SSS4Rec

This repository contains the code used in our SIGIR 2026 paper Pay Attention to Sequence Split: Uncovering the Impacts of Sub-Sequence Splitting on Sequential Recommendation Models.

---

1. Repository Structure

The codebase is organized into two parallel frameworks corresponding to two dominant sub-sequence splitting (SSS) paradigms in sequential recommendation.

SSS/
├── Multi_Target_Implementation/
│   ├── data/                     # Raw and preprocessed datasets
│   └── src/                      # Multi-target training framework
│
├── Single_Target_Implementation/
│   ├── data/                     # Raw datasets
│   └── src/                      # Single-target training framework
│
└── README.md

Design Rationale

• Single-Target Implementation
  Sub-sequences are generated on-the-fly during training, and each training instance predicts one next item.

• Multi-Target Implementation
  Sub-sequences are generated offline before training and saved as processed dataset files.

This separation is intentional: it reflects two different implementation paradigms adopted by prior sequential recommendation codebases, and allows us to study the effect of data splitting without conflating it with other training-pipeline differences.

---

2. Multi-Target Data Preprocessing

Important: The multi-target framework requires offline preprocessing before training. Raw data should not be used directly for training in this framework.

Raw Data Location

Raw dataset files should be placed in:

Multi_Target_Implementation/data/

Preprocessing Script

The preprocessing script is:

preprocess_data.py

It reads a raw user-item interaction file and generates an augmented dataset according to the specified splitting strategy.

Supported Augmentation Types

augment_type	Description
none	Use the raw sequence without splitting
pre	Prefix-based splitting
suffix	Suffix-based splitting
slide	Sliding-window splitting
slide_plus_full	Sliding-window splitting plus the original full sequence

Sliding-related Arguments

Argument	Description
slide_window	Window size of each sub-sequence
slide_step	Step size between adjacent windows
slide_keep_tail	Whether to keep tail sequences shorter than one window

Example Preprocessing Commands

python preprocess_data.py \
  --data_dir ../data/ \
  --data_name Beauty \
  --augment_type slide_plus_full \
  --slide_window 3 \
  --slide_step 1

python preprocess_data.py \
  --data_dir ../data/ \
  --data_name CDs \
  --augment_type slide \
  --slide_window 5 \
  --slide_step 1

python preprocess_data.py \
  --data_dir ../data/ \
  --data_name ML-1M \
  --augment_type pre

python preprocess_data.py \
  --data_dir ../data/ \
  --data_name Douyin \
  --augment_type suffix

Generated Output Files

The processed files are saved in the same data directory. Typical filenames include:

Beauty_pre.txt
Beauty_suffix.txt
Beauty_slide_win3_step1.txt
Beauty_slide_plus_full_win3_step1.txt
CDs_slide_win5_step1.txt
ML-1M_pre.txt

Recommended Sliding Window Sizes

Because sequence-length distributions differ substantially across datasets, we do not use a unified sliding-window size. The following settings correspond to the configurations used in our experiments.

Dataset	Recommended Window Size
Beauty	3
Sports_and_Outdoors	3
CDs	5
ML-1M	10
Douyin	5
LastFM	5

In general, shorter and denser sequence datasets prefer smaller windows, while datasets with longer user histories can benefit from larger windows.

---

3. Single-Target Implementation

Overview

The single-target framework follows the on-the-fly splitting paradigm:

• each training instance predicts exactly one next item
• sub-sequence splitting, if enabled, is performed during training-time data loading
• no preprocessing is required

Raw Data Usage

Single-target training uses the raw dataset directly. There is no need to run any preprocessing script in advance.

Supported Models

The following models are supported in the single-target framework:

• SASRec
• BSARec
• FMLPRec
• GRU4Rec

Supported Arguments

Argument	Options / Meaning
model_type	SASRec, BSARec, FMLPRec, GRU4Rec
loss_type	BCE, CE
augment_type	ori, pre, suffix, slide
data_name	Dataset name
model_idx	Run index / random seed
train_name	Log and checkpoint name

Augmentation Types

augment_type	Description
ori	Use the original training sequence without sub-sequence splitting
pre	Generate prefix-based training samples on-the-fly
suffix	Generate suffix-based training samples on-the-fly
slide	Generate sliding-window training samples on-the-fly

Choosing Sliding Window Size

For slide augmentation, the recommended window sizes are:

Dataset	Recommended Window Size
Beauty	3
Sports_and_Outdoors	3
CDs	5
ML-1M	10
Douyin	5
LastFM	5

Example Training Commands

python main.py \
  --model_type SASRec \
  --loss_type BCE \
  --augment_type ori \
  --model_idx 1 \
  --data_name Beauty \
  --train_name SASRec_Beauty_ori

python main.py \
  --model_type SASRec \
  --loss_type BCE \
  --augment_type pre \
  --model_idx 1 \
  --data_name Beauty \
  --train_name SASRec_Beauty_pre

python main.py \
  --model_type SASRec \
  --loss_type BCE \
  --augment_type slide \
  --model_idx 1 \
  --data_name Beauty \
  --train_name SASRec_Beauty_slide

python main.py \
  --model_type GRU4Rec \
  --loss_type BCE \
  --augment_type ori \
  --model_idx 1 \
  --data_name Beauty \
  --train_name GRU4Rec_Beauty_ori

python main.py \
  --model_type BSARec \
  --loss_type CE \
  --augment_type ori \
  --model_idx 1 \
  --data_name ML-1M \
  --train_name BSARec_ML1M_ori

python main.py \
  --model_type FMLPRec \
  --loss_type CE \
  --augment_type suffix \
  --model_idx 1 \
  --data_name CDs \
  --train_name FMLPRec_CDs_suffix

---

4. Multi-Target Implementation

Overview

The multi-target framework follows the offline preprocessing paradigm:

• sub-sequences are generated before training
• training is performed on the generated files
• different augmentation strategies correspond to different processed datasets

Supported Models

The following models are supported in the multi-target framework:

• SASRec
• BSARec
• FMLPRec
• GRU4Rec

Representative Training Commands

BSARec

python main.py \
  --model_name BSARec \
  --loss_type CE \
  --c 5 \
  --alpha 0.7 \
  --lr 0.001 \
  --num_attention_heads 1 \
  --augment_type None \
  --model_idx 1 \
  --data_name CDs \
  --gpu_id 3

python main.py \
  --model_name BSARec \
  --loss_type CE \
  --c 9 \
  --alpha 0.3 \
  --lr 0.0005 \
  --num_attention_heads 4 \
  --augment_type None \
  --model_idx 1 \
  --data_name ML-1M \
  --gpu_id 0

python main.py \
  --model_name BSARec \
  --c 3 \
  --alpha 0.9 \
  --lr 0.001 \
  --num_attention_heads 1 \
  --augment_type None \
  --model_idx 2 \
  --data_name Douyin \
  --gpu_id 0

python main.py \
  --model_name BSARec \
  --c 3 \
  --alpha 0.9 \
  --lr 0.001 \
  --num_attention_heads 1 \
  --augment_type None \
  --model_idx 2 \
  --data_name LastFM \
  --gpu_id 3

python main.py \
  --model_name BSARec \
  --c 5 \
  --alpha 0.3 \
  --lr 0.001 \
  --num_attention_heads 4 \
  --augment_type None \
  --model_idx 2 \
  --data_name Sports_and_Outdoors \
  --gpu_id 0

python main.py \
  --model_name BSARec \
  --c 5 \
  --alpha 0.7 \
  --lr 0.0005 \
  --num_attention_heads 1 \
  --augment_type None \
  --model_idx 2 \
  --data_name Beauty \
  --gpu_id 3

FMLPRec

python main.py \
  --model_name FMLPRec \
  --num_hidden_layers 4 \
  --loss_type CE \
  --augment_type slide_plus_full_win3_step1 \
  --model_idx 1 \
  --data_name Beauty \
  --gpu_id 0

SASRec

python main.py \
  --model_name SASRec \
  --loss_type CE \
  --augment_type slide_plus_full_win5_step1 \
  --model_idx 1 \
  --data_name Douyin \
  --gpu_id 0

python main.py \
  --model_name SASRec \
  --loss_type CE \
  --augment_type slide_plus_full_win3_step1 \
  --model_idx 1 \
  --data_name Beauty \
  --gpu_id 0

Hyperparameter Notes

• For BSARec, we follow the original paper setup and tune c, alpha, learning rate, and the number of attention heads according to each dataset.
• For SASRec, GRU4Rec, and FMLPRec, the key hyperparameters also follow the original paper settings as closely as possible.
• Dataset-specific tuning is applied when necessary, and the representative commands above reflect the configurations used in the paper.

---

5. Experimental Protocol

To ensure fair and leakage-free evaluation, we use the following protocol in both frameworks.

No Information Leakage

• Sub-sequence splitting / augmentation is applied only to the training set
• Validation and test sets are not augmented
• Future interactions are never included in training inputs

Evaluation on Complete Sequences

Although training may use split sub-sequences, validation and test are always performed on the complete original user sequence under the standard evaluation protocol.

Concretely:

• the validation target is taken from the held-out interaction near the end of the original sequence
• the test target is taken from the final held-out interaction
• the model is evaluated using the user's full observable history before the target

Consistency Across Frameworks

Although the single-target and multi-target frameworks differ in how training data is generated, they share:

• the same original dataset split
• the same validation / test protocol
• the same ranking-based evaluation setting

Therefore, differences in results are attributable to the training-time effect of sub-sequence splitting, rather than inconsistencies in evaluation.

---

6. Acknowledgement

This codebase is built upon and inspired by the following excellent works:

• SASRec
• BSARec

We thank the authors for releasing high-quality, reproducible implementations.

---

7. Citation

@inproceedings{dang2026pay,
  title={Pay Attention to Sequence Split: Uncovering the Impacts of Sub-Sequence Splitting on Sequential Recommendation Models},
  author={Dang, Yizhou and Wu, Yifan and Huang, Minhan and Zhao, Chuang and Ma, Lianbo and Guo, Guibing and Wang, Xingwei and Sun, Zhu},
  journal={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2026}
}