1. Installation
2. Running experiments
   1. Training
   2. Evaluation
3. Results
4. General information about the framework
5. NOTICE
   1. Models
   2. DGL

This project contains the model GInRec, and inductive recommender for both users and items.

It was presented at the KARS workshop under the CC 4.0 license and can be found here: GInRec: A Gated Architecture for Inductive Recommendation using Knowledge Graphs

Python version. 3.8 all other requirements are located in requirements.txt. We highly recommend using a virtual environment.

In order to use the code you should first download datasets. As some query changing knowledge bases results are bound to differ. Exact datasets and results will be uploaded after submission.

For local run, install all required packages (possibly install wheel before):

pip3 install -r requirements.txt

The methods can be run using:

python3 train/dgl_trainer.py --out_path ./results --data ./datasets --experiments ml_mr_1m_warm_start --include ginrec --test_batch 1024 --workers 1

For help with commands write:

python3 train/dgl_trainer.py --help

The method will automatically tune hyperparameters, otherwise the method can be set to gridsearch in the under configuration/models.py, where setting the keyword 'gridsearch' to True as in PPR. All models can be found in the configuration/models.py file, all datasets in the configuration/datasets.py file, and all experiments in the configuration/experiments.py file. Note that a dataset is a single instance of a dataset, while there might be multiple experiments for a single dataset. E.g., we have the full MovieLens dataset ml-mr and the subsampled MovieLens dataset ml-mr-1m. We then have multiple experiments for each of these datasets both warm and cold start.

We can use the --parameter option to use a parameter tuned at another dataset. E.g, if we have tuned a model on the ml-mr-1m dataset and want to train on the ab dataset, use --experiments ab_warm_start --parameter ml_mr_1m_warm_start.

Evaluation requires two steps, first running evaluate/dgl_evaluator.py and then metric_calculator.py. Similarly to the trainer you can use the --help argument if in doubt. Examples:

--data ./datasets --results ./results --experiments ml_mr_1m_warm_start --include ginrec --test_batch 1024

The ml-mr-1m dataset is a subset of the ml-mr dataset. It is therefore possible to train on the subset and evaluate on the full dataset. When running the evaluator script there are multiple arguments. For the coldstart we have:

--state - use state from another experiment - useful for coldstart methods
--parameter - use parameters from another experiment - useful for training with pre-determined parameters
--parameter_usage - when parameter argument was used. 0) it was not passed during training (default) and 1) it was passed during training.
--other_model - use parameters from other method i.e. change one parameter
--other_model_usage - where other model was used: 0) location for parameters, 1) location of state, 2) both. E.g., if you want to load parameters from another model but use state from the model itself use 1.

For example, lets say we have tuned a model on the ml-mr-1m dataset, trained it on the ab-s dataset using those parameter (see last example of the training section) and want to evaluate in the coldstart setting of ab. We would then have to pass the arguments as:

--experiment ab_user_cold_start --state ab_warm_start --parameter ml_mr_1m_warm_start --parameter_usage 1

If we do not set parameter usage as 1 we might end up using the wrong state as it only uses the parameters for evaluation, e.g. tuned on the ml-mr-1m dataset and evaluating on the ml-mr dataset. We wouldn't need to retrain on the ml-mr dataset in the cold-start setting and therefore did not use the parameter option during training.

The other_model and other_model_usage are similar but are primarily used for ablation studies.

Under configuration you will find all parameters for the models as well as dataset and experiment configuration. For datasets you can define a mapping of ratings, such as explicit to implicit and apply filters, such as minimal number of ratings using CountFilter:

ml_mr_1m = DatasetConfiguration('ml-mr-1m', lambda x: {x > 3: 1}.get(True, 0),
                                filters=[CountFilter(lambda c: c >= 5, sentiment=Sentiment.POSITIVE)],
                                max_users=12500)

All methods, experiments and datasets must be defined in these folders before use. All methods are located in the models folder and inherit from the RecommenderBase class in dgl_recommender_base.py

Look in datsets folder for the construction of datasets.

Section describing modifications to documents under the Apache License or for recognition. Both IGMC and NGCF are originally under the Apache License.

The IGMC model is a modified version of Based on https://github.com/LspongebobJH/dgl/blob/pr1815_igmc/examples/pytorch/igmc/ For download: https://github.com/LspongebobJH/dgl/archive/refs/heads/pr1815_igmc.zip

Specifically, it has been implemented with the RecommenderBase and uses a DGL collator for training.

Heavily inspired by the work of jennyzhang0215 in https://github.com/jennyzhang0215/DGL-KGAT.

Based on the example of NGCF in Based on https://github.com/dmlc/dgl/tree/master/examples/pytorch/NGCF modified to include LightGCN and using blocks.

Under models/shared/dgl_dataloader.py, you will find changes of the DGL EdgeCollator classes and EdgeDataloader to skip certain graph building steps.