Garuda - Garba Rujukan Digital

Auditable Automated Essay Scoring and Formative Feedback: A Rubric-Grounded Pipeline for Secondary and Higher Education Qi Xin
Journal of Applied Artificial Intelligence in Education Vol 2, No 1 (2026): July 2026
Publisher : Academic Bright Collaboration

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.66053/jaaie.v2i1.348

Automated essay scoring in education is increasingly expected to do more than reproduce human holistic scores; classroom use also demands rubric-aligned feedback, transparent evidence, and a way to route uncertain cases to teachers. In this study, “LLM-ready” refers to a system that outputs structured score evidence, weak-trait signals, and document-level anchors that can later be verbalized by a language model without changing the underlying decision trace. This study aimed to evaluate whether a rubric-grounded, LLM-ready pipeline can achieve competitive scoring accuracy while also generating auditable formative feedback and a teacher-controllable review signal. The evaluation used the public ASAP corpus of 12,976 essays across eight prompts and prompt-wise five-fold cross-validation. Four holistic scorers were compared: length-only, rubric forest, prompt-adaptive centroid regressor (PACR), and the final RG-Score ensemble with trait grounding, isotonic calibration, and audit control. Auxiliary analytic scoring was examined on Prompts 2 and 7–8, and feedback experiments were conducted on all 2,292 essays from Prompts 7 and 8. PACR obtained the highest macro QWK of 0.739, while RG-Score reached 0.738 and provided a calibrated, auditable path to feedback. The prompt-level QWK for RG-Score ranged from 0.66 to 0.82, with particularly strong gains on Prompts 6 and 7. Auxiliary analytic scoring yielded QWK values of 0.623 for Prompt 2 domain2, 0.604 on average for Prompt 7 traits, and 0.506 on average for Prompt 8 traits. The rubric-grounded evidence feedback template achieved a Trait Recall@2 of 0.829, a valid evidence rate of 0.912, and an auditability index of 0.893 on Prompts 7 and 8. These findings support rubric-grounded AES as a practical assessment-support approach for secondary-school writing and as a structured foundation for higher-education formative feedback workflows, while also indicating that weaker trait models should be treated as advisory rather than fully autonomous

Log Anomaly Detection with Conformal Alert Control and Evidence-Grounded Incident Ticket Generation Qi Xin
Aviation Electronics, Information Technology, Telecommunications, Electricals, and Controls (AVITEC) Vol 8, No 2 (2026): August
Publisher : Institut Teknologi Dirgantara Adisutjipto

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.28989/avitec.v8i2.3974

Operational logs are a primary source of evidence for reliability engineering, incident response, and security operations, but log anomaly detection is useful only when scores can be translated into controlled alerts and auditable incident evidence. This paper presents a reproducible end-to-end AIOps pipeline that normalizes raw logs into templates, aggregates them into sliding windows, scores anomalies with representative detectors, calibrates alerts with conformal prediction, and generates evidence-grounded incident tickets. The revised evaluation includes BGL_2k and two additional public sequence benchmarks, HDFS and OpenStack, and adds representative LogAnomaly-style and LogBERT-lite baselines to the original TF-IDF+LR, Isolation Forest, DeepLog-style LSTM, and Transformer comparisons. On BGL_2k, Isolation Forest provides the best ranking performance among the original four detectors (test PR-AUC = 0.750), while the additional HDFS experiment shows that the masked-context LogBERT-lite baseline obtains the strongest sequence-level result (PR-AUC = 0.947, F1 = 0.905). OpenStack remains difficult because the available normal training sample is very small, producing low F1 across all added baselines. We also report inference latency, throughput, memory footprint, conformal alpha sensitivity, window-size sensitivity, model-strategy ablations, and structured false-positive/false-negative patterns. The results should be interpreted as reproducible operational validation of the detection-calibration-ticket workflow rather than a claim of state-of-the-art detector accuracy. The pipeline demonstrates how calibrated scores and template-level evidence can support practical alert control and ITSM-ready ticket generation.

LLM-Agent-Style Automated Usability Testing on MiniWoB++: A Reproducible Chunked Full-Run with ReAct, Plan-Execute, and Self-Reflect Policies Qi Xin
Blockchain, Artificial Intelligence, and Future Research Vol. 2 No. 1 (2026): May 2026
Publisher : WISE Pendidikan Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.70211/bafr.v2i1.384

Automated usability testing can reduce the cost of repeatedly checking whether a web interface supports reliable and efficient task completion, but existing scripted tests are brittle and many agent evaluations report benchmark scores without translating failures into usability diagnostics. This study asks how three LLM-agent-style strategies—ReAct, Plan-Execute, and Self-Reflect—differ in effectiveness, efficiency, and failure modes when applied to MiniWoB++ tasks, and whether their logged traces can support actionable UI analysis. We conducted a controlled experimental benchmark on 130 MiniWoB++ web tasks, running each strategy once under a fixed seed with the same deterministic DOM-grounded controller, headless Chromium harness, 10-step limit, and 2.0 s episode budget, producing 390 episodes. We analyzed task success, steps, wall-clock time, interaction category, difficulty bins, and failure categories using paired per-task comparisons and descriptive aggregation. Plan-Execute achieved the highest success rate (14.6%, 19/130), compared with 10.8% (14/130) for both ReAct and Self-Reflect; its advantage was most evident in form/transaction and selection tasks, while all strategies performed similarly on simple click/button tasks and failed on drag/scroll tasks. Failure analysis showed that wrong outcomes and element-grounding errors were the dominant bottlenecks, indicating that explicit planning improves coverage only when target elements can be reliably grounded. The findings contribute a reproducible baseline and a usability-oriented failure taxonomy for automated web-agent testing, suggesting that future frameworks should prioritize semantic grounding, plan validation, richer action primitives, and designer-facing diagnostics.

Host-Based Intrusion Detection with System Call Sequences: Window Localization and Forensic Narratives Qi Xin
Aviation Electronics, Information Technology, Telecommunications, Electricals, and Controls (AVITEC) Vol 8, No 2 (2026): August
Publisher : Institut Teknologi Dirgantara Adisutjipto

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.28989/avitec.v8i2.3973

Host-based intrusion detection systems (HIDS) and endpoint detection and response platforms increasingly rely on high-volume host telemetry such as system-call and API-call sequences. Accurate trace classification is useful, but operational response also requires localization of the suspicious fragment and a readable evidence narrative. This paper presents an analyst-oriented HIDS workflow with three stages: window-based detection using 3-gram logistic regression, 1D CNN, BiLSTM with attention pooling, and Transformer encoder models; suspicious-window localization using maximum-probability window selection and attention attribution; and evidence-grounded forensic narrative generation from the same localized evidence. The main ADFA-LD evaluation uses 1,263 training traces and 316 test traces; the best detector, 3-gram logistic regression with W=100, achieves precision 0.958, recall 0.913, F1 0.935, and AUROC 0.980. To address dataset-size and cross-telemetry concerns, this study adds two bounded additional evaluations: an eight-family Windows API-call sequence experiment on Mal-API-2019 and a large NetFlow stress test on NF-ToN-IoT-v3 with 10,000, 20,000, 50,000, and 100,000 training records. The NetFlow experiment shows that deep models become more stable as data volume increases, but the linear baseline remains strongest at 100,000 records (F1=0.888), while the Transformer improves to F1=0.856. These results support a careful conclusion: larger datasets help deep models, but they do not automatically make a Transformer outperform a strong shallow baseline when the signal is dominated by compact local motifs or well-separated tabular flow features.

Early-warning analytics with LLM intervention rationales for student retention decisions: Classroom interaction modeling with xAPI-edu-data and dropout/success prediction Qi Xin
Interdisiplinary Journal of Pedagogy and Research in Media Technology Vol. 2 No. 1 (2026): Interdisciplinary Journal of Pedagogy and Research in Media Technology
Publisher : CV. SPDFHarmony

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.64268/inspire.v2i1.117

Background: Student-retention early-warning systems have substantially improved predictive performance, yet their outputs often remain limited to risk scores that provide little guidance for educational intervention.Aims: This study proposes and empirically evaluates a reproducible two-stage framework that integrates classroom interaction modeling, institutional retention prediction, and structured LLM-ready intervention rationales for student-retention decision support.Methods: The framework was evaluated using two complementary benchmark datasets. xAPI-Edu-Data was used for classroom interaction modeling, whereas Predict Students' Dropout and Academic Success was used for institutional retention prediction. The datasets were analyzed independently rather than merged. Stratified train-test splits, ensemble machine learning models, feature ablation, feature-importance analysis, probability calibration, and rationale-quality evaluation were employed.Result: LightGBM achieved the highest Macro-F1 (0.7775) on the xAPI-Edu-Data benchmark, while XGBoost produced the best overall performance on the institutional retention dataset, achieving a multiclass accuracy of 0.7672, a Macro-F1 of 0.6964, and a ROC-AUC of 0.8889. In binary dropout prediction, XGBoost achieved a ROC-AUC of 0.9360 and an Average Precision of 0.9101. Behavioral engagement, attendance, academic progression, and tuition-related variables consistently emerged as the most informative predictors. The structured rationale layer achieved complete evidence alignment, actionability, and monitoring specificity while generating 392 unique intervention rationales.Conclusion: The proposed framework demonstrates that early-warning analytics can move beyond risk prediction by integrating predictive analytics, explainable AI, and structured intervention rationale generation into a transparent, evidence-grounded decision-support workflow for improving student-retention decisions in higher education.

Self-Supervised Log Anomaly Detection with LogBERT-Style Transformers: Full Empirical Evaluation on a Reproducible SynHDFS Benchmark Qi Xin
JEECS (Journal of Electrical Engineering and Computer Sciences) Vol. 11 No. 1 (2026): June
Publisher : Fakultas Teknik Universitas Bhayangkara

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.54732/jeecs.v11i1.3

Log-based anomaly detection is a core problem in AIOps because system logs provide fine-grained evidence of failures, performance regressions, and security incidents. Recent work has shown that self-supervised sequence modeling substantially improves generalization compared with purely frequency-based detectors, especially when labeled anomalies are scarce. This paper presents a LogBERT-style transformer framework for session-level log anomaly detection and reports a complete, reproducible experimental evaluation. Due to download constraints of large archived log datasets in this environment, we construct a faithful fallback benchmark, SynHDFS-6k, which mimics HDFS-style block workflows by composing normal execution patterns and injecting five realistic anomaly types. SynHDFS-6k contains 6000 sessions with a fixed 5.0% anomaly rate and a vocabulary of 20 event templates. We train a two-layer transformer encoder with masked language modeling on normal sessions only and derive an anomaly score using pseudo log-likelihood (PLL) computed by masking each token position once. We compare against unigram and bigram probabilistic models, PCA reconstruction error, one-class SVM, isolation forest, a DeepLog-style GRU next-event predictor, and a supervised logistic regression upper bound. On the SynHDFS-6k test split, the proposed LogBERT-PLL achieves Precision=0.615, Recall=0.533, F1=0.571, ROC-AUC=0.898, and PR-AUC=0.594. We additionally analyze transformer scoring strategies (PLL mean, PLL top-k, PLL max, random masking, and CLS Mahalanobis), report runtime and model capacity trade-offs, and quantify per-anomaly-type detection behavior. The study provides an end-to-end blueprint for transformer-based self-supervised log anomaly detection under a fully specified protocol, and it highlights strengths and limitations that inform deployment on real-world HDFS logs.

Explainable and Fair Credit Risk Scoring with Counterfactual Explanations: A Reproducible Evaluation on the German Credit Dataset (HELOC-Motivated) Qi Xin
J-INTECH ( Journal of Information and Technology) Vol 14 No 02 (2026): Journal of Information and Technology
Publisher : LPPM Universitas Bhinneka Nusantara

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.32664/j-intech.v14i02.2228

Credit risk scoring requires models that are accurate, fair, and able to provide actionable explanations for adverse decisions. Motivated by the HELOC explainable lending benchmark, but using the publicly downloadable Statlog German Credit dataset as a fully reproducible proxy when direct HELOC access is constrained, this study evaluates explainable and fair credit scoring on 1,000 applicants. We train five common models—logistic regression (LR), decision tree (DT), random forest (RF), XGBoost (XGB), and LightGBM (LGBM)—on a fixed 600/200/200 train/validation/test split with a consistent preprocessing pipeline. Thresholds are selected on the validation set to maximize F1 for bad-risk detection. LR achieves the best test AUC (0.7888), LGBM the highest test accuracy (0.6550), and RF the best calibration (ECE=0.0473), showing that discrimination, thresholded accuracy, and calibration do not align. Fairness is audited by a derived sex attribute using demographic parity and equalized odds. Baseline LGBM shows an approval-rate difference of −0.0846 (female minus male) and an equalized-odds gap of 0.1000. Reweighing reduces the approval-rate difference to −0.0642 while preserving AUC (0.7804), and equal-opportunity thresholding reduces the equalized-odds gap to 0.0750. For individual explainability, we generate counterfactual recourse for rejected applicants using six actionable features. Feasible recourse is defined as the existence of at least one action-constrained counterfactual that changes the decision from reject to approve; 78.64% of rejected applicants receive such recourse, with mean cost 1.5298 measured as standardized numeric change plus categorical steps. Across five retraining seeds, LGBM AUC is stable (mean 0.7742, std 0.0022), but fairness gaps vary. The study provides a reproducible template for jointly evaluating performance, calibration, fairness, mitigation, and recourse in credit scoring

Self-Supervised Customer Representation Learning for Segmentation and Next-Purchase Prediction on UCI Online Retail Qi Xin
J-INTECH ( Journal of Information and Technology) Vol 14 No 01 (2026): Journal of Information and Technology
Publisher : LPPM Universitas Bhinneka Nusantara

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.32664/j-intech.v14i01.2229

Customer analytics in financial retail, payments, and bank marketing frequently relies on segmentation and propensity prediction, but transactional logs are sparse, high-dimensional, and only weakly labeled. This paper presents a fast and reproducible self-supervised learning pipeline that converts raw e-commerce transactions into customer representations and evaluates them on two downstream tasks: customer segmentation and next-purchase prediction. We conduct full experimental evaluation on the UCI Online Retail dataset (541,909 invoice-line transactions from 2010-12-01 to 2011-12-09). After deterministic cleaning (removing cancellations and non-positive prices/quantities), 397,884 valid line items remain, spanning 4,338 customers, 18,532 invoices, 3,665 products, and 37 countries. For each customer we construct an ordered invoice sequence and define a canonical item per invoice (the item with the largest aggregated quantity). For each invoice transition we build a dual-view customer state vector that concatenates a lifetime purchase count view and a recent-window view (30 days), then learn embeddings via TF-IDF reweighting and truncated SVD. To increase robustness we introduce a denoising ridge projection (DRP) objective: a linear denoising model trained to map corrupted TF-IDF state vectors back to clean SVD embeddings without using labels, which yields denoised customer embeddings for downstream models. Our main contribution is an applied, computationally light integration of TF-IDF+SVD embeddings with a denoising linear projection for reuse across segmentation and next-purchase prediction, rather than a fundamentally new learning paradigm. In next-purchase prediction restricted to the 200 most frequent target items, a multinomial logistic model trained on DualDRP embeddings achieves Hit@20=0.587, outperforming MostPopular (Hit@20=0.327) and Markov (Hit@20=0.291). In segmentation we apply k-means clustering and analyze cluster-level RFM statistics and dominant products, showing that the learned embeddings recover actionable segments such as high-value frequent buyers and low-activity long-tail customers. All results, tables, and figures are generated with fixed random seeds and are reproducible in this environment

Title

Found 8 Documents
Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title Search

Found 8 Documents Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title

Found 8 Documents
Search