• Contents
• TBNet Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• Script Description
• • Script and Sample Code
  • Script Parameters
  • Inference Process
  • • Export MindIR
    • Infer on Ascend310
    • Result
• Model Description
• • Performance
  • • Training Performance
    • Evaluation Performance
    • Inference and Explanation Performance
• Description of Random Situation
• ModelZoo Homepage

Contents

• Contents
  • TBNet Description
  • Model Architecture
  • Dataset
  • Environment Requirements
  • Quick Start
  • Script Description
    • Script and Sample Code
    • Script Parameters
    • Inference Process
      • Export MindIR
      • Infer on Ascend310
      • Result
  • Model Description
    • Performance
      • Training Performance
      • Evaluation Performance
      • Inference and Explanation Performance
  • Description of Random Situation
  • ModelZoo Homepage

TBNet 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

Interaction of users and games, and the games' feature data on the game platform Steam are public on Kaggle.

Dataset directory: ./data/{DATASET}/, e.g. ./data/steam/.

• train: train.csv, evaluation: test.csv

Each line indicates a <user>, an <item>, the user-item <rating> (1 or 0), and PER_ITEM_NUM_PATHS paths between the item and the user's <hist_item> (<hist_item> is the item whose the user-item <rating> in historical data is 1).

#format:user,item,rating,relation1,entity,relation2,hist_item,relation1,entity,relation2,hist_item,...,relation1,entity,relation2,hist_item  # module [relation1,entity,relation2,hist_item] repeats PER_ITEM_NUM_PATHS times

• infer and explain: infer.csv

Each line indicates the <user> and <item> to be inferred, <rating>, and PER_ITEM_NUM_PATHS paths between the item and the user's <hist_item> (<hist_item> is the item whose the user-item <rating> in historical data is 1).
Note that the <item> needs to traverse candidate items (all items by default) in the dataset. <rating> can be randomly assigned (all values are assigned to 0 by default) and is not used in the inference and explanation phases.

#format:user,item,rating,relation1,entity,relation2,hist_item,relation1,entity,relation2,hist_item,...,relation1,entity,relation2,hist_item  # module [relation1,entity,relation2,hist_item] repeats PER_ITEM_NUM_PATHS times

We have to download the data package and put it underneath the current project path。

wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/tbnet_data.tar.gz
tar -xf tbnet_data.tar.gz

Environment Requirements

• Hardware（GPU）
  • Prepare hardware environment with GPU processor.
• Framework
  • MindSpore
• For more information, please check the resources below：
  • MindSpore Tutorials
  • MindSpore Python API

Quick Start

After installing MindSpore via the official website, you can start training and evaluation as follows:

• Data preprocessing

Download the data package(e.g. 'steam' dataset) and put it underneath the current project path.

wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/tbnet_data.tar.gz
tar -xf tbnet_data.tar.gz
cd scripts

and then run code as follows.

• Training

bash run_standalone_train_ascend.sh [DATA_NAME] [DEVICE_ID]

GPU device:

bash run_standalone_train_gpu.sh [DATA_NAME] [DEVICE_ID]

Example:

bash run_standalone_train_ascend.sh steam 0

• Evaluation

Evaluation model on test dataset.

bash run_eval_ascend.sh [CHECKPOINT_ID] [DATA_NAME] [DEVICE_ID]

GPU device:

bash run_eval_gpu.sh [CHECKPOINT_ID] [DATA_NAME] [DEVICE_ID]

Argument [CHECKPOINT_ID] is required.

Example:

bash run_eval_ascend.sh 19 steam 0

• Inference and Explanation

Recommende items to user acrodding to user, the number of items is determined by items.

python infer.py \
  --dataset [DATASET] \
  --checkpoint_id [CHECKPOINT_ID] \
  --user [USER] \
  --items [ITEMS] \
  --explanations [EXPLANATIONS] \
  --csv [CSV]

Arguments --checkpoint_id and --user are required.

Example:

python infer.py \
  --dataset steam \
  --checkpoint_id 19 \
  --user 2 \
  --items 1 \
  --explanations 3 \
  --csv test.csv

Script Description

Script and Sample Code

.
└─tbnet
  ├─README.md
  ├── scripts
      ├─run_infer_310.sh                  # Ascend310 inference script
      ├─run_standalone_train_ascend.sh    # Ascend NPU training script
      ├─run_eval_ascend.sh                # Ascend NPU evaluation script
      ├─run_standalone_train_gpu.sh       # NVIDIA GPU training script
      └─run_eval_gpu.sh                   # NVIDIA GPU evaluation script
  ├─data
    ├─steam
        ├─config.json               # data and training parameter configuration
        ├─src_infer.csv             # inference and explanation dataset
        ├─src_test.csv              # evaluation dataset
        ├─src_train.csv             # training dataset
        └─id_maps.json              # explanation configuration
  ├─src
    ├─aggregator.py                 # inference result aggregation
    ├─config.py                     # parsing parameter configuration
    ├─dataset.py                    # generate dataset
    ├─embedding.py                  # 3-dim embedding matrix initialization
    ├─metrics.py                    # model metrics
    ├─steam.py                      # 'steam' dataset text explainer
    └─tbnet.py                      # TB-Net model
  ├─export.py                         # export mindir script
  ├─preprocess_dataset.py           # dataset preprocess script
  ├─preprocess.py                         # inference data preprocess script
  ├─postprocess.py                         # inference result calculation script
  ├─eval.py                         # evaluation
  ├─infer.py                        # inference and explanation
  └─train.py                        # training

Script Parameters

• preprocess_dataset.py parameters

--dataset         'steam' dataset is supported currently
--device_target   run code on GPU or Ascend NPU
--same_relation   only generate paths that relation1 is same as relation2

• train.py parameters

--dataset         'steam' dataset is supported currently
--train_csv       the train csv datafile inside the dataset folder
--test_csv        the test csv datafile inside the dataset folder
--device_id       device id
--epochs          number of training epochs
--device_target   run code on GPU or Ascend NPU
--run_mode        run code by GRAPH mode or PYNATIVE mode

• eval.py parameters

--dataset         'steam' dataset is supported currently
--csv             the csv datafile inside the dataset folder (e.g. test.csv)
--checkpoint_id   use which checkpoint(.ckpt) file to eval
--device_id       device id
--device_target   run code on GPU or Ascend NPU
--run_mode        run code by GRAPH mode or PYNATIVE mode

• infer.py parameters

--dataset         'steam' dataset is supported currently
--csv             the csv datafile inside the dataset folder (e.g. infer.csv)
--checkpoint_id   use which checkpoint(.ckpt) file to infer
--user            id of the user to be recommended to
--items           no. of items to be recommended
--reasons         no. of recommendation reasons to be shown
--device_id       device id
--device_target   run code on GPU or Ascend NPU
--run_mode        run code by GRAPH mode or PYNATIVE mode

Inference Process

Export MindIR

python export.py --config_path [CONFIG_PATH] --checkpoint_path [CKPT_PATH] --device_target [DEVICE] --file_name [FILE_NAME] --file_format [FILE_FORMAT]

• CKPT_PATH parameter is required.
• CONFIG_PATH is config.json file, data and training parameter configuration.
• DEVICE should be in ['Ascend', 'GPU'].
• FILE_FORMAT should be in ['MINDIR', 'AIR'].

Example：

python export.py \
  --config_path ./data/steam/config.json \
  --checkpoint_path ./checkpoints/tbnet_epoch19.ckpt \
  --device_target Ascend \
  --file_name model \
  --file_format MINDIR

Infer on Ascend310

Before performing inference, the mindir file must be exported by export.py script. We only provide an example of inference using MINDIR model.

# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID]

• MINDIR_PATH specifies path of used "MINDIR" model.
• DATA_PATH specifies path of test.csv.
• DEVICE_ID is optional, default value is 0.

Example：

bash run_infer_310.sh ../model.mindir ../data/steam/test.csv 0

Result

Inference result is saved in current path, you can find result like this in acc.log file.

auc: 0.8251359368836292

Model Description

Performance

Training Performance

Parameters	GPU	Ascend NPU
Model Version	TB-Net	TB-Net
Resource	NVIDIA RTX 3090	Ascend 910
Uploaded Date	2022-07-14	2022-06-30
MindSpore Version	1.6.1	1.6.1
Dataset	steam	steam
Training Parameter	epoch=20, batch_size=1024, lr=0.001	epoch=20, batch_size=1024, lr=0.001
Optimizer	Adam	Adam
Loss Function	Sigmoid Cross Entropy	Sigmoid Cross Entropy
Outputs	AUC=0.8573，Accuracy=0.7733	AUC=0.8592，准确率=0.7741
Loss	0.57	0.59
Speed	1pc: 90ms/step	单卡：80毫秒/步
Total Time	1pc: 297s	单卡：336秒
Checkpoint for Fine Tuning	686.3K (.ckpt file)	671K (.ckpt 文件)
Scripts	TB-Net scripts

Evaluation Performance

Parameters	GPU	Ascend NPU
Model Version	TB-Net	TB-Net
Resource	NVIDIA RTX 3090	Ascend 910
Uploaded Date	2022-07-14	2022-06-30
MindSpore Version	1.3.0	1.5.1
Dataset	steam	steam
Batch Size	1024	1024
Outputs	AUC=0.8487，Accuracy=0.7699	AUC=0.8486，Accuracy=0.7704
Total Time	1pc: 5.7s	1pc: 1.1秒

Inference and Explanation 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.

ModelZoo Homepage

Please check the official homepage.