ywjawmw/clepr

This is our Pytorch implementation for SDPR model

SDPR: Prescription Recommendation with Syndrome Differentiation

This is our Pytorch implementation for the paper:

Introduction

We propose a new four-partite graph modeling paradigm that enables prescription recommendation models to represent the four steps of SD. Based on this paradigm, we design a prescription recommendation model, SDPR, which includes four modules corresponding to the steps of SD. We introduce a graph neural network-based entity representation module to enhance symptom and herb representation. To address the challenges of modeling syndrome and therapeutic method information, we propose a symptom set aggregator and an herb set aggregator. Our pre-training strategy for inducing syndrome and the therapeutic method-aware contrastive learning framework can model the complex relationships among symptoms, syndromes, therapeutic methods, and herbs. Finally, we integrate these modules into a multi-task learning framework to complete the SD analysis. Experiments on a public prescription recommendation dataset demonstrate that our model accurately recommends herbs and retrieves existing prescriptions, showing that the four-partite graph paradigm enhances precise herb prescribing.

Requirement

The code has been tested running under Python 3.7.0. The required packages are as follows:

• torch == 1.10.0
• numpy == 1.17.4
• scipy == 1.21.6
• temsorboardX == 2.0

Usage

The hyperparameter search range and optimal settings have been clearly stated in the codes (see the 'utils' dict in parser.py).

• Train

bash CLEPR_train.sh 

• Pretrain Train File

\weights-CLEPR\Herb\CLEPR_norm_CLEPR_l2\64-128\32\date_2022-09-18_60_ori_emb_seed1234\model.pkl

Retraining:

bash CLEPR_train.sh

the parameters should be notice:

• --attention 0

• --use_S1 0

• --alg_type CLEPR(S1)

• --batch_size 1024

• Test

bash CLEPR_test.sh 

Some important hyperparameters:

• lrs

  • It indicates the learning rates.
  • The learning rate is default as 2e-5.

• mess_dropouts

  • It indicates the message dropout ratio, which randomly drops out the outgoing messages.
  • The message dropout is default as '[0.0,0.0]'.

• steps

  • It indicates the length of the hard-negative sample.
  • We search it in {2, 4, 5, 6, 8, 10, 16, 24}.

• max_step_lens

  • It indicates the search length of the hard-negative sample.
  • We search it in {16, 24, 25, 26, 27, 28, 29}.

• ts

  • temperature parameter.
  • We search it in {0.05, 0.1, 0.2, 0.3, 0.5, 0.7, 1.0}.

• hard_neg

  • It indicates that whether utilizing the hard-negative sample.

• use_S1

  • It indicates that whether using the first stage to pretrain.

Dataset

We provide one public TCM dataset: TCM. \data\Herb

• train_id.txt

  • Train file.
  • Each line is 'symptom ID1 symptom ID2 ... \t herb ID1 herb ID2 ...\n'.
  • Every observed interaction means symptom s once interacted herb h, symptom set sc once interacted prescription p

• valid_id.txt

  • Validation file.
  • Each line is 'symptom ID1 symptom ID2 ...\t herb ID1 herb ID2 ...\n'.

• test_id.txt

  • Test file.
  • Each line is 'symptom ID1 symptom ID2 ...\t herb ID1 herb ID2 ...\n'.

• x.npz

• entity interaction matrix