1. Create Environment

conda env create -f train.yml
conda activate dprec_train

Installs the CUDA build of PyTorch (torch==2.12.1+cu130) from the internal conda/pip channels. Tested with PyTorch 2.12 + CUDA 13.0.

---

2. Data Download

Dataset	Download
ML-1M	https://files.grouplens.org/datasets/movielens/ml-1m.zip
ML-10M	https://files.grouplens.org/datasets/movielens/ml-10m.zip
KuaiRand-27K	https://zenodo.org/records/10439422/files/KuaiRand-27K.tar.gz

Download and extract into a local directory, e.g. ~/datasets/.

# ML-1M
curl -L -O https://files.grouplens.org/datasets/movielens/ml-1m.zip
unzip ml-1m.zip

# ML-10M
curl -L -O https://files.grouplens.org/datasets/movielens/ml-10m.zip
unzip ml-10m.zip

# KuaiRand-27K
curl -L -O https://zenodo.org/records/10439422/files/KuaiRand-27K.tar.gz
tar -xzf KuaiRand-27K.tar.gz

---

3. Data Pre-process

Processed files are written to the data/ directory as <dataset>.txt with format user item timestamp per line.

MovieLens (ML-1M and ML-10M)

python src/data_process_mlm.py ml-1m/ratings.dat
python src/data_process_mlm.py ml-10M100K/ratings.dat --dataset-name ml-10m

ml-10m.zip extracts to a folder named ml-10M100K/, so pass --dataset-name ml-10m to keep the training convention (dataset=ml-10m).

Users with fewer than --min-seq-len interactions are dropped (default: 5). To build a long-sequence variant (e.g. ML-10M-L), raise the threshold:

python src/data_process_mlm.py ml-10M100K/ratings.dat --dataset-name ml-10m --min-seq-len 500

KuaiRand-27K

# Point at the KuaiRand-27K data directory — the log_standard_* parts are
# merged automatically (the log_random_* exposure log is excluded).
python src/data_process_kuairec.py KuaiRand-27K/data

# Or pass a single pre-merged CSV
python src/data_process_kuairec.py KuaiRand-27K/data/full_data.csv

---

4. Train DP-Rec

DP-Rec requires a pre-trained patcher (boundary detector) checkpoint. Train it first.

All training is launched via main.py using Hydra configuration. Model configs live in configs/model/.

Step 1 — Train the patcher

python main.py \
  model=dp_rec_patcher \
  experiment_folder=my_experiment \
  model_suffix=patcher \
  dataset=ml-1m \
  local_data_dir=data/

The checkpoint is saved to exp/my_experiment/dp_rec/dp_rec_patcher_ml-1m_<hash>/.

Step 2 — Train DP-Rec

python main.py \
  model=dp_rec \
  experiment_folder=my_experiment \
  model_suffix=dprec \
  dataset=ml-1m \
  local_data_dir=data/ \
  patcher_model_path=exp/my_experiment/dp_rec/dp_rec_patcher_ml-1m_<hash>

Overriding config parameters

Any parameter can be overridden on the command line with ++:

python main.py \
  model=dp_rec \
  experiment_folder=ablation \
  model_suffix=maxlen50 \
  dataset=ml-1m \
  local_data_dir=data/ \
  patcher_model_path=exp/my_experiment/dp_rec/dp_rec_patcher_ml-1m_<hash> \
  ++model_args.maxlen=50 \
  ++training_args.num_epochs=200

---

5. Evaluate Model

Evaluation runs automatically at the end of training. To run inference on a saved checkpoint without retraining:

python main.py \
  dataset=ml-1m \
  local_data_dir=data/ \
  task=inference_only \
  model_dir=exp/my_experiment/dp_rec/dp_rec_dprec_ml-1m_<hash>

Results are written to results.csv inside the experiment directory. Each row corresponds to one evaluation epoch with columns such as Valid_NDCG@10, Test_NDCG@10, Test_HR@10, etc.

---

Configuration Reference

Parameter	Description	Default
model	Config file in configs/model/	dp_rec
experiment_folder	Top-level directory grouping related runs	required
model_suffix	Suffix identifying this run	required
dataset	Dataset name, must match a .txt file in local_data_dir	required
local_data_dir	Path to processed data files	../data/
local_output_dir	Path where checkpoints and logs are saved	exp/${experiment_folder}/
task	train or inference_only	train
model_dir	Path to saved checkpoint (required for inference_only)	null
patcher_model_path	Path to pre-trained patcher checkpoint (required for dp_rec training)	required
training_args.num_epochs	Number of training epochs	model-specific
model_args.maxlen	Maximum sequence length	model-specific
training_args.batch_size	Batch size	model-specific
training_args.lr	Learning rate	model-specific

Output files per run

Each run produces a directory under local_output_dir containing:

• config.yaml — full Hydra config used for the run
• logs.txt — training logs
• results.csv — per-epoch validation and test metrics
• state_dict.pth — best model checkpoint
• params.json — model architecture metadata (used by patcher loading)