mindspore/model_zoo/official/gnn/bgcf

• Bayesian Graph Collaborative Filtering
• Model Architecture
• Dataset
• Features
• • Mixed Precision
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Training Process
  • • Training
  • Evaluation Process
  • • Evaluation
• Model Description
• • Performance
  • • Evaluation Performance
    • Inference Performance
• Description of random situation
• ModelZoo Homepage

• Bayesian Graph Collaborative Filtering
• Model Architecture
• Dataset
• Features
  • Mixed Precision
• Environment Requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
    • Training
  • Evaluation Process
    • Evaluation
• Model Description
  • Performance
• Description of random situation
• ModelZoo Homepage

Bayesian Graph Collaborative Filtering

Bayesian Graph Collaborative Filtering(BGCF) was proposed in 2020 by Sun J, Guo W, Zhang D et al. By naturally incorporating the
uncertainty in the user-item interaction graph shows excellent performance on Amazon recommendation dataset.This is an example of
training of BGCF with Amazon-Beauty dataset in MindSpore. More importantly, this is the first open source version for BGCF.

Paper: Sun J, Guo W, Zhang D, et al. A Framework for Recommending Accurate and Diverse Items Using Bayesian Graph Convolutional Neural Networks[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 2030-2039.

Model Architecture

Specially, BGCF contains two main modules. The first is sampling, which produce sample graphs based in node copying. Another module
aggregate the neighbors sampling from nodes consisting of mean aggregator and attention aggregator.

Dataset

Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.

• Dataset size:
  Statistics of dataset used are summarized as below:

  	Amazon-Beauty
  Task	Recommendation
  # User	7068 (1 graph)
  # Item	3570
  # Interaction	79506
  # Training Data	60818
  # Test Data	18688
  # Density	0.315%

• Data Preparation

  • Place the dataset to any path you want, the folder should include files as follows(we use Amazon-Beauty dataset as an example)"

  .
  └─data
      ├─ratings_Beauty.csv
  

  • Generate dataset in mindrecord format for Amazon-Beauty.

  cd ./scripts
  # SRC_PATH is the dataset file path you download.
  sh run_process_data_ascend.sh [SRC_PATH]
  

Features

Mixed Precision

To ultilize the strong computation power of Ascend chip, and accelerate the training process, the mixed training method is used. MindSpore is able to cope with FP32 inputs and FP16 operators. In BGCF example, the model is set to FP16 mode except for the loss calculation part.

Environment Requirements

• Hardward (Ascend)
• Framework
  • MindSpore
• For more information, please check the resources below:
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website and Dataset is correctly generated, you can start training and evaluation as follows.

• running on Ascend

  # run training example with Amazon-Beauty dataset
  sh run_train_ascend.sh
  
  # run evaluation example with Amazon-Beauty dataset
  sh run_eval_ascend.sh 
  

Script Description

Script and Sample Code

.
└─bgcf      
  ├─README.md
  ├─scripts 
  | ├─run_eval_ascend.sh          # Launch evaluation
  | ├─run_process_data_ascend.sh  # Generate dataset in mindrecord format
  | └─run_train_ascend.sh         # Launch training   
  |
  ├─src
  | ├─bgcf.py              # BGCF model
  | ├─callback.py          # Callback function
  | ├─config.py            # Training configurations
  | ├─dataset.py           # Data preprocessing
  | ├─metrics.py           # Recommendation metrics
  | └─utils.py             # Utils for training bgcf
  |
  ├─eval.py                # Evaluation net
  └─train.py               # Train net

Script Parameters

Parameters for both training and evaluation can be set in config.py.

• config for BGCF dataset

  "learning_rate": 0.001,            # Learning rate
  "num_epoch": 600,                 # Epoch sizes for training
  "num_neg": 10,                     # Negative sampling rate 
  "raw_neighs": 40,                  # Num of sampling neighbors in raw graph
  "gnew_neighs": 20,                 # Num of sampling neighbors in sample graph
  "input_dim": 64,                   # User and item embedding dimension
  "l2": 0.03                   # l2 coefficient
  "neighbor_dropout": [0.0, 0.2, 0.3]# Dropout ratio for different aggregation layer
  

  config.py for more configuration.

Training Process

Training

• running on Ascend

 sh run_train_ascend.sh 

Training result will be stored in the scripts path, whose folder name begins with "train". You can find the result like the
followings in log.

Epoch 001 iter 12 loss 34696.242
Epoch 002 iter 12 loss 34275.508
Epoch 003 iter 12 loss 30620.635
Epoch 004 iter 12 loss 21628.908

...
Epoch 597 iter 12 loss 3662.3152
Epoch 598 iter 12 loss 3640.7612
Epoch 599 iter 12 loss 3654.9087
Epoch 600 iter 12 loss 3632.4585
...

Evaluation Process

Evaluation

• Evaluation on Ascend

 sh run_eval_ascend.sh 

Evaluation result will be stored in the scripts path, whose folder name begins with "eval". You can find the result like the
followings in log.

epoch:020,      recall_@10:0.07345,     recall_@20:0.11193,     ndcg_@10:0.05293,    ndcg_@20:0.06613,   
sedp_@10:0.01393,     sedp_@20:0.01126,    nov_@10:6.95106,    nov_@20:7.22280
epoch:040,      recall_@10:0.07410,     recall_@20:0.11537,     ndcg_@10:0.05387,    ndcg_@20:0.06801,   
sedp_@10:0.01445,     sedp_@20:0.01168,    nov_@10:7.34799,    nov_@20:7.58883
epoch:060,      recall_@10:0.07654,     recall_@20:0.11987,     ndcg_@10:0.05530,    ndcg_@20:0.07015,   
sedp_@10:0.01474,     sedp_@20:0.01206,    nov_@10:7.46553,    nov_@20:7.69436

...
epoch:560,      recall_@10:0.09825,     recall_@20:0.14877,     ndcg_@10:0.07176,    ndcg_@20:0.08883,   
sedp_@10:0.01882,     sedp_@20:0.01501,    nov_@10:7.58045,    nov_@20:7.79586
epoch:580,      recall_@10:0.09917,     recall_@20:0.14970,     ndcg_@10:0.07337,    ndcg_@20:0.09037,   
sedp_@10:0.01896,     sedp_@20:0.01504,    nov_@10:7.57995,    nov_@20:7.79439
epoch:600,      recall_@10:0.09926,     recall_@20:0.15080,     ndcg_@10:0.07283,    ndcg_@20:0.09016,   
sedp_@10:0.01890,     sedp_@20:0.01517,    nov_@10:7.58277,    nov_@20:7.80038
...

Model Description

Performance

Evaluation Performance

Parameter	BGCF
Model Version	Inception V1
Resource	Ascend 910
uploaded Date	09/23/2020(month/day/year)
MindSpore Version	1.0.0
Dataset	Amazon-Beauty
Training Parameter	epoch=600,steps=12,batch_size=5000,lr=0.001
Optimizer	Adam
Loss Function	BPR loss
Training Cost	25min
Scripts	bgcf script

Inference Performance

Parameter	BGCF
Model Version	Inception V1
Resource	Ascend 910
uploaded Date	09/23/2020(month/day/year)
MindSpore Version	1.0.0
Dataset	Amazon-Beauty
Batch_size	5000
Output	probability
Recall@20	0.1534
NDCG@20	0.0912

Description of random situation

BGCF model contains lots of dropout operations, if you want to disable dropout, set the neighbor_dropout to [0.0, 0.0, 0.0] in src/config.py.

ModelZoo Homepage

Please check the official homepage.