Library of Recommender System based on PyTorch

Introduction

A Python package to integrate the pipeline of recommender systems for simple model designing and fast idea verification.

Installation

bash install.sh

Requirement

0. Python >= 3.8
1. PyTorch >= 1.2.0
2. visdom == 0.1.8.9 (for visualization)
3. nvidia_ml_py3 == 7.352.0 (for auto-selection of GPU)
4. Negative sampling recommender systems

Advantage

1. Separate model design from other dirty things like CLI, logging, GPU selection, etc.
2. Provide two general methods to save and visualize key values in the pipeline, TrainHook and Metric.
3. Provide widely used loss functions and model design patterns in recommendation, supporting customization easily.

Submodules

analysis

The submodule is to automatically analyse the dataset or the model results for paper presentation. Now the submodule only supports:

• multimodel.py: organize models' metrics and report them as a csv table

argument

data

utils

logger.py

The submodule is to easily record the hyperparameters setting, the environment setting (like optimizer) and all related values of each model into specific path.

• Logger: the class to implement the function above

loss.py

The submodule is to contain some commonly used loss functions in recommendation, including L2loss, BPR, MSE with mask, BCE.

metric.py

The submodule is to provide a general way to save metrics in model evaluation and visualize them by Visdom. We provide Precision Recall NDCG MRR for fully-ranking and leave-one-out mode (in leave-one-out mode, Precision Recall->HR)

model.py

The submodule is to provide some base class for models.

• Model: the abstract class for all models in our framework
• EmbeddingBasedModel: the abstract class for embedding based models, which provide embedding generation
• DotBasedModel: the abstract class for dot based models, which simplify forward, before_evaluation and evaluation and provide propagate interface

pipeline.py

test.py

The submodule is to evaluate model in fully-ranking or leave-one-out mode by fully_ranking_test or leave_one_out_test, respectively.

train.py

The submodule is to train model.

trainhook.py

The submodule is to provide a general way to save key values in model forward phase and visualize them as metrics. The trainhooks will be found in model.trainhooks as dict, whose key is the title of trainhook (__loss__ is used for recording loss of each epoch).

• Trainhook: the base class for trainhook, providing interface start() stop() __call__() title value

• ValueMeanHook: the implementation of trainhook, which calculate the meaning of each value

Example

Configuration

config.json

{
    "user": "name_for_setproctitle",
    "visdom": {
        "server": "visdom_ip",      # use empty string to disable visdom visualization
        "port": {
            "dataset_name1": 10001,
            "dataset_name2": 10002
        }
    },
    "training": {
        "test_interval": 5,
        "early_stop": 50,
        "overfit": {
            "protected_epoch": 10,
            "threshold": 0.1
        }
    },
    "dataset": {
        "path": "path_to_dataset",
        "seed": 123,
        "use_backup": true
    },
    "logger": {
        "path": "path_to_log",
        "policy": "best"
    },
    "metric": {
        "target": {
            "type": "NDCG",
            "topk": 10
        },
        "metrics": [
            {
                "type": "Recall",
                "topk": 5
            },
            {
                "type": "Recall",
                "topk": 10
            },
            {
                "type": "NDCG",
                "topk": 5
            },
            {
                "type": "NDCG",
                "topk": 10
            }
        ]
    }
}

• user: One part of processes name when running the codes.

Visdom

visdom part in config.json

• server: The IP of visdom server. Set it to "" to disable visualization.
• port: Key-value pairs, which key is dataset name and value is the port of visdom server.

Training

training part in config.json

• test_interval: The interval epoch for testing.
• early_stop: Stop training when the target metric drops early_stop times.
• overfit: overfitting detection
  • protected_epoch: Disable overfitting detection for the first protected_epoch epochs.
  • threshold: Stop training when the target metric is less than threshold, which means the model is overfitting.

Dataset

dataset part in config.json

• path: The root folder path of all datasets.
• seed: The seed for negative sampling.
• use_backup: Save/load the negative sampling results into/from file or not.

Logger

logger part in config.json

• path: The root folder path for log files.
• policy: [Do not modify] When to save the model as checkpoint.

Metric

metric part in config.json

Each metric is described by

{
    "type": "MetricName",
    "topk": X
}

, which means MetricName@X (e.g. NDCG@10, Recall@5, MRR@40).

• target: Metric used for early stopping, overfitting detection and so on.
• metrics: The list of metrics which should be computed and recorded.

Pretrain Configuration

pretrain.json

{
    "BeiBei": {
        "GBMF": "path-to-GBMF-pretrain-model"
    }
}

The key of the first layer is dataset name.

The key of the second layer is pretrain model name and the value is the path.

According to dataset used for training or testing, the second layer will be sended into model.load_pretrain as python dict pretrain_info.

Roadmap

• Automatic analysis
  • Performance evaluation for data sparsity issue
  • Dataset profile
• Pass static check by complete type hints
• Support more activation functions
• Add DGL module support
• Convergence judgment based on linear fitting instead of early-stopping
• Replace Visdom to Tensorboard for visualization of training
• Avoid same-id error in metric and speed up by sorting top-k, providing context manager