Yes, you need to use the the pre-trained sentence embedding models to get feature_embedding. We applied the pre-trained paraphrase-distilroberta-base-v1 sentence embedding model in a public transformer implementation: https://github.com/UKPLab/sentence-transformers

OK, thank you very much!

My apology for disturbing.
I have read the experiment implementation details of this paper, which mentioned that (for the Amazon dataset as an example): The final feature_embedding is the average of embedding of the 'title', 'description', and 'feature'.
I have a question about that: Is it pre_train model would finish the average job? Or I should compute the average manually?

Mr.Xu, apology again. Would it be possible for you to share the source code of computing feature_embedding used in your research?
I am very new to this area and try to study further. I try to program the codes but I am not sure if it's the correct method for getting feature_embedding. That would be a very great help if you want to share.

The feature_embedding is the average of embedding of the 'title', 'description' and 'feature'. I used pre-trained model to encode sentences separately and manually compute the average.

That's nice. I have achieved the function.
But there is a new problem that I need to compute IPS_EXPO_PROB.😂
大佬，能给一份代码吗，嘤嘤嘤

It does not need special code. We first train a IPSBiasedMF model, and then save the predicted user-item matrix as the exposure probability into IPS_EXPO_PROB file. I have updated the README to provide more details.

OK, thank you very much!
Oh my god, I just have found the IPSBiasedMF model in the project!😂

The propensity score is calculated based on the method introduced in "Unbiased recommender learning from missing-not-at-random implicit feedback". Or you can refer the equation 18 in our paper. After calculating the propensity score, save the score in propensity.npy file. Use the calculated score to train a IPSBiasedMF model.