IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 3, June 2026, pp. 2385 ∼ 2397 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i3.pp2385-2397 ❒ 2385 Enhancing fak e news detection: a h ybrid BER T -XGBoost model f or impr o v ed perf ormance and inter pr etability Nishant V asantkumar Hegde 1 , Suneesh Bar e 1 , Namruth Reddy 2 , Rajat Gondkar Ara vinda 1 , Minal Moharir 1 , Aamir Ibrahim 1 1 Department of Computer Science and Engineering, R V Colle ge of Engineering, Beng aluru, India 2 Senior Security Engineer , NVIDIA Corporation, Santa Clara, United States Article Inf o Article history: Recei v ed Sep 29, 2025 Re vised Mar 6, 2026 Accepted Apr 22, 2026 K eyw ords: BER T Deep learning F ak e ne ws detection Machine learning Model interpretability Natural language processing XGBoost ABSTRA CT The widespread spread of f ak e ne ws poses a serious threat to the inte grity of information. The dominant approach to detection in v olv es end-to-end ne-tuning of lar ge transformer models lik e bidirectional encoder representations from transformers (BER T), which, despite achie ving high accurac y , often function as opaque “black box es” with limited interpretability . This paper proposes and v alidates a h ybrid, decoupled architecture that pro v es to be a more practical and po werful alternati v e. W e rst ne-tune a DistilBER T model on the full WELF ak e dataset of 71,537 articles after cleaning to create domain-specic embeddings. These high-dimensional v ectors are then used as input features to train a r ob us t e xtreme gradient boosting (XGBoost) classier . The result s demonstrate that the h ybrid model achie v es a state-of-the-art accurac y of 99.76%, slightly surpassing the already high performance of a standard end-to-end ne-tuned model. Crucially , this approach pro vides this top-tier performance while of fering signicant adv antages in model interpretability through feat ure importance analysis. This w ork establishes that a decoupled architecture is not just a viable alternati v e b ut a superior practical strate gy for combating misinformation, successfully balancing state-of-the-art accurac y with essential model transparenc y . This is an open access article under the CC BY -SA license . Corresponding A uthor: Nishant V asantkumar He gde Department of Computer Science and Engineering, R V Colle ge of Engineering Beng aluru, India Email: he gde.nishant2005@gmail.com 1. INTR ODUCTION In today’ s digital era, the swift circulation of misinformation and f ak e ne ws has become a major societal concern. Maliciously crafted content, designed to decei v e and manipulate public opinion, can destabilize democratic processes, erode public trust, and cause signicant harm [1], [2]. The increasing sophistication of such content, often le v eraging sensationalism and emotional language, mak es manual detection unfeasible at scale, necessitating the de v elopment of adv anced automated systems [3]. This has spurred e xtensi v e research into computational techniques capable of discerning f als e information with high accurac y and reliability [4], [5]. Initial methods for detecting f ak e ne ws depended on con v entional machine learning models t hat le v eraged basic linguistic features. Approaches using term frequenc y–in v er se document frequenc y (TF-IDF) J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
2386 ❒ ISSN: 2252-8938 and n-grams combined with classiers such as support v ector machines (SVMs) and nai v e Bayes established the foundational frame w ork [6], [7]. Ho we v er , these methods often struggle to capture the subtle conte xtual and semantic nuances that dif ferentiate sophisticated f ak e ne ws from le gitimate reporting, as the y treat te xt as a mere “bag of w ords” [8]. The adv ent of deep learning, particularly transformer -based models lik e bidirectional encoder representations from transformers (BER T), mark ed a paradigm shift [9]–[11]. As demonstrated by numerous studies [12], [13], models lik e BER T e xcel because of their pre-training on v ast te xt corpora, which endo ws them with a deep understanding of language conte xt, a critical capability where pre vious models f ailed. The current state-of-the-art methodology typically in v olv es taking a pre-trained BER T -lik e model and ne-tuning it end-to-end on a specic f ak e ne ws dataset. This approach has pro v en highly ef fecti v e, achie ving impressi v e accurac y by adapting the model’ s millions of parameters to classication task [12], [13]. Ho we v er , this performance comes at a signicant cost: interpretabilit y . A ne-tuned transformer operates as a “black box”, making it nearly impossible to unders tand or e xplain wh y it classied a particular ne ws article as f ak e [14]. This lack of transparenc y is a major barrier to trust and adoption in critical applications, where accountability and the ability to audit model decisions are paramount [15]. In elds lik e journalism and polic y-making, a model’ s prediction is often insuf cient without a corresponding e xplanation. Recent w ork has highlighted concerns about adv ersarial rob ustness [16] and the need for source credibility assessment [17] alongside content-based detection. T o address this critical trade-of f between performance and transparenc y , this study proposes and in v estig ates a h ybrid, decoupled architecture that syner gizes the strengths of both deep learning and classical machine learning. The contrib utions are threefold. First, demonstrate that decoupled h ybrid architecture achie v es a state-of-the-art accurac y of 99.76%, slightly s urpassing the already high accurac y of a fully ne-tuned DistilBER T model. Second, establish that this top-tier performance is achie v ed without the “black-box” trade-of f, pro viding full model interpretability through e xtreme gradient boosting (XGBoost)’ s feature importance analysis, a crucial step to w ards e xplainable articial intelligence (XAI) i n this domain [18], [19]. Finally , mak e the case that the h ybrid model is a more practical and trustw orth y architecture for real-w orld deplo yment, as it deli v ers superior accurac y , transparenc y , and greater inference ef cienc y , representing a pragmatic and generalizable templat e for XAI i n other high-stak es te xt classication tasks where transparenc y is crucial, such as hate speech detection, sentiment analysis in sensiti v e conte xts, and the identication of medical misinformation. 2. RELA TED W ORK The scholarly e xploration of automated f ak e ne ws detection has adv anced considerably , shif ting from con v entional statistical techniques to adv anced deep learning models. This e v olution can be broadly cate gorized into se v eral k e y phases, each b uilding upon the last to address the increasing comple xity of misinformation. These adv ancements highlight a continuous search for models that are not only accurate b ut also rob ust and understandable. 2.1. Early machine lear ning appr oaches Initial research ef forts focused on manual feature engineering, e xtracting le xical, syntactic, and content-based features from ne ws articles. Models such as nai v e Bayes, logistic re gression, and SVMs were frequently emplo yed using features lik e TF-IDF v ectors, w ord frequencies, and readability metrics [6], [7]. These approaches often incorporated a rich set of handcrafted features, including stylistic attrib utes lik e punctuation frequenc y and capitalization, as well as psychological features deri v ed from sentiment analysis. While these models pro vided a strong baseline and were computationally ef cient, their primary limitation w as the reliance on surf ace-le v el features. This made them vulnerable to simple adv ersarial attacks (e.g., minor te xt alterations) and fundame n t ally unable to grasp deeper semantic meanings or conte xtual nuances, a core challenge identied in foundational surv e ys of the eld [8]. 2.2. Deep lear ning and sequential models The adv ent of deep learning brought forth models that could automatically learn feature representations, thereby minimizing the reliance on e xtensi v e manual feature engineering. A signicant step forw ard came with the application of con v olutional neural netw orks (CNNs) for te xt classication, which pro v ed ef fecti v e at capturing local patterns and n-gram-lik e features from te xt. Subsequently , recurrent neural netw orks (RNNs) and their adv anced v ariants, long short-term memory (LSTM) and g ated recurrent units Int J Artif Intell, V ol. 15, No. 3, June 2026: 2385–2397 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2387 (GR U), g ained prominence [20]. These models processed te xt in sequence, enabling them to capture w ord order and short-term dependencies—an e vident adv ancement o v er traditional bag-of-w ords approaches [21]. Ho we v er , these architectures struggled with long-range dependencies—dif culty in connecting information across long passages of te xt—and were often computationally intensi v e to train on full-length ne ws articles, limiting their ef fecti v eness on comple x narrati v es [22]. 2.3. The transf ormer r e v olution A signicant breakthrough emer ged with the introduction of the transformer architecture and i ts self-attention mechanism, rst proposed by V asw ani et al. [23]. The transformer’ s capability to assess the rele v ance of all w ords in a sequence simul taneously—irrespecti v e of their distance—ef fecti v ely addressed the long-range dependenc y problem. This led to the de v elopment of pre-trained language models lik e BER T [9], which re v olutionized natural language processing (NLP). BER T’ s pre-training on tw o unsupervised tasks—the mask ed language model (MLM) and ne xt sentence prediction (NSP)—allo ws it to de v elop a deep, bidirectional understanding of language conte xt. As empirically demonstrated by Ramzan et al . [12], this gi v es it a distinct adv antage o v er both traditional models and unidirectional models lik e LSTMs, especially when generalizing to ne w , unseen data where conte xt is paramount. F ollo wing this shift, the dominant paradigm has been the de v elopment and end-to-end ne-tuning of increasingly comple x BER T -based architectures. V ariants lik e RoBER T a [24], which optimized BER T’ s training methodology , and ALBER T [25], which introduced parameter -reduction techniques for ef cienc y , pushed performance e v en higher . Researchers quickly adapted these models for f ak e ne ws detection. F or instance, Kaliyar et al . [26] proposed F ak eBER T , combining BER T with a CNN to enhance feature e xtraction, while Jw a et al . [27] introduced e xB AKE, which augments BER T’ s pre-training with a lar ge corpus of ne ws articles to impro v e domain-specic kno wledge. The v ersatility of BER T is further highlighted by its use as a core te xtual analysis component in multimodal systems, which analyze the coherence between a ne ws article’ s te xt and its accompan ying images to detect inconsistencies [28]. In parallel, other research a v enues ha v e e xplored stance detection [29], [30], rumor propag ation on social media [31], and adv anced architectures lik e graph neural netw orks (GNNs) to model relationships within ne ws content [32], [33]. 2.4. Hybrid models and inter pr etability Despite impressi v e performance of these end-to-end models, a signicant and widely ackno wledged dra wback is their inherent lack of interpret ability . A ne-tuned transformer with millions of parameters functions as a “black box”, making its decision-making process opaque to human users [13], [14]. This “interpretability crisis” has become a major focus of the XAI mo v ement in NLP [15]. Recent w ork by Li et al. [34] further emphasizes that while models lik e BER T are po werful, their black-box nature can hinder trust and adoption in critical journalistic applications. This challenge has moti v ated tw o parallel lines of research. The rst focuses on de v eloping pos t-hoc e xplainability methods to probe these comple x models, using techniques lik e local i nterpretable model-agnostic e xplanations (LIME), which approximates the model’ s beha vior locally , or Shaple y additi v e e xplanations (SHAP), which applies g ame-theoretic principles to allocate importance scores to features [18], [19]. F or e xample, Szczepa ´ nski et al . [35] de v eloped a ne w method specically to pro vide e xplanations for BER T -based f ak e ne ws classiers after the y ha v e made a prediction. The second line of research, which this w ork contrib utes to, in v olv es designing h ybrid systems that are more transparent by design. This approach decouples the feature e xtraction from the classication stage. The po werful b ut opaque transformer is used solely to generate high-quality semantic embeddings, which are then fed into an inherently more transparent and ef cient classier . Studies ha v e sho wn that combining deep learning embeddings with tree-based models lik e XGBoost or light gradient boosting machine (LightGBM) can yield competiti v e or e v en superior performance in v arious te xt classication tasks, often with a fraction of the inference time [36]–[38]. This w ork b uilds upon this principle of decoupled architectures, applying it to f ak e ne ws detection to create a system that retains the state-of-the-art performance of modern transformers while pro viding a clearer , more auditable, and ultimately more trustw orth y decision-making process. 3. METHODOLOGY T o conduct a f air and rigorous com parison between a standard end-to-end transformer model and our proposed h ybrid architecture, a unied e xperime ntal w orko w is designed. This process be gins with Enhancing fak e ne ws detection: a hybrid BERT -XGBoost model ... (Nishant V asantkumar He gde) Evaluation Warning : The document was created with Spire.PDF for Python.
2388 ❒ ISSN: 2252-8938 data preparation and culminates in a comparati v e e v aluat ion of tw o distinct modeling architectures b uilt upon the sam e foundational language model. The architectural o v ervie w of our proposed h ybrid system is depicted in Figure 1. The diagram illustrates the decoupled tw o-stage process: rst, deep feature e xtraction using a ne-tuned DistilBER T model to generate semantic embeddings, follo wed by a transparent classication stage using an XGBoost model. Figure 1. An architectural o v ervie w of the proposed h ybrid f ak e ne ws detection system 3.1. Dataset and pr epr ocessing This study utilizes the complete WELF ak e dataset [39], a lar ge and balanced corpus for f ak e ne ws research containing articles from v arious sources. The dataset’ s primary features are the ra w te xt of the articles, contained in ‘title’ and ‘te xt’ columns, alongside a binary ‘label’. This approach uses only this te xtual data, fore going an y reliance on handcrafted or metadata features. The initial dataset contains 72 ,134 ne ws articles. This preprocessing pipeline w as designed to ensure data quality and prepare the te xt for transformer -based analysis. First, all ro ws containing null v alues in cr itical elds such as the title or te xt were remo v ed, resulting in a cleaned dataset of 71,537 articles, with 34,704 labeled as real and 36,833 as f ak e. Ne xt, to pro vide the model with maximum cont e xt , the article title and te xt elds were concatenated into a single full te xt input, separated by a special [SEP] tok en. This allo ws the model to le v erage signals from both the headline and the body of the articl e. Finally , the full, cleaned dataset w as partitioned into a training set (80%) and a nal, held-out test set (20%). This resulted in a training corpus of 57,229 articles and a test set of 14,308 articles. Stratication w as emplo yed during this spl it to ensure that the original distrib ution of real and f ak e ne ws labels w as preserv ed in both partitions. This test set w as k ept entirely separate and w as only used for the nal e v aluation of the trained models to guarantee an unbiased assessment. Int J Artif Intell, V ol. 15, No. 3, June 2026: 2385–2397 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2389 3.2. Model ar chitectur es f or comparison Both architectures are b uilt upon the distilbert-base-uncased pre-traine d language model, a light er and f aster v ersion of BER T that retains most of its performance [40]. This choice allo ws for a direct and f air comparison of the architectural approaches. By using the same distilled model as a foundation, the performance and interpretability dif ferences attrib utable solely to the architectural choice—end-to-end ne-tuning v ersus our decoupled h ybrid system—can be isolated. 3.2.1. Baseline model: end-to-end ne-tuning The baseline represent s the standard high-performance approach in modern NLP . The DistilBertF orSequenceClassication model are utilized from the Hugging F ace T ransformers library , which appends a c lassication head (a linear layer) to the core DistilBER T model. This model contains approximately 66 million trainabl e parameters. This entire model, including the transformer body and the ne w classication layer , w as trained end-to-end on the 57,229-sample training set. T raining w as conducted for a maximum of 10 epochs, with an early stopping callback monitoring the v alidation loss on a subset of the training data. This mechanism ensures that selection of the optimal model checkpoint, ef fecti v ely pre v enting o v ertting. 3.2.2. Pr oposed model: h ybrid ar chitectur e The proposed model is a decoupled, three-stage h ybrid architecture des igned to achie v e high performance while enhancing model transparenc y and interpretability . The rst stage mirrors the baseline’ s ne-tuning process. The objecti v e here is not to creat e the nal classier , b ut to adapt the int ernal parameters of the DistilBER T model. This is accomplished through the self-attention mechanism, enabling the model to e v aluate the relati v e importance of v arious w ords within the input te xt. The core of this is the scaled dot-product attention, gi v en by (1). Attention ( Q, K , V ) = softmax QK T √ d k V (1) Where Q (query), K (k e y), and V (v alue) are matrices deri v ed from the input embeddings, and d k is the dimension of the k e ys. This process transforms the model from a general-purpose language model into a domain-specic e xpert. After ne-tuning, the classication head of the transformer is discarded. The specialized DistilBER T encoder is t hen used as a po werful feature e xtractor [41]. All articles are processed through this model, and for each article, the 768-dimensional embedding from the nal hidden state of the special [CLS] tok en is e xtracted. This v ector serv es as a rich, dense feature repres entation. Finally , an XGBoost classier is trained on these embeddings [42]. XGBoost optimizes an objecti v e function that combines a loss term and a re gularization term, as dened in (2). Obj (Θ) = n X i =1 l ( y i , ˆ y i ) + K X k =1 Ω( f k ) (2) Where l is the loss function, and Ω is a re gularization term that penalizes model comple xity . 3.3. Experimental setup All e xperiments were carried out in the Kaggle notebook en vironment using a T esla T4 GPU to f acilitate model training and acceleration. The implementation w as carried out using Python with the PyT orch, transformers, XGBoost, and Scikit-learn libraries. F or DistilBER T ne-tuning, the model w as trained for a maximum of 10 epochs with a batch size of 16, using the AdamW optimizer and an early stopping callback with a patience of 1 to pre v ent o v ertting. F or the XGBoost classier , which is signic antly less comple x, a model with n estimators=200 and a maximum tree depth of 5 w as trained, pro viding a rob ust conguration. T o ensure the reproducibility of the results, a consistent random seed (SEED=42) w as used throughout all stages of data partitioning and model training. 3.4. Ev aluation framew ork The nal performance of both models w as assessed on the held-out test set. A suite of s tandard classication metrics are emplo yed: o v erall accurac y , precision, recall, and F1-score. These are dened as (3) to (5). Precision = T P T P + F P (3) Enhancing fak e ne ws detection: a hybrid BERT -XGBoost model ... (Nishant V asantkumar He gde) Evaluation Warning : The document was created with Spire.PDF for Python.
2390 ❒ ISSN: 2252-8938 Recall = T P T P + F N (4) F1-score = 2 × Precision × Recall Precision + Recall (5) Where true positi v es (TP), f alse positi v es (FP), and f alse ne g ati v es (FN), respecti v ely . The confusion matrix and the feature importance scores also analyzed from the XGBoost model to e v aluate its interpretability [43]. 4. RESUL TS AND DISCUSSION The e xperimental results, deri v ed from training on the full cleaned WELF ak e dataset, demonstrate the e xceptional performance of the h ybrid BER T -XGBoost model. The ndings conrm that this architecture achie v es state-of-the-art accurac y while of fering signicant, tangible adv antages in model analysis and interpretability . The follo wing sections will detail these ndings, be ginning with a quantitati v e performance comparison, follo wed by an analysis of the training dynamics, and culminating in an in-depth e xamination of the model’ s interpretability . 4.1. Quantitati v e perf ormance comparison The detailed performance metrics for our proposed h ybrid model are presented in T able 1. On the held-out test set of 14,308 articles, the model achie v ed an outstanding o v erall accurac y of 99.76%. The precision, recall, and F1-scores are e xceptionally high and well-balanced at 0.9975 or higher for both the “Real” and “F ak e” classes. This balance is a strong indicator of a rob ust class ier that is not biased to w ards one class and performs reliably on both positi v e and ne g ati v e samples. The high precision for the “F ak e” class, in particular , means that when the model ags an article as f ak e, its judgment is highly trustw orth y , while the high recall demonstrates its ef fecti v eness in capturing the v ast majority of misinformation. T able 1. Detailed performance of the h ybrid model on the full test set Class Precision Recall F1-score Support Real (0) 0.9973 0.9977 0.9975 7006 F ak e (1) 0.9978 0.9974 0.9976 7302 Accurac y 0.9976 Macro A vg 0.9975 0.9976 0.9976 14308 W eighted A vg 0.9976 0.9976 0.9976 14308 T o conte xtualize this performance, T able 2 compares our results ag ainst se v eral other notable BER T -based approaches from the literature. This model signicantly outperforms baseline BER T implementations [12] and demonstrates a notable impro v ement o v er other ne-tuned models on dif ferent datasets [13]. Most importantly , the h ybrid model’ s accurac y of 99.76% is highly competiti v e with, and e v en surpasses, state-of-the-art specialized architectures lik e F ak eBER T [26], which reached 98.90% accurac y . This result is particularly compelling as our model achie v es this top-tier performance on the full WELF ak e dataset—which is signicantly lar ger and more di v erse than the corpora used in se v eral baseline studies—while also of fering the crucial benets of transparenc y . Furthermore, the end-to-end ne-tuned DistilBER T model achie v ed a v alidation accurac y of 99.72% on the same data. This conrms that our decoupled h ybrid approach sacrices no discernible predicti v e po wer; in f act, the slight performance edge suggests a syner gistic benet, where the specialized XGBoost classier is able to le v erage the rich embeddings more ef fecti v el y than a standard linear classication head. T able 2. Comparati v e analysis of f ak e ne ws detection models with dataset and model conte xt Model/Approach Dataset Articles Class balance (R/F) Model size Acc. (%) BER T (Baseline) [12] Custom small ˜2,000 Approx. Bal. ˜110M 84.00 RoBER T a (Fine-T uned) [13] NELA-GT -2022 ˜8,000 Unbalanced ˜125M 89.68 (F1) F ak eBER T (BER T+CNN) [26] Kaggle/ISO T ˜45,000 Bal. (53/47) > 110 M 98.90 End-to-End DistilBER T WELF ak e 71,537 Bal. (48/52) ˜66M 99.72 Hybrid BER T+XGBoost WELF ak e 71,537 Bal. (48/52) ˜66M+XGB 99.76 T o pro vide qualitati v e e vidence of the model’ s performance, T able 3 sho wcases the headlines from se v eral articles that were correctly classied by the h ybrid system. Whil e the model made its predictions using Int J Artif Intell, V ol. 15, No. 3, June 2026: 2385–2397 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2391 the full concatenated title and te xt, the headlines al one often re v eal the stark contrast in linguistic style between the tw o classes. The f ak e ne ws e xamples are characterized by sensationalism, including the use of all-caps (e.g., PRICELESS!), clickbait framing (e.g., W A TCH:), and emoti o na lly char ged, non-journalistic language. In contrast, the real ne ws headlines, which are traceable to credible outlets, maintain a f act ual and objecti v e tone. This demonstrates the model’ s capability to lear n and generalize based on these crucial stylistic cues, which are often most prominent in the headline. T able 3. Sample headlines from correctly classied articles on the test set Headline T rue label Predicted label Examples of correctly classied f ak e ne ws PRICELESS! ANTI-TR UMP RIO TER THR O WS T ANTR UM When Arrested: “I w ant. . . I w ant. . . I w ant!” [V ideo] F ak e F ak e W A TCH: embattled GOP senator just killed his campaign with this racist remark F ak e F ak e Examples of correctly classied real ne ws Hillary Clinton Dines with Her ’SNL ’ Impersonator Kate McKinnon Real Real Islamist militants kill six soldiers in southern Philippines Real Real Snap shares leap 44% in deb ut as in v estors doubt v alue will v anish Real Real Note: The model’ s prediction w as based on the full article te xt; headlines are sho wn here for concise illustration of stylistic dif ferences 4.2. T raining dynamics and err or analysis The high performance is underpinned by a rob ust and ef cient ne-tuning process. The v al idation loss w as monitored across epochs, as sho wn in Figure 2. The loss reached its mini mum after the third epoch, and our early stopping callback correctly selected this checkpoint for the nal feature e xtraction model. This rapid con v er gence demonstrates the po wer of transfer learning; the pre-trained DistilBER T model required only a fe w epochs to specialize for the task. The subsequent slight rise in v alidation loss in the fourth epoch conrms that the early stopping mechanism w as crucial in pre v enting o v ertting and selecting the most generalizable model. Figure 2. V alidation loss per epoch during DistilBER T ne-tuning A granular analysis of the h ybrid model’ s error prole is pro vided by the confusion matrix in Figure 3. The model made a total of only 35 misclassications out of 14,308 test samples. Critically , the number of f alse ne g ati v es—the most dangerous error type where a f ak e article is misclas sied as real—w as e xtremely lo w at just 19 instances out of o v er 7,300 f ak e articles. Thi s demonstrates the model’ s high sensiti vity in catching misinformation. Simil arly , the model produced only 16 f alse positi v es, where le gitimate ne ws is incorrectly agged as f ak e. While less harmful, minimizing this error is important for maintaining trust in credible sources. F ollo wing this, the classier’ s calibration is illustrated in Figure 4, whi ch sho ws that the optimal F1-score is achie v ed at a threshold of approximately 0.59. The atness of the curv es across a wide range of thresholds i ndicates a v ery rob ust and well-separated classier . This signies that the feature embeddings for Enhancing fak e ne ws detection: a hybrid BERT -XGBoost model ... (Nishant V asantkumar He gde) Evaluation Warning : The document was created with Spire.PDF for Python.
2392 ❒ ISSN: 2252-8938 the “Real” and “F ak e” classes are distinctly clustered, allo wing the XGBoost model to nd a clear and stable decision boundary , making its performance not o v erly sensiti v e to the e xact threshold choice. Figure 3. Confusion matrix for the proposed h ybrid model on the full test set Figure 4. Precision, recall, and F1-score as a function of the classication threshold for the h ybrid model 4.3. Model inter pr etability with Shapley additi v e explanations The h ybrid architecture’ s primary adv antage is its inherent interpretability . Unlik e “black-box” end-to-end models, the XGBoost classier allo ws detailed inspection of features pro vided by the ne-tuned transformer . Figure 5 presents a SHAP summary plot, pro viding a f ar richer vie w than standard feature importance charts [19]. Each dot in the gure represents a sample from the test set; its color indicates the feature’ s v alue (red is high, blue is lo w), and its position on the x-axis sho ws its impact on the prediction score. Crucially , the features sho wn (e.g., Feature 512) are not pre-dened linguistic inputs, b ut dimensions within the 768-dimensional embedding space generated by DistilBER T . The SHAP plot conrms the importance of features lik e Feature 512 and Feature 15, re v ealing their directional impact: high v alues (red dots) push predictions to w ards “F ak e”, while lo w v alues (blue dots) f a v or “Real”. Instance-le v el e xplanations, sho wn in Figure 6, bridge abstract features to concrete analysis . At this gure, features in red (e.g., Feature 512, Feature 15) increase the lik elihood of a “F ak e” prediction, pushing the output v alue higher from the base v alue. Features in blue push it lo wer . Qualitati v e inspection of articles Int J Artif Intell, V ol. 15, No. 3, June 2026: 2385–2397 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 2393 with high positi v e Feature 512 v alues re v eals common misinformation traits, such as sensationalist headlines, emotionally char ged adjecti v es (e.g., outrageous, shocking), and conspiratorial framing. This suggests the dimension acts as a high-le v el detector for inammatory language. While full linguistic mapping remains future w ork, this demonstrates ho w our h ybrid model enables genuinely e xplainable f ak e ne ws detection [43]. Figure 5. SHAP summary plot sho wing the impact of the top 20 features on the model’ s output Figure 6. SHAP force plot for a single f ak e ne ws article While SHAP identies prediction dri v ers, ensuring e xplanation f aithfulness remains a core XAI challenge [43]. Furthermore, despite e xce ptional accurac y , the model’ s long-term production utility depends on rob ustness ag ainst adv ersarial attacks tar geting transformer lters [44]. T o transition to a comprehensi v e Enhancing fak e ne ws detection: a hybrid BERT -XGBoost model ... (Nishant V asantkumar He gde) Evaluation Warning : The document was created with Spire.PDF for Python.
2394 ❒ ISSN: 2252-8938 trust-v erication system, future iterations could inte grate multimodal analysis techniques [45], [46], automated claim v erication frame w orks [47]–[49], and broader rumor detection datasets [50], creating a multi-layered defense ag ainst misinformation. 5. CONCLUSION This paper conduct ed a comprehensi v e e v aluation of a h ybrid BER T -XGBoost model, achie ving a state-of-the-art accurac y of 99.76% on the WELF ak e dataset. The results conrm that decoupling feature e xtraction from classication pro vides a vital layer of transparenc y via SHAP analysis without sacricing predicti v e po wer , addressing the inherent “black-box” limitations of standard transformers. Ho we v er , this study is limited by its v alidation on a single dataset and that the model’ s rob ustness ag ainst sophisticated adv ersarial attacks remains to be fully in v estig ated. Future w ork will focus on enhancing this rob ustness, mapping embedding dimensions to concrete linguistic patterns lik e sensationalism, and v alidating this h ybrid template in other critical domains such as medical misinformation and hate speech detection. Ultimately , this decoupled architecture establishes a superior and more trustw orth y strate gy for high-stak es te xt classication where both accurac y and accountability are paramount. A CKNO WLEDGMENTS The authors w ould lik e to e xpress their gratitude to the Management of Rashtree ya Sikshana Samithi T rust (RSST), as well as the principal and vice principal of R V Colle ge of Engineering, Beng aluru, India, for their continuous support and encouragement throughout this research. FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Nishant V asantkumar He gde ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Suneesh Bare ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Namruth Reddy ✓ ✓ ✓ ✓ ✓ Rajat Gondkar Ara vinda ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Minal Moharir ✓ ✓ ✓ ✓ ✓ ✓ Aamir Ibrahim ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT The authors declare that the y ha v e no kno wn competing nancial interests or personal relat ionships that could ha v e appeared to inuence the w ork reported in this paper . D A T A AND CODE A V AILABILITY The data that support the ndings of this study are from the WELF ak e dataset, which is publicly a v ailable with the original study cited as reference [39]. The complete source code for data preprocessing and Int J Artif Intell, V ol. 15, No. 3, June 2026: 2385–2397 Evaluation Warning : The document was created with Spire.PDF for Python.