python train.py
Time elapsed = 6.89 mins, Training: loss = 389.51047, mrr = 0.63130, ndcg = 0.71369, hr1 = 0.50939, hr3 = 0.69945, hr5 = 0.78027, hr10 = 0.87522 | Val:loss = 2172.41870, mrr = 0.25467, ndcg = 0.40172, hr1 = 0.15110, hr3 = 0.25807, hr5 = 0.33136, hr10 = 0.45893
0 lr=0.0001,lamb=0.55,batch_size=400,numNegative=100,featEmbedDim=64,idenEmbedDim=64,outputDim=128,pathNum=7 Test loss:2033.5934; Test mrr:0.25339168; Test ndcg:0.3976466; Test hr1:0.14939758; Test hr3:0.2633283; Test hr5:0.34176204; Test hr10:0.46430722
These variant models below had been supported:
- ReGCN
- ReGCN_{MP}
- RecoGCN
- python == 3.6
- tensorflow == 1.13.1
- numpy == 1.16.3
- h5py == 2.9.0
- GPUtil ==1.4.0
- setproctitle == 1.1.10
You can download the experiment data from Here. An example loading code is provided as follow.
adj = {0:{}, 1:{}, 2:{}, 3:{}}
with h5py.File(dataset, 'r') as f:
adj[0][1] = f['adj01'][:]
adj[1][0] = f['adj10'][:]
adj[0][2] = f['adj02'][:]
adj[2][0] = f['adj20'][:]
adj[0][3] = f['adj03'][:]
adj[3][0] = f['adj30'][:]
train_sample = f['train_sample'][:]
val_sample = f['val_sample'][:]
test_sample = f['test_sample'][:]
item_freq = f['item_freq'][:]
user_feature = f['user_feature'][:]
agent_feature = f['agent_feature'][:]
item_feature = f['item_feature'][:]
userCnt = f['userCnt'][()]
agentCnt = f['agentCnt'][()]
itemCnt = f['itemCnt'][()]
The data structure is explained as follow.
adj[x][y] denotes the adjancy relationship from x to y. Here, 0 stands for user, 1 is selling agent, 2 and 3 are two kinds of items. The shape of adj[x][y] is [Num_of_node_x ,maximum_link]. Each line stores the node ids of type y who are linked with node x. Note that maximum_link should be the same for each of these relations.
train_sample, val_sample, test_sample are triplet of [user, selling_agent, item] pairs. Each type of node is encoded from 0.
item_freq is [item_id, item_frequency] matrix denotes the occur frequency of each item in train set.
user_feature, agent_feature, item_feature are three featrue matrix of shape [node_num, feature_num]. Here features for each node are multi-hot encoded, and different type of node can have different feature numbers.
If you use our code or dataset in your research, please cite:
@inproceedings{xu2019relation,
title={Relation-aware graph convolutional networks for agent-initiated social e-commerce recommendation},
author={Xu, Fengli and Lian, Jianxun and Han, Zhenyu and Li, Yong and Xu, Yujian and Xie, Xing},
booktitle={Proceedings of the 28th ACM International Conference on Information and Knowledge Management},
pages={529--538},
year={2019}
}