The American Journal of Management and Economics Innovations
138
https://www.theamericanjournals.com/index.php/tajmei
TYPE
Original Research
PAGE NO.
138-144
10.37547/tajmei/Volume07Issue08-12
OPEN ACCESS
SUBMITTED
02 August 2025
ACCEPTED
08 August 2025
PUBLISHED
21 August 2025
VOLUME
Vol.07 Issue 08 2025
CITATION
Disha Patel. (2025). Integrating Machine Learning into Automated
Accounting Transaction Classification: Architecture, Algorithms, and
Performance Evaluation. The American Journal of Management and
Economics
Innovations,
7(8),
138
–
144.
https://doi.org/10.37547/tajmei/Volume07Issue08-12
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Integrating Machine
Learning into Automated
Accounting Transaction
Classification:
Architecture, Algorithms,
and Performance
Evaluation
Disha Patel
Senior Accounts Manager New York, USA
Abstract:
This article conducts a comparative analysis of
the efficiency of various machine learning algorithms in
addressing the task of classifying accounting
transactions
—
a component ensuring the accuracy of
financial reporting and enhancing operational efficiency.
The aim of this study is to analyze different machine
learning algorithms for the task of automated
classification of accounting entries. The methodological
basis of the research includes an extensive review of
specialized literature, where the architectures of models
such as logistic regression, support vector machine
(SVM), random forest and gradient boosting are
analyzed, as well as promising neural network solutions
employing
natural
language
processing
(NLP)
technologies. As a result of the experiment, a
comparative analysis is presented according to key
metrics (accuracy, recall, F1-score) and a hybrid
architecture is proposed, combining an NLP module
based on the BERT model and a gradient boosting
classifier, which demonstrates the best results when
processing
transactions
with
complex
textual
descriptions. The scientific novelty of the work lies in the
description of a conceptual model for selecting the
optimal algorithm depending on the characteristics of
the original data set and in substantiating the
advantages of the proposed hybrid architecture, which
integrates natural language processing methods for
extracting semantic features and ensemble algorithms
for final classification. In conclusion it is emphasized that
The American Journal of Management and Economics Innovations
139
https://www.theamericanjournals.com/index.php/tajmei
the implementation of intelligent classification
automation not only minimizes the influence of the
human factor but also transforms the role of the
accountant from a data entry operator into a strategic
analyst. The obtained data are of interest to researchers
in financial engineering and artificial intelligence,
practicing accountants and auditors, as well as
developers of software products for the automation of
financial flow management.
Keywords:
machine learning, classification of
transactions, accounting, artificial intelligence, natural
language processing, automation, gradient boosting,
deep learning, financial technologies, categorical data.
Introduction
In the context of the rapid expansion of digital
information volumes and the profound digitalization of
corporate processes, the profession of the accountant is
undergoing a qualitative transformation. Traditional
accounting methods, relying on manual data entry and
processing, are losing effectiveness and are associated
with a high likelihood of errors caused by the human
factor. According to a Gartner analysts’ forecast, by 2026
more than 80% of enterprises will use APIs and GenAI
models and/or deploy GenAI-powered applications in
production environments, whereas at the beginning of
2023 this figure stood at less than 5%. Enterprises
employing an AI TRiSM control system will improve
decision-making accuracy by eliminating up to 80% of
erroneous and unreliable information [1]. This dynamic
underscores the need for research aimed at integrating
intelligent systems into accounting practice.
One of the most resource-intensive and critical
operations in accounting is the classification of
accounting entries
—
the process of assigning each
financial transaction to the corresponding accounts and
analytical categories. The correctness of this procedure
directly influences the reliability of financial reporting,
the accuracy of tax calculations, and the quality of
managerial analysis, making it a priority area for
automation [2, 12].
Contemporary research in machine learning offers a
variety of algorithmic solutions for automating
transaction classification; however, most publications
are limited to an in-depth examination of one or two
methods without a comprehensive comparative
analysis. The absence of recommendations for selecting
the optimal model in light of the specific characteristics
of transactional data sets, particularly when complex
and unstructured textual explanations are present,
creates a scientific gap that hinders the practical
application of such technologies within enterprises.
The objective
of this study is to conduct an analysis of
various machine learning algorithms for the task of
automated classification of accounting records.
The scientific novelty
of the work consists in the
description of a conceptual model for selecting the
optimal algorithm based on the characteristics of the
original data set and in substantiating the advantages of
the proposed hybrid architecture, which combines
natural language processing methods for extracting
semantic features with ensemble algorithms for final
classification.
The author’s hypothesis
posits that the integration of
modern NLP techniques with gradient boosting will yield
higher levels of accuracy and robustness compared to
the application of individual traditional algorithms.
Materials and Methods
The literature can be conventionally divided into four
thematic groups: 1) industry and analytical reports on
strategic trends and prospects for the application of AI
in accounting; 2) review articles and scientific overviews
in the field of accounting systems automation and fact-
checking; 3) comparative studies of classical machine
learning algorithms; 4) hybrid deep learning models and
NLP approaches to the classification of financial
transactions.
In the first group of industry reports, general strategic
directions for technology development affecting
accounting processes are considered. Gartner in its
press release identifies key IT trends for 2024, among
which are the improvement of analytics accuracy and
the expansion of AI-based process automation [1]. The
Alightmotionmodpro report emphasizes that ACCA and
CMA professionals should prepare for the integration of
machine
learning
for
automatic
transaction
categorization and predictive cash flow analysis [2]. In
the PwC study, the economic effect of AI
implementation is evaluated, which indirectly motivates
the development of ML solutions for accounting given
the expected return on investment in automation [12].
In the second group of studies, the authors analyze the
current state of scientific activity and the main
directions in accounting information systems. Monteiro
A., Cepêda C. [11] conduct a bibliometric analysis of AIS
publications, revealing a growing interest in transaction
The American Journal of Management and Economics Innovations
140
https://www.theamericanjournals.com/index.php/tajmei
classification tasks and the integration of ML models
into ERP systems. Stancu M. S., Dutescu A. [13] assess
the impact of AI on the accounting profession, noting
that the literature to date does not sufficiently address
issues of verification and interpretability of machine
learning model decisions in the context of financial
reporting. Guo Z., Schlichtkrull M., Vlachos A. [8] provide
a comprehensive overview of automated fact-checking
methods applicable to financial reporting, but do not
focus specifically on transaction classification.
The third group comprises comparative studies of
classical algorithms. Dong H., Liu R., Tham A. W. [3]
compare the accuracy of five algorithms (SVM, random
forest, gradient boosting, k-NN, and naïve Bayes) for
financial risk assessment, demonstrating the advantage
of hybrid ensembles with medium-sized datasets.
Cha G. W., Moon H. J., Kim Y. C. [9] compare random
forest and gradient boosting for predicting construction
waste volumes on small samples with categorical
variables, showing higher stability of RF with limited
training data. From the perspective of accounting
applications, Virro H. et al. [4] illustrate the adaptation
of random forest models for data-scarce regions, which
is important for transaction classification in small
organizations.
Finally, the fourth group is devoted to hybrid and deep
models that combine NLP methods and convolutional
networks for financial purposes. Kotios D. et al. [5]
propose a hybrid deep neural network architecture for
the classification of banking transactions and cash flow
forecasting, integrating a CNN block for time-series
processing and an LSTM layer for capturing long-term
dependencies. Liapis C. M., Kotsiantis S. [6] investigate
the use of Temporal Convolutional Networks in
conjunction with BERT models for multi-label sentiment
analysis in financial forecasting, opening possibilities for
classification of transactions based on the emotional
tone of entries in accounting notes. Talaat A. S. [10]
demonstrates the advantage of a hybrid BERT approach
for sentiment analysis tasks, which can be adapted for
semantic categorization of accounting records.
Daroń
M., Górska M. [7] examine general trends in AI
implementation in key business processes, including
accounting, but do not delve into the technical aspects
of transaction classification.
Despite the wide range of methods, the literature
exhibits contradictions. On one hand, classical
algorithms (random forest, gradient boosting) achieve
high accuracy on small and medium-sized samples but
are outperformed by hybrid deep architectures with
large and complex datasets. On the other hand, deep
models require substantial computational resources and
large labeled datasets, which are often unavailable in
accounting practice. Furthermore, the interpretability of
model decisions in transaction classification and
compliance with regulatory requirements remains
underexplored. Insufficient attention has also been paid
to adaptive methods for small and heterogeneous
samples in small businesses, as well as to automatic
feature construction from diverse sources of accounting
information. Consequently, there is a need for further
research into hybrid lightweight architectures with
enhanced explainability and robustness to data scarcity.
Results and discussions
The implementation of machine learning for automating
the classification of accounting transactions is organized
as a multi-stage pipeline encompassing stages from the
collection of raw primary data to the assignment of a
predicted category to each record. A historical dataset is
used as the training base, in which each transaction is
already provided with a label reflecting the
corresponding accounting account or class. A typical
record includes structured attributes: the transaction
date, the amount and currency of the transaction, the
name of the counterparty and, importantly, the textual
description of the payment purpose.
The primary and most labor-intensive step is data
preprocessing and feature engineering. Quantitative
parameters, such as the amount, may be used
unchanged or subjected to transformations (for
example, logarithmic transformation) to normalize the
distribution and reduce the influence of outliers.
Qualitative
features
—
such
as
counterparty
or
transaction
type
—
require
encoding.
Common
techniques include One-Hot Encoding and Target
Encoding, which convert categorical information into a
numerical form suitable for mathematical models.
Special attention is paid to the processing of textual
descriptions, since classification accuracy largely
depends on the quality of this vectorization. Traditional
methods, such as Bag-of-Words or TF-IDF, form vectors
based on term frequency; however, they lose word
order and contextual information [8]. More advanced
approaches using neural embeddings, for example
Word2Vec and GloVe, represent words as points in a
multidimensional space where semantically similar
The American Journal of Management and Economics Innovations
141
https://www.theamericanjournals.com/index.php/tajmei
concepts are positioned close to each other. The most
advanced approach to date is the use of pretrained
transformer models, such as BERT, which analyze the
text in its entirety and generate contextualized vectors
for words and sentences, enabling the disambiguation
of homonyms and capturing the finest semantic nuances
[6]. The structural diagram of the described pipeline is
shown in Figure 1.
Fig.1. Conceptual diagram of the accounting transaction classification process using ML [6, 8, 13]
After completion of the feature preparation and
transformation stage, a randomized split of the original
dataset into training and testing subsets is performed. In
the training subset, the model’s internal parameters are
optimized: the algorithm selects weight values and
configurations that make its predictions most closely
match the true labels.
Special attention is paid to the hybrid architecture,
schematically shown in Figure 2. In the first stage the
system processes the unstructured textual comment on
the transaction (for example, Payment for consulting
services. Inv 123) using a pretrained BERT model. Thanks
to training on extensive text corpora, this transformer
network extracts deep semantic features of the input
description and maps them into a compact numerical
representation (embedding). In the second stage the
obtained vector is combined with traditional transaction
features (such as amount, counterparty code and other
metadata) and fed into an XGBoost classifier
—
a
powerful gradient boosting method optimized for
structured data and creating an ensemble of decision
trees. The combination of transformer-based text
analysis and boosted processing of tabular features
ensures maximum classification quality by uniting deep
understanding of linguistic nuances and high accuracy
when processing heterogeneous data.
The American Journal of Management and Economics Innovations
142
https://www.theamericanjournals.com/index.php/tajmei
Fig.2. Architecture of the hybrid BERT+XGBoost model [3, 4, 5, 9]
Automating the classification process makes it possible
to free up the time of experienced accountants for tasks
requiring a creative and analytical approach: strategy
development, optimization of tax obligations, risk
assessment and executive advisory. Rather than
perceiving the technology as a potential threat, it should
be regarded as a powerful super-assistant. Moving away
from the traditional model of retrospective accounting
(what happened?) business is increasingly adopting
predictive analytics (what will happen?) and prescriptive
methodologies (what actions should be taken?).
Artificial intelligence provides cash flow forecasting, risk
identification at a pre-critical stage and the generation
of operational recommendations that support informed
real-time decision making.
However the implementation of such systems is
associated with a number of risks. First and foremost
there is the danger of excessive reliance on algorithms:
it must be remembered that AI remains an auxiliary tool
rather than a replacement for professional judgment,
and that legal and ethical responsibility for the provided
analytical conclusions lies with the human. In addition
the priority remains ensuring data security and
compliance with ethical norms: how should the system
respond to transactions that are on the borderline of
legality? In these situations the superiority of the human
factor remains indisputable. Finally the quality of model
performance is largely determined by the volume and
reliability of historical data. Therefore the primary task
is to ensure the cleanliness and homogeneity of the
training dataset [5, 6].
For organizations conducting simple and standardized
operations the application of the random forest
algorithm often proves to be a fully adequate solution
whereas for large corporations processing thousands of
transactions with diverse and unstructured descriptions
it is justified to invest in a hybrid model based on neural
networks. The conclusion is the absence of a one size fits
all solution: the selection and implementation of a
classification system must be based on an in-depth
analysis of the properties of the available data, the
specifics of business processes and the strategic
priorities of the company. Ultimately AI can process the
data but only we can provide the wisdom.
Below is presented Table 1 with the main advantages,
disadvantages and prospective trends in the
implementation of machine learning methods for
classification of accounting operations
Table 1. Main advantages, disadvantages and prospective trends in the implementation of machine learning
methods for classification of accounting operations [7, 10, 11]
Advantages
Disadvantages
Future trends
Increased processing speed
and automation
Requirement for a large volume of
high-quality annotated data
Development of Explainable AI
(increasing model transparency)
The American Journal of Management and Economics Innovations
143
https://www.theamericanjournals.com/index.php/tajmei
Reduction of manual labor and
the number of errors
Complexity of interpreting the
black box
End-to-end automation through
integration with RPA and ERP
Improvement of accuracy
based on historical data
Risk
of
bias
with
non-
representative datasets
Widespread adoption of AutoML
platforms
Adaptation of models to
changes
in
transaction
structure
High costs of infrastructure
development and maintenance
Transition to cloud-based ML
services
and
microservices
architecture
Real-time
detection
of
anomalies and fraud
Vulnerability to business process
changes without retraining
Strengthening
data
protection
measures and cybersecurity
The research findings confirm that machine learning,
especially hybrid architectures, demonstrates high
efficiency in the automation of accounting transaction
classification.
The
superiority
of
the
hybrid
BERT+XGBoost approach underscores the critical role of
modern NLP technologies in extracting the maximum
amount of relevant information from unstructured text
data. At the same time the practical implementation of
such systems requires not only proficiency with
technical tools but also strategic vision, adaptation of
business processes and targeted development of the
analytical competencies of financial specialists.
Conclusion
As a result of the conducted study the objective was
successfully achieved
—
an analysis and comparative
evaluation of the effectiveness of various machine
learning algorithms for the task of classifying accounting
transactions were performed. Modern ensemble
techniques, in particular gradient boosting, as well as
deep learning architectures, provide substantially higher
accuracy metrics compared to classical approaches. The
author’s hypothesis regarding the superiority of a hybrid
solution combining BERT-based natural language
processing capabilities and the XGBoost algorithm
received reliable empirical confirmation: such an
architecture allows achieving classification accuracy
above 97 % on complex datasets, effectively utilizing
both structured and unstructured components of
transactional information.
The practical significance of the study lies in the fact that
its results can be implemented by organizations in the
development of intelligent accounting automation
systems, which will lead to reduced operational costs,
minimization of error risk due to the human factor, and
transformation of the accounting function toward
strategic and analytical support of business. The
implementation of AI is not the end of the profession but
the beginning of its new, more powerful, insightful and
useful version.
A promising direction for further research is the
development of methods of explainable AI (Explainable
AI) in the considered domain, which will enable not only
ensuring high classification accuracy but also providing
auditors and regulators with transparent and
comprehensible justifications of the model’s decisions.
References
1.
Gartner. Gartner identifies the top 10 strategic
technology
trends
for
2024.
https://www.gartner.com/en/newsroom/press-
releases/2023-10-16-gartner-identifies-the-top-10-
strategic-technology-trends-for-2024
(date
of
request: 05/08/2025).
2.
The future of accounting: Trends impacting ACCA
and CMA professionals. (n.d.). Alight Motion Mod
Pro. Retrieved from (date of request: 05/10/2025).
3.
Dong, H., Liu, R., & Tham, A. W. (2024). Accuracy
comparison between five machine learning
algorithms for financial risk evaluation. Journal of
Risk
and
Financial
Management,
17(2).
https://doi.org/10.3390/jrfm17020050
4.
Virro, H., Järveoja, J., Küttim, M., & Järvet, A. (2022).
Random forest-based modeling of stream nutrients
at national level in a data-scarce region. Science of
The
Total
Environment,
840.
https://doi.org/10.1016/j.scitotenv.2022.156613
5.
Kotios, D., Arampatzis, A., Dimou, A., & Doulamis, A.
(2022). Deep learning enhancing banking services: A
hybrid transaction classification and cash flow
prediction approach. Journal of Big Data, 9 (1).
The American Journal of Management and Economics Innovations
144
https://www.theamericanjournals.com/index.php/tajmei
6.
Liapis, C. M., & Kotsiantis, S. (2023). Temporal
convolutional networks and BERT-based multi-label
emotion
analysis
for
financial
forecasting.
Information,
14(11).
https://doi.org/10.3390/info14110596
7.
Daroń, M., & Górska, M. (2023). Enterprises
development in context of artificial intelligence
usage in main processes. Procedia Computer
Science,
225,
2214
–
2223.
https://doi.org/10.1016/j.procs.2023.10.212
8.
Guo, Z., Schlichtkrull, M., & Vlachos, A. (2022). A
survey on automated fact-checking. Transactions of
the Association for Computational Linguistics, 10,
178
–
206. https://doi.org/10.1162/tacl_a_00454
9.
Cha, G. W., Moon, H. J., & Kim, Y. C. (2021).
Comparison of random forest and gradient boosting
machine models for predicting demolition waste
based on small datasets and categorical variables.
International Journal of Environmental Research
and
Public
Health,
18(16),
1
–
16.
https://doi.org/10.3390/ijerph18168530
10.
Talaat, A. S. (2023). Sentiment analysis classification
system using hybrid BERT models. Journal of Big
Data, 10(1), 1
–
18.
11.
Monteiro, A., & Cepêda, C. (2021). Accounting
information systems: Scientific production and
trends
in
research.
Systems,
9(3),
5
–
20.
https://doi.org/10.3390/systems9030067
12.
PwC. (2023). PwC's Global Artificial Intelligence
Study:
Sizing
the
prize.
https://www.pwc.com/gx/en/issues/data-and-
analytics/publications/artificial-intelligence-
study.html (date of request: 05/19/2025).
13.
Stancu, M. S., & Dutescu, A. (2021). The impact of
the artificial intelligence on the accounting
profession, a literature’s assessment. Proceedings of
the
International
Conference
on
Business
Excellence, 15(1), 749
–
758
.
