PreTPP

Installation | Data preparation | Usage | HT-Transformer | HT-Transformer (paper) | Citing

Advanced pretraining for TPP, MTPP and Event Sequences.

The code is highly dependent on and compatible with HoTPP.

Installation

pip install --no-build-isolation .

HT-Transformer

The code for HT-Transformer can be found at:

pretpp/nn/encoder/history_token_transformer.py
pretpp/nn/encoder/history_token_strategy.py

Data preparation

Some datasets are inherited from HoTPP. For them just make a symlink to the data folder:

cd experiments/DATASET
ln -s <hotpp>/experiments/DATASET/data .

To make datasets, specific to PreTPP, use the following command:

cd experiments/DATASET
spark-submit --driver-memory 16g -c spark.network.timeout=100000s --master 'local[12]' scripts/make-dataset.py

Parameters

All configs are placed at experiments/DATASET/configs.

Results

All results are stored in experiments/DATASET/results.

Usage

Example training of HT-Transformer on the Churn dataset:

cd experiments/transactions-rosbank-full-3s
CUDA_VISIBLE_DEVICES=0 python3 -m hotpp.train_multiseed --config-dir configs --config-name next_item_hts_transformer

Fine-tune:

CUDA_VISIBLE_DEVICES=0 python3 -m hotpp.train_multiseed --config-dir configs --config-name htl_transformer_ft_multi base_name=next_item_hts_transformer

Example training of NTP-Transformer on the Taobao dataset:

cd experiments/taobao
CUDA_VISIBLE_DEVICES=0 python3 -m hotpp.train_multiseed --config-dir configs --config-name next_item_transformer

Fine-tune:

CUDA_VISIBLE_DEVICES=0 python3 -m hotpp.train_multiseed --config-dir configs --config-name transformer_ft_multi base_name=next_item_transformer

Citation

@article{karpukhin2025httransformer,
  title={HT-Transformer: Event Sequences Classification by Accumulating Prefix Information with History Tokens},
  author={Karpukhin, Ivan and Savchenko, Andrey},
  journal={arXiv preprint arXiv:2508.01474v1},
  year={2025},
  url ={https://arxiv.org/abs/2508.01474v1}
}