GInRec: A Gated Architecture for Inductive Recommendation using Knowledge Graphs

📝 Paper Summary

Inductive Recommendation Graph Neural Networks (GNNs) Knowledge Graph (KG) Integration

GInRec is an inductive recommender system that uses Graph Neural Networks with relation-specific gates to effectively aggregate knowledge graph information for new users and items without retraining.

Core Problem

Existing recommender systems struggle with inductive settings (new users/items) because they rely on fixed IDs or user metadata that isn't always available, or they fail to exploit Knowledge Graph structure effectively.

Why it matters:

Transductive methods require heavy retraining whenever new items or users are added, which is computationally expensive
Existing inductive methods often ignore Knowledge Graphs, relying only on interactions, making recommendations impossible for new items with no interactions
Current methods relying on subgraph extraction for every user-item pair scale poorly for ranking tasks

Concrete Example: In a movie system, if a new user rates 'Inception' but has no other history, standard Collaborative Filtering fails. GInRec uses the Knowledge Graph to link 'Inception' to 'Christopher Nolan' and 'Sci-Fi', inferring the user might like 'The Prestige' without needing retrained user embeddings.

Key Novelty

Gated Inductive Recommendation (GInRec) with Relation-Specific Gates

Replaces standard GNN aggregation with relation-specific gates that learn distinct weights for different edge types (e.g., 'directed_by' vs 'has_genre') in the Knowledge Graph
Uses an autoencoder jointly trained with the ranking loss to compress rich initial feature vectors (like BERT embeddings of movie plots) into manageable dense representations
Operates purely inductively by deriving user representations from their current interactions and item features, requiring no user metadata or ID-based embeddings

Architecture

The GInRec architecture, detailing the flow from initial entity features to the final user-item score.

Evaluation Highlights

Outperforms state-of-the-art inductive baselines (PinSAGE, IDCF) by statistically significant margins on MovieLens-20m and Amazon-Book datasets
+33% improvement in NDCG@20 over PinSAGE on the Amazon-Book Subsampled dataset (0.06472 vs 0.05398)
Demonstrates high scalability: generates embeddings for all nodes in MovieLens-Subsampled in 0.48s, compared to PinSAGE's 0.446s, while achieving better ranking performance

Breakthrough Assessment

7/10

Strong engineering contribution combining GNNs, KGs, and gating for inductive tasks. Exposes significant flaws in previous evaluation protocols (negative sampling) and offers a scalable solution.

⚙️ Technical Details

Problem Definition

Setting: Top-k recommendation in a Collaborative Knowledge Graph (CKG) setting

Inputs: Interaction matrix I, Knowledge Graph G, feature function X (textual/structural features)

Outputs: Ranking score F(u, i) representing likelihood of user u interacting with item i

Pipeline Flow

Feature Initialization (Sentence-BERT on text + structural data)
Embedding Layer (Autoencoder compression)
Gated Propagation (Multi-layer GNN with relation-specific gates)
Prediction (Dot product of final user/item embeddings)

System Modules

Embedding Layer

Compress high-dimensional features into dense embeddings

Model or implementation: Autoencoder (Encoder-Decoder)

Gated Propagation Layer

Aggregate neighborhood information using relation-specific gates

Model or implementation: GNN with Custom Gating

Prediction Layer

Generate ranking scores

Model or implementation: Dot Product

Novel Architectural Elements

Relation-specific gating mechanism within the GNN aggregation step, differentiating information flow based on edge type (e.g., actor vs. genre)
Integration of an autoencoder directly into the end-to-end training pipeline to handle high-dimensional text features for inductive learning

Modeling

Base Model: Custom GNN architecture (GInRec)

Training Method: End-to-end training with combined Ranking and Reconstruction loss

Objective Functions:

Purpose: Optimize ranking of positive items over negative items.

Formally: L_CF = Sum -ln(sigmoid(y_ui - y_uj)) [BPR Loss]
Purpose: Ensure compressed embeddings retain original feature information.

Formally: L_AE = MSE(X, AE_de(AE_en(X)))
Purpose: Combined objective.

Formally: L = L_CF + lambda * L_AE + gamma * ||Theta||^2

Key Hyperparameters:

batch_size: 1024
optimizer: Adam
ego_network_size: 10
+ 1 more
initial_feature_dim: >756

Compute: Training/Inference on RTX 2070 Super + Intel i9-9900. ML-S embedding generation: 0.48s.

Comparison to Prior Work

vs. PinSAGE: GInRec uses relation-specific gates rather than simple aggregation, allowing finer control over KG signals
vs. IDCF: GInRec uses BPR ranking loss and KG structure, whereas IDCF uses Cross-Entropy (pointwise) and focuses on user reconstruction
vs. IGMC [not cited in paper]: IGMC extracts subgraphs for every user-item pair (slow), whereas GInRec computes node embeddings once (scalable)

Limitations

Requires textual descriptions/features for items to initialize the inductive process
Autoencoder compression adds a hyperparameter (lambda) to tune
Performance depends on the quality and completeness of the underlying Knowledge Graph

Reproducibility

Code: https://github.com/theisjendal/kars2023-recommendation-framework

Source code publicly available at https://github.com/theisjendal/kars2023-recommendation-framework. Datasets (MovieLens-20m, Amazon-Book) are public. Hyperparameters tuned via ASHA.

📊 Experiments & Results

Evaluation Setup

Inductive cold-start recommendation (predicting for users not seen during training)

Benchmarks:

MovieLens-20m (ML-S) (Movie Recommendation (Cold-Start))
Amazon-Book (AB-S) (Book Recommendation (Cold-Start))

Metrics:

NDCG@20
Recall@20
Precision@20
Coverage@20
Statistical methodology: Student t-test used for statistical significance (* indicates significant increase)

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
GInRec consistently outperforms inductive baselines on ranking metrics across multiple datasets and sparsity settings.
MovieLens Subsampled	NDCG@20	0.14129	0.18085	+0.03956
MovieLens Subsampled	Recall@20	0.18745	0.24618	+0.05873
Amazon Book Subsampled	NDCG@20	0.05398	0.06472	+0.01074
Ablation studies confirm the necessity of relation-specific information and the gating mechanism.
MovieLens Subsampled + 1250 users	NDCG@20	0.12482	0.18085	+0.05603
MovieLens Subsampled + 1250 users	NDCG@20	0.17720	0.18085	+0.00365

Experiment Figures

Performance (NDCG@20) breakdown by user groups based on activity level and popularity bias.

Main Takeaways

Evaluated utilizing standard ranking protocols (ranking all items) rather than sampled metrics (which bias results towards popularity), showing GInRec's true superiority.
Relational gates are critical: The 'Inner Product' gate (affinity) and 'Concatenation' gate (feature selection) allow the model to filter KG noise effectively.
Highly scalable: The complexity is bounded by edge aggregation using MapReduce logic, making it faster than subgraph-based methods and competitive with PinSAGE.
Diversity: GInRec achieves significantly higher coverage (0.185 vs 0.029 for PinSAGE on ML-S), indicating it recommends a broader variety of items rather than just popular ones.

📚 Prerequisite Knowledge

Prerequisites

Graph Neural Networks (GNNs) and Message Passing
Knowledge Graphs (KGs) and Collaborative Filtering
Inductive vs. Transductive Learning
Autoencoders

Key Terms

Inductive Learning: A learning paradigm where the model can generalize to unseen nodes (users/items) not present during training, utilizing their features rather than learned IDs

Transductive Learning: A learning paradigm where all nodes must be present during training; the model learns specific embeddings for specific IDs and requires retraining for new nodes

Collaborative Knowledge Graph (CKG): A graph combining a Knowledge Graph (entities/relations) with user-item interaction data (users connected to items they liked)

NDCG: Normalized Discounted Cumulative Gain—a ranking metric that values correct recommendations higher when they appear at the top of the list

BPR: Bayesian Personalized Ranking—a pairwise loss function optimizing the relative order between observed (positive) and unobserved (negative) items

Cold-start: The scenario where the system must make recommendations for users or items with very few or no historical interactions

GraphSAGE: An inductive framework for generating node embeddings by sampling and aggregating features from a node's local neighborhood

Gating Mechanism: A neural network component (like in GRUs/LSTMs) that regulates the flow of information, deciding how much of a neighbor's signal to pass forward