基于复杂网络的社会化标签语义相似度分析

Complex Network Based Semantic Similarity Measure for Social Tagging Systems

针对社会化标签系统所对应的标签共现复杂网络,引入标签语义相似度权值和抽象权值算子,建立了标签语义相似度计算模型。相比基于"用户-对象-标签"三元组的统计性计算公式或基于复杂网络拓扑结构的节点相似性计算公式,本模型可以在标签语义相似度计算中将标签标注行为的统计特性与复杂网络的拓扑特性有机地结合起来,形成一个具有良好数学性质的形式化系统。仿照模糊逻辑中T范数、S范数给出了抽象权值算子的具体化实现,形成具体化算子簇,可以通过调节参数(如参数h和阶数l)形成不同类型或不同全局性的具体化算子。设计实验方案,利用复杂网络链路预测的AUC指标、Precision指标对典型算子及算子簇进行了综合分析。分析结果表明,这些具体化算子同时具有"语义补充"、"语义破坏"两种相反作用,在算子阶数较低(如2≤l≤5)时能明显提高标签语义相似度计算的准确性,在社会化标签系统的高精确性个性化推荐算法设计中具有应用价值。

Regarding to the complex network composed of the vast amount of tags in social tagging systems in Internet with their co-occurrences, the weights as the statistical semantic similarity of tag-tag edges and two abstract operators for weights computation were introduced, and a model of tag semantic similarity measurement is established. Comparing with traditional ＂users-items-tags＂ tripartite graph based statistic measures or network topology focused nodes similarity measures, this model provides a well defined formal system, which explicitly addresses both the statistical influential factors and the topological influential factors in computation of tag semantic similarities. A cluster of concrete implementations of the abstract operators are devised, which have similar format with T norms and S norms in fuzzy logics. In this cluster, concrete operators of different types or addressing different scopes of network topological factors are configurated with particular parameters （e.g., parameter h and order/）. By incorporating the AUC index and precision index in link prediction of complex network, an experiment is conducted to analyze the effectiveness and feasibility of these concrete operators. The experimental results show that these concrete operators introduce the effects of ＂semantic complementation＂ as well as the effects of ＂semantic destruction＂ when they are applied, but lower ordered calculations （e.g., 2~〈 l~ 5 in the model） with these operators are helpful for precise analysis of tag semantic similarities, therefore they are useful in devising high accurate tag-aware recommendation algorithms for social tagging systems.