MS/CAS PENS Dataset Overview

• MS/CAS PENS Dataset is a collection of multi-modal pen sensor data capturing kinematic, pressure, and inertial signals for cognitive and handwriting analysis.
• It supports benchmark tasks using deep-learning architectures like CNN+BiLSTM, Transformers, and TCNs, enhancing online handwriting recognition and neurocognitive evaluation.
• The dataset also enables personalized summarization by modeling user interaction trajectories with augmentation techniques that boost diversity metrics and improve model performance.

The MS/CAS PENS Dataset denotes a collection of datasets, experimental protocols, and benchmark tasks stemming from motion sensor and computer-aided systems (MS/CAS), most prominently for applications in pen-based cognitive assessment, handwriting recognition using sensor-enhanced pens, and personalized content summarization based on longitudinal user preference data. The dataset has found application in neurocognitive analytics, pen-on-paper handwriting recognition, and, most recently, in the modeling and augmentation of user interaction histories for personalized text summarization. This entry synthesizes the structure, properties, methodologies, and use cases of the MS/CAS PENS dataset family, drawing on technical reports and peer-reviewed analyses.

1. Origin and Scope of MS/CAS PENS

The MS/CAS PENS Dataset initially emerged from research on signal-level digital pen analytics, focusing on pen-on-paper interfaces that capture detailed kinematic, pressure, and inertial measurements in clinical and engineering scenarios. Early deployments targeted semi-automatic dementia assessment, employing digital pens to digitize traditional tests (e.g., Clock Drawing, MMSE, CERAD), with over 160 features per stroke recorded, including time-stamped signals of pressure, tremor, in-air time, and duration. Later expansions of the dataset included multimodal pen signals for online handwriting recognition, captured with embedded IMUs (accelerometer, gyroscope, magnetometer, force sensor), producing high-frequency (100 Hz), 13-dimensional time series spanning triaxial motion, magnetic field, and force data (Sonntag, 2018, Ott et al., 2022).

Subsequently, the PENS framework was adapted to user modeling and personalized summarization, aggregating detailed user preference trajectories—click-skip/update sequences over textual items—but without corresponding reference summaries. Thus, over time, the MS/CAS PENS datasets developed into a multi-purpose resource for signal analytics and user-centric modeling in both neurocognitive assessment and natural language personalization pipelines (Chatterjee et al., 11 Oct 2025).

2. Data Acquisition and Signal Channels

The physical instantiation of the MS/CAS PENS Dataset typically relies on specially instrumented ballpoint pens (such as the DigiPen by STABILO), which record the following sensor modalities:

Sensor Type	Channel Count	Signal Characteristics
Accelerometer (dual)	6	Triaxial, front/rear configuration for fine-grained motion
Gyroscope	3	Angular velocity along three principal axes
Magnetometer	3	Magnetic field intensity in three spatial dimensions
Force sensor	1	Scalar normal force at pen tip (pen-surface interaction)

Sensor streams are captured at 100 Hz, enabling high-resolution time-series reconstruction of pen kinematics. The force sensor is critical for accurate pen stroke segmentation, as pen-up/pen-down transitions provide consistent stroke delimiters for both character- and sequence-level recognition tasks (Ott et al., 2022). The resulting raw data supports the derivation of both stroke-level features (>100 per stroke in clinical neurocognitive applications) and global features for sequence learning in handwriting recognition.

In user modeling and summarization, the dataset encodes user-document interactions, capturing click/skip events, timestamps, document identifiers, topic labels, and, in the case of augmentation experiments, derived or perturbed summaries (Chatterjee et al., 11 Oct 2025).

3. Benchmark Tasks and Model Architectures

3.1 Handwriting Recognition

A principal benchmark established on the MS/CAS PENS Dataset involves online handwriting recognition (OnHWR) from pen-based IMU data—distinct from classical tablet or vision-based methods. Recognition tasks span:

• Sequence-to-sequence learning: Given a multivariate time series (MTS), predict the entire sequence (equation, word string).
• Single character-based classification: Segment MTS into individual characters, each classified independently.

Benchmarked architectures include:

• CNN+BiLSTM: Convolutional modules for feature extraction, followed by bidirectional LSTM layers for temporal context. Achieves superior Character Error Rates (CER) relative to alternatives.
• Transformer-based models: Incorporate multi-head self-attention. Standard attention: $$\mathrm{softmax}\left( \frac{QK^T}{\sqrt{d_k}} \right) V$$.
• Temporal convolutional networks (TCNs), and ensemble CNNs (InceptionTime).

Loss functions vary: sequence tasks leverage Connectionist Temporal Classification (CTC) loss, eliminating explicit alignment; for classification, categorical cross-entropy (CE) and its variants (Focal Loss, label smoothing, bootstrapping, Symmetric Cross Entropy) are used to mitigate class imbalance and overconfidence.

3.2 User Trajectory Modeling for Personalization

In the context of personalized summarization, the dataset defines each user by a trajectory (sequence) of interactions with textual documents, capturing behavioral signals over time. However, native MS/CAS PENS data lacks gold-reference summaries, limiting end-to-end supervised learning. Solutions such as PerAugy introduce augmentation schemes operating over the user interaction graph, thereby supporting training of user-encoders (e.g., NAML, NRMS, EBNR, TrRMIo) and text summarization frameworks (Chatterjee et al., 11 Oct 2025).

4. Data Augmentation and Diversity Enhancement

Modern applications of the MS/CAS PENS Dataset, particularly in personalized summarization, have exposed intrinsic limitations: the absence of ground-truth summaries and low diversity of topic transitions in user trajectories hinder model generalization. The PerAugy augmentation framework addresses both:

• Double Shuffling (DS): Cross-trajectory substitution of segments (with offset and gap-length parameters), preserving natural starts and bootstrapping mid-trajectory shifts from other users’ histories.
• Stochastic Markovian Perturbation (SMP): For each substituted summary node ($s$-node), SMP minimizes a weighted sum of Root Mean Square Distance (RMSD) over $k$ preceding steps:

$$\mathrm{SMP}(s^{(t_i)}) = \arg\min_{st_p \in d^{(t_{i-1})}} \left[ \sum_{q=1}^k e^{-\lambda \cdot \mathrm{pos}(c_q)} \cdot \sigma(e_{st_p}, e_{c_q}) \right],$$

where $\sigma$ is RMSD, $\lambda$ a decay constant, and $\mathrm{pos}(\cdot)$ the positional weight.

The composite effect is a synthetically diversified dataset with increased thematic transitions, empirically shown to correlate with substantial gains in user-encoder and summarizer metrics (e.g., up to 0.13–0.14 AUC improvement, average 61.2% increase in PSE-SU4 personalization score) (Chatterjee et al., 11 Oct 2025).

5. Diversity Metrics and Empirical Correlations

To quantify and analyze dataset diversity post-augmentation, three primary metrics are introduced:

Metric	Definition	Empirical Correlate
Topics per Trajectory (TP)	Count of unique discrete topics visited per trajectory	Positive with user-encoder accuracy
Rate of Topic Change (RTC)	Frequency of topic label changes in a trajectory	Positively associated, less nuanced
Degree-of-Diversity (DegreeD)	Ratio of “thematic divergence” in document and summary space (embedding-based, RMSD)	Strongest correlation ($r = 0.68$ Pearson, $0.73$ Spearman) with user-encoder performance

Higher TP and DegreeD reliably predict improved personalization and model generalization. These metrics supplement basic trajectory statistics for rigorous dataset diagnostics and future tuning (Chatterjee et al., 11 Oct 2025).

6. Applications and Domain Generalization

The versatility of the MS/CAS PENS Dataset is evidenced by its expansion across domains:

• Neurocognitive Testing: Early work confirms digital pen features as discriminative biomarkers for dementia and mild cognitive impairment, with >160 temporal/kinematic features per stroke enabling analytics that go beyond content to the “how” of writing (Sonntag, 2018).
• Handwriting Recognition: Sensor-enhanced pen data bridges the gap between real-world pen-and-paper input and traditional touch/tablet datasets, providing realism in signature verification, writer identification, and robust text recognition on uninstrumented surfaces (Ott et al., 2022).
• Personalized Content Summarization: User preference histories are leveraged to inform recommender and summarizer models, with augmented PENS trajectories (via PerAugy) supporting improved personalization. This framework generalizes to open-domain data (e.g., Reddit), suggesting applicability beyond the benchmark corpus (Chatterjee et al., 11 Oct 2025).
• Potential Extended Uses: A plausible implication is transfer to affective computing (e.g., mood inference from pen features), adaptive educational systems (monitoring engagement or stress), and continuous user modeling.

7. Future Directions and Open Challenges

The principal limitations of the MS/CAS PENS Dataset include the absence of user-generated target summaries and the relatively low topic-transition diversity in user trajectories. Augmentation frameworks such as PerAugy have demonstrated efficacy in addressing these gaps through strategic cross-trajectory sampling and summary perturbation. Theoretical extensions—such as continuous-time stochastic user preference diffusion (e.g., Itô processes)—are identified as future enhancements, potentially improving the temporal realism of simulated user behavior. The strong empirical role of diversity metrics (TP, DegreeD) in predicting model gains underscores their utility for broader dataset engineering.

A plausible implication is that these types of augmentation and evaluation protocols, validated on PENS, could be profitably adapted to related tasks in personalized recommendation, cognitive analytics, and adaptive user interfaces, further broadening the scope and impact of MS/CAS data-driven research.