Attentive Aspect-based Recommendation Model

This is our implementation for the paper:

Xinyu Guan, Zhiyong Cheng, Xiangnan He, Yongfeng Zhang, Zhibo Zhu, Qinke Peng and Tat-Seng Chua. Attentive Aspect Modeling for Review-aware Recommendation[J]. ACM Transactions on Information Systems (TOIS).

Please cite our TOIS paper if you use our codes.

Environment Settings

• tensorflow 1.2.1
• gensim 2.2.0

Dataset

We provide the processed Amazon Beauty core-5 dataset. The original dataset can be found in here.

Beauty dataset.

users.txt:

• The users' names in the orignal dataset.
• Line n is the orginal name of user whose id is n-1

product.txt:

• The products' names in the orignal dataset.
• Line n is the orginal name of product whose id is n-1

aspect.txt:

• The aspects' names.
• We use Sentires to extract aspects from user reviews. The tool is available at here.
• Line n is the orginal name of aspect whose id is n-1

user_aspect_rank.txt:

• The aspect set of each user.
• Line n is the aspect set of user whose id is n-1: aspect1,aspect2,...

item_aspect_rank.txt:

• The aspect set of each product.
• Line n is the aspect set of product whose id is n-1: aspect1,aspect2,...

emb128.vector:

• The word embedding pretrained with Word2vec model (implemented with gensim).
• Use gensim.models.KeyedVectors.load_word2vec_format to load the embedding matrix.

train_pairs.txt:

• Positive (user, item) pairs in training set.
• Each line is a training instance: userId,itemId

valid_pairs.txt

• Positive (user, item) pairs in validation set.
• Each line is an instance: userId,itemId

test_pairs.txt:

• Positive (user, item) pairs in test set.
• Each line is a test instance: userId,itemId

Example to run the codes.

The instruction of commands has been clearly stated in the codes (see the parse_args function).

Run aarm:

python running.py --productName Beauty --is_l2_regular 1 --lamda_l2 0.1 --is_out_l2 0 --dropout 0.5 --learning_rate 0.003 --num_aspect_factor 128 --num_mf_factor 128