LianHaiMiao / Attentive-Group-Recommendation

Attentive Group Recommendation
131 stars 50 forks source link

Attentive Group Recommendation

This is our implementation for the paper:

Da Cao, Xiangnan He, Lianhai Miao, Yahui An, Chao Yang, and Richang Hong. 2018. Attentive Group Recommendation. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval (SIGIR '18). ACM, New York, NY, USA, 645-654.

In order to learn the group interest, we use attention mechanism to learn the aggregation strategy from data in a dynamic way.

Please cite our SIGIR'18 paper if you use our codes. Thanks!

BibTeX:

@inproceedings{Cao2018Attentive,
 author = {Cao, Da and He, Xiangnan and Miao, Lianhai and An, Yahui and Yang, Chao and Hong, Richang},
 title = {Attentive Group Recommendation},
 booktitle = {The 41st International ACM SIGIR Conference on Research \&\#38; Development in Information Retrieval},
 series = {SIGIR '18},
 year = {2018},
 isbn = {978-1-4503-5657-2},
 location = {Ann Arbor, MI, USA},
 pages = {645--654},
 numpages = {10},
 url = {http://doi.acm.org/10.1145/3209978.3209998},
 doi = {10.1145/3209978.3209998},
 acmid = {3209998},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {atention mechanism, cold-start problem, group recommendation, neural collaborative filtering, recommender systems},
}

Environment Settings

We use the framework pytorch.

It's better to use Pytorch 1.x to run the code.

Thanks zanshuxun for providing the newest pytorch version code.

Example to run the codes.

Run AGREE:

python main.py

After training process, the value of HR and NDCG in the test dataset will be printed in command window after each optimization iteration.

Output:

AGREE at embedding size 32, run Iteration:30, NDCG and HR at 5
...
User Iteration 10 [449.8 s]: HR = 0.6216, NDCG = 0.4133, [1.0 s]
Group Iteration 10 [471.9 s]: HR = 0.5910, NDCG = 0.4005, [23.0 s]

Parameter Tuning

we put all the papameters in the config.py

Dataset

We provide one processed dataset: CAMRa2011.

Because we have another paper use the MaFengWo dataset are under reviewing, so we can't release MaFengWo dataset now.

group(user) train.rating:

test.rating:

test.negative