Sequential web user classification from system logs using BiLSTM and TCN
Artykuł w czasopiśmie
MNiSW
140
Lista 2024
| Status: | |
| Autorzy: | Rak Tomasz, Greń Piotr, Charytanowicz Małgorzata |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 16 |
| Numer artykułu: | 35144 |
| Strony: | 1 - 16 |
| Impact Factor: | 4,9 |
| Web of Science® Times Cited: | 0 |
| Scopus® Cytowania: | 0 |
| Bazy: | Web of Science | Scopus |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | TAK |
| Licencja: | |
| Sposób udostępnienia: | Witryna wydawcy |
| Wersja tekstu: | Ostateczna wersja opublikowana |
| Czas opublikowania: | W momencie opublikowania |
| Data opublikowania w OA: | 5 września 2026 |
| Abstrakty: | angielski |
| Modern transaction cloud platforms generate large amounts of system log data during customer operations. The analysis of these data can be employed for automatic classification of application user types (e.g., active investor, passive user, analyst) without the need to track direct personal data. The specific problem is that feature-centric classifiers discard the temporal order of requests. While the same stock-trading environment had not previously been evaluated using a controlled comparison of recurrent and convolutional sequence encoders. The objective of this study is therefore to compare BiLSTM and TCN for assigning fixed-length log sequences to three predefined behavioral classes. This article extends previous research on user classification (WUC - Web User Classification) based on stock exchange system performance, utilizing modern deep sequential architectures: the Bidirectional Long Short-Term Memory network (BiLSTM) and the Temporal Convolutional Network (TCN). The experimental data consists of system logs from load testing, annotated with user class labels. The process of preparing sequential data from logs is presented, along with two model architectures that are capable of learning user behavior patterns over time. The effectiveness of the BiLSTM and TCN models is evaluated in the task of user class prediction. The dataset contains 138,480 sequences of up to 10 encoded API endpoints and three associated timing measurements. Both models were trained for 10 epochs and evaluated using the same stratified 80/20 sequence-level split. BiLSTM achieved an accuracy of 0.9888 and macro- of 0.9908, whereas TCN achieved 0.9883 and 0.9904, respectively. These results demonstrate high class separability in the controlled dataset, but the logs contain no stable user or session identifier and the sequences were formed after grouping events by class. The reported scores should therefore be interpreted as proof-of-concept evidence for sequence-class assignment rather than an independent user-level generalization. |
