This is the implementation of our two recent models for recommending scientific articles. The first model, CATA, utilizes articles’ titles and abstracts via a single attentive autoencoder, while the second model, CATA++, utilizes additional articles' information (such as tags and citations between articles) via two parallel attentive autoencoders.

1. Collaborative Attentive Autoencoder for Scientific Article Recommendation (ICMLA 2019)
2. CATA++: A Collaborative Dual Attentive Autoencoder Method for Recommending Scientific Articles (Access 2020)

• Python 3
• Tensorflow
• Keras

You can evaluate our models with different settings in terms of:

1. data_name -d: 'a' for citeulike-a and 't' for citeulike-t. [Optional] [Default=citeulike-a]
2. sparse -s: '0', 'no', 'false', or 'f' for dense. [Optional] [Default=sparse]
3. pretrain -pretrain: '1', 'yes', 'true' or 't' for pretrain the attentive autoencoder. [Optional] [Default=No-pretraining]
4. epochs -e: Number of epochs to pretrain the attentive autoencoder. [Optional] [Default=150]
5. lambda_u -u: value of lambda_u. [Optional] [Default=10]
6. lambda_v -v: value of lambda_v. [Optional] [Default=0.1]
7. pmf_epochs -pe: Number of iterations for PMF. [Optional] [Default=100]
8. output_name -o : Name of the output file. [Optional]
9. latent_size -l :Size of latent space. [Optional] [Default=50]

• To run CATA model, use test_CATA.py file.

• To run CATA++ model, use test_CATA++.py file.

• Example 1: Run CATA++ model for citeulike-t dataset with the dense setting. Also, always pretrain the two autoencoders if you run the code for the first time by setting -pretrain 1:

  • python3 test_CATA++.py -pretrain 1 -d 't' -s 0

• Example 2: Run CATA++ model for citeulike-t dataset with the sparse setting:

  • python3 test_CATA++.py -pretrain 0 -d 't' -s 1

• Example 3: Run CATA model for citeulike-a dataset with the sparse setting:

  • python3 test_CATA.py -pretrain 0 -d 'a' -s 1