Introduction to Hopwise

Welcome to hopwise, an advanced extension of the RecBole library designed to enhance recommendation systems with the power of knowledge graphs.

What is Hopwise?

Hopwise integrates knowledge embedding models, path-based reasoning methods, and path language modeling approaches to support both recommendation and link prediction tasks with a focus on explainability.

../_images/hopwise.png

Key Features

Hopwise extends the popular RecBole framework with several unique capabilities:

  1. Path Reasoning Methods: Novel recommendation models that leverage knowledge graph paths to provide explainable recommendations (KGGLM, PEARLM, PLM).

  2. Knowledge Graph Embeddings: 14 KGE methods for both recommendation and link prediction tasks (TransE, TransH, TransR, TransD, RotatE, ComplEx, DistMult, etc.).

  3. Path Quality Metrics: New evaluation metrics for assessing the quality of explanation paths (LIR, SEP, SED, PTD, PTC, etc.).

  4. Path Sampling Utilities: Tools for sampling and managing paths from knowledge graphs.

  5. Link Prediction Support: Full support for knowledge graph completion tasks alongside recommendation.

Model Categories

Hopwise supports six categories of recommendation models:

  • General Recommendation: Collaborative filtering models (BPR, NeuMF, LightGCN, etc.)

  • Sequential Recommendation: Models that consider interaction sequences (SASRec, BERT4Rec, etc.)

  • Context-aware Recommendation: Models that use contextual features (DeepFM, xDeepFM, etc.)

  • Knowledge-based Recommendation: Models that leverage knowledge graphs (KGAT, KGIN, RippleNet, etc.)

  • Path Reasoning Recommendation: Explainable models using KG paths (KGGLM, PEARLM, PLM)

  • Knowledge Graph Embeddings: KGE models for recommendation and link prediction

Quick Example

Here’s a simple example to train a knowledge-aware model:

from hopwise.quick_start import run_hopwise

# Train a knowledge-aware model
run_hopwise(model='KGAT', dataset='ml-100k')

For path-based models with language modeling:

from hopwise.quick_start import run_hopwise

# Train a path language model
run_hopwise(model='KGGLM', dataset='ml-100k')

Command Line Interface

Hopwise provides a convenient CLI for common tasks:

# Train a model
hopwise train --model KGAT --dataset ml-100k

# Evaluate from checkpoint
hopwise evaluate --model KGAT --dataset ml-100k --checkpoint saved/model.pth

# Hyperparameter tuning
hopwise tune --model KGAT --dataset ml-100k --params_file hyper.test

Next Steps