“YANGI O‘ZBEKISTONDA ZAMONAVIY PSIXOLOGIYA VA PEDAGOGIKAGA DOIR
MUAMMOLARNI TADQIQ ETISHNING TRANSFORMATSION IMKONIYATLARI”
Xalqaro ilmiy - amaliy konferensiyasi, 2025-yil 24-aprel
67
PERFORMANCE SPEED AND ACCURACY METRICS OF FEATURE
CLASSIFICATION MODELS
Shukrulloev Bektosh
Head of the Department of Applied Mathematics and Informatics,
TMC Institute
https://doi.org/10.5281/zenodo.15268088
Abstract. The efficiency of feature classification models in machine learning and
artificial intelligence is primarily evaluated through two essential dimensions: performance
speed and classification accuracy. This paper investigates the trade-off between these two
aspects across different classification algorithms, including K-NN, SVM, Naïve Bayes, Decision
Trees, and Deep Neural Networks. Through empirical evaluation on benchmark datasets
(MNIST, CIFAR-10, UCI), we analyze training time, inference time, memory consumption, and
accuracy-related metrics such as precision, recall, and F1-score. The results provide insight into
selecting optimal models based on application-specific constraints such as real-time
requirements or accuracy sensitivity.
Keywords: Feature classification, accuracy metrics, inference speed, machine learning,
model performance, F1-score, real-time classification.
Introduction.
Feature classification models are at the core of decision-making systems
powered by artificial intelligence and machine learning. They enable machines to distinguish
between categories — such as identifying whether an email is spam, recognizing human faces, or
detecting medical anomalies in radiographic images. However, as these models become more
deeply embedded in high-stakes domains like autonomous driving or diagnostics, a critical
challenge arises: achieving high classification accuracy without compromising performance
speed.
While accuracy ensures reliable predictions, speed — both in training and inference —
determines real-time applicability. For instance, a highly accurate model that takes several
seconds to infer results is impractical for real-time video surveillance. Conversely, a faster model
might lack the nuanced understanding required in complex visual contexts. Thus, evaluating
models across both dimensions is essential for selecting optimal algorithms under specific
conditions.
This study expands the understanding of classification models by comparing their real-
world accuracy metrics and speed parameters on standardized datasets using consistent
benchmarking tools.
Methods and Evaluation Criteria
Datasets.
The datasets were chosen to represent varying levels of complexity:
•
MNIST
offers grayscale digit images (28x28), ideal for benchmarking lightweight
classifiers.
•
CIFAR-10
presents more complex, colored images (32x32) across 10 classes,
testing image processing depth.
“YANGI O‘ZBEKISTONDA ZAMONAVIY PSIXOLOGIYA VA PEDAGOGIKAGA DOIR
MUAMMOLARNI TADQIQ ETISHNING TRANSFORMATSION IMKONIYATLARI”
Xalqaro ilmiy - amaliy konferensiyasi, 2025-yil 24-aprel
68
•
UCI Iris Dataset
, while small, is ideal for quick prototyping and baseline metric
comparison.
Models Analyzed.
A range of both traditional and deep learning models was selected to
evaluate the spectrum of complexity and performance:
•
Naïve Bayes (NB)
for statistical simplicity.
•
K-NN,
known for its non-parametric nature and clarity.
•
SVM
, well-regarded for handling nonlinear separable data.
•
Decision Tree,
due to its interpretability.
•
CNN,
representing modern deep learning models with state-of-the-art accuracy.
Performance Metrics.
Evaluation covered:
•
Training Time:
Time to fully train the model.
•
Inference Time:
Delay introduced during prediction on a single instance.
•
Memory Usage:
RAM consumed during inference.
•
Accuracy, Precision, Recall, and F1-score:
Derived from the confusion matrix,
capturing both correctness and robustness of classification.
Each model was executed under identical computational conditions (Intel Core i7, 16GB
RAM, NVIDIA GTX 1650) to ensure fair comparison.
Results.
The results, visualized in the provided table, show marked differences:
•
CNN dominated accuracy with 97.8% and F1-score of 0.96, confirming its
superiority in handling image data with complex patterns. However, it demanded the highest
training time (120.3 s) and RAM (780 MB), limiting its use in embedded systems.
•
SVM achieved a strong balance, with 91.5% accuracy, 0.90 F1-score, and
moderate training/inference costs, making it ideal for structured but high-dimensional data.
•
K-NN, while accurate (86.7%), suffered from slower inference (1.5 ms) due to
runtime distance calculations — making it unsuitable for low-latency applications.
•
Naïve Bayes offered minimal computational overhead but compromised on
precision. Nonetheless, its sub-second training time (0.5 s) makes it ideal for fast-deploy
scenarios like spam filters.
•
Decision Trees provided a solid middle ground — interpretable, quick, and
reasonably accurate.
These differences demonstrate the importance of choosing a model not just for accuracy,
but for how and where it will be deployed.
Discussion.
The trade-off between speed and accuracy is highly dependent on the
intended application:
•
In mobile health diagnostics, lightweight models like Decision Trees or Naïve
Bayes are preferred due to limited processing power.
•
For autonomous navigation, latency must be extremely low, pushing the need for
optimized deep learning models or even edge-computing strategies.
•
Industrial automation might benefit from models like SVM which deliver high
accuracy with reasonable computational cost.
“YANGI O‘ZBEKISTONDA ZAMONAVIY PSIXOLOGIYA VA PEDAGOGIKAGA DOIR
MUAMMOLARNI TADQIQ ETISHNING TRANSFORMATSION IMKONIYATLARI”
Xalqaro ilmiy - amaliy konferensiyasi, 2025-yil 24-aprel
69
Moreover, developments such as model pruning, quantization, and knowledge distillation
are becoming essential in maintaining accuracy while reducing computational burden — a must
for real-time AI on edge devices.
This study also suggests the value of hybrid approaches, where fast classifiers are used
for pre-screening, and deep models validate critical decisions.
Conclusion.
There is no universally “best” model in feature classification; each
algorithm thrives under different constraints. This study showed that while CNNs deliver
cutting-edge accuracy, their latency and hardware cost limit applicability. Meanwhile, traditional
algorithms still provide viable, efficient alternatives.
For developers and data scientists, this analysis provides a performance-based framework
to select models not only by accuracy benchmarks but by full-system efficiency — a decisive
factor in real-world deployment of intelligent systems.
References:
1.
Asadov, Q. U., & Shukrulloyev, B. R. (2025). Amaliy matematika fanini o'qitishda
iqtisodiy masalalarning qo‘llanilishi. Modern Science and Research, 4(2), 305-307.
2.
Zakhidov, D., & Bektosh, S. (2023). Division of heptagonal social networks into two
communities by the maximum Likelihood method. Horizon: Journal of Humanity and
Artificial Intelligence, 2, 641-645.
3.
Останов, К., Абсаломов, Ш. К., & Шукруллоев, Б. Р. О. (2018). О методических
особенностях изучения квадратичных неравенств. Вопросы науки и образования,
(11 (23)), 43-44.
4.
Shukrulloyev, B. (2025). BELGILARNI TASNIFLASHDA ISHLATILADIGAN
ASOSIY O ‘LCHOV MEZONLARI. Modern Science and Research, 4(2), 107-109.
5.
Shukrulloyev, B., & Abdujabborov, M. (2025). BELGILARNI TANLASH VA
OPTIMALLASHTIRISH USULLARI. Modern Science and Research, 4(2), 51-53.
6.
Shukrulloyev, B., & Abdujabborov, M. (2025). BELGILARNI TANLASH VA
OPTIMALLASHTIRISH USULLARI. Modern Science and Research, 4(2), 51-53.
7.
Zhang, J., et al. (2021). Trade-off Between Speed and Accuracy in AI Models. IEEE
Access, 9, 12245–12256.
8.
Han, S. et al. (2016). Deep compression: Compressing DNNs with pruning, trained
quantization and Huffman coding. ICLR.
9.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.
10.
Pedregosa, F. et al. (2011). Scikit-learn: Machine Learning in Python. JMLR, 12, 2825–
2830.
