• Contents
• TB-Net Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Arguments
• Model Description
• • Performance
  • • Training Performance
    • Evaluation Performance
    • Inference and Explaining Performance
• Description of Random Situation

Contents

• Contents
  • TB-Net Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
    • Script and Sample Code
    • Script Arguments
  • Model Description
    • Performance
      • Training Performance
      • Evaluation Performance
      • Inference and Explanation Performance
  • Description of Random Situation

TB-Net Description

TB-Net is a knowledge graph based explainable recommender system.

Paper: Shendi Wang, Haoyang Li, Xiao-Hui Li, Caleb Chen Cao, Lei Chen. Tower Bridge Net (TB-Net): Bidirectional Knowledge Graph Aware Embedding Propagation for Explainable Recommender Systems

Model Architecture

TB-Net constructs subgraphs in knowledge graph based on the interaction between users and items as well as the feature of items, and then calculates paths in the graphs using bidirectional conduction algorithm. Finally we can obtain explainable recommendation results.

Dataset

We provide an example dataset that created from
Interaction of users and games and games' feature data of the game platform Steam that are publicly available on Kaggle.

Please refer to Downloading Data Package for the way of obtaining the example dataset and the file format descriptions.

Environment Requirements

• Hardware
  • Supports GPU.
• Framework
  • MindSpore
  • MindSpore XAI
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

Please refer to Using TB-Net for the quick start guide.

Script Description

Script and Sample Code

.
└─tbnet
  ├─README.md
  ├─README_CN.md
  ├─data
  │ └─steam
  │   ├─LICENSE
  │   ├─config.json                 # hyper-parameters and training configuration
  │   ├─src_infer.csv               # source datafile for inference
  │   ├─src_test.csv                # source datafile for evaluation
  │   └─src_train.csv               # source datafile for training
  ├─src
  │ ├─dataset.py                    # dataset loader
  │ ├─embedding.py                  # embedding module
  │ ├─metrics.py                    # model metrics
  │ ├─path_gen.py                   # data preprocessor
  │ ├─recommend.py                  # result aggregator
  │ └─tbnet.py                      # TB-Net architecture
  ├─export.py                       # export MINDIR/AIR script
  ├─preprocess.py                   # data pre-processing script
  ├─eval.py                         # evaluation script
  ├─infer.py                        # inference and explaining script
  └─train.py                        # training script

Script Arguments

• preprocess.py arguments

--dataset <DATASET>        'steam' dataset is supported currently (default 'steam')
--same_relation            only generate paths that relation1 is same as relation2

• train.py arguments

--dataset <DATASET>        'steam' dataset is supported currently (default 'steam')
--train_csv <TRAIN_CSV>    the train csv datafile inside the dataset folder (default 'train.csv')
--test_csv <TEST_CSV>      the test csv datafile inside the dataset folder (default 'test.csv')
--epochs <EPOCHS>          number of training epochs (default 20)
--device_id <DEVICE_ID>    device id (default 0)
--run_mode <MODE>          run model in 'GRAPH' mode or 'PYNATIVE' mode (default 'GRAPH')

• eval.py arguments

--checkpoint_id <CKPT_ID>    the id of the checkpoint(.ckpt) to be used
--dataset <DATASET>          'steam' dataset is supported currently (default 'steam')
--csv <CSV>                  the test csv datafile inside the dataset folder (default 'test.csv')
--device_id <DEVICE_ID>    device id (default 0)
--run_mode <MODE>          run model in 'GRAPH' mode or 'PYNATIVE' mode (default 'GRAPH')

• infer.py arguments

--checkpoint_id <CKPT_ID>    the id of the checkpoint(.ckpt) to be used
--dataset <DATASET>          'steam' dataset is supported currently (default 'steam')
--csv <CSV>                  the infer csv datafile inside the dataset folder (default 'infer.csv')
--items <ITEMS>              no. of items to be recommended (default 3)
--explanations <EXP>         no. of recommendation reasons to be shown (default 3)
--device_id <DEVICE_ID>      device id (default 0)
--run_mode <MODE>            run model in 'GRAPH' mode or 'PYNATIVE' mode (default 'GRAPH')

• export.py arguments

--config_path <CFG_PATH>        config (config.json) file path
--checkpoint_path <CKPT_PATH>   checkpoint (.ckpt) file path
--file_name <FILENAME>          export filename
--file_format <FORMAT>          export format 'MINDIR' or 'AIR' (default 'MINDIR')
--device_id <DEVICE_ID>         device id (default 0)
--run_mode <MODE>               run model in 'GRAPH' mode or 'PYNATIVE' mode (default 'GRAPH')

Model Description

Performance

Training Performance

Parameters	GPU
Model Version	TB-Net
Resource	Tesla V100-SXM2-32GB
Uploaded Date	2021-08-01
MindSpore Version	1.3.0
Dataset	steam
Training Parameter	epoch=20, batch_size=1024, lr=0.001
Optimizer	Adam
Loss Function	Sigmoid Cross Entropy
Outputs	AUC=0.8596, Accuracy=0.7761
Loss	0.57
Speed	1pc: 90ms/step
Total Time	1pc: 297s
Checkpoint for Fine Tuning	104.66M (.ckpt file)

Evaluation Performance

Parameters	GPU
Model Version	TB-Net
Resource	Tesla V100-SXM2-32GB
Uploaded Date	2021-08-01
MindSpore Version	1.3.0
Dataset	steam
Batch Size	1024
Outputs	AUC=0.8252, Accuracy=0.7503
Total Time	1pc: 5.7s

Inference and Explaining Performance

Parameters	GPU
Model Version	TB-Net
Resource	Tesla V100-SXM2-32GB
Uploaded Date	2021-08-01
MindSpore Version	1.3.0
Dataset	steam
Outputs	Recommendation Result and Explanation
Total Time	1pc: 3.66s

Description of Random Situation

• Initialization of embedding matrix in tbnet.py and embedding.py.