[Paper Review] Graph DNA: Deep Neighborhood Aware Graph Encoding for Collaborative Filtering

Liwei Wu, Hsiang‐Fu Yu+3 more|arXiv (Cornell University)|May 29, 2019

Caching and Content Delivery43 references4 citations

TL;DR

This paper proposes Graph DNA, a deep neighborhood-aware graph encoding method that uses Bloom filters to efficiently approximate multi-hop neighborhood information in collaborative filtering. By encoding higher-order graph structures in linear time with logarithmic space, Graph DNA enables significant performance gains in matrix factorization and GCN-based models with minimal computational overhead, outperforming direct higher-order graph methods by up to 6x in speed and 15x in performance on some datasets.

ABSTRACT

In this paper, we consider recommender systems with side information in the form of graphs. Existing collaborative filtering algorithms mainly utilize only immediate neighborhood information and have a hard time taking advantage of deeper neighborhoods beyond 1-2 hops. The main caveat of exploiting deeper graph information is the rapidly growing time and space complexity when incorporating information from these neighborhoods. In this paper, we propose using Graph DNA, a novel Deep Neighborhood Aware graph encoding algorithm, for exploiting deeper neighborhood information. DNA encoding computes approximate deep neighborhood information in linear time using Bloom filters, a space-efficient probabilistic data structure and results in a per-node encoding that is logarithmic in the number of nodes in the graph. It can be used in conjunction with both feature-based and graph-regularization-based collaborative filtering algorithms. Graph DNA has the advantages of being memory and time efficient and providing additional regularization when compared to directly using higher order graph information. We conduct experiments on real-world datasets, showing graph DNA can be easily used with 4 popular collaborative filtering algorithms and consistently leads to a performance boost with little computational and memory overhead.

Motivation & Objective

• To address the challenge of efficiently leveraging deep neighborhood information in graph-based collaborative filtering, where higher-order neighbors are typically too costly to compute.
• To reduce the computational and memory complexity of incorporating multi-hop graph information in recommendation systems.
• To develop a scalable encoding method that preserves structural information from deeper neighborhoods without exponential growth in resource usage.
• To integrate the encoding into both feature-based and graph-regularized collaborative filtering frameworks for broad applicability.
• To theoretically and empirically validate the effectiveness and efficiency of the proposed Graph DNA encoding.

Proposed method

• Graph DNA uses Bloom filters to encode the set of nodes within a specified number of hops from each node, approximating deep neighborhood structure.
• Each node's encoding is a c-dimensional binary vector with c = O(log n), enabling logarithmic space representation of neighborhood sets.
• The method computes approximate neighborhood information in linear time by iteratively applying Bloom filter operations over the graph adjacency matrix.
• The resulting node encodings are used as input features or regularization terms in collaborative filtering models, such as graph-regularized matrix factorization and GCN-based models.
• The approach provides implicit regularization by smoothing node representations through the Bloom filter's probabilistic membership encoding.
• The method supports both explicit and implicit feedback settings and can be applied to various collaborative filtering algorithms without architectural changes.

Experimental results

Research questions

• RQ1 Can we efficiently approximate deep neighborhood information in large sparse graphs for collaborative filtering without incurring prohibitive computational costs?
• RQ2 Does using Bloom filter-based encoding of multi-hop neighborhoods lead to better generalization and performance in recommendation systems compared to shallow or direct higher-order methods?
• RQ3 How does Graph DNA compare in efficiency and accuracy to directly computing higher-order adjacency matrices (e.g., G^4) in terms of time, memory, and prediction performance?
• RQ4 In what scenarios is deeper neighborhood information more beneficial than immediate neighborhood information, and can Graph DNA adaptively capture this?
• RQ5 What is the theoretical relationship between Graph DNA's encoding and shortest-path distance metrics in graphs?

Key findings

• Graph DNA achieved up to a 15x performance gain over direct higher-order graph methods (e.g., G^4) when using 4th-order information, with a 6x speedup in computation.
• On the Flixster dataset, GRMF_DNA-3 (6-hop) achieved a 10x performance gain over 1-hop GRMF_G, demonstrating the value of deeper information.
• On the Douban dataset, GRMF_DNA-1 outperformed all baselines, achieving a 1.38x gain over baseline MF, indicating that shallow information is more critical in dense graphs.
• Graph DNA reduced memory usage significantly: the Bloom filter matrix B in GRMF_DNA-1 contained 20% fewer non-zeros than G^2, while still preserving rich structural information.
• In the synthesis dataset, GRMF_DNA-3 achieved a 19.1% relative improvement over baseline MF, outperforming even full G^6 with 17% gain.
• On implicit feedback tasks, WMF_DNA-3 improved MAP by 7.1% on Douban and 6.6% on Flixster compared to WMF_G, with minimal overhead.