Information Technology Reference
In-Depth Information
In recent years, NNSC [5] regarded as a kind of neural code performs well
in HCDT. Given a potentially large set of input patterns, NNSC, using a con-
stant parameter so-called ”sparseness”, can find a few representative patterns
combined in the right non-negative proportions and then reproduce the original
input patterns.
Most methods above have good performance on decomposing multivariate
data and detecting hidden components, but they still cant implement an ecient
hierarchical expressiveness.
However, it is a consensus and common behavior that people always tend
to find the inner relationship among the objects and then arrange them in a
hierarchical structure, exp. in the patent system and library system. Thus, its
desired to develop an automatic approach to learn the topics or clusters of the
data, additionally the subtopics or sub-clusters in multiple layers.
In this paper, we propose a non-negative mutative-sparseness coding towards
hierarchical representation. The whole organization of the data has a tree-like
structure with the entire collection situated at the root level. Then, the subse-
quent levels of the tree function as the further expanded analysis of the data.
Particularly for a sublevel, the sparseness of each data for each basis is adjusted
according to the corresponding hidden components of the data in upper level.
And the natural data non-negativity is also preserved. Hence, keeping the ac-
curate reconstruction, there are connections among the data sparseness in each
layer. Our experimental evaluations show that the proposed method performs
excellent in document clustering and owns great eciency of hierarchical repre-
sentation.
Once again, the research motivation comes mainly from exploring bases, inter-
action and hierarchy of dataset. The analysis adding mutative sparseness serves
to show a clearer data framework and different significance of each data, mean-
while bring forth ideas to information dissecting and representing.
2 Non-negative Mutative Sparseness Coding (NMSC)
The proposed algorithm provides a useful framework to implement the hierar-
chical representation of the data. There are three main steps as follows:
2.1 Learning the Basis of the Root Level
Generally, data tend to be combined with a set of sparse response on the basic
elements in neural system, text information processing models, etc. Meanwhile,
according to the objective law of cognition, these data and their elements should
be non-negative, which has been also widely recognized.
Whats more, at the beginning of analyzing a mono-level structure or the
root level in a hierarchical structure, each data shares equal importance and
contribution. So the sparseness of all the data on a specific basis should be the
same.
Search WWH ::
Custom Search