Awesome

Learning an Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems

This is our experimental code for RecSys 2021 paper "Learning an Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems".

The paper is available here.
The video is available here.
The slide is available here.

Requirements

tensorflow 1.4.0
pandas
numpy

GPUs with memory >= 10GB

Data Preprocessing

The raw data can be obtained from:
Tmall Data data_format1
Sobazaar Data Data > Sobazaar-hashID.csv.gz
MovieLens Data ml-25m

To preprocess the above raw data, save them in the raw_data folder under the root directory, and do

cd preproc
python tmall_preproc.py
python soba_preproc.py
python ml_preproc.py

The preprocessed datasets will be saved in the datasets folder for later use.

Pretraining

To simulate the real-world applications, the first 10 periods of dataset are used to pretrain an initial Embedding&MLP base model, and all the compared model updating methods will restore from the same pretrained model.

To pretrain a model for Tmall/Sobazaar/MovieLens, do

cd Tmall/pretrain
python train_tmall.py

cd Sobazaar/pretrain
python train_soba.py

cd MovieLens/pretrain
python train_ml.py

The pretrained base model will be saved in Tmall/pretrain/ckpts, Sobazaar/pretrain/ckpts and MovieLens/pretrain/ckpts respectively.

All the hyper-parameters can be easily configured in train_config at the beginning of each entry file (i.e., train_xxx.py).

Note: pretraining must be done before conducting any model updating method.

Baselines and Variants

All the compared model updating methods for a specific dataset are contained in the folder named by that dataset.

Our proposed method:
ASMGgru_multi

Baseline methods:
IU
BU
SPMF
IncCTR
SML
SMLmf

Variants of ASMGgru_multi:
ASMGgru_zero
ASMGgru_full
ASMGgru_single
(we do not create a separate folder for ASMGgru_uniform, as it can be easily implemented in ASMGgru_multi, see the code for more details)

To perform any of the ASMGgru methods, we need to first conduct a run of IU to generate the input model sequence.

For example, to perform a run of IU experiment for Tmall, do

cd Tmall/IU
python train_tmall.py

Then we can proceed to perform any of the ASMGgru methods

cd Tmall/ASMGgru_multi
python train_tmall.py

Other model updating methods can be conducted on their own without any pre-requisite.

Note that for SMLmf, since it is based on a different base model (i.e., Matrix Factorization), additional pretraining needs to be performed for this method.

cd Tmall/SMLmf/pretrain
python train_tmall.py

Then

cd Tmall/SMLmf/SML
python train_tmall.py

All the hyper-parameters can be easily configured in train_config at the beginning of each entry file (i.e., train_xxx.py).

The evaluation results can be found from the path with the following format:

<Dataset_Name>/<Method_Name>/ckpts/<Model_Alias>/<Last_Period_Alias>/test_metrics.txt

where <Model_Alias> is configured in train_config of the entry file, containing some essential hyper-parameter settings, and <Last_Period_Alias> by default is date20141030 for Tmall and period30 for MovieLens and Sobazaar.

Here are some examples of the possible paths that the evaluation results may reside:

Tmall/ASMGgru_multi/ckpts/ASMGgru_multi_linear_train11-23_test24-30_4emb_4mlp_1epoch_3_0.01/date20141030/test_metrics.txt

MovieLens/IU/ckpts/IU_train11-23_test24-30_1epoch_0.001/period30/test_metrics.txt

Citation

If you find this repo useful in your research, please cite the following:

@inproceedings{peng2021learning,
  title={Learning an Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems},
  author={Peng, Danni and Pan, Sinno Jialin and Zhang, Jie and Zeng, Anxiang},
  booktitle={Fifteenth ACM Conference on Recommender Systems},
  pages={411--421},
  year={2021}
}