Lightweight n-grams rival neural nets for event-log prediction
A recent preprint on arXiv (2604.21629) evaluates lightweight automata-based models, specifically n-grams, against neural architectures such as LSTM and Transformer for predicting the next activity in streaming event logs. Testing on synthetic patterns and five actual process mining datasets reveals that n-grams, when utilizing suitable context windows, can achieve accuracy levels similar to those of neural models while consuming significantly fewer resources. In contrast to windowed neural architectures, which exhibit inconsistent performance, n-grams maintain stable accuracy. Although traditional ensemble techniques like voting enhance n-gram performance, they necessitate the parallel operation of multiple agents during inference, leading to increased memory use and latency. The authors introduce the promotion algorithm, an ensemble method that smartly selects between two active models during inference, minimizing overhead compared to conventional voting methods. On real-world datasets, these ensembles either match or surpass the accuracy of neural models.
Key facts
- arXiv:2604.21629 compares n-grams with LSTM and Transformer for event-log prediction.
- Experiments used synthetic patterns and five real-world process mining datasets.
- N-grams with appropriate context windows achieve accuracy comparable to neural models.
- N-grams require substantially fewer resources than neural architectures.
- Windowed neural architectures show unstable performance patterns.
- N-grams provide stable and consistent accuracy.
- Classical ensemble methods like voting improve n-gram performance but increase memory and latency.
- The proposed promotion algorithm dynamically selects between two active models during inference.
Entities
Institutions
- arXiv