Accessing civil engineering project documentation can be challenging when information needs to be retrieved quickly and conveniently in the field. This study aims to implement a mobile Knowledge Management System (KMS) application incorporating a Hybrid Search mechanism to support project documentation retrieval. The system was developed using the Waterfall method, which consists of requirements analysis, design, implementation, testing, and maintenance. The Hybrid Search mechanism combines an AI-based recommendation approach, which relies on rule matching and civil engineering domain-specific terminology, with keyword-based searching using SQL LIKE. The recommendation process involves query preprocessing, including case folding, tokenization, stopword removal, and synonym/domain expansion, followed by ranking based on project information matching scores. The recommendation results are then integrated with keyword search and given priority in the ranking process. Black-box testing results indicate that the search functionality and related features operate as expected. This study delivers a search mechanism that helps users retrieve civil engineering project documentation based on relevant terminology and project information matching.
Academic information systems require performance evaluation to determine how responsiveness changes under concurrent access. This study compares measurements obtained using Apache JMeter and Locust for a Laravel-based academic information system testbed under equivalent authenticated workloads. Experiments involved 10, 50, and 100 virtual users, with three replications per tool and load level, yielding 18 runs. Each virtual user authenticated once using a randomly selected student account and subsequently accessed Dashboard, KRS, Nilai, and KHS services. The workload used an approximately 10-second ramp-up, a 300-second duration, and a 1–5-second think time. Response time, percentile latency, throughput, error rate, and between-run variability were analyzed. At 100 virtual users, mean response time reached 1,205.33 ms for JMeter and 1,745.74 ms for Locust, while mean P95 reached 1,343.33 ms and 2,833.33 ms, respectively. Between 50 and 100 virtual users, throughput increased from 15.507 to 23.445 requests/s for JMeter and from 15.444 to 19.934 requests/s for Locust, indicating diminishing gains relative to the increase in concurrency. All runs recorded a 0% error rate, although responsiveness deteriorated substantially at the highest tested load. Locust exhibited greater between-run variability at 100 virtual users. The study contributes a controlled comparison of authenticated academic workloads combining endpoint-level analysis, percentile latency, and repeated measurements. These findings characterize the tested environment and do not establish production capacity, identify a specific server-side bottleneck, or demonstrate universal tool superiority.
Indonesian mobile application reviews are informal free-text feedback. Their sentiment and emotion labels are costly to obtain and only moderately reliable. This study determines which pre-trained encoder, IndoBERT or IndoBERTweet, classifies emotion and sentiment more reliably in Indonesian app reviews, and tests whether combining or routing between the two encoders improves on the reference single model. IndoBERT was pre-trained on general Indonesian text and IndoBERTweet on Indonesian tweets. Both were fine-tuned for six-class emotion and three-class sentiment classification on 21,065 de-duplicated reviews. Five random seeds, McNemar tests, and paired bootstrap confidence intervals evaluated confidence-based and task-based routing, probability averaging, sentiment derived from emotion, and a lexical register analysis. IndoBERTweet outperformed IndoBERT on emotion, reaching a macro-F1 of 0.627 against 0.596, with the advantage concentrated on reviews containing slang. Because IndoBERTweet performed at least as well on both formal and informal reviews, confidence-based routing gave no additional gain. Averaging the probabilities of ten models raised emotion macro-F1 significantly to 0.652, an improvement of 0.024 over the reference single model. Sentiment derived from the predicted emotion was never worse than a dedicated sentiment head. It performed significantly better for a single IndoBERTweet model, reaching 0.807 against 0.796, eliminated 10.2 to 14.6% of reviews with contradictory separate heads, and halved inference cost. Single-model modifications that did not degrade performance converged near 0.62 to 0.63, consistent with annotation reliability limiting attainable performance. These findings give practitioners an evidence-based default configuration: IndoBERTweet as the encoder, with sentiment read directly from the predicted emotion label.
Smart home networks connect many low-cost IoT devices that exchange traffic continuously. This is a very attractive target for attackers and a difficult setting for intrusion detection. In this paper we investigate a hierarchical dependent ensemble approach in which one dependent ensemble algorithm is nested inside another. Three algorithms, AdaBoostM1, MultiBoostAB, and Stacking, were combined in all six possible outer-inner configurations and evaluated in WEKA on the DS2OS smart home traffic dataset using stratified 3-fold cross-validation, with the dataset's naturally imbalanced class distribution left intact. Five of the six configurations reached near-perfect results, with accuracy between 99.9964% and 99.9997%, Kappa between 0.9993 and 0.9999, and a false alarm rate of effectively zero. The AdaBoostM1→MultiBoostAB configuration, which relied on OneR rather than J48 as its weak learner, achieved 99.11% accuracy and a Kappa of 0.837, and it failed to detect the Data Type Probing and Wrong Setup classes altogether. Confusion matrix analysis showed that the nesting order of the ensembles had little effect on detection once the base classifier was held constant, whereas the choice of weak learner had a large one. Configurations with Stacking as the outer ensemble were also the fastest to train. Finally, we show that the near-ceiling scores are likely inflated by device-identifier attributes in DS2OS and discuss what this means for how intrusion detection results on this dataset should be reported and interpreted.
Social media is an important communication channel for higher education institutions to share information and build institutional identity. This study analyzes Instagram content from a selected institutional account, including collaborative posts involving other accounts. A descriptive quantitative approach was applied to 100 valid posts collected through web scraping. The analysis combined Exploratory Data Analysis (EDA) to describe content formats, posting activity, likes, and comments with Latent Dirichlet Allocation (LDA) to identify recurring themes in captions. The results showed that image posts were the most common format, followed by carousel posts and videos. LDA identified three main themes: student achievement and development, religious values and character building, and institutional collaboration and community engagement. Religious values and character building accounted for the largest proportion of posts and had the highest average numbers of likes and comments among the identified topics. Content specifically focused on student admission was less prominent in the identified themes. These findings provide descriptive insights for future content planning; however, they do not establish causal relationships between content characteristics, engagement, or admission outcomes.
