The swift diffusion of misinformation online is a great threat to public trust and credibility of information. This

paper compares four supervised machine learning models: Logistic Regression, Support Vector Machine (SVM),

Random Forest, and eXtreme Gradient Boosting (XGBoost) in binary classification of real and fake news on the

news detection when used on high-quality, balanced datasets. The results confirm the use of ensemble and

kernel-based techniques in interpretable, scalable, and high-accuracy misinformation detection systems. This

study contributes to the literature on computational journalism, big data analytics, and artificial intelligence by

demonstrating the effectiveness of non-deep learning methods in curbing digital disinformation.

Keywords: Machine Learning, Fake news detection, XGBoost, Random Forest, ISOT dataset, Text

classification, NLP, Supervised learning, big data analytics, Misinformation, social media

The 21st century has been a time of unpresented change in the way information is being created, shared, and

consumed. With over 5 billion individuals online and over 4.9 billion active social media accounts worldwide

that is, deliberately designed content with the aim of misleading, shaping opinion, damaging reputations, or

generating profit (Shu et al., 2019a; Bryanov & Vziatysheva, 2021). Unlike spontaneous reporting errors or

satirical observation, fake news is intended to deceive, often aping the stylistic features of authoritative reporting

in order to seem more real. The consequences of unchecked fake news are profound and multidimensional:

i. Political manipulation – influencing elections, referendums, and policy debates.

ii. Public health crises – fueling misinformation during pandemics, such as anti-vaccine narratives.

iii. Social unrest – inciting violence, hate speech, and conspiracy-driven mobilization.

iv. Economic disruption – enabling market manipulation, brand damage, and advertising fraud.

High-profile cases illustrate its global scope. The 2016 US presidential election and the 2019 Indian general

election were both marred by allegations of foreign interference through coordinated disinformation campaigns

of today's information dissemination. Social networking websites like Twitter, Facebook, and YouTube generate

huge volumes of end-user content every second, which makes it virtually impossible to detect and control the

propagation of misinformation based solely on human surveillance. This menace has provided a way for the use

of big data analytics to fight fake news. Big data are sets of data that are too huge, too swift, or too complex to

be handled by traditional data processing systems. It is commonly defined by four key dimensions, known as the

Four Vs:

i. Volume: The sheer quantity of data, ranging from terabytes to petabytes, produced continuously across

digital platforms.

ii. Velocity: The rapid rate at which data is generated, transmitted, and must be processed—often in real-

time.

information. To counter fake news in such an environment requires intelligent systems that have the ability to

ingest and process enormous data streams, tease out salient features, and carry out accurate classification in real

time.

Machine learning lies at the core of this task. Supervised machine learning algorithms, trained on labeled data

comprising both true and fake news, can learn to recognize characteristic linguistic patterns, structural cues,

sentiment markers, and propagation dynamics that mark credible information apart from deceptive narratives.

Trained models can then be employed to detect suspect content automatically, rank items for human fact-

checking, or even counter the viral spread of misinformation prior to it going viral. In effect, ML-based big data

analytics presents a scalable, adaptive, and data-driven approach to fake news detection. By converting the issue

of information overload into an opportunity for smart intervention, such systems enable platforms and

offer computational efficiency, interpretability, and noise robustness, yet they have been largely overlooked in

favor of more complex architectures. Recent comparison works demonstrate that with the assistance of strong

preprocessing, feature engineering, and ensemble techniques, traditional classifiers can attain competitive almost

state-of-the-art performance on benchmark datasets such as ISOT (Dwivedi et al., 2023; Patel & Parsania, 2024;

Shivhare et al., 2024). This raises a crucial question: Is deep learning truly necessary for high-performance fake

news detection, or would highly optimized traditional models perform equally or even better at a fraction of cost

and complexity? This paper is motivated by the need to question widely-held assumptions, to provide guidance

on the trade-offs between deep learning and traditional machine learning, and to produce practical, resource-

efficient solutions for misinformation detection in diverse application domains.

The overarching objective of this study is to evaluate the effectiveness of traditional machine learning classifiers

To achieve this, the study pursues the following specific objectives:

3) to identify top-performing models to establish benchmarks for future research and applications.

Fake news is defined as intentionally created information aimed at misleading individuals for political, social,

or economic advantage (Shu et al., 2019a), thereby differentiating it from unintentional misinformation. The

phenomenon frequently emulates legitimate journalism in its tone and structure, yet displays distinct

characteristics: sensational headlines, emotionally charged language, unverifiable sources, and clickbait

strategies (Bryanov & Vziatysheva, 2021). These features leverage cognitive biases, including confirmation

bias, truth bias, and naïve realism, to enhance virality. Research indicates that false news disseminates more

rapidly and extensively than accurate information on platforms such as Twitter (Vosoughi et al., 2018; Khan et

Three primary methodological strands have emerged over time.

1) Content-based methods examine linguistic and stylistic indicators, such as sentiment and lexical

diversity, through techniques like Bag of Words, TF-IDF, or embeddings (Castillo et al., 2011; Rashkin

et al., 2017). Although effective on curated data, they encounter difficulties when fake news mimics

authentic reporting styles.

2) Context-based approaches integrate propagation patterns, user credibility, and network dynamics (Shu

et al., 2019b). Despite their capabilities, they encounter limitations in data access and are susceptible to

organized manipulation.

3) Hybrid and ensemble models integrate textual, social, and metadata signals to enhance robustness (Wang

et al., 2018; Nguyen et al., 2024). These provide advanced performance; however, they entail increased

complexity and computational demands.

i. Ahmed et al. (2017): SVM on ISOT → 92% accuracy

These experiments show that ensemble and kernel methods can keep pace with deep learning on high-quality,

balanced datasets at the advantage of efficiency, interpretability, and deployability in low-resource settings.

Due to its class balance, real-world source (Reuters vs sites reported by PolitiFact), and temporal relevance

(2016–2017), the ISOT dataset has attained the status of a benchmark for the identification of false news.

Previous research has repeatedly shown good performance with conventional models; however, none of these

models have achieved flawless classification. This leaves open for investigation into whether or not improved

how fabricated narratives concentrate around large or politically sensitive events, reflecting real-world

deception. This investigation is necessary to evaluate model performance since topic variation may impair

classification accuracy. Politically trained models may perform poorly when given with disinformation in

underrepresented fields like health or international affairs. Due to its subject-level variation, the ISOT dataset

may assess fake news classifier generalizability and topic sensitivity.

The typical corpus includes 21,417 real items from Reuters.com and 23,481 fake pieces from Wikipedia and

PolitiFact. Title, content, subject, and publication date are listed for each entry, which covers January 2016 to

December 2017, covering the U.S. election and Brexit. The preparation pipeline included binary labeling, text

cleaning (lowercasing, tokenization, stop word deletion, stemming), feature building (TF–IDF, one-hot

encoding, temporal parsing), and normalizing (MinMax, StandardScaler). Class balance was achieved by

stratifying the data into 80% training and 20% testing.

East. Fake news used emotional or conspiratorial language, whereas legitimate news used impartial facts.

Table 5: Overview of the ISOT fake News Dataset

Dataset spans January 2016–December 2017. Class distribution: Fake = 52.3%, Real = 47.7%.

Four supervised classifiers were selected to evaluate their effectiveness in fake news detection:

i. Logistic Regression (LR) – a linear, interpretable, and computationally efficient baseline.

spaces.