Downloads · 30 days
19
18% of all-time downloads
EmanuelePiemontese/SENSEI
SENSEI is a machine learning model from EmanuelePiemontese. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-sa-4.0.
This is the SENSEI model introduced in A Foundation Activity Trace Encoder for Predictive Process Monitoring.
Downloads · 30 days
19
18% of all-time downloads
All-time downloads
107
Public
Parameters
13.5M
54.1 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors54.1 MB · 100%
From the Hugging Face model README
This is the SENSEI model introduced in A Foundation Activity Trace Encoder for Predictive Process Monitoring.
We show that a bidirectional Transformer architecture, specifically the BERT-Small configuration, is highly effective for Predictive Process Monitoring (PPM) tasks. SENSEI is trained from scratch on over 13 million synthetic process traces, avoiding the limitations associated with adapting Natural Language Processing models and textual abstractions to business event logs.
SENSEI represents business activities using an Abstract Symbol Mapping (ASM) alphabet, allowing the model to learn the underlying topology, routing patterns, and long-range dependencies of event logs directly from activity sequences.
The goal of SENSEI is to support research in process-aware information systems and encourage the development of approaches that go beyond the direct application of Natural Language Processing techniques to business event logs.
Model developers:
Vincenzo Pasquadibisceglie, Emanuele Piemontese
Paper authors:
Vincenzo Pasquadibisceglie, Emanuele Piemontese, Annalisa Appice, Donato Malerba
The model architecture corresponds to the BERT-Small (4/512) configuration:
BertForMaskedLMSENSEI is intended for research in Predictive Process Monitoring (PPM) and can be fine-tuned on real-world event logs for operational prediction tasks, including:
SENSEI is most effective when raw event logs are pre-processed using the ASM vocabulary, in order to preserve the structural representation learned during pre-training.
Business activities should therefore be mapped to the ACT_XXX format before tokenization. Using arbitrary textual activity names may result in sub-word fragmentation and can negatively affect the model's ability to exploit the learned activity structure.
The model and tokenizer can be loaded using the transformers library:
from transformers import AutoTokenizer, AutoModelForMaskedLM
model_name = "EmanuelePiemontese/SENSEI"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForMaskedLM.from_pretrained(model_name)
# Example: replace business activities with SENSEI's
# Abstract Symbol Mapping (ASM) representation.
# ACT_001 and ACT_002 are known events.
# [MASK] is the event to be predicted.
sequence = "[CLS] ACT_001 ACT_002 [MASK] ACT_005 [SEP]"
# Tokenize the input
inputs = tokenizer(sequence, return_tensors="pt")
# Generate predictions for the masked activity
outputs = model(**inputs)
SENSEI was pre-trained on over 13 million synthetic process traces, corresponding to more than 337 million incremental prefixes.
The synthetic processes cover a wide range of structural complexities, including:
SENSEI uses Masked Activity Structure Modeling (MASM) as its pre-training objective.
During pre-training, 15% of the activities in each trace are stochastically masked, and the model is trained to reconstruct the masked activities from their surrounding context.
If you use SENSEI in your research, please cite the following paper:
@article{pasquadibisceglie2026sensei,
title={SENSEI: A Foundation Activity Trace Encoder},
author={Pasquadibisceglie, Vincenzo and Piemontese, Emanuele and Appice, Annalisa and Malerba, Donato},
year={2026}
}
This model is released under the CC BY-NC-SA 4.0 license.
Please refer to the license terms before using or redistributing the model.