euler GraphSage的所有node的embedding都相似度非常接近

1亿规模的节点，30亿左右的边，训练GraphSage模型。训练参数如下：

--max_id 90170696
--feature_idx 0 
--feature_dim 1
--learning_rate 0.00005
--num_epochs 1
--dim 128
--batch_size 128

首先，模型的收敛性并不好，loss没怎么下降

INFO:tensorflow:loss = 437.71362, mrr = 0.39257812, step = 1488562
INFO:tensorflow:loss = 413.18103, mrr = 0.46171877, step = 1488919 (2.109 sec)
INFO:tensorflow:loss = 449.76367, mrr = 0.4251302, step = 1489304 (2.003 sec)
INFO:tensorflow:loss = 422.31223, mrr = 0.43528646, step = 1489705 (2.045 sec)
....
INFO:tensorflow:loss = 439.37802, mrr = 0.4332031, step = 1601393 (2.210 sec)
INFO:tensorflow:loss = 495.66333, mrr = 0.47460938, step = 1601788 (2.124 sec)
INFO:tensorflow:loss = 411.83685, mrr = 0.47942704, step = 1602187 (2.181 sec)
INFO:tensorflow:loss = 434.9126, mrr = 0.4126302, step = 1602586 (2.162 sec)
.....
INFO:tensorflow:loss = 432.3944, mrr = 0.45585936, step = 3520258 (2.242 sec)
INFO:tensorflow:loss = 419.18842, mrr = 0.4451823, step = 3520649 (2.169 sec)
INFO:tensorflow:loss = 439.15546, mrr = 0.3761719, step = 3521049 (2.237 sec)
INFO:tensorflow:loss = 420.41013, mrr = 0.40924478, step = 3521450 (2.228 sec)
INFO:tensorflow:loss = 425.3105, mrr = 0.4282552, step = 3521840 (2.208 sec)

其次，获得的Embedding，任意两个node的Embedding相似度超过0.96。