Optimizing Python Code Metrics Feature Reduction Through

Meta-Analysis and Swarm Intelligence

Marina Ivanova

, Zamira Kholmatova

and Nikolay Pavlenko

Innopolis University, Innopolis, Universitetskaya str., 1 bldg, Russia

{m.ivanova, z.kholmatova, n.pavlenko}@innopolis.ru

Keywords:

Code Metrics, Feature Reduction, Meta-Analysis, Artiﬁcial Bee Colony.

Abstract:

Feature selection plays an important role in reducing the complexity of datasets while preserving the integrity

of data for analysis and predictive tasks. This study tackles this problem in the context of optimizing metrics

for Python source code quality assessment. We propose a combination of meta-analysis with the Modiﬁed

Discrete Artiﬁcial Bee Colony (MDisABC) algorithm to identify an optimal subset of metrics for evaluating

code repositories. A systematic preprocessing step using correlation-based thresholds (0.7, 0.8, 0.9) through

random-effects meta-analysis effectively reduces redundancy while retaining relevant metrics. The MDisABC

algorithm is then employed to minimize Sammon error, ensuring the preservation of structural properties in the

reduced feature space. Our results demonstrate signiﬁcant error reductions, faster convergence, and consistent

identiﬁcation of key metrics that are critical for assessing code quality. This work highlights the utility of

integrating meta-analysis and nature-inspired algorithms for feature selection and establishes a foundation for

scalable, accurate, and interpretable models in software quality assessment. Future research could expand this

methodology to other programming languages and explore alternative algorithms or cost functions for more

comprehensive evaluations. All the relevant code can be found on our GitHub repository

∗

1 INTRODUCTION

Software metrics play a crucial role in project eval-

uations, planning, risk estimation, and quality assur-

ance (Jabborov et al., 2023). However, the variety

of the software metrics causes a number of issues re-

lated to computation complexity, time, and resources

needed, and the possibility of diverting excessive at-

tention to irrelevant metrics. Therefore, the ability to

choose a minimal set of relevant metrics that could

precisely represent and evaluate the source code of

the product could help to efﬁciently estimate the qual-

ity of the software (Bugayenko et al., 2024). This

can be used for determining the choice of libraries,

tools, and dependencies during situations like devel-

opment planning, product contests, and tenders. An

optimized method for swiftly and efﬁciently estimat-

ing software product quality can be integrated into

CI/CD pipelines, enabling real-time quality tracking

and predictive insights into project success by moni-

toring quality improvements and ensuring compliance

https://orcid.org/0009-0002-6841-3736

https://orcid.org/0000-0003-1688-1183

https://orcid.org/0009-0008-0066-5252

∗

https://github.com/Daru1914/ABCFeatureSelector

with predeﬁned requirements and standards.

Addressing the challenge of feature selection in

software metrics requires balancing competing ob-

jectives: reducing dimensionality while maintaining

accuracy and computational feasibility. Traditional

methods often fall short when faced with the high-

dimensional, non-linear, and heterogeneous nature of

source code datasets. Nature-inspired optimization

algorithms, such as Genetic Algorithms (GAs), Par-

ticle Swarm Optimization (PSO), and Artiﬁcial Bee

Colony (ABC) algorithms, offer a promising solution

to these challenges. These algorithms mimic biolog-

ical or physical processes to efﬁciently explore vast

solution spaces and are particularly well-suited to the

combinatorial nature of feature selection problems.

For example, ABC algorithms, inspired by the for-

aging behavior of bees, excel at optimizing binary so-

lutions like selecting subsets of metrics.

While nature-inspired algorithms can effectively

identify optimal subsets of features, their success is

highly dependent on the quality of the input data. This

is where meta-analysis plays a critical role. By sys-

tematically aggregating and analyzing metric correla-

tions across multiple datasets, meta-analysis provides

an efﬁcient framework for identifying redundant met-

338

Ivanova, M., Kholmatova, Z. and Pavlenko, N.

Optimizing Python Code Metrics Feature Reduction Through Meta-Analysis and Swarm Intelligence.

DOI: 10.5220/0013377200003896

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

In Proceedings of the 13th International Conference on Model-Based Software and Systems Engineering (MODELSWARD 2025), pages 338-345

ISBN: 978-989-758-729-0; ISSN: 2184-4348

rics and reducing dataset dimensionality before opti-

mization. This preprocessing step not only enhances

the performance of optimization algorithms but also

ensures that the selected features capture the most

informative and independent aspects of source code

quality.

Despite their potential, the combined use of meta-

analysis and nature-inspired algorithms for feature se-

lection in software metrics remains underexplored.

To bridge this gap, this study aims to integrate these

approaches, leveraging meta-analysis for data-driven

dimensionality reduction and employing a Modiﬁed

Discrete Artiﬁcial Bee Colony (MDisABC) algorithm

for optimal feature subset selection. The ultimate ob-

jective is to develop a scalable and accurate model for

assessing Python code repositories, enabling efﬁcient

quality assurance and comparative analysis.

The remainder of this paper is organized as fol-

lows. In Section 2, we provide a detailed review

of existing literature on feature selection methods,

with a focus on nature-inspired algorithms and their

application to software metrics. Section 3 outlines

the methodology employed in this study, including

dataset collection, preprocessing steps using meta-

analysis, and the implementation of the Modiﬁed Dis-

crete Artiﬁcial Bee Colony algorithm. Section 4

presents the experimental results and evaluates the

performance of the proposed approach. A discussion

of the ﬁndings, their implications, and potential lim-

itations is provided in Section 5. Finally, Section 6

concludes the paper by summarizing the key contri-

butions and suggesting directions for future research.

2 LITERATURE REVIEW

Feature selection is the process of choosing a subset

of features of the dataset for a model. It is a task

that has been explored by researchers for decades,

with many approaches emerging as a result. Re-

cently, more than a hundred meta-heuristics from the

ﬁeld of nature-inspired algorithms were developed to

solve this task (Sharma and Kaur, 2021)(Abu Khurma

et al., 2022). Many of the nature-inspired algorithms

are similar in performance to the genetic algorithm

(GA) (Holland, 1975)(Goldberg, 1988)(Michalewicz

and Schoenauer, 1996), which is one of the most

often-used algorithms for feature selection (Colaco

et al., 2019). In (Meiri and Zahavi, 2006) the au-

thors performed a comparison between Simulated

Annealing (SA)(Pincus, 1970)(Khachaturyan et al.,

1979)(Kirkpatrick et al., 1983) and Stepwise Regres-

sion (SWR) (Efroymson, 1960) for feature selection.

They proved that SA is more stable in handling com-

plex data, while SWR performed comparably in sim-

pler datasets. SWR was faster but more sensitive

to parameter choices. In (Peng et al., 2018a), the

authors used Ant Colony Optimization (ACO) algo-

rithm, termed FACO, for feature selection. They en-

hanced the pheromone updating mechanism and de-

signed a ﬁtness function to prevent the algorithm from

falling into local optima. FACO was then compared to

existing algorithms. FACO outperformed the existing

ACO-based algorithms by converging faster, achiev-

ing higher classiﬁcation accuracy, and reducing false

positive rates. The algorithm also improved classi-

ﬁcation performance across different classiﬁers, par-

ticularly KNN and SVM (Peng et al., 2018b). In

(Schiezaro and Pedrini, 2013), researchers explored

a binary version of the Artiﬁcial Bee Colony (ABC)

algorithm for feature selection, which operates using

forward search strategies. They applied this method

to multiple UCI datasets and compared its perfor-

mance to PSO, ACO, and GA. The proposed ABC

method consistently reduced the number of selected

features while maintaining or improving classiﬁca-

tion accuracy compared to other methods and outper-

formed other nature-inspired algorithms in terms of

feature reduction and accuracy on most datasets.

In (Hancer et al., 2015) authors introduce the

Modiﬁed Discrete ABC (MDisABC) improvement

of the binary ABC algorithm for feature reduction,

and compare it to other algorithms, such as DisABC

(Kashan et al., 2012), AMABC (Pampar

a and En-

gelbrecht, 2011), and MRABC (Akay and Karaboga,

2012). MDisABC achieved better accuracy and clas-

siﬁcation performance after feature optimization in

the latter two cases and generalized better to large

datasets compared to the ﬁrst one, proving it is more

effective in feature selection problems.

Insights from feature selection algorithms can be

enhanced by a prior meta-analysis (Borenstein et al.,

2021), which identiﬁes consistent patterns across

multiple repositories. By aggregating correlation data

using ﬁxed-effects and random-effects models, meta-

analysis reveals signiﬁcant relationships that persist

across different repositories, offering robust criteria

for identifying redundant metrics. This systematic

approach allows for the removal of highly corre-

lated features while ensuring that important variabil-

ity across repositories is considered, ultimately im-

proving the efﬁciency and accuracy of the feature se-

lection process. Additionally, this step reduces the di-

mensionality of the dataset, enabling optimization al-

gorithms to operate more effectively on a reﬁned set

of features.

In the end, our ﬁndings identiﬁed the lack of re-

search in the optimization of feature selection in soft-

Optimizing Python Code Metrics Feature Reduction Through Meta-Analysis and Swarm Intelligence

339

ware development, in particular, source code and its

quality properties assessment.

3 METHODOLOGY

3.1 Task

The primary objective of this study is to determine

an optimal subset of metrics that quantify or evaluate

properties of individual methods (or functions) within

a software codebase while minimizing the dimension-

ality of the data. Speciﬁcally, starting with a dataset

represented as an n×m matrix of real values (n - num-

ber of methods and m - number of metrics), the goal

is to reduce the number of metrics m as much as pos-

sible while preserving the pairwise distances between

norms, measured using the ℓ

norm.

To achieve this, the task involves:

1. Preprocessing and Meta-Analysis ensured that

only the most independent and informative met-

rics were retained for further analysis, thereby

reducing computational overhead and improving

the efﬁciency of the optimization algorithm.

2. Optimization Algorithm Selection and Imple-

mentation. For feature selection, we employed a

MDisABC algorithm with the Sammon error as an

objective function. By minimizing this error, the

algorithm ensures that the reduced dataset retains

the structural properties of the original.

3. Application of Feature Optimization. The

MDisABC algorithm was applied separately to

each repository, considering the unique character-

istics of codebases (e.g., differences in structure

and complexity). This ensured that the selected

subsets of metrics were optimally tailored to each

dataset while maintaining consistency in evalua-

tion criteria.

4. Performance Evaluation and Validation: The

effectiveness of the selected metric subsets was

evaluated by calculating the average and best

Sammon error for each repository. Additionally,

we compared results across different levels of di-

mensionality reduction achieved through meta-

analysis.

3.2 Dataset Collection

In order to collect Python source code metrics, we

selected repositories for the dataset using PyGithub

(pyg, 2024). PyGithub is a Python library, aim-

ing to provide optimized access to the GitHub API

(git, 2024; Kalliamvakou et al., 2015) from a Python

script, overcoming the complexity and maintainabil-

ity issues of handling HTTP API requests and re-

sponses. Through PyGitHub we selected code repos-

itories, written in Python, that satisfy the quality rules

that we set:

• the repository is not archived, open access and

public;

• the repository is not a dataset;

• the main language of the repository code is

Python;

• the repository contains at least 10 commits, 10

stars, and 10 forks;

• the repository has at least 500000 lines of code.

We selected 100 repositories, corresponding to these

criteria, in descending order, starting from the largest

amount of stars.

In the research, we focused on the methods met-

rics of the Python source code, as the most frequent

and meaningful unit in Python. In order to extract

and calculate the metrics, we used the SourceMeter

tool (sou, 2024), as it offers a good variety of metrics

and presents well-organized datasets, and Radon (rad,

2024), as it covers complexity metrics of the code,

such as Cyclomatic Complexity and Halstead metrics.

We computed metrics for each repository using the

tools of our choice. The ﬁnal datasets contains 35

method metrics, described in our Github repository.

Further, we split the largest datasets to increase the

performance and decrease the memory requirements

of the algorithm. Finally, we rescaled the metrics to

the [0,1] range in order to allow Sammon error as the

cost function to have some meaning.

3.3 Meta-Analysis for

Correlation-Based Dimensionality

Reduction

Some of the metrics extracted by SourceMeter failed

to gather any results, leading to the existence of sev-

eral zero columns across all parts of our dataset, as

well as a few nulls. We handled the missing values by

replacing them with zeros where necessary and en-

sured only metrics with non-zero variance were con-

sidered further.

For each repository, we computed pairwise Pear-

son’s correlation coefﬁcients (Eq. 1) between all

metrics. Pearson’s correlation is suitable for meta-

analysis as it quantiﬁes the strength and direction

of linear relationships. These coefﬁcients were later

converted to Fisher’s z-scores for stability:

MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering

340

r =

∑

i=1

−

X)(Y

−

Y )

∑

i=1

−

∑

i=1

−

Y )

(1)

Given the heterogeneity inherent in our dataset,

we relied on the random-effects model of meta-

analysis to identify metrics with consistently high cor-

relations and statistical signiﬁcance across reposito-

ries as candidates for removal.

High-correlation thresholds (0.7, 0.8, 0.9) were

applied to ﬁlter out redundant metrics, creating dif-

ferent sub-datasets retaining only the most informa-

tive and independent metrics to reduce the computa-

tional burden and compare the results of the applica-

tion of the feature selection algorithm on them and on

the original dataset.

3.4 Modiﬁed Discrete Artiﬁcial Bee

Colony (MDisABC)

The ABC is a Swarm Intelligence algorithm primarily

used for solving optimization problems. In this study,

we employed a MDisABC algorithm tailored for the

feature selection task (Schiezaro and Pedrini, 2013)

(Hancer et al., 2015). This approach is inspired by

the behavior of foraging bees, where solutions (food

sources) are explored, exploited, and reﬁned collabo-

ratively by employed, onlooker, and scout bees.

1. Initialization: Create initial solutions as binary

vectors, each representing a subset of features. To

constrain the number of features, each binary vec-

tor includes exactly m active bits (1s), correspond-

ing to selected features.

2. Fitness Evaluation: Calculate the Sammon error

for each feature subset deﬁned by the binary vec-

tors and store it as the ﬁtness of the solution.

3. Exploitation Stage: Generate a new neighbor so-

lution for each food source through mutation and

crossover using the MDisABC algorithm:

(a) Select three random neighbors X

for

the current food source X.

(b) Compute the Jaccard dissimilarity score (Eq. 2)

between X

and X

. The Jaccard metric is

particularly suited for binary representations,

as it measures the proportion of differing ele-

ments between two subsets relative to their to-

tal unique elements. Scale this dissimilarity by

a factor φ to set a target diversity for the mutant

solution:

J(A, B) = 1 −

|A ∩ B|

|A ∪ B|

(2)

difference between the target dissimilarity and

the mutant’s dissimilarity with X

(d) Generate the ﬁnal neighbor solution through

crossover, probabilistically combining ele-

ments from the parent and mutant solutions.

4. Neighbor Evaluation: Calculate the Sammon er-

ror for the newly generated neighbor solution and

compare it with the original food source:

• If the neighbor’s ﬁtness is better, it replaces the

original solution.

• Otherwise, increment the exploitation count

(LIMIT) for the original food source. If the

LIMIT exceeds the MAX LIMIT parameter,

the food source is replaced with a new random

solution (exploration stage).

5. Best Solution Update: After processing all food

sources, update the current global best solution.

6. Termination: Repeat steps 2-5 until the maxi-

mum number of iterations is reached or another

termination condition is satisﬁed.

3.5 Sammon Error

In our feature selection task, we use the Sammon Er-

ror as the ﬁtness function when evaluating the solu-

tion to preserve the structure of data when project-

ing it from a high-dimensional space into a lower-

dimensional space. The Sammon error measures the

difference between pairwise distances in the original

high-dimensional space and the reduced space:

E =

∑

i< j

i j

∑

i< j

i j

− d

′

i j

)

i j

(3)

where d

i j

represents the Euclidean distance between

points i and j in the high-dimensional space, and d

′

i j

is the distance in the lower-dimensional space.

4 RESULTS

In Fig. 1 we have depicted the progression of average

error reduction as the number of features in the dataset

is reduced to are incrementally increased. The aver-

age error for each feature count was calculated across

all repositories by aggregating the errors for a given

feature count and dividing them by the total number

of repositories.

A consistent trend is observed: adding more fea-

tures initially reduces the error signiﬁcantly, but the

rate of improvement slows beyond a certain number

of features (10 in 1a, 6 in 1b, 4 in 1c). Since the

case 1d removes most of the metrics from the analy-

sis, the reduction only happens to 2, 3, and 4 metrics,

and the effect cannot be readily observed.

Optimizing Python Code Metrics Feature Reduction Through Meta-Analysis and Swarm Intelligence

341

(a) Error Reduction without Meta-Analysis.

(b) Error Reduction with 0.9 Threshold.

(d) Error Reduction with 0.7 Threshold.

Figure 1: Comparison of Error Reductions Across Different

Meta-Analysis Thresholds.

In all cases except the one without meta-analysis,

the best error reduction happens signiﬁcantly faster

than average, indicating a better convergence rate to

the best solution.

In Fig. 2 we can observe the metrics that ap-

pear most often in all the subsets of reduced fea-

tures. In the ﬁrst dataset ( 2a), features like HDIF,

NUMPAR, NOI, Anti Pattern, HPV, and Complexity

Metric Rules dominate, with frequencies signiﬁcantly

higher than others. These features represent cogni-

tive complexity, code interaction, and structural de-

sign properties, reﬂecting their critical role in evaluat-

ing maintainability, readability, and functional clarity.

However, a long tail of less frequently selected fea-

tures suggests redundancy and potential noise in the

dataset.

After applying a meta-analysis threshold of 0.9

( 2b), the focus on the top metrics remains, with

only NOS increasing in importance. NOS (Number

of Statements) is a simple size metric that quantiﬁes

the functionality of code but lacks the depth to assess

complexity, interaction, or structural design compared

to the other mentioned metrics.

At a stricter 0.8 threshold ( 2c), the dominance of

NUMPAR, NOI, and Anti Pattern persists, but met-

rics like HDIF and HPV are no longer in the dataset,

leading to an increase in the importance of purely nu-

merical metrics like LLOC and CLOC.

With the most aggressive 0.7 reductions ( 2d), the

remaining metrics are selected almost as equally of-

ten, suggesting that among the subsets NUMPAR,

NOI, Complexity Metric Rules, CLOC, LLOC, and

Coupling Metric Rules none is signiﬁcantly more im-

portant than the other.

By analyzing the subset proportions across the

four scenarios in Table 1, we can conclude the fol-

lowing:

• Scenario 1 (No Meta-Analysis): Without prior

meta-analysis reduction, subsets with higher pro-

portions (e.g., k = 31/35: (Anti Pattern, CD,

CLOC, Complexity Metric Rules, ...) at

52.9%) dominate. The lack of dimensionality

reduction results in multiple overlapping subsets

(e.g., k = 30/35: (Anti Pattern, CD, CLOC,

Complexity Metric Rules, ...) at 23.5%),

indicating potential redundancy and complexity

in feature selection. The presence of subsets

like k = 2/35: (Anti Pattern, HDIF) at 17.6%

suggests that smaller subsets play a limited role

compared to larger, redundant subsets.

• Scenario 2 (Threshold = 0.9): Applying a

0.9 threshold reduces redundancy, with smaller

subsets achieving higher proportions. For in-

stance, k = 11/13: (Anti Pattern, CLOC,

MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering

342

Table 1: Comparison of Subset Proportions Across Different Scenarios.

Scenario k Subset Proportion

1 31/35 Anti Pattern, CD, CLOC, Complexity Metric Rules, ... 0.529

1 30/35 Anti Pattern, CD, CLOC, Complexity Metric Rules, ... 0.235

1 2/35 Anti Pattern, HDIF 0.176

2 11/13 Anti Pattern, CLOC, Complexity Metric Rules, ... 0.424

2 10/13 Anti Pattern, CLOC, Complexity Metric Rules, ... 0.251

2 9/13 Anti Pattern, CLOC, Complexity Metric Rules, ... 0.172

2 2/13 Anti Pattern, HDIF 0.136

3 6/8 Anti Pattern, CLOC, Complexity Metric Rules, ... 0.259

3 5/8 Anti Pattern, CLOC, Complexity Metric Rules, ... 0.151

3 3/8 Anti Pattern, NOI, NUMPAR 0.129

3 5/8 Anti Pattern, CLOC, LLOC, NOI, NUMPAR 0.122

4 2/6 Complexity Metric Rules, NOI 0.137

4 3/6 Complexity Metric Rules, NOI, NUMPAR 0.129

4 3/6 LLOC, NOI, NUMPAR 0.115

4 4/6 CLOC, Complexity Metric Rules, NOI, NUMPAR 0.107

Scenario Deciphering:

Results without any prior meta-analysis reduction.

Results with a meta-analysis threshold of 0.9 applied.

Results with a meta-analysis threshold of 0.8 applied.

Results with a meta-analysis threshold of 0.7 applied.

Complexity Metric Rules, ...) achieves

42.4%, while subsets like k = 10/13 and k = 9/13

maintain proportions of 25.1% and 17.2%, re-

spectively. Even smaller subsets, such as k =

2/13: (Anti Pattern, HDIF) at 13.6%, show a

balance between reducing redundancy and main-

taining predictive capacity.

• Scenario 3 (Threshold = 0.8): A stricter

threshold further narrows the focus, with sub-

sets like k = 6/8: (Anti Pattern, CLOC,

Complexity Metric Rules, ...) achieving

25.9%. Smaller subsets such as k = 3/8:

(Anti Pattern, NOI, NUMPAR) (12.9%) and

k = 5/8: (Anti Pattern, CLOC, LLOC, NOI,

NUMPAR) (12.2%) gain prominence. These results

suggest reduced redundancy but also indicate a

shift towards smaller and more diverse subsets.

• Scenario 4 (Threshold = 0.7): At the strictest

threshold, even smaller subsets dominate, but

with reduced proportions. For example, k = 2/6:

(Complexity Metric Rules, NOI) achieves

13.7%, and k = 3/6: (Complexity Metric

Rules, NOI, NUMPAR) achieves 12.9%. The

consistent inclusion of subsets like (CLOC,

LLOC) and (CLOC, LLOC, NUMPAR) across

scenarios highlights their critical importance,

even in highly reduced datasets.

To conclude, the results highlight the shifting impor-

tance of speciﬁc metrics and subsets across different

meta-analysis thresholds. Features such as Anti Pat-

tern, Complexity Metric Rules, and CLOC emerge

as consistently important across scenarios, particu-

larly in smaller subsets under stricter thresholds, un-

derscoring their relevance in predictive tasks. How-

ever, lower thresholds like 0.7 result in substantially

reduced proportions, indicating a trade-off between

aggressive dimensionality reduction and consistency

in feature importance. While Scenario 1 (no meta-

analysis) demonstrates higher proportions in larger

subsets, the lack of redundancy reduction limits its

practicality. Scenarios 2 (0.9 threshold) and 3 (0.8

threshold) achieve a better balance, with Scenario 2

providing more focused subsets and Scenario 3 main-

taining diversity while reducing redundancy, making

them the most promising approaches for practical ap-

plications depending on the desired balance between

compactness and feature diversity.

Additionally, the computational intensity of work-

ing with the full dataset highlights the practical bene-

ﬁts of meta-analysis-based dimensionality reduction.

For instance, applying a threshold of 0.9 to remove

highly correlated metrics reduced the dataset size

signiﬁcantly and sped up computations by approxi-

mately 10 times compared to the full dataset.

5 DISCUSSION

The results of this study demonstrate the utility of

combining meta-analysis with feature selection al-

gorithms to enhance the efﬁciency and accuracy

of source code quality assessment. By systemati-

cally applying correlation-based thresholds through

random-effects meta-analysis, we were able to iden-

tify and remove redundant metrics, resulting in

Optimizing Python Code Metrics Feature Reduction Through Meta-Analysis and Swarm Intelligence

343

(a) Metric Selection Frequency with-

out Meta-Analysis.

(b) Metric Selection Frequency with

0.9 Threshold.

0.8 Threshold.

(d) Metric Selection Frequency with

0.7 Threshold.

Figure 2: Comparison of Metric Selection Frequencies

Across Different Meta-Analysis Thresholds.

smaller, more focused subsets. This reduction in di-

mensionality improved the performance of the ABC

algorithm, as evident from the faster convergence

to lower errors and more consistent feature subsets

across repositories.

The consistent selection of metrics, such as

NUMPAR, NOI, Complexity Metric Rules, CLOC,

and LLOC, underscores their indispensability for

quality assessment across all scenarios. Metrics like

HDIF and HPV played smaller but still relevant roles

in reduced feature sets, indicating potential dataset-

speciﬁc contributions.

The MDisABC algorithm’s ability to optimize

subsets effectively, even in reduced datasets, conﬁrms

its robustness as a feature selection method for binary

solutions. However, the reliance on hyperparameters

such as MAX LIMIT and iteration count highlights

the algorithm’s sensitivity to tuning, which could be

explored in future work.

An important limitation of the current study is the

exclusive use of the Sammon error as the cost func-

tion. While the Sammon error is effective for preserv-

ing the structure of high-dimensional data in reduced

feature sets, it may not fully capture other important

aspects of the datasets or the quality of the selected

features.

6 CONCLUSIONS

This study addresses the gap in feature selection op-

timization for software source code assessment by

integrating meta-analysis with nature-inspired algo-

rithms. The proposed methodology balances dimen-

sionality reduction and predictive accuracy, achieving

signiﬁcant improvements in error reduction and fea-

ture subset selection.

Key ﬁndings include:

1. Metrics such as NUMPAR, NOI, Complexity

Metric Rules, CLOC, and LLOC are consistently

important across all scenarios, indicating their

strong predictive relevance for Python code qual-

ity assessment.

2. Applying some meta-analysis threshold is neces-

sary in order to ﬁnd an optimal balance between

reducing redundancy and maintaining accuracy.

3. The MDisABC algorithm demonstrates robust

performance, effectively leveraging reduced fea-

ture sets to minimize Sammon error while explor-

ing and exploiting the solution space.

In this work, we highlight the potential of combin-

ing theoretical frameworks such as meta-analysis with

MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering

344

practical optimization algorithms to address the prob-

lem of feature selection. Moreover, the proposed

technique can account for metrics deemed mandatory

according to the regulatory or domain-speciﬁc con-

straints by considering them as ﬁxed inputs while se-

lection process. For future work, we propose the fol-

lowing directions, that could extend our approach:

1. Algorithm Tuning - investigation of the MDis-

ABC sensitivity to different hyperparameters val-

ues;

2. Alternative Cost Functions - incorporation of

other metrics alternatives instead of Sammon er-

ror, to evaluate feature subsets;

3. Generalization - evaluation of the scalability and

adaptability of this method for other programming

languages or domains in order to further validate

its efﬁciency and extend its applicability.

REFERENCES

Github rest api. https://docs.github.com/en/rest?

apiVersion=2022-11-28. Accessed Nov. 27, 2024.

[Online].

Pygithub. https://pygithub.readthedocs.io/en/stable/. Ac-

cessed Nov. 27, 2024. [Online].

Radon. https://pypi.org/project/radon/. Accessed Nov. 27,

2024. [Online].

Sourcemeter. https://sourcemeter.com/. Accessed Nov. 27,

2024. [Online].

Abu Khurma, R., Aljarah, I., Sharieh, A., Abd Elaziz, M.,

Dama

sevi

cius, R., and Krilavi

cius, T. (2022). A re-

view of the modiﬁcation strategies of the nature in-

spired algorithms for feature selection problem. Math-

ematics, 10(3):464.

Akay, B. and Karaboga, D. (2012). A modiﬁed artiﬁcial

bee colony algorithm for real-parameter optimization.

Information sciences, 192:120–142.

Borenstein, M., Hedges, L. V., Higgins, J. P., and Rothstein,

H. R. (2021). Introduction to meta-analysis. John

Wiley & Sons.

Bugayenko, Y., Kholmatova, Z., Kruglov, A., Pedrycz, W.,

and Succi, G. (2024). Selecting optimal software code

descriptors—the case of java. PLOS ONE, 19(11):1–

23.

Colaco, S., Kumar, S., Tamang, A., and Biju, V. G. (2019).

A review on feature selection algorithms. Emerging

Research in Computing, Information, Communication

and Applications: ERCICA 2018, Volume 2, pages

133–153.

Efroymson, M. A. (1960). “multiple regression analysis.

Goldberg, D. E. (1988). Genetic algorithms in search opti-

mization and machine learning.

Hancer, E., Xue, B., Karaboga, D., and Zhang, M. (2015).

A binary abc algorithm based on advanced similarity

scheme for feature selection. Applied Soft Computing,

36:334–348.

Holland, J. (1975). Adaptation in Natural and Artiﬁcial Sys-

tems: An Introductory Analysis with Applications to

Biology, Control, and Artiﬁcial Intelligence. Univer-

sity of Michigan Press.

Jabborov, A., Kharlamova, A., Kholmatova, Z., Kruglov,

A., Kruglov, V., and Succi, G. (2023). Taxonomy

of quality assessment for intelligent software sys-

tems: A systematic literature review. IEEE Access,

11:130491–130507.

Kalliamvakou, E., Gousios, G., Blincoe, K., Singer, L., Ger-

man, D., and Damian, D. (2015). The promises and

perils of mining github (extended version). Empirical

Software Engineering.

Kashan, M. H., Nahavandi, N., and Kashan, A. H.

(2012). Disabc: a new artiﬁcial bee colony algo-

rithm for binary optimization. Applied Soft Comput-

ing, 12(1):342–352.

Khachaturyan, A., Semenovskaya, S., and Vainstein, B.

(1979). Statistical-thermodynamic approach to deter-

mination of structure amplitude phases. Sov. Phys.

Crystallography, 24(5):519–524.

Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983).

Optimization by simulated annealing. Science,

220(4598):671–680.

Meiri, R. and Zahavi, J. (2006). Using simulated anneal-

ing to optimize the feature selection problem in mar-

keting applications. European journal of operational

research, 171(3):842–858.

Michalewicz, Z. and Schoenauer, M. (1996). Evolution-

ary algorithms for constrained parameter optimization

problems. Evolutionary computation, 4(1):1–32.

Pampar

a, G. and Engelbrecht, A. P. (2011). Binary artiﬁcial

bee colony optimization. In 2011 IEEE Symposium on

Swarm Intelligence, pages 1–8. IEEE.

Peng, H., Ying, C., Tan, S., Hu, B., and Sun, Z. (2018a).

An improved feature selection algorithm based on ant

colony optimization. Ieee Access, 6:69203–69209.

Peng, H., Ying, C., Tan, S., Hu, B., and Sun, Z. (2018b).

An improved feature selection algorithm based on ant

colony optimization. Ieee Access, 6:69208.

Pincus, M. (1970). Letter to the editor—a monte carlo

method for the approximate solution of certain types

of constrained optimization problems. Operations Re-

search, 18(6):1225–1228.

Schiezaro, M. and Pedrini, H. (2013). Data feature selection

based on artiﬁcial bee colony algorithm. EURASIP

Journal on Image and Video processing, 2013:1–8.

Sharma, M. and Kaur, P. (2021). A comprehensive anal-

ysis of nature-inspired meta-heuristic techniques for

feature selection problem. Archives of Computational

Methods in Engineering, 28:1103–1127.

Optimizing Python Code Metrics Feature Reduction Through Meta-Analysis and Swarm Intelligence

345