Study of Data Mining Algorithm in Social Network Analysis

Chang Zhang, Yanfeng Jin, Wei Jin, Yu LiuAdvances 2015 · DOI

Abstract

Some mining methods and algorithms will be introduced to describe the social network. Method based on similarity measurement and inductive logic programming are useful here to analyze social networks, moreover, some specific datasets are used to analyze the characters of social networks by Ucinet which is very powerful data mining software. The Zachary club social network's property can be found from analysis.

Introduction

Traditional data mining typically relies on "property-value" tables where each data point is treated as an independent vector. However, social networks are inherently relational; the most critical information often lies in the links between individuals rather than the attributes of the individuals themselves.

The challenge of "link mining"—the process of predicting missing connections or identifying subgroups when data points are fundamentally dependent on one another—is central to this study. Ignoring these relational dependencies leads to poor knowledge discovery, necessitating a move toward structural analysis.

Methods

The study highlights two primary algorithmic approaches for Social Network Analysis (SNA):

Inductive Logic Programming (ILP)

ILP represents a crossover between machine learning and logic programming. Unlike propositional logic, ILP uses first-order logic to represent relationships, allowing researchers to incorporate "background knowledge" into the learning process. This makes the resulting models more expressive and easier for humans to interpret, addressing the "black box" problem of many machine learning models.

Similarity Measurements

These metrics are used for link prediction, determining the likelihood of a future connection between two nodes. The authors compare three specific metrics:

  • Common Neighbors: A simple count of shared connections.
  • Jaccard Coefficient: A normalized measure of shared neighbors relative to the total number of neighbors.
  • Adamic/Adar Method: A sophisticated approach that weights shared neighbors based on their uniqueness. It assumes that sharing a "rare" friend is a stronger predictor of a relationship than sharing a "social butterfly."

Additionally, the study utilizes K-cores analysis to identify maximal subgroups and In-degree statistics to measure the "prestige" or power of individual actors.

Datasets and Visualization

The methods are evaluated using the Zachary Karate Club dataset, a benchmark in SNA mapping social ties between 34 members of a club that eventually split into two factions. Software tools like Ucinet and NetDraw are employed to visualize and partition these social groups.

Key Results

Using the Zachary dataset, the network was visualized using both standard and "circle graph" layouts. The circle layout makes high-density connection areas immediately apparent.

Visualization of the Zachary Karate Club network using circle and standard layouts to identify subgroups.

Subgroup Discovery

Through K-core analysis, the study identified four distinct subgroups within the club, represented by different colors and shapes in the visualization. This structural partitioning reflects the underlying social dynamics that led to the club's eventual split.

Influence Mapping

Quantitative analysis of "in-degree" (the number of incoming connections to a node) was used to identify powerful or prestigious actors. The descriptive statistics for these nodes provide a numerical basis for assessing influence.

Descriptive statistics for network nodes, used to calculate in-degree and identify prestigious actors.

The results showed that while high in-degree often correlates with prestige, it can also lead to "information overload" or "noise interference" for that specific actor.

Discussion and Conclusion

The core insight of this research is that in social networks, the link itself is a primary data object. SNA succeeds because it embraces relational dependence rather than assuming independence. ILP provides a transparent way to discover these relationships, constructing logic statements that researchers can easily understand.

While this study demonstrates the efficacy of ILP and similarity measures on classic datasets, future work involves scaling these relational methods to larger, more dynamic datasets and further exploring how positionality affects social probability in evolving networks.

Summary updated

Abstract

Some mining methods and algorithms will be introduced to describe the social network. Method based on similarity measurement and inductive logic programming are useful here to analyze social networks, moreover, some specific datasets are used to analyze the characters of social networks by Ucinet which is very powerful data mining software. The Zachary club social network's property can be found from analysis.
BibTeX
@inproceedings{zhang2015study,
  author = {Chang Zhang and Yanfeng Jin and Wei Jin and Yu Liu},
  title = {{Study of Data Mining Algorithm in Social Network Analysis}},
  booktitle = {Advances},
  year = {2015},
  doi = {10.2991/icmra-15.2015.231},
}
Figure