Social Network Analysis using RLVECN: Representation Learning via Knowledge-Graph Embeddings and Convolutional Neural-Network

Bonaventure C. Molokwu, Ziad KobtiIJCAI 2020 · DOI

TL;DR

This paper introduces RLVECN, a novel hybrid model that combines knowledge graph embeddings and convolutional neural networks to learn rich representations of social networks. This approach aims to improve performance on crucial social network analysis tasks like node classification and link prediction, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Social Network Analysis (SNA) is a critical area in Artificial Intelligence, dealing with the complex interactions within systems. However, tasks like classifying actors (nodes) into groups or predicting future connections (links) remain challenging. Existing methods often struggle to capture the nuanced dynamics and inherent representations within these networks. This paper tackles this by proposing a new method to extract meaningful features from social graphs, enabling more accurate classification and prediction.

Methods

The core of this work is the RLVECN (Representation Learning via Knowledge-Graph Embeddings and Convolutional Neural-Network) model. It's a two-stage approach:

  1. Representation Learning:
    • Embedding Layer: This layer projects the social graph's features into a lower-dimensional real-number space. It assigns a vector to each actor, where the similarity between vectors reflects the correlation between actors.
  2. Convolutional Neural Network (ConvNet) Layer: This layer takes the embeddings from the first layer and further extracts inherent features from the social graph. This is done using an edge sampling strategy to capture the context of actors within their neighborhood.
  3. Classification/Prediction Layer:
    • A supervised learning layer, typically a multi-layer perceptron (MLP), takes the features extracted by the ConvNet. This layer is trained to map the extracted features to the desired output (e.g., node labels for classification, or link existence for prediction).

The training process is largely semi-supervised for representation learning and supervised for the final classification/prediction.

Datasets

The paper mentions evaluating RLVECN against various benchmarks for node classification and link prediction tasks. However, specific dataset names or details about newly introduced benchmarks are not provided in the text.

Key Insights and Ideas

  • Hybrid Approach Synergy: Combining knowledge graph embeddings with ConvNets proves effective. The embedding layer captures initial actor relationships, while the ConvNet refines these by considering local network structures.
  • Edge Sampling for Context: The use of an edge sampling strategy is crucial for capturing the contextual information of actors within their local network neighborhoods, leading to richer representations.
  • Dual Representation Learning: The two distinct representation-learning layers (embedding and ConvNet) work in tandem to extract and learn sufficient features, which is highlighted as a key reason for RLVECN's superior performance.
  • Addressing Both Classification and Prediction: The framework is designed to be versatile, tackling both node classification (identifying clusters) and link prediction (understanding tie formation) within social networks.

Key Results

The paper states that RLVECN demonstrates "surpassing performance" compared to several state-of-the-art methodologies and Representation Learning (RL) models. It specifically mentions outperforming baselines for node-classification tasks, including DeepWalk, GCN, LINE, Node2Vec, and SDNE. For link prediction, it implicitly suggests superiority over ComplEx, ConvKB, DistMult, and HolE. The authors attribute this success to RLVECN's dual RL layers.

Limitations and Future Work

  • Dataset Specificity: While baselines are mentioned, the paper does not detail the specific datasets used for evaluation, making it difficult to assess the generalizability of the results.
  • Computational Cost: The paper does not discuss the computational complexity or scalability of RLVECN, which can be a significant factor for large-scale social networks.
  • Future Work: Not explicitly discussed in the provided text.

Summary updated

BibTeX
@inproceedings{molokwu2020social,
  author = {Bonaventure C. Molokwu and Ziad Kobti},
  title = {{Social Network Analysis using RLVECN: Representation Learning via Knowledge-Graph Embeddings and Convolutional Neural-Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/739},
}
Figure