Skip to content

Latest commit

 

History

History
68 lines (47 loc) · 2.95 KB

README.md

File metadata and controls

68 lines (47 loc) · 2.95 KB
Raw

ASLI: Attentive Sequential model of Latent Intent

This is a TensorFlow implementation for the paper:

Attentive Sequential Models of Latent Intent for Next Item Recommendation. In Proceedings of The Web Conference (WWW'20)

Please cite our paper if you use the code or dataset:

@inproceedings{tanjim2020attentive,
    title={Attentive sequential models of latent intent for next item recommendation},
    author={Tanjim, Md Mehrab and Su, Congzhe and Benjamin, Ethan and Hu, Diane and Hong, Liangjie and McAuley, Julian},
    booktitle={Proceedings of The Web Conference 2020},
    pages={2528--2534},
    year={2020}
}

The code is tested with TensorFlow 1.15 and Python 3.6.

Dataset

Tmall Dataset

This is a dataset from Alibaba e-commerce platform. The preprocessed dataset is included in the repo (data/tmall.pickle), where each line contains a sequence of user's past interactions (sorted by timestamp). In each sequence, we include the following tuple: (item id, action+category id, action id, category id). There are four actions: click, collect, add-to-cart, and payment. The action+category id concatenates each unique combination action types along with the category id into a new id (not currently used in the model)

For downloading the original dataset:

Model Training

To train our model on Tmall (with default hyper-parameters), simply run:

python main.py

If you want to experiment with changing other parameters, following are the important parameters to change:

'--dataset': Name of the dataset
'--train_dir': Name of the directory for experimentation
'--attention_type': Type of the attention: latent_intent, self, item, action, category, action_category (please see the paper for more details, default=latent_intent)
'--batch_size': Default=128
'--maxlen': Maximum length of the sequence (default=300)
'--hidden_units': Embedding dimension (default=200)
'--dropout_rate': Dropout in FFN (default=0.3) 
'--num_heads': Number of attention heads (default=1)
'--num_blocks': Number of attention layers (default=2) 
'--kernel_size': Kernel size for the convolution layer (default = 10)

Results

On the Tmall dataset, below is an ablation study of different attention types:

Attention Type NDCG@5 HR@5
Seq-Seq (SASRec) 0.4588 0.5358
Seq-Item 0.3958 0.4573
Seq-Action 0.3928 0.4682
Seq-Category 0.5029 0.5791
Seq-Action+Category 0.4928 0.5687
Seq-Latent Intent 0.5107 0.5979

As can be seen from the above table, paying attention to the proposed latent intent of a user significantly improves the recommendation for the e-commerce platform.