenics, 7(3):240–250.

Smith, O., Hallauer, A., and Russell, W. (1981). Use of in-

dex selection in recurrent selection programs in maize.

Euphytica, 30:611–618.

Spindel, J., Begum, H., Akdemir, D., Virk, P., Collard, B.,

Redona, E., Atlin, G., Jannink, J.-L., and McCouch,

S. R. (2015). Genomic selection and association map-

ping in rice (oryza sativa): effect of trait genetic archi-

tecture, training population composition, marker num-

ber and statistical model on accuracy of rice genomic

selection in elite, tropical rice breeding lines. PLoS

genetics, 11(2):e1004982.

Westhues, C. C., Mahone, G. S., da Silva, S., Thorwarth,

P., Schmidt, M., Richter, J.-C., Simianer, H., and

Beissinger, T. M. (2021). Prediction of maize pheno-

typic traits with genomic and environmental predic-

tors using gradient boosting frameworks. Frontiers in

plant science, 12:699589.

Meta-Ensemble Learning for Multi-Trait Optimization in Maize Breeding: Combining Gradient Boosting, Random Forests, and Deep

Learning with SVM Integration

883