Leveraging the Metapath and Entity Aware Subgraphs for Recommendation

Introduction

Integrating content information for user preference prediction remains a challenging task in the development of recommender systems. Due to the shallow structure, classic graph neural networks (GNNs) failed in modelling high-order graph structures that deliver critical insights of task relevant relations. A natural solution of capturing the inter- and intra-relations between content features and user-item pairs is to explore the high-order information encoded by metapaths, a set of composite relations designed for representing multi-hop structure and sequential semantics.

## PEAGNN Framework  We propose Meta Path- and Entity-Aware Graph Neural Network (PEAGNN), a unified GNN framework tailored for recommendation tasks, which is capable of exploiting the rich semantics in metapaths. PEAGNN trains multilayer GNNs to perform metapath-aware information aggregation on collaborative subgraphs, h-hop subgraphs around the target user-item pairs. After the attentive fusion of aggregated information from different metapaths, a graph-level representation is then extracted for matching score prediction. To leverage the local structure of collaborative subgraphs, we present entity-awareness that regularizes node embedding with the presence of features in a contrastive manner. Further analysis indicates that trained PEAGNN automatically derives meaningful metapath combinations from the given metapaths. ## Results Our model is able to outperform the other competitive baselines on three public datasets, namely: MovieLens-small, MovieLens-25M and Yelp. We use Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG) metrics. The performances of PEAGNN variants and baselines are presented below:  ## Requirements and Installation * Python 3.6 * [PyTorch](http://pytorch.org/) 1.5.1 * [PyTorch Geometric](https://github.com/rusty1s/pytorch_geometric) 1.5.0 * Install all dependencies run ``` pip3 install -r requirements.txt ``` * our code can be installed by ``` python3 setup.py install ``` ## Directory structure The basic skeleton of our source code will look like this : ```bash ├── datasets │ └── Yelp │ └── yelp_dataset.tar ├── experiments │ ├── checkpoint │ │ ├── data │ │ │ ├── Movielenslatest-small │ │ │ ├── Movielenslatest-25m │ │ │ └── Yelp │ │ ├── loggers │ │ │ ├── Movielenslatest-small │ │ │ ├── Movielenslatest-25m │ │ │ └── Yelp │ │ └── weights │ │ ├── Movielenslatest-small │ │ ├── Movielenslatest-25m │ │ └── Yelp │ ├── scripts │ │ ├── **/*.ps1 │ └── **/*.py ├── graph_recsys_benchmark │ ├── datasets │ │ ├── **/*.py │ ├── models │ │ ├── **/*.py │ ├── nn │ │ ├── **/*.py │ ├── parser │ │ ├── **/*.py │ ├── utils │ │ ├── **/*.py │ └── **/*.py ├── images │ └── **/*.png ├── license.txt ├── README.md ├── requirements.txt └── setup.py ``` ## Description of the Code The code is based on [PyTorch Geometric](https://github.com/rusty1s/pytorch_geometric) documentation. - [`experiments`](experiments): contain experiment files for PEAGNN and baseline models - [`checkpoint`](experiments/checkpoint): contain processed data, logs and model weights - [`scripts`](experiments/scripts): scripts to reproduce the results for each dataset - [`datasets`](graph_recsys_benchmark/datasets): creates Heterogenous Information network for the datasets - [`models`](graph_recsys_benchmark/models): creates PEAGNN and baseline models - [`nn`](graph_recsys_benchmark/nn): contains convolutional networks for GNN based models - [`parser`](graph_recsys_benchmark/parser): functions to parse the raw dataset files - [`utils`](graph_recsys_benchmark/utils): functions to save, load models and compute evaluation metrics ## Running the experiments ### Download the datasets #### MovieLens dataset Dataset will be downloaded during experiment execution and saved in [``experiments/checkpoint/data``](experiments/checkpoint/data) folder. No need to explicity download the dataset. Weights of pre-trained Movielens-small dataset can be found [here](experiments/checkpoint/weights/Movielenslatest-small). #### Yelp dataset Download the dataset via this [link](https://www.yelp.com/dataset/download) and save ``yelp_dataset.tar`` in the ``datasets/Yelp`` folder. Like MovieLens dataset, the extracted raw files will be saved in [``experiments/checkpoint/data``](experiments/checkpoint/data) folder during execution. ## Running the Code For training and testing PEAGNN and baseline models, pass the appropriate arguments to the [``experiments``](experiments) files. To reproduce the benchmark results for particular dataset, use the arguments as mentioned in [``experiments/scripts``](experiments/scripts) folder. For instance, for training PEAGAT model on MovieLens-small dataset run the following command: ``` python3 peagat_solver_bpr.py --dataset=Movielens --dataset_name=latest-small --num_core=10 --num_feat_core=10 --sampling_strategy=unseen --entity_aware=false --dropout=0 --emb_dim=64 --repr_dim=16 --hidden_size=64 --meta_path_steps=2,2,2,2,2,2,2,2,2 --entity_aware_coff=0.1 --init_eval=true --gpu_idx=0 --runs=5 --epochs=30 --batch_size=1024 --save_every_epoch=26 --metapath_test=false ``` For instance, for training PEAGAT model with entity awareness on MovieLens-small dataset run the following command: ``` python3 peagat_solver_bpr.py --dataset=Movielens --dataset_name=latest-small --num_core=10 --num_feat_core=10 --sampling_strategy=unseen --entity_aware=true --dropout=0 --emb_dim=64 --repr_dim=16 --hidden_size=64 --meta_path_steps=2,2,2,2,2,2,2,2,2 --entity_aware_coff=0.1 --init_eval=true --gpu_idx=0 --runs=5 --epochs=30 --batch_size=1024 --save_every_epoch=26 --metapath_test=true ``` For training other baselines e.g. KGAT on MovieLens-small dataset run the following command: ``` python3 kgat_solver_bpr.py --dataset=Movielens --dataset_name=latest-small --num_core=10 --num_feat_core=10 --sampling_strategy=unseen --entity_aware=false --dropout=0.1 --emb_dim=64 --hidden_size=64 --entity_aware_coff=0.1 --init_eval=false --gpu_idx=0 --runs=5 --epochs=30 --batch_size=1024 --save_every_epoch=26 ``` # Pre-Trained Models available for download The pre-trained models for all three datasets used in this paper can be accessed via this [link](https://drive.google.com/drive/folders/1Y9sTnS_TXk3d8Hv5_I061w-iA3lmdU0e?usp=sharing).