inconsistent prediction results using HinSage

getting inconsistent prediction results after fitting a HinSage model

I am using the HinSage-link-prediction demo script.

After model fitting, I ran model.predict(test_gen) and got results

array([[3.9639022],
       [3.1461685],
       [3.5686603],
       ...,
       [4.0026994],
       [3.790706 ],
       [3.9910824]], dtype=float32)

ran model.predict(test_gen) again, this time got results

array([[4.011217 ],
       [3.3255208],
       [3.701041 ],
       ...,
       [4.098828 ],
       [3.6664836],
       [3.870615 ]], dtype=float32)

I have also set random seeds as

import os 
import tensorflow as tf 
import random as python_random  

seed_value= 0
np.random.seed(123)
tf.random.set_seed(1234)
python_random.seed(123)
os.environ['PYTHONHASHSEED']="3"

I am using Stellagraph 1.2.1.

Please advise how to fix this issue. Thank you!

[mikhail247]

I found that a possible workaround is to set batch_size to 1 in HinSAGENodeGenerator, which is used to call model.predict() method. Other seed parameters need to be set as well. The results become deterministic across different runs, however, this hurts prediction performance.

[enjieli]

@huonw thank you for your input!

Are there ways to get node feature importance in the HinSage Link Prediction model? like the GCN saliency_maps .