仅根据Proximity数据构建向量空间模型的方法

An Approach to Constructing Vector Space Models from Proximity Data Alone

在实际应用中，许多研究对象都是抽象的，难以用某种特征向量的形式表示，这使得许多成熟的数据挖掘和机器学习方法难以被采用。不过，通常可将其转化成一个Proximity数据矩阵，使得矩阵中的元素表示两个对象间某种“比较”关系。针对该问题，本文提出仅根据Proximity数据矩阵利用多维尺度分析法（MDS）将研究对象进行向量化表示，即构建了一种向量空间模型。最后，对汉语科技词系统中的词语进行了聚类分析，结果表明，向量空间模型构建后再聚类的结果明显优于直接针对Proximity数据进行聚类分析的结果，从而验证了该方法的可行性和有效性。

In real-world applications, there are lots and lots of abstract research objects that cannot be represented as feature vectors, therefore many mature data mining and machine learning methods cannot be utilized directly. Nevertheless, it is often not difficult to obtain a proximity matrix, which indicates some ＂comparison＂ relationship between objects. To overcome this problem, this study puts forward to obtain corresponding feature vectors for objects only from proximity data matrix by multidimensional scaling （MDS）, that is, to construct a vector space model. Finally, the clustering analysis is conducted on words from Chinese Scientific ＆ Technical Vocabulary System. Experimental results show that the clustering performance from vector space model construction is obviously better than that from clustering analysis directly on proximity data, which verifies the feasibility and efficiency of our approach.