クラスタリング手法Dynamic Coalescence Modelに基づく新しい離散クラスタリング手法

摘要

本論文では場の概念に基づく新しい離散クラスタリング手法を提案する．特徴空間内の点を質点と仮定し，その位置を格子点に量子化した後，点をその加速度に応じて高々隣接する格子点に動かすことを繰り返し，1個の格子点に集めた点の集合をクラスタとする．点の加速度が，全ての点の質量を1と仮定する場合にその点が他の点から受ける引力を表す引力関数と各格子点にある点の質量を表す点分布関数との畳み込み和で表され，その点から見た点分布の概略を表しかつその点を中心とし半径が一定の超球内にある他の点から受ける引力で決まることを明らかにする．2次元2クラス混合正規，半円弧分布で表されるパターン分布において，据案手法で所属クラスが正しく推測されるパターンの割合は連続的なDynamiccd傍αnCeMethodの135．0％と高く，計算時間は0．0747％と極めて短い．We propose a new discrete clustering method based on concept of the field. It is supposed that a point in a feature space is a mass point, and location of a point is quantized to a lattice point. Points are moved to a lattice point of at most its adjacent iteratively, and it is speculated that a set of points gathered in a lattice point is a cluster as a class to which the points belong to. It is clarified that acceleration of a point is expressed by convolution of attractive force function that expresses the force of the point subject to other point under the supposition of mass of each point is 1. It is also clarified that the acceleration is determined by attractive force from points that expresses outline of point distribution observed from the location of the point and also in a sphere with certain radius of which center is the location of the point. On pattern distribution that is expressed by mixture normal and semicircular distributions with two classes in two dimension, rate of points of which classes are accurately speculated is 30.5% greater than that of continuous method Dynamic Coalescence Method, and duration is 0.0747%.