oijcalled
the stress
2
function:
S=
ij
d
ij
* dij
dij
ij
Note that the observed dissimilarities
oij do
not appear in this formula. Instead, the
discrepancy between the predicted distances
*
dij and the target distances dij are measured.
as a single point in multidimensional space
(e.g., Borg & Groenen, 1997) whereas in probabilistic MDS (MacKay, 1989), each object is represented as a probability distribution in multidimensional space. In understanding the mental representation of objects, this last approach is useful when representation of objects is assumed to be noisy (i.e., the presentation of the same object on every trial gives rise to different internal representations).
An important issue in MDS is choosing the number of dimensions for the scaling solution. A configuration with a high number of dimensions achieves very low stress values but cannot easily be comprehended by the human eye, and is apt to be determined more by noise than by the essential structure in the data. On the other hand, a solution with too few dimensions might not reveal enough of the structure in the data.
A well known method to select the dimensionality is the scree test (also known as the elbow test) where stress (or other lack of fit measure) is plotted against the dimensionality. Ideally, this choice is visually obvious from the “elbow” in the scree plot where after a certain number of dimensions, the stress is not reduced substantially. However, in many data-sets, stress decreases smoothly with increasing dimensionality making the choice of appropriate dimensionality very difficult with this method. In Figure 1b, the filled circles shows the scree plot for the face similarity dataset. Note that a slight elbow is present at two dimensions which suggests that a two dimensional configuration might be appropriate. A more salient indicator for the appropriate dimensionality can be obtained by cross-validation. The idea is to test how the configuration optimised to model the proximity data for one group of subjects can generalize to the proximity data of a different group of subjects. In Figure 1b, the open circles show the stress value for a second group of subjects with a clear rise in stress value after two dimensions while the stress continues to decrease for the first group of subjects. In this case, it seems reasonable to conclude that a two dimensional configuration is appropriate because it can best generalize to other subjects. Lee (2001) has explored other techniques to determine dimensionality based on balancing the trade-off between model fit and model complexity.
Another important issue is the interpretation of the scaling solution resulting from MDS procedures. If the proximity data were
Based on the current configuration of points, the target distances
*dij are found by
monotonic regression and represent the
distances that are monotonically related to the observed dissimilarities oij. Several iterative minimization algorithms exist to move the object points in a multidimensional space in order to minimize stress (see Borg & Groenen, 1997). In the face similarity example, Figure 1d displays what is known as the Shepard plot. It shows the relationship between predicted distances dij(for the two-dimensional scaling solution in Figure 1c) and observed dissimilarities as filled circles and can serve to understand what metric transformation would be appropriate to relate one to the other. The line in the plot shows the relationship between the target distances
*dij found by monotonic
regression and observed dissimilarities.
Kruskal stress essentially is a measure based on the sum of the squared deviations between the filled circles and the line along the abscissa.
Another distinction in MDS is between weighted MDS, replicated MDS and MDS on a single matrix (Young, & Hamer, 1994). In replicated MDS, several matrices of similarity data can be analysed simultaneously. The matrices are provided by different subjects or by a single subject tested at multiple times and a single scaling solution captures the similarity data of all matrices through separate metric or nonmetric relationships for each matrix. This approach can take individual differences in response bias into account. In weighted replicated MDS (e.g., INDSCAL, Carroll & Chang, 1970), the dimensions in the scaling solution can be weighted differently for each subject or subject replication to model differences in attention or sensitivity for the different dimensions.
Finally, there is the distinction between deterministic and probabilistic MDS. In deterministic MDS, each object is represented
多维标度法(MDS)的改进及其在国际市场研究中的应用多维标度法(MDS)的改进及其在国际市场研究中的应用隐藏>> ? 1995-2004 Tsinghua Tongfang Optical Disc Co.,...
本章仅介绍常用的古典多维标度法和权重多维 标度法。 第二节 古典多维标度法 (Classical MDS) 一 相似与距离的概念 二 古典多维标度分析的思想及方法三 度量...
23页 免费 MDS多维标度 5页 免费 市场分析方法LV11 多维尺... 暂无评价 9页 免费 多维尺度分析 14页 1下载券喜欢此文档的还喜欢...
第十章 多维标度法第一节 第二节 第三节 第四节 引言 古典多维标度法(Classical 古典多维标度法 MDS) 权重多维标度(WMDS) 权重多维标度 实例分析与计算实现...
多维标度_经济/市场_经管营销_专业资料。第十章第一节 引言 多维标度法 第二节 第三节第四节 古典多维标度法(Classical MDS) 权重多维标度(WMDS) 实例分析...
本章仅介绍常用的古典多维标度法和权重多维 标度法。 第二节 古典多维标度法 (Classical MDS) 一 相似与距离的概念 二 古典多维标度分析的思想及方法三 度量...
多维标度法汇报学习_教育学/心理学_人文社科_专业资料。多维标度法 ? 第一节 第二节 第三节 引言 古典多维标度法(Classical MDS) 权重多维标度(WMDS) ? ...
第二章 多维标度法简介多维标度法(multidimensional scaling MDS)是一种把高维降为低维,在低维空间展 示距离数据结构的多元数据分析技术,简称 MDS。多维标度法...
多维标度法的聚类分析_问题与解法_工学_高等教育_教育专区 暂无评价0人阅读0次下载举报文档 多维标度法的聚类分析_问题与解法_工学_高等教育_教育专区。MDS ...
满足非度量多 维标度法对相异性数据的要求。因此本文提出的RSSICMDS 算法, 直接使用 RSSI 值进行定位。 基于 RSSI 构建多维标度法中的相异性矩阵 P , 计算出...
热门文档