[Paper Review] Unsupervised Meta-path Reduction on Heterogeneous Information Networks

Xiaokai Wei, Zhiwei Liu+2 more|arXiv (Cornell University)|Oct 30, 2018

Neural Networks and Reservoir Computing4 citations

TL;DR

This paper proposes SPMR, an unsupervised meta-path reduction method for heterogeneous information networks that preserves semantic information by optimizing transition probability similarity among meta-paths. It achieves superior clustering performance with fewer meta-paths, outperforming both full meta-path sets and random selection on real-world datasets like BlogCatalog and DBLP.

ABSTRACT

Heterogeneous Information Network (HIN) has attracted much attention due to its wide applicability in a variety of data mining tasks, especially for tasks with multi-typed objects. A potentially large number of meta-paths can be extracted from the heterogeneous networks, providing abundant semantic knowledge. Though a variety of meta-paths can be defined, too many meta-paths are redundant. Reduction on the number of meta-paths can enhance the effectiveness since some redundant meta-paths provide interferential linkage to the task. Moreover, the reduced meta-paths can reflect the characteristic of the heterogeneous network. Previous endeavors try to reduce the number of meta-paths under the guidance of supervision information. Nevertheless, supervised information is expensive and may not always be available. In this paper, we propose a novel algorithm, SPMR (Semantic Preserving Meta-path Reduction), to reduce a set of pre-defined meta-paths in an unsupervised setting. The proposed method is able to evaluate a set of meta-paths to maximally preserve the semantics of original meta-paths after reduction. Experimental results show that SPMR can select a succinct subset of meta-paths which can achieve comparable or even better performance with fewer meta-paths.

Motivation & Objective

• To address the challenge of redundant and noisy meta-paths in heterogeneous information networks (HINs), which degrade downstream task performance.
• To develop an unsupervised method for meta-path reduction that preserves the semantic structure of the original HIN without relying on costly supervision.
• To identify a succinct subset of meta-paths that maintains the essential connectivity and semantic relationships present in the full set.
• To reveal intrinsic network characteristics through the reduced meta-path set, aiding human interpretation of HIN structure.

Proposed method

• SPMR formulates meta-path reduction as an optimization problem that preserves transition probabilities between node pairs across meta-paths.
• It models the semantic similarity between meta-paths using transition probability matrices derived from random walks on the HIN.
• The method employs a projected Quasi-Newton optimization technique to select a minimal subset of meta-paths that best preserve the original transition probability distribution.
• The objective function balances semantic preservation and meta-path set size, favoring compact, informative subsets.
• The approach is unsupervised, relying solely on the network structure and not on external labels or feedback.
• It evaluates the quality of reduced meta-path sets via downstream clustering tasks to validate semantic preservation.

Experimental results

Research questions

• RQ1 Can an unsupervised method effectively reduce the number of meta-paths in a heterogeneous information network while preserving their semantic meaning?
• RQ2 How does the performance of downstream tasks like clustering compare when using a reduced set of meta-paths versus all meta-paths?
• RQ3 What characteristics of the network can be revealed through the selection of a minimal, informative subset of meta-paths?
• RQ4 To what extent can a small number of meta-paths achieve performance comparable or better than using all meta-paths?

Key findings

• On the BlogCatalog dataset, using only 3 meta-paths selected by SPMR improved clustering accuracy by 42.4% compared to using all 15 meta-paths.
• SPMR achieved a clustering NMI of 0.3777 on BlogCatalog with 3 meta-paths, outperforming the baseline of 0.1861 using all paths.
• On the DBLP dataset, SPMR with 3 meta-paths improved NMI by 17.0% over the full set and by 77.3% over random selection.
• The top-ranked meta-path on DBLP, 'Author - Paper - Author - Paper - Term - Paper - Author', preserved most semantic information, outperforming longer, redundant variants.
• Meta-paths involving user and tag relationships were prioritized in BlogCatalog, suggesting their centrality in the network’s structure.
• The elimination of 'Blog - User - User - User - Blog' meta-path indicated that friend-of-friend relationships are sparse and uninformative in BlogCatalog.