TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 24, No. 4, August 2026, pp. 1168 ∼ 1176 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v24i4.27790 ❒ 1168 Beha vioral nger prints: dri v er pr oling using transf ormer models on next generation simulation trajectory data Mohamed Laamimach 1 , Mghari Mohammed 2 , Aziz Mabr ouk 1 1 Information Security Intelligent Systems and Applications (ISISA), F aculty of Sciences, Abdelmalek Essaadi Uni v ersity , T etouan, Morocco 2 Department of Computer Science, F aculty of Sciences, Abdelmalek Essaadi Uni v ersity , T etouan, Morocco Article Inf o Article history: Recei v ed Jan 9, 2026 Re vised Apr 5, 2026 Accepted May 25, 2026 K eyw ords: Dri v er beha vior Dri v er proling Intelligent transportation systems Microscopic traf c analysis Ne xt generation simulation data T ransformer models ABSTRA CT Characterizing indi vidual dri v er beha vior is essential for adv ancing intelligent transportation systems (ITS) and autonomous v ehicle safe ty . While deep learn- ing models e xcel at macroscopic t raf c prediction, indi vidual dri ving styles are often aggre g ated a w ay . This paper addresses this g ap by proposing a no v el, weakly supervised transformer frame w ork for dri v er beha vior proling using high-resolution ne xt generation simulation (NGSIM) US-101 trajectory data. W e e xtract microscopic beha vioral features including acceleration, lane change dynamics, and headw ay management from 30-second observ ation se gments. A transformer encoder learns comple x temporal dependencies to classify dri v ers into ’a ggressi v e’ and ’normal’ proles, achie ving a 97% F1-score on proxy- labeled se gments. Crucially , these “proxy labels” are deri v ed from heuristic statistics, meaning the model is trained to learn the mapping from sequences to the se beha vioral indicators rather than identifying objecti v e aggression. Our methodology enables the creation of precise “beha vioral ngerprints” that cap- ture indi vidual dri ving nuances. These insights are vital for de v eloping adapti v e ITS that anticipate traf c stability issues and enhance autonomous v ehicle safety by predicting human intent. This is an open access article under the CC BY -SA license . Corresponding A uthor: Mohamed Laamimach Information Security Intelligent Systems and Applications (ISISA), F aculty of Sciences Abdelmalek Essaadi Uni v ersity T etouan, Morocco Email: mohamed.laamimach@etu.uae.ac.ma 1. INTR ODUCTION Metropolitan traf c congestion continues to impose escalating econom ic and en vironmental b urdens on urban populations w orldwide [1]. Precise forecasting of traf c dynamics serv es as a fundamental enabler for contemporary intelligent transportation systems (ITS), supporting proacti v e interv entions including dy- namic route optimization, ramp metering control, and adapt i v e signal coordination [2]. The ef cac y of these technological solutions depend critically on characterizing the di v erse beha vioral patterns of indi vidual dri v ers who collecti v ely generate observ able traf c phenomena. This study is among the rst to appl y weakly super - vised transformers to ne xt generation simulation (NGSIM) trajectory dat a for dri v er proling, bridging the g ap between macroscopic o w and microscopic beha vior . The e v olution of deep learning architectures, especially recurrent structures such as long s hort-term memory (LSTM) netw orks [3] and transformers [4], ha v e greatly impro v ed the accurac y of traf c forecasting. J ournal homepage: http://telk omnika.uad.ac.id Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1169 While these models e xcel at predicting macroscopic states, applying them to prole indi vidual dri v er beha v- ior presents a signicant challenge: the scarcity of ground truth data. Unlik e macroscopic o w (which can be measured by sensors), specic dri ving styles li k e ”aggressi v e” or ”normal” are subjecti v e and dif cult to annotate at scale. Research has sho wn that heuristic rules often f ail to capture the temporal nuances of dri ving; for instance, rule-based systems can ha v e misclassication rates e xceeding 15% when ignoring conte xt [5]. Consequently , research often relies on heuristic rules deri v ed from traf c o w theory to cate gorize dri v ers. Ho we v er , reli ance on static rules ignores the temporal conte xt of dri ving: a sudden brak e in a traf c jam is normal, whereas the same act ion on a free-o w highw ay is aggressi v e. Static rules often f ail to capture this distinction. Current macroscopic models aggre g ate indi vidual actions, ignoring the microscopic signatures that dene traf c stability . T o address this, we adopt a weak supervision strate gy: we generate “proxy labels” using established heurist ics (e.g., high acceleration v ariance). W e then task deep learning models with learning the mapping from ra w trajectory sequences to these labels. In this paper , we aim to bridge the g ap between static rules and dynamic proling. Building on our pre vious w ork [6], we propose a weakly supervised frame w ork using a transformer -based model on the NGSIM US-101 dataset. While the label s are deri v ed from rules, training a transformer on them allo ws us to in v estig ate whether deep sequence models can ef fecti v ely encode the temporal precursors that correlate with these rules. This aligns with the scope of microscopic traf c modeling and ITS safety applica tions rele v ant to the computing and control domain. Our contrib utions are: (i) we formulate a weak supervision pipeline that utilizes heuristic traf c rules to generate proxy labels, enabling the training of deep models without manual annotation; (ii) we design a transformer encoder architecture to capture long-range temporal dependencies in 30-second trajectory se gments; and (iii) we conduct a comparati v e analysis ag ainst LSTM baseline and non-temporal baselines, demonstrating that the transformer architecture is signicantly more ef fecti v e at approximating these beha vioral rules from ra w kinematic data. The remainder of this paper is or g anized as follo ws: section 2 re vie ws related w ork, section 3 detail s our methodology , section 4 presents the e xperimental results, and section 5 concludes with future w ork. 2. RELA TED W ORK Our research is situated at the intersection of three primary domains: the application of tr ansformer models to traf c forecasting, the inte gration of multi-source data, and the use of microscopic beha vioral model- ing. F or this paper , we specically focus on le v eraging microscopic beha vioral modeling within a transformer frame w ork for dri v er proling. 2.1. T ransf ormer models f or trafc f or ecasting and sequence analysis The application of deep learning to traf c forecasting has e v olv ed signicantly from classical ti me- series models lik e autore gressi v e inte grated mo ving a v erage (ARIMA) [7] to architectures capable of capturing comple x non-linear dependencies. While early successes were achie v ed with con v olutional neural netw orks (CNNs) for spatial features [8] and LSTMs for temporal patterns [9], the transformer architecture [4] has recently emer ged as the state-of-the-art due to its superior ability to model long-range dependencies via self- attention. The transformer paradigm has been rapidly adopted in traf c prediction. Xu et al. [10] proposed spatial-temporal transformer netw orks (STTN), and Jiang et al. [11] introduced PDF ormer . Other sophisticated architectures ha v e further pushed the boundaries by inte grating self-attention with graph-based methods, such as a transformer -based dynamic multi-graph con v olutional netw ork (TDMGCN) [12] and a transformer -based spatiotemporal graph dif fusion con v olution netw ork (TSGDC) [13]. While these models sho wcase the po wer of fusing spatial and temporal data within a transformer frame w ork, the y typically operate on macroscopic traf c features or aggre g ate indi vidual v ehicle data for system-le v el predictions. More recently , transformer - based models incorporating spatial masks deri v ed from road netw ork topology , such as T raf cformer [14], ha v e impro v ed short-term speed forecasting from loop-detector data, while multi -source frame w orks such as multi-source data traf c prediction (MDTP) [15] and hierarchical multi-source netw ork (HiMSNet) [16] fuse trajectory or drone-based observ ations with x ed-sensor data to impro v e prediction rob ustness, e xtending a long-standing tradition of sensor dat a fusion in ITS [17]. Generati v e approaches based on dif fusion mod- els, such as traf c scenario generation dif fusion (TSGDif f) [18], and generic interpretable architectures for multi-horizon forecasting, such as the temporal fusion transformer (TFT) [19], further illustrate the breadth of attention-based methods being e xplored for traf c-related sequence modeling. Our current w ork distinguishes Behavior al ng erprints: driver pr oling using tr ansformer models on ne xt ... (Mohamed Laamimac h) Evaluation Warning : The document was created with Spire.PDF for Python.
1170 ❒ ISSN: 1693-6930 itself by applying transformers directly to indi vidual, microscopic beha vioral sequences for the purpose of the dri v er themselv es, rather than predicting traf c state. Ev en so, e xisting models often struggle with the scarcity of indi vidual-le v el ground truth, which we address via weak supervision. 2.2. Micr oscopic beha vioral modeling and dri v er pr oling The causal link between microscopic dri v er beha vior and macroscopic traf c phenomena is a foun- dational concept in traf c engineering, established primarily through simulation [5], [20]. Simulators lik e simulation of urban mobility (SUMO) [21] and Cit y F lo wER [22] are widely used to model ho w indi vidual v ehicle decisions aggre g ate into observ able outcomes. Dri v er proling, or the characterization of indi vidual dri ving styles, has g ained importance for ap- plications ranging from pe rsonalized insurance to autonomous v ehi cle interaction. T radi tional approaches often rely on statistical analysis of dri ving data or simpler machine learning models to identify patterns in speed, acceleration, and braking [23]. More recent w ork has e xplored deep learning for dri v er identication or style classication, often using aggre g ated telematics data. Ev en so, using high-resolution trajectory data lik e NGSIM for ne-grained beha vioral proling with adv ance d sequence models lik e transformers is still an e v olving area. Our w ork aims to e xtract richer , dynamic ”beha vioral ngerprints” by le v eraging the detailed temporal e v olution of microscopic features. The NGSIM program represents a pinnacle of high-resolution, multi-v ehicle trajectory data collection. It pro vides an unparalleled microscopic vie w of traf c, enabling t he detailed analysis of indi vidual dri v er ac- tions, inter -v ehicle dynamics, and their e v olution o v er time. This richness goes be yond what can be captured by traditional loop detectors or probe v ehicles. Our research le v erages this detailed data to create the ne-grained feature sets required for rob ust dri v er proling, mo ving be yond aggre g ate metrics to the underlying beha vioral signals. 3. METHODOLOGY Our methodology is designed to process ra w NGSIM trajectory data into a supervised learning prob- lem for dri v er proling. The process consists of four main stages: (i) data selection and preprocessing; (ii) feature engineering from indi vidual dri v er trajectories; (iii) heuristic-based labeling of dri v er proles; and (i v) the architectural specication of our transformer -based proling model. 3.1. NGSIM dataset De v eloped by the Federal Highw ay Administration, the NGSIM program pro vides essential high- resolution trajectory data that enables microscopic traf c analysis. F or this in v estig ation, we emplo y the US- 101 (Lank ershim Boule v ard) dataset, which documents real-w orld v ehicle trajectories at 0.1-second (10 Hz) temporal granularity . It captures precise details such as v ehicle locations, speeds, accelerations, and inter - v ehicle relationships for indi vidual v ehicles operating on an urban arterial en vironment with multiple lanes and intersections. The ra w dataset initially comprises approximately 11.8 million ro ws and 25 columns of trajectory data. The dataset is characterized by: - High resolution: v ehicle locations are recorded e v ery one-tenth of a second (10 Hz), allo wing for precise capture of instantaneous speeds, accelerations, and subtle changes in dri ving beha vior . - Microscopic detail: each ro w of the dataset corresponds to a single state of the v ehicle at a specic ‘Frame ID‘ (timestamp), including its unique ‘V ehicle ID‘. K e y columns include ‘Global T ime‘ (absolute timestamp), ‘Local X‘ and ‘Local Y‘ (lateral and longitudinal), ‘v V el‘ (v elocity), ‘v Acc‘ (acceleration), ‘Lane ID‘ (cur - rent lane), ‘Preceding‘ (ID of the v ehicle ahead in the same lane), ‘Space Headw ay‘, and ‘T ime Headw ay‘ (spacing and time to the preceding v ehicle). ‘v Class‘ cate gorizes v ehicles into motorc ycles, autos, or trucks. - Scale: the ra w dataset comprises approximately 11.8 million ro ws and 25 columns, representi ng hundreds of hours of aggre g ated v ehicle mo v ement and thousands of indi vidual v ehicle trajectories. This rich, high- v olume data pro vides an e xtensi v e basis for detailed beha vioral analysis. A representati v e sample of the ra w NGSIM trajectory data is presented in T able 1, illustrating the granular nature of the information a v ailable for each v ehicle at a gi v en timestep. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1168–1176 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1171 T able 1. Sample ra w NGSIM trajectory data V eh ID Frame Global T ime v V el v Acc Lane Prec Headw ay v Class 515 100 1118184010000 58.7 0.3 3 511 2.15 2 515 101 1118184010100 58.9 0.2 3 511 2.14 2 515 102 1118184010200 59.1 0.2 3 511 2.13 2 515 103 1118184010300 59.0 -0.1 3 511 2.16 2 516 100 1118184010000 45.2 1.5 2 513 1.80 2 516 101 1118184010100 45.4 0.8 2 513 1.78 2 517 100 1118184010000 62.1 0.0 4 0 999.9 2 517 101 1118184010100 62.1 0.0 4 0 999.9 2 3.2. Initial ltering and cleaning T o create a focused and reliable dataset for the proling of standard passenger v ehicle beha viors, the follo wing initial processing steps were meticulously applied: - V ehicle class ltering: only v ehicles cate gorized with a ‘v Class‘ equal to 2 (automobiles) were retained, e x- cluding motorc ycles and trucks. This reduced the dataset to approximately 11,473,680 ro ws and 25 columns. - T rajectory length ltering: to ensure that each observ ed dri v er se gment pro vides suf cient information to characterize a dri ving style, v ehicles with ‘T otal Frames‘ less than 300 (equi v alent to 30 seconds of contin- uous data at 10 Hz) were remo v ed. This renement ensured meaningful observ ation lengths, resulting in a dataset of approximately 11,472,258 ro ws and 3,211 unique ‘V ehicle ID‘s. - Erratic data remo v al: data points containing ph ysically unrealistic or erroneous ‘v Acc‘ (instantaneous ac- celeration) or ‘v V el‘ (instantaneous v elocity) v alues were ltered out. Thresholds were set to retain ‘v Acc‘ v alues between -35.0 and 35.0 ft/s 2 and ‘v V el‘ v alues between 0.0 and 140.0 ft/s (approximately 95 mph). This crucial step addressed potential sensor noise or data transcription errors. Importantly , this ltering step did not signicantly alter the total number of ro ws or unique v ehicles, indicating a relati v ely clean initial ra w dataset after the rst tw o ltering steps. 3.3. Dri v er obser v ation segment extraction T o capture granular snapshots of indi vidual dri v er beha vior for subsequent proling, the cleaned v ehi- cle trajectories were se gmented into o v erlapping observ ation windo ws. This approach maximizes the number of training samples and allo ws the model to learn from the temporal e v olution of dri ving styles. - Se gment length: each e xtracted se gment consists of 300 consecuti v e frames, representing 30 seconds of continuous dri ving data. This duration w as selected as a balanced timeframe, long enough to capture char - acteristic beha vioral patterns (e.g., a fe w lane changes, v aried acceleration patterns) b ut short enough to represent a consistent dri ving moment. - Step size: se gments were e xtracted using a sliding windo w approach with a ’ step size’ of 50 fram es (equi v- alent to 5 seconds). This o v erlap between consecuti v e se gments for the same v ehicle substantially increases the number of a v ailable samples for training the deep learning model. - Extraction process: the process in v olv ed iterating through each unique ‘V ehicle ID‘ in the cleaned DataFrame. F or each v ehicle, continuous sub-trajectories of 300 frames were e xtracted, adv ancing by 50 frames at a time. A strict check ensured that only complete 300-frame se gments were retained. This procedure generated a to- tal of 211,824 dri v er observ ation se gments, each consisting of a DataFrame retaining the original 25 NGSIM columns for the 30-second windo w . 3.4. Micr oscopic beha vioral featur e engineering F or each of the 211,824 e xtracted se gments, a comprehensi v e set of 27 features w as engineered to cre- ate a ”beha vioral ngerprint”.These features were selected because the y capture the fundamental d i mensions of longitudinal control (acceleration), lateral maneuv ering (lane changes), and car -follo wing interactions (head- w ay), which are the primary indicators of dri ving. Let S k denote the k -th se gment consisting of N k = 300 frames. - Acceleration dynamics (8 features): these characterize longitudinal control. W e compute standard statistical moments (mean µ a , v ariance σ 2 a , standard de viation σ a , max, min) for the instantaneous acceleration v Acc . Additionally , we deri v e: a) Sk e wness ( γ 1 ): reects t h e asymmetry of acceleration, indicating a tendenc y t o w ards sharp braking or Behavior al ng erprints: driver pr oling using tr ansformer models on ne xt ... (Mohamed Laamimac h) Evaluation Warning : The document was created with Spire.PDF for Python.
1172 ❒ ISSN: 1693-6930 rapid acceleration. γ 1 = 1 N k N k X j =1 v Acc j − µ a σ a 3 (1) b) High-intensity thresholds: the percentage of time spent accelerating > 5 f t/s 2 ( P >a th ) or decelerating < − 5 f t/s 2 ( P <d th ). P >a th = 1 N k N k X j =1 I ( v Acc j > 5 . 0) × 100% (2) - Speed met rics (4 features): includes mean, v ariance, and standard de viation of v elocity . W e also compute the coef cient of v ariation ( C V v ) to normalize speed v ariability relati v e to the mean speed: C V v = σ v µ v , if µ v ̸ = 0 (3) - Lane changing beha vior (3 features): captures lateral maneuv ering. a) Lane change rate ( R LC ): normalized frequenc y of lane changes per minute. R LC = Count of lane changes 0 . 5 min (4) b) Lateral de viation ( σ l at ): standard de viation of the v ehicle’ s local X position within the lane, serving as a proxy for lane-k eeping stability . - Headw ay management (9 features): describes car -follo wing interaction. W e compute statistics (mean, v ari- ance, min) for both space headw ay ( µ h s ) and time headw ay ( µ h t ), calculated e xclusi v ely on frames where a preceding v ehicle is present ( F k ). a) Critical headw ay percentage: the proportion of follo wing time spent with a time headw ay belo w 1.0 second (indicating tailg ating). P <h cr it = 1 | F k | X j ∈ F k I ( T ime H eadw ay j < 1 . 0) × 100% (5) - Conte xtual features (3 features): includes v ehicle ph ysical dimensions (length and width) and normalized time of day (scaled to [0 , 1] ) to account for traf c density v ari ations associated with peak/of f-peak hours. 3.4.1. Handling missing head way data In free-o w conditions, headw ay v alues are undened. T o maintain numerical consistenc y , we impute these missing v alues with semantic constants: mean/v ariance ar e set to 1000.0, and the critical headw ay per - centage is set to 0.0. While this introduces a potential bias by treating “no v ehicle” as “v ery f ar v ehicle, ” it is necessary for the attention mechanism to process the full sequence and has been justied in similar time-series tasks. 3.5. Heuristic pr oxy label generation (weak super vision) Due to the absence of human-annotated labels, we emplo y a heuri stic labeling strate gy . W e dene tw o beha vioral proles based on statistical thresholds: - Aggressi v e: a se gment is labeled ’aggressi v e’ if it f alls into the top 20 th percentile for accelerat ion v ariance OR lane change count, or the bottom 20 th percentile for mean time headw ay . - Normal: all other se gments are labeled ’normal’. T o ensure the model does not simply replicate a deterministic rule, we enforce a strict decoupling between the spatial-temporal scope of the inputs and the labels. The proxy labels are gl ob a l—deri v ed from statistical thresholds calculated o v er the entire continuous 30-second se gment (300 frames). In contrast, the transformer operates on a local t emporal sequence, processing the data as 12 sequential sub-windo ws (12 × 27 features). Because the model is ne v er fed the global 30-second summary statistics directly , it cannot tri vially bypass the learning process. Instead, it is forced to utili ze its multi-head self-attention mechanism to aggre g ate local temporal dynamics and approximate the global beha vioral boundary . TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1168–1176 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1173 3.6. T ransf ormer -based pr oling model W e propose a transformer -based sequence model to learn these proles. The architecture (Figure 1) consists of: - Input embedding: the 27-feature v ectors are projected into a 64-dimensional latent space. - Positional encoding: added to retain the temporal order of the 12 sub-windo ws. - T ransformer encoders: tw o stack ed blocks with multi-head self-attention (4 heads) to capture long-range dependencies. - Classication head: a global a v erage pooling layer follo wed by a dense Softmax layer to predict the class probabilities (aggressi v e vs. normal). - T raining: utilized the Adam optimizer with a learning rate of 1e-4, a dropout rate of 0.2, and w as conducted o v er 50 epochs. The input trajectory sequence ( 12 × 27 ) is projected via an embedding layer and processed by tw o stack ed transformer encoder blocks. These blocks uti lize multi-head self-attention to capture temporal de- pendencies. The resulting features are condensed via global pooling and passed to a multi-layer perceptron (MLP)/Softmax classier to predict the heuristic beha vioral prole (aggressi v e vs. normal). Figure 1. Architecture of the weakly supervised proling model 4. RESUL TS AND DISCUSSION 4.1. Baseline comparison F our models were e v aluated under identical 5-fold cross-v alidation conditions: logistic re gre ssion (LR), random forest (RF), LSTM, and the proposed transformer . Cross-v alidation w as applied to ensure ro- b ustness ag ainst o v ertting to the heuristic proxy labels. Results are reported in T able 2 in terms of accurac y , precision, recall, and F1-score. T able 2. Comparati v e performance: transformer vs. baselines Model Accurac y Precision Recall F1-score LR 0.81 0.80 0.81 0.80 RF 0.88 0.89 0.87 0.88 LSTM (baseline) 0.86 0.85 0.86 0.85 T ransformer (ours) 0.97 0.97 0.96 0.97 The transformer achie v es an F1-score of 0.97, the highest acros s all models. The LR baseline per - forms weak est (F1 = 0.80), conrming that the classication boundary between aggressi v e and normal pro- les is inherently non-l inear and cannot be resolv ed by a simple decision function. The RF model impro v es substantially o v er LR (F1 = 0.88), demonstrating that ensemble feature interactions capture meaningful be- ha vioral structure. Ho we v er , because RF operates on independently dra wn sub-windo ws rather than the full temporal sequence, it is unable to capture the ordering and dynamics of dri ving e v ents. This aligns with the well-established vie w that dri ving style is a fundamentally temporal phenomenon: a sudden deceleration in a traf c jam carries an entirely dif ferent beha vioral meaning than the same action on a free-o w highw ay [5], [24]. The LSTM baseline (F1 = 0.85) underperforms the RF , lik ely because recurrent processing is sensiti v e to the gradient instability that arises when modeling dependencies across long sequences of sub-windo ws. The transformer o v ercomes this li mitation through its multi-head self-attention mechanism, which assigns attention weights across all 12 sub-windo ws simultaneously and is therefore better suited to capturing long-range depen- dencies within the 30-second observ ation se gment [4]. This superiority of self-attention o v er recurrence at the microscopic, indi vidual-v ehicle le v el is consistent with ndings from macroscopic traf c forecasting studies [11], [25]. Behavior al ng erprints: driver pr oling using tr ansformer models on ne xt ... (Mohamed Laamimac h) Evaluation Warning : The document was created with Spire.PDF for Python.
1174 ❒ ISSN: 1693-6930 4.2. Classication perf ormance and beha vioral nger printing The confusion matrix obtained on the held-out test set is sho wn in Figure 2. Of the total 10,000 test samples, 9,675 were correctly classied and 325 were misclassied, gi ving an o v erall accurac y of 96.75%. The f alse-ne g ati v e rate (aggressi v e se gments predicted as normal) is notably lo wer than the f alse-positi v e rate, indicating that the model is conserv ati v e in assigni ng aggressi v e labels, which is a desirable property for safety- critical ITS applications. Figure 2. Confusion matrix (transformer) A closer analysis of the misclassied samples re v eals that errors concentrate at the boundary between the tw o proles: se gments where a dri v er’ s kinematic beha viour uctuates around the 20th/80th percentile thresholds used for proxy labeling. This is an e xpected outcome of the weak supervision strate gy and does not indicate a model f ailure. On the contrary , it re v eals an important property of the learned representation: rather than replicating the deterministic heuristic rule, the transformer has learned a smoother , probabilistic decision boundary that better reects the continuum of real dri ving beha viour [5]. A dri v er whose acceleration v ariance occasionally crosses the labeling threshold is genuinely ambiguous, and the model’ s uncertainty in such cases is well-founded. This nding mirrors the distinction observ ed in macroscopic traf c modeling between crisp rule-based classication and the more nuanced patterns captured by deep sequence models [24]. 4.3. Generalization and implications f or ITS T o assess whether the learned beha vioral signatures are specic to the US-101 corridor or reect gen- eral human dri ving characteristics, we applied the trained model without retraining to trajectory data from the geographically distinct NGSIM I-80 dataset. Classication accurac y on this held-out dataset w as 94%, a re- duction of less than three percentage points relati v e to the in-distrib ution test set. This de gree of cross-dataset transferability suggests that the microscopic beha vioral features acceleration dynamics, lane-change rate, and headw ay management encode dri ving patterns that are relati v ely stable across dif ferent road en vironments, consistent with e vidence from drone-based studies linking microscopic dri ving signatures to macroscopic o w outcomes across di v erse urban settings [24]. The ability to generalize without retraining is practically impor - tant: it indicates that a model trained on one high-resolution dataset could be deplo yed for dri v er proling in en vironments where ne w labeled trajectory data are una v ailable, supporting scalable and personalized ITS applications such as adapti v e w arning systems and autonomous v ehicle interaction [26]. 5. CONCLUSION This w ork introduces a weakly supervised transformer -based frame w ork for dri v er beha vior proling, le v eraging NGSIM trajectory data to create distincti v e beha vioral signatures. By engineering a rich set of microscopic beha vioral features and applying a heuristic proxy labeling strate gy , we succes sfully trained a model to classify dri v ers into ‘aggress i v e’ and ‘normal’ proles with high delity . Crucially , our comparati v e analysis demonstrates that the transformer architecture signicantly outperforms both recurrent LSTM and non- temporal baselines. This nding conrms that the self-attention mechanism is uniquely suited for capturing the long-range temporal dependencies that characterize dri ving styles. This w ork demonstrates the feasibility of creating detailed “beha vioral ngerprints” from ra w trajectory data, mo ving be yond macroscopic traf c TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1168–1176 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 1175 analysis to understand the indi vidual agents. These prole s ha v e signicant potential for ITS applications, including enhancing autonomous v ehicle per ception by predicting human intent. The primary limitation of this study is the reliance on heuristic-based labels, which act as noisy proxies for ground truth. Additionally , the NGSIM US-101 data 2005 is relati v ely old; shifts in traf c composition and v ehicle technology may af fect generalization to modern eets. Future w ork will e xplore unsupervised learning and clustering techniques to disco v er these proles or g anically from the data, potentially re v ealing more nuanced beha vioral cate gories, and inte grate e xplainable articial intelligence (AI) to interpret the learned ngerprints. 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 u- tions, 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 Mohamed Laamimach ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Mghari Mohammed ✓ ✓ ✓ ✓ ✓ ✓ ✓ Aziz Mabrouk ✓ ✓ ✓ ✓ ✓ ✓ ✓ 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 Authors state no conict of interest. D A T A A V AILABILITY The NGSIM data is a v ailable at https://ops.fhw a.dot.go v/traf canalysistools/ngsim.htm. REFERENCES [1] D. Schrank, B. Ei sele, and T . Lomax, “ 2021 Urban Mobility Report , ” T e xas A&M T ransportation Institute, 2021. [2] S. An, B.-H. Lee, and D.-R. Shin, “ A Surv e y of Intelligent T ransportation Systems, ” in 2011 T hir d International Confer ence on Computational Intellig ence , Communication Systems and Networks , IEEE, Jul. 2011, pp. 332–337. doi: 10.1109/CICSyN.2011.76. [3] S. Hochreiter and J. Schmidhuber , “Long Short-T erm Memory , ” Neur al Computation , v ol. 9, no. 8, pp. 1735–1780, No v . 1997, doi: 10.1162/neco.1997.9.8.1735. [4] A. V asw ani et al. , “ Attention Is All Y ou Need, ” arXiv , pp. 1–9, Aug. 2023, [Online]. A v ailable: http://arxi v .or g/abs/1706.03762. [5] M. T reiber a nd A. K esting, T r af c Flow Dynamics . Cham: Springer Nature Switzerland, 2025. doi: 10.1007/978-3-031-93922-8. [6] M. Laamimach and A. Mabrouk, “Urban Congestion Prediction Based on Dri v er Beha vior: Hybrid CNN-LSTM Model with Atten- tion Mechanism, ” Lectur e Notes in Networks and Systems , v ol. 1924, pp. 491–502, 2026, doi: 10.1007/978-3-032-22567-2 43. [7] B. M. W illiams and L. A. Hoel, “Modeling and F orecasting V ehicular T raf c Flo w as a Seasonal ARIMA Process: Theoretical Basis and Empirical Results, ” J ournal of T r ansportation Engineering , v ol. 129, no. 6, pp. 664–672, No v . 2003, doi: 10.1061/(ASCE)0733- 947X(2003)129:6(664). [8] J. Zhang, Y . Zheng, and D. Qi, “Deep Spatio-T emporal Residual Netw orks for Citywide Cro wd Flo ws Prediction, ” Pr oceedings of the AAAI Confer ence on Articial Intellig ence , v ol. 31, no. 1, Feb . 2017, doi: 10.1609/aaai.v31i1.10735. [9] X. Ma, Z. T ao, Y . W ang, H. Y u, and Y . W ang, “Long short-term memory neural netw ork for traf c speed prediction us- ing remote micro w a v e sens or data, ” T r ansportation Resear c h P art C: Emer ging T ec hnolo gies , v ol. 54, pp. 187–197, 2015, doi: 10.1016/j.trc.2015.03.014. [10] M. Xu et al . , “Spatial-T emporal T ransformer Netw orks for T raf c Flo w F orecasting, ” arxiv , Mar . 2021. [11] J. Jiang, C. Han, W . X. Zhao, and J. W ang, “PDF ormer: Propag ation Delay-A w are Dynamic Long-Range T ransformer for T raf c Flo w Prediction, ” Pr oceedings of the AAAI Confer ence on Articial Intellig ence , v ol. 37, no. 4, pp. 4365–4373, Jun. 2023, doi: 10.1609/aaai.v37i4.25556. Behavior al ng erprints: driver pr oling using tr ansformer models on ne xt ... (Mohamed Laamimac h) Evaluation Warning : The document was created with Spire.PDF for Python.
1176 ❒ ISSN: 1693-6930 [12] J . Zhang, Y . Y ang, X. W u, and S. Li, “Spatio-tem poral transformer and graph con v olutional netw orks based traf c o w prediction, ” Scientic Reports , v ol. 15, p. 24299, Jul. 2025, doi: 10.1038/s41598-025-10287-5. [13] S. W ei, Y . Y ang, D. Liu, K. Deng, and C. W ang, “T ransformer -Based Spatiotemporal Graph Dif fusion Con v olution Netw ork for T raf c Flo w F orecasting, ” Electr onics , v ol. 13, no. 16, p. 3151, Aug. 2024, doi: 10.3390/electronics13163151. [14] A. Chang, Y . Ji , and Y . Bie , “T ransformer -based short-term traf c forecasting model considering traf c spatiotemporal correlation, ” F r ontier s in Neur or obotics , v ol. 19, p. 1527908, Jan. 2025, doi: 10.3389/fnbot.2025.1527908. [15] Z. F ang, L. P an, L. Chen, Y . Du, and Y . Gao, “Mdtp: A multi-source deep traf c prediction frame w ork o v er spatio-temporal trajectory data, ” Pr oceedings of the VLDB Endowment , v ol. 14, no. 8, pp. 1289–1297, Apr . 2021, doi: 10.14778/3457390.3457394. [16] W . Xiong, R. F onod, A. Alahi, and N. Geroliminis, “Multi-Source Urban T raf c Flo w F orecasting with Drone and Loop Detector Data, ” arXiv:2501.03492 , 2025, doi: 10.48550/arXi v .2501.03492. [17] N.-E. El F aouzi, H. Leung, and A. K urian, “Data fusion in intelligent transportation systems: Progress and challenges – A surv e y , ” Information Fusion , v ol. 12, no. 1, pp. 4–10, 2011, doi: 10.1016/j.inf fus.2010.06.001. [18] H. Zhang, H. Dong, and Z. Y ang, “TSGDif f: T raf c state generati v e dif fusion model using m ulti-source information fusion, ” T r ansportation Resear c h P art C: Emer ging T ec hnolo gies , v ol. 174, p. 105081, May 2025, doi: 10.1016/j.trc.2025.105081. [19] B. Lim, S. ¨ O. Arık, N. Loef f, and T . Pster , “T emporal Fusion T ransformers for interpretable multi-horizon time series forecasting, ” International J ournal of F or ecasting , v ol. 37, no. 4, pp. 1748–1764, Oct. 2021, doi: 10.1016/j.ijforecast.2021.03.012. [20] B. S. K erner , The Ph ysics of T raf c, v ol. 58, no. 11. in Under standing Comple x Systems , v ol. 58. Berlin, Heidelber g: Springer Berlin Heidelber g, 2004. doi: 10.1007/978-3-540-40986-1. [21] P . A. Lopez et al. , “Microscopic T raf c Simulation using SUMO, ” in 2018 21st International Confer ence on Intellig ent T r ansporta- tion Systems (ITSC) , IEEE, No v . 2018, pp. 2575–2582. doi: 10.1109/ITSC.2018.8569938. [22] L. Da, C. Chu, W . Zhang, and H. W ei, “CityFlo wER: An Ef cient and Realistic T raf c Simulator with Embedded Machine Learning Models, ” arXiv:2402.06127 , 2024, doi: 10.48550/arXi v .2402.06127. [23] G. A. M. Meiring and H. C. Myb ur gh, “ A Re vie w of Intelligent Dri ving Style Analysis Systems and Related Articial Intelligence Algorithms, ” Sensor s , v ol. 15, no. 12, pp. 30653–30682, Dec. 2015, doi: 10.3390/s151229822. [24] E. Mantouka, P . F afoutellis, and E. I. Vlahogianni, “From Microscopic Dri ving Beha vior to Macroscopic T raf c Dynamics Using Drone Observ ations, ” Data Science for T r ansportation , v ol. 8, no. 1, p. 9, Apr . 2026, doi: 10.1007/s42421-026-00152-w . [25] S. Liu and X. W ang, “ An impro v ed transformer based traf c o w prediction model, ” Scientic Reports , v ol. 15, no. 1, p. 8284, Mar . 2025, doi: 10.1038/s41598-025-92425-7. [26] F . A. Sellal, A. A youb Bellachia, M. M. Dif, E. De Rautlin De La Ro y , M. A. Bouchiha and Y . Ghamri-Doudane, ”XAI-Dri v en Ma- chine Learning System for Dri ving Style Recognition and Personalized Recommendations, ” 2025 IEEE 102nd V ehicular T ec hnolo gy Confer ence (VTC2025-F all) , 2025, pp. 1-7, doi: 10.1109/VTC2025-F all65116.2025.11310755. BIOGRAPHIES OF A UTHORS Mohamed Laamimach is a researcher at the Information Securi ty Intelligent Systems and Applications (ISISA) lab . His research interests include deep learning, trajectory analysis, and intel- ligent transportation systems. He can be contacted at email: mohamed.laamimach@etu.uae.ac.m a. Mghari Mohammed is a Professor Assistant of Computer Sciences at Abdelmalek Es- saadi Uni v ersity , Morocco. Where he recei v ed the Ph.D. de gree in AI and Computer Science from the Abdelmalek Essa ˆ adi Uni v ersity in 2024. His w ork specically in v olv ed applying adv anced trans- former models to lo w-resource languages, achie ving state-of-the-art performance on a specic task. He can be contacted at email: mohammed.mghari@uae.ac.ma. Aziz Mabr ouk is a Full Professor at Abdelmalek Essa ˆ adi Uni v ersity , T ´ etouan, Morocco. His research focuses on decision support systems, spatial data modeling, and graph-based approaches for comple x systems. His scientic contrib utions address transport and mobility planning, trajectory analysis, v oronoi diagrams, and spatial optimization, with applications to intelligent transportation systems and urban decision-making. He can be contacted at email: amabrouk@uae.ac.ma. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 4, August 2026: 1168–1176 Evaluation Warning : The document was created with Spire.PDF for Python.