The American Journal of Engineering and Technology
202
https://www.theamericanjournals.com/index.php/tajet
TYPE
Original Research
PAGE NO.
202-210
10.37547/tajet/Volume07Issue05-20
OPEN ACCESS
SUBMITED
24 March 2025
ACCEPTED
26 April 2025
PUBLISHED
31 May 2025
VOLUME
Vol.07 Issue 05 2025
CITATION
Rinat Sharipov. (2025). Analysis and Reduction of Errors in AI Models. The
American Journal of Engineering and Technology, 7(05), 202
–
210.
https://doi.org/10.37547/tajet/Volume07Issue05-20
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Analysis and Reduction of
Errors in AI Models
Rinat Sharipov
Founder of UpLook AI
Abstract:
The issue of errors in artificial intelligence (AI)
models is a critical aspect that requires systematic
analysis and the application of effective methods for
their reduction. Errors in AI models can occur at various
stages of development and deployment, including data
collection, model training, and operation phases. The
key tasks in this field involve identifying error sources
and applying approaches aimed at eliminating them.
Methods such as cross-validation, regularization, and
the use of ensemble models play a significant role in
reducing errors and improving prediction accuracy.
Therefore, for the successful use of AI technologies in
various domains, continuous attention to model
monitoring,
parameter
adjustment,
and
the
implementation of innovative methods to minimize risks
is necessary.
Keywords:
artificial intelligence, AI model errors, cross-
validation, regularization, error reduction, data analysis.
Introduction:
Artificial intelligence (AI) models have
become an integral part of modern technologies and are
widely used in various industries, including healthcare,
finance, transportation, and education. However,
despite significant progress in the development and
application of these technologies, AI models still face the
challenge of errors, which can significantly impact their
effectiveness and reliability. Errors can occur both
during the data collection phase and throughout the
model training process, ultimately leading to a reduction
in prediction quality and the generalization ability of
models. The relevance of this issue is heightened by the
widespread adoption of AI in critical areas where
prediction accuracy and reliability are key factors in
decision-making.
There are many sources of errors in AI models, including
incorrectly represented data, overfitting, unbalanced
datasets, and insufficient model interpretability. Errors
The American Journal of Engineering and Technology
203
https://www.theamericanjournals.com/index.php/tajet
not only reduce prediction accuracy but can also lead to
user mistrust in AI technologies. Therefore, the analysis
and reduction of errors are crucial for improving model
performance and ensuring the safe use of AI. Recently,
active research has been conducted to develop methods
aimed at identifying and addressing various types of
errors in AI models.
The goal of this work is to analyze the sources of errors
in AI models and explore existing methods for their
reduction.
1. CLASSIFICATION OF ERRORS IN AI MODELS
Artificial intelligence represents a set of algorithms and
systems designed to train computers to perform
complex tasks in data processing and analysis,
mimicking human perception processes. Through these
models, machines can extract information from data,
identify patterns, and make decisions with minimal
human involvement.
Automation and digitalization are made possible
through the development of artificial intelligence. The
essence of AI is to enable computers and other devices
to act based on principles similar to human thinking. By
programming machines to imitate human cognitive
processes, significant improvements in productivity and
accuracy can be achieved in many tasks. While this may
seem intimidating to some, as it evokes images of robots
from science fiction, AI helps optimize workflows and
enhance their efficiency.
There are many different approaches to creating
artificial intelligence models. AI models are a
combination of methods and algorithms designed to
teach machines to work with data in a way similar to
how humans do. Machine learning is an important
branch of AI in which computers can independently
create new algorithms by analyzing large volumes of
data. Some AI models require programming of specific
algorithms, after which they can adjust their actions
based on the experience gained. There are also models
that lack the ability to learn independently and operate
solely on pre-programmed algorithms, requiring regular
human intervention.
An example of an AI model application is Google Maps
and other navigation systems, which help users plot
routes. These systems use AI algorithms to process data,
such as information on traffic and road changes,
improving the accuracy of predictions and route
suggestions. Each new user experience is integrated into
the system, making it more reliable and precise.
However, the question remains: does artificial
intelligence contribute to the improvement of society,
or does it pose a threat by rendering human labor
unnecessary? There are differing viewpoints on this
issue.
Stephen Hawking, the renowned theoretical physicist,
expressed concern that the creation of fully
autonomous AI could lead to the extinction of humanity,
as such systems would evolve at an increasing pace,
leaving humans behind due to the slow process of
biological evolution. In contrast, Ginni Rometty, the CEO
of IBM, believes that AI will not replace humans but
rather help enhance our intellectual capabilities,
suggesting that AI should be viewed not as a
replacement but as a means of augmenting human
intelligence [1].
Artificial intelligence (AI) models play a crucial
role across a wide range of fields, enhancing workflows,
automating tasks, and addressing complex challenges.
Below are several examples of key AI applications (Table
1).
Table 1. Examples of AI capabilities [2]
AI application
Description
Image analysis
AI is used for image recognition and classification, which is particularly useful in areas such
as e-commerce, security, and medical diagnostics.
Natural
language AI can analyze and interpret human speech, widely applied in text analysis systems, virtual
The American Journal of Engineering and Technology
204
https://www.theamericanjournals.com/index.php/tajet
processing (NLP)
assistants, and chatbots.
Recommendation
systems
AI is actively used to develop recommendations for products, services, and content based
on users' preferences and actions.
Generative models
AI can create original solutions in situations with no clear answers. These technologies are
in demand for creative processes, marketing material development, and even coding.
Personalization of user
experience
AI enables the creation of personalized offers for users, whether it be individualized
learning plans or customized marketing messages.
Prediction
AI is used to analyze data and predict future events and trends, with applications in
finance, healthcare, and marketing analytics.
Fraud detection
AI technologies can detect fraudulent activities such as data theft, insurance fraud, and
credit card fraud.
Autonomous
transportation
systems
AI is being developed to create autonomous vehicles, which can be integrated into logistics
and transportation systems.
Chatbots
for
user
interaction
AI enables the creation of intelligent chatbots that can not only respond to standard
queries but also provide customer support and facilitate online commerce.
Optimization
of
production processes
AI is applied in industries to automate and improve production cycles, particularly in
sectors such as automotive manufacturing, electronics, and metalworking [2].
It is also important to note that the metrics for evaluating models depend on the type of task to be solved and
are divided into two main categories (Fig. 1).
The American Journal of Engineering and Technology
205
https://www.theamericanjournals.com/index.php/tajet
Fig. 1. Metrics used in evaluating models depending on the tasks [3].
For each of these tasks, various metrics are used to
assess the quality of the models. In this context, special
attention will be given to classification.
One of the simplest and most popular metrics, often
mentioned by non-experts, is accuracy. To calculate it, a
formula based on the confusion matrix results is used:
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
1
𝑁
∑
𝑁
𝑖=1
𝐼[𝑦
𝑖
= 𝑦
𝑖
̂] =
𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
(1)
Where:
- TP: true positive (correctly classified positive cases),
- TN: true negative (correctly classified negative cases),
- FP: false positive (incorrectly classified positive cases,
type I error),
- FN: false negative (incorrectly classified negative cases,
type II error).
However, accuracy is not always a reliable measure of
model performance, especially when classes are
imbalanced. For instance, let’s assume we have 100
images of cats and only 10 images of dogs. For
convenience, w
e’ll consider cats as class 0 and dogs as
class 1. Since the number of cats is ten times higher than
the number of dogs, the dataset is considered
imbalanced.
Let’s assume the model correctly classified 90 out of 100
cats, meaning True Negative is 90, and False Negative is
10. Additionally, the model correctly identified 5 out of
10 dogs, so True Positive is 5, and False Positive is 5.
Substituting these values into the formula, we get an
accuracy of 86.4%. However, if the model simply "said"
that all the images were of cats, the accuracy would be
90%, even though this result would be achieved without
the model's actual involvement. This demonstrates that
high accuracy does not always guarantee good model
performance [3].
Therefore, the Precision metric, which characterizes the
proportion of correctly predicted positive classes among
all samples that the model predicted as positive,
becomes an interesting alternative:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃𝑅 =
𝑇𝑃
𝑇𝑃+𝐹𝑃
(2)
Where:
- TP: true positive,
- FP: false positive (type I error),
- FN: false negative (type II error).
Given the wide range of available methods, selecting the
optimal algorithm for solving a specific task can be
challenging. It is necessary to compare different models
Classification
This method involves predicting which category
a particular observation falls into. For example,
when analyzing an image, the task is to
determine whether the image depicts a dog, a
cat, or a mouse. An important subspecies of
classification is binary classification, where the
result is one of two values - e.g., 1 or 0. This
may mean that the system must determine
whether the picture depicts a cat or another
animal.
Regression
In this case, the model's task is
to predict a quantitative
indicator based on the available
data. For example, if bitcoin
was worth $32,000 yesterday,
the model can predict that
tomorrow its value will be
$34,533. Thus, the goal is to
predict a specific value.
The American Journal of Engineering and Technology
206
https://www.theamericanjournals.com/index.php/tajet
and choose the one that best meets the task's
requirements. It is important to note that a model's
accuracy is not always the primary criterion for its
selection, so other factors must also be considered,
which will be discussed in further studies.
The application of the sklearn library provides the ability
to evaluate the performance of machine learning
models, helping to choose the algorithm with the best
metrics for forecasting. One of the evaluation methods
is calculating errors, such as mean absolute error (MAE)
and mean squared error (MSE). These metrics allow for
the minimization of deviations in model predictions,
thereby improving its efficiency.
The man absolute error (MAE) is the average of all
absolut errors, showing how much the model's
predictions deviate from actual data.
The mean squared error (MSE) is the average of the
squared errors, accounting for both the magnitude and
frequency of deviations.
For example, let's consider a classification task using the
Titanic dataset. In this case, we will build a model based
on two algorithms
–
logistic regression and k-nearest
neighbors (KNN). Other approaches could also be used
for data analysis.
# Importing libraries
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn import metrics
# Loading data
data = pd.read_csv("gfg_data")
# Selecting features and target variable
x = data[['Pclass', 'Sex', 'Age', 'Parch', 'Embarked', 'Fare', 'Has_Cabin', 'FamilySize', 'title', 'IsAlone']]
y = data[['Survived']]
# Splitting data into training and testing sets
X_train, X_test, Y_train, Y_test = train_test_split(x, y, test_size=0.3, random_state=None)
# Logistic regression
lr = LogisticRegression()
lr.fit(X_train, Y_train)
Y_pred_lr = lr.predict(X_test)
# Logistic regression model evaluation
logreg_score = round(lr.score(X_test, Y_test), 2)
mae_lr = round(metrics.mean_absolute_error(Y_test, Y_pred_lr), 4)
mse_lr = round(metrics.mean_squared_error(Y_test, Y_pred_lr), 4)
# KNN
knn = KNeighborsClassifier(n_neighbors=2)
knn.fit(X_train, Y_train)
Y_pred_knn = knn.predict(X_test)
The American Journal of Engineering and Technology
207
https://www.theamericanjournals.com/index.php/tajet
# KNN model evaluation
knn_score = round(knn.score(X_test, Y_test), 2)
mae_knn = metrics.mean_absolute_error(Y_test, Y_pred_knn)
mse_knn = metrics.mean_squared_error(Y_test, Y_pred_knn)
# Model comparison
compare_models = pd.DataFrame({
'Model': ['Logistic Regression', 'KNN'],
'Accuracy': [logreg_score, knn_score],
'MAE': [mae_lr, mae_knn],
'MSE': [mse_lr, mse_knn]
})
print(compare_models)
Thus, based on the above, it should be noted that in the
scientific literature there is an opinion regarding the fact
that the logistic regression model, compared to KNN,
has low MAE and MSE values. What is the preferred
choice for building a model [4].
2. METHODS FOR ERROR ANALYSIS AND MODEL
DIAGNOSTICS
Error analysis and model diagnostics methods play a key
role in the development and improvement of machine
learning. These methods aim to identify the sources of
inaccuracies
and
shortcomings
in
a
model’s
performance, as well as to enhance its overall efficiency.
Several approaches exist that allow for effective
diagnosis of errors and the correction of model
performance.
The first method for error analysis is decomposing the
error into its components. This helps to understand
what contributes to the model's reduced accuracy. The
main components of error include bias, variance, and
data noise. For example, high bias error indicates a
systematic underestimation of true values, which
requires a revision of the model structure. On the other
hand, high variance points to overfitting, which can be
addressed by increasing the dataset size or applying
regularization techniques. Below in Figure 2, the
elements that make up this method will be presented.
The American Journal of Engineering and Technology
208
https://www.theamericanjournals.com/index.php/tajet
Fig.2. The elements that make up the method of decomposing errors into their component parts (compiled
independently)
The next method is residual analysis, i.e., the difference
between predicted models and actual values. This
method helps identify where and why the model is
making errors. Visualizing residuals using graphs can
help uncover systematic errors, such as underestimation
or overestimation of certain classes or objects [5].
Below, Figure 3 presents the components of the residual
analysis method used to analyze and reduce errors in
artificial intelligence models.
The American Journal of Engineering and Technology
209
https://www.theamericanjournals.com/index.php/tajet
Fig.3. Components of the residual analysis method used to analyze and reduce errors in
Another significant approach is cross-validation, which
helps assess a model's generalization ability. It prevents
overfitting by splitting the data into training and testing
sets. Moreover, techniques such as K-fold cross-
validation provide a more accurate evaluation of model
performance through repeated testing on various data
subsets.
Sensitivity analysis and feature importance analysis are
also essential diagnostic tools. They help determine
which features in the data contribute most to the
model’s predictions and which ones may be noise.
Eliminating less important features can improve both
the model’s performance and its interpretability.
Additionally, data visualization methods, such as
heatmaps and feature importance charts, help visually
represent data relationships and contribute to
understanding the causes of model errors. This
approach is especially useful for complex models, such
The American Journal of Engineering and Technology
210
https://www.theamericanjournals.com/index.php/tajet
as deep neural networks [6].
3. APPROACHES TO REDUCING AND PREVENTING
ERRORS
Analyzing the results of McKinsey’s “The State of AI in
2021” report, released in December 2021, several key
areas can be identified for mitigating risks in the
development and application of artificial intelligence
(AI) solutions.[7].
Both groups of respondents also emphasized the
importance of the interpretability of AI models. This
term refers to the need to explain the decisions made by
AI algorithms. Business clients demand transparency
because the responsibility for decisions made by the
system still lies with people, not machines.
Organizations that have successfully integrated AI into
their processes have already established methods and
procedures to ensure model transparency.
Ways to mitigate risks when implementing AI. McKinsey
identified three key areas where companies can reduce
risks when working with AI:
1. Working with data for training and testing models. To
mitigate risks in creating and testing AI systems,
companies use several methods. One of them is
checking data samples for insufficient representation
(according to 47% of companies with highly effective AI
implementation). The quality of models directly
depends on data quality, so protecting against errors
and omissions is crucial for improving AI prediction
accuracy.
Another important step is checking data for bias and
distortions, also supported by 47% of respondents.
Despite automatic verification procedures, the
involvement of data engineers is still necessary to clean
and correct the data.
A further method includes protecting data from
incorrect input as volumes increase (36% of companies).
As data accumulates, ensuring its correctness is
essential for maintaining model quality.
2. Evaluating accuracy and model adjustments. AI
models are dynamic systems that require regular
updates and retraining. Companies that have
successfully applied AI use the following approaches:
retraining when issues are detected, monitoring data
deviations and model concepts, and reviewing decisions
with experts. Additionally, training users to identify
errors in AI operations is a key aspect.
3. Documenting processes. Documentation remains an
integral part of any IT infrastructure. For AI, this includes
systematically
recording
model
performance,
architecture, the data used, and the training process. It
is also important to keep records of issues and trade-offs
that arise during the operation of AI solutions [7].
However, there are problematic aspects. Thus, the
problem of overfitting in machine learning presents a
significant challenge for specialists developing
intelligent systems. This phenomenon occurs when a
model becomes overly adapted to the data on which it
was trained, greatly reducing its ability to generalize to
new, previously unseen data. While the model may
show excellent results on the training data, its
performance in real-world conditions can be ineffective.
A deep understanding of the mechanisms behind
overfitting helps identify its causes and develop
strategies to mitigate this effect. Overfitting should be
considered within the framework of model risk
management, as it impacts the quality of decision-
making.
Overfitting is closely related to the concept of the trade-
off between bias and variance in a model. A model with
high bias may be too simple, leading to underfitting and
an inability to capture important patterns in the data. In
contrast, high variance indicates that the model is overly
complex, picking up random variations in the data. The
optimal solution in machine learning lies in finding a
balance between these two extremes, allowing for the
creation of a model with strong generalization
capabilities.
One of the key methods for detecting overfitting is
cross-validation. This approach splits the data into
several subsets for training and testing the model,
providing a more reliable assessment of its
generalization ability. By using different parts of the data
for validation at each stage, cross-validation helps
determine how well the model will perform with new
data. Techniques such as k-fold cross-validation
significantly reduce the risk of overfitting, improving the
model's generalization capabilities.
Regularization methods are aimed at reducing the risk of
overfitting by introducing constraints on model
complexity. Regularization helps prevent the model
The American Journal of Engineering and Technology
211
https://www.theamericanjournals.com/index.php/tajet
from becoming too adapted to the data, making it more
robust to noise and outliers. The most popular
regularization methods include L1 and L2 regularization,
as well as dropout and early stopping when training
neural networks. These approaches ensure the model
performs more reliably on new data, avoiding
overfitting.
Feature selection and engineering play a crucial role in
combating overfitting. Sometimes reducing the number
of features or developing new, more informative ones
can significantly improve the model. Reducing
unnecessary features simplifies the model, making it less
prone to overfitting. This is especially important in tasks
involving large datasets, such as text or image
classification.
Ensemble learning methods, which involve combining
multiple models, play a key role in increasing resilience
to overfitting. Approaches such as random forests or
gradient boosting combine the outputs of several
models, helping to offset errors caused by overfitting in
individual models. Averaging predictions from different
models significantly reduces the risk of overfitting and
improves the generalization ability of the final model [8].
In conclusion, understanding the nature of overfitting
and applying various methods to prevent it are crucial
aspects of developing high-quality machine learning
models. Maintaining a balance between model
complexity and generalization ability enables the
creation of systems that will perform reliably in real-
world conditions.
CONCLUSION
The analysis has shown that errors in AI models
represent a significant issue that impacts the accuracy
and reliability of their predictions. For the successful
application of AI models, it is crucial to consider the full
range of possible error sources and implement methods
to reduce them at all stages of development and
deployment. Cross-validation, regularization, and the
use of ensemble models have proven effective in
combating overfitting and other types of errors.
Therefore, systematic analysis and adjustment of model
parameters remain essential for the successful
application of AI in real-world conditions, improving
system performance and reinforcing user trust in AI
technologies.
REFERENCES
Moor M. et al. Foundation models for generalist medical
artificial intelligence //Nature.
–
2023.
–
vol. 616.
–
No.
7956.
–
pp. 259-265.
Sarker I. H. AI-based modeling: techniques, applications
and research issues towards automation, intelligent and
smart systems //SN Computer Science.
–
2022.
–
Vol. 3.
–
No. 2.
–
p. 158.
Zinoviev V. A., Zinoviev D. V. On the generalized cascade
construction of codes in modular metrics and Lie metrics
//Problems of information transmission.
–
2021.
–
vol.
57.
–
No. 1.
–
pp. 81-95.
Models Score and Error. [Electronic resource] Access
mode:
https://www.geeksforgeeks.org/ml-models-
score-and-error / (accessed 09/06/2024).
Yablokov A. E., Blagoveshchensky I. G. Scientific and
practical foundations for the creation of automated
systems for technical monitoring and diagnostics of
grain processing enterprises' equipment based on
neural network data analysis methods //Yablokov AE,
Blagoveshchenskiy IG
–
M., MGUPP.
–
2022.
Grachev V. V. et al. The method of synthesis of neural
network diagnostic models of complex technical objects
//Automation in transport. - 2020.
–
vol. 6.
–
No. 4.
–
pp.
466-483.
Ways to reduce the risks of AI solutions. [Electronic
resource]
Access
mode:
https://datanomics.ru/artciles/sposoby-snizheniya-
riskov-ai-reshenij / (accessed 06.09.2024).
Retraining: Preventing model risk by using methods to
prevent overfitting. [Electronic resource] Access mode:
https://fastercapital.com/ru/content/%.html (accessed
06.09.2024).
