@article{4655,
  author = {Pit Pichappan},
  title = {Modeling User Navigation with Pattern Mining, Markov Chains, and LSTM Networks},
  journal = {International Journal of Web Applications},
  year = {2026},
  volume = {18},
  number = {1},
  doi = {https://doi.org/10.6025/ijwa/2026/18/1/1-12},
  url = {https://www.dline.info/ijwa/fulltext/v18n1/ijwav18n1_1.pdf},
  abstract = {This study investigates the comparative effectiveness of three modeling paradigms frequent pattern mining,
Markov chains, and Long Short Term Memory (LSTM) networks in predicting user navigation behavior on
the MSNBC Anonymous Web Navigation dataset. The dataset, comprising nearly one million anonymized
browsing sessions, exhibits strong self transition dominance and low entropy, reflecting "sticky" user behavior
where individuals tend to remain within a single content category rather than explore diverse sections
of the site. Using next category prediction accuracy and perplexity as evaluation metrics, we find that
while LSTM models achieve marginally higher performance, the improvement over first and higher order
Markov models is not statistically significant (p>0.05). In contrast, frequent pattern mining despite its
interpretability underperforms due to limited generalization beyond observed subsequences. Our results
demonstrate that for structured, low entropy clickstream data, classical probabilistic models like Markov
chains offer a compelling balance of predictive power, computational efficiency, robustness, and interpretability.
These findings challenge the prevailing assumption that deep learning architectures inherently outperform
simpler models in all sequential prediction tasks. Instead, they underscore the importance of aligning
model complexity with the intrinsic characteristics of behavioral data a core tenet of domain driven
actionable knowledge discovery. The study concludes that in many real world applications involving navigational
or transactional logs, parsimonious models can deliver actionable insights without the overhead of
deep neural networks, provided they respect the underlying behavioral dynamics. Future work may explore
hybrid approaches or incorporate contextual signals to better capture rare but meaningful cross category
transitions.},
}