<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-2002">
  <Title>Intelligent patent analysis through the use of a neural network: experiment of multi-viewpoint analysis with the MultiSOM model</Title>
  <Section position="3" start_page="0" end_page="2" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> The digital maps are not only tools of visualization. They also represent an analysis tool. Appropriate display of class points can give the analyst an insight that it is impossible to get from reading tables of output or simple summary statistics. For some tasks, appropriate visualization is the only tool needed to solve a problem or confirm a hypothesis, even though we do not usually think of maps as a kind of analysis, as for patent analysis. There is many ways to create digital maps. The one we consider here is based on Artificial Neural Networks (ANNs). ANNs are a useful class of models consisting of layers of nodes. The power of ANNs is derived from their learning capability defined as a change in the weight matrix (W), which represents the strength of the links among nodes. Moreover, both their relationships with multivariate data analysis and their non-linear capabilities represent added-values for classing and mapping. The Kohonen self-organizing map (SOM) model is a specific kind of ANN which implements in only one step the tasks of classing and mapping a data set. In the SOM case, the learning is competitive and unsupervised and the approach gives central attention to spatial order in the classing of data. The purpose is to compress information by forming reduced representations of the most relevant features, without loss of information about their interrelationships. The main advantages of the SOM model are its robustness and its very good illustrative power. Conversely, the fact that original model he his only able to deal with one classification of the data at a time might be considered as a serious bottleneck for exploiting it for fine mining tasks.</Paragraph>
    <Paragraph position="1"> In this article we shall be dealing with an innovation that was firstly introduced for the information retrieval purposes [13]. It has also been successfully tested for multimedia mining and browsing tasks, exploiting both the multi-map concept and the synergy between images and text on the same maps [14]. It is the multi-map extension of the Kohonen SOM algorithm. This will be from now signified by the name of MultiSOM. As we shall notice, the MultiSOM introduces the concepts of viewpoints and dynamics into the information analysis concept with its multi-map displays and its inter-map communication process. The dynamic information exchange between maps can be exploited by an analyst in order to perform cooperative deduction between several different analyzes that have been performed on the same data. The principal intent of this article is to propose the MultiSOM model as an ANN implementation of the information analysis concept.</Paragraph>
    <Paragraph position="2"> We will mainly focuses on the study of the contribution of the viewpoint's oriented data analysis proposed by the MultiSOM model as compared to the global analysis proposed by the other models. An attempt will be made to define a protocol and to design a platform for this comparison. As soon as the MultiSOM model can be used either in a global way or in a viewpoint-oriented way, it will be used as the reference model for our comparison. The section 2 of the article presents the Kohonen self-organizing maps (SOM) and their main applications in mapping of science and technology. Sections 3 deals with MultiSOM, the multi-map innovation of the SOM algorithm.</Paragraph>
    <Paragraph position="3"> The context of the experiment on the oil engineering patents and the preprocessing of these latter will be described in the section 4. The Section 5 describes the protocol of comparison which has been set up along with its results. The conclusions are finally exposed.</Paragraph>
    <Paragraph position="4"> 2. The self-organizing map (SOM) The basic principle of the SOM is that our knowledge organization at higher levels is created during learning by algorithms that promote self-organization in an spatial order (see [5],[6],[7],[8],[9],[10],[11],[12],[28]). Thus, the architecture form of the SOM network is based on the understanding that the representation of data features might assume the form of a self-organizing feature map that is geometrically organized as a grid or lattice. In the pure form, the SOM defines an &amp;quot;elastic net&amp;quot; of points (parameter, reference, or codebook vectors) that are fitted to the input data space to approximate its density function in an ordered way. The algorithm takes thus a set of N-dimensional objects as input and maps them onto nodes of a two-dimensional grid, resulting in an orderly feature map [9]. A layer of two-dimensional array of competitive output nodes is used to form the feature map. The lattice type of array can be defined to be square, rectangular, hexagonal, or even irregular. Every input is connected to every output node via a variable connection weight. It is the self-organizing property. The SOM belongs to the category of the unsupervised competitive learning networks [4],[11],[13]. It is called competitive learning because there is a set of nodes that compete with one another to become active. To this category belongs also the adaptive resonance theory (ART) model of Grossberg and Carpenter, as well as the self-organizing maps discussed in this paper. In the SOM, the competitive learning means also that a number of nodes is comparing the same input data with their internal parameters, and the node with the best match (say, &amp;quot;winner&amp;quot;) is then tuning itself to that input, in addition the best matching node activates its topographical neighbors in the network to take part in tuning to the same input. More a node is distant from the winning node the learning is weaker. It is also called unsupervised learning because no information concerning the correct classes is provided to the network during its training. Like any unsupervised clustering method, the SOM can be used to find classes in the input data, and to identify an unknown data vector with one of the classes. Moreover, the SOM represents the results of its classing process in an ordered two-dimensional space (R  ). A mapping from a high-dimensional data space R n onto a two dimensional lattice of nodes is thus defined. Such a mapping can effectively be used to visualize metric ordering relations of input data. As Kohonen [9] says: &amp;quot;The main applications of the SOM are in the visualization of complex data in a two dimensional display, and creation of abstractions like in many classing techniques.&amp;quot; The SOM algorithm is presented in details in ([2],[9],[12],[13],[19]). It consists of two basic procedures: (1) selecting a winning node and (2) updating weights of the winning node and its neighboring nodes. This preliminary learning phase is not straightforward process [9]. It necessitates several different learning steps, single map evaluations, and comparisons between a lot of generated maps in order to find at least a reliable map, at most an optimal one [13],[32].</Paragraph>
    <Paragraph position="6"> (t)} be the input vector selected at time t, and W</Paragraph>
    <Paragraph position="8"> (t)} the weights for node k at time t.</Paragraph>
    <Paragraph position="9"> The smallest of the Euclidean distances ||x(t) -</Paragraph>
    <Paragraph position="11"> After the winning node s thus selected, the weights of s and the weights of the nodes in a defined neighborhood (for example all nodes within a square or a cycle around the winning node) are adjusted so that similar input patterns are more likely to select this node again. This is achieved through the following computation:</Paragraph>
    <Paragraph position="13"> for 1 [?] i [?] N where a(t) is a gain term (0 [?] a(t) [?] 1) that decreases in time and converges to 0, and h(t) is the neighborhood function.</Paragraph>
    <Paragraph position="14"> Once the SOM algorithm is achieved, the data can be set to the nodes of the map. For each input data vector, the winning node is selected according to the algorithm first step presented above, and the data are affected to this selected node.</Paragraph>
    <Paragraph position="15"> In the quantitative studies of science, the Kohonen self-organizing maps have been successfully used for mapping scientific journal networks [2], and also author co-citation data [33]. Maps have been also successfully used for several other applications in the general area of data analysis like for classifying meeting output [30], for classing socio-economic data [32] and for documentary database contents mapping and browsing [13],14]. Kaski et al. have implemented a specific adaptation of SOM, named WEBSOM, for the analysis of important document collections [6]. WEBSOM main characteristic is to include strategies for reducing the dimension of the entry data descriptions by using random projection techniques applied on word histograms extracted from the document contents. WEBSOM method has been tested for patents abstract analysis [7].</Paragraph>
    <Paragraph position="16"> Nevertheless, as this method only manages such an analysis in a global way, it can only provide the analyst with general overview of the topics covered by the patents along with their interactions. A more exhaustive description of all the SOM applications might be found in [32].</Paragraph>
    <Paragraph position="17"> After the map building, the main characteristics of the classes resulting from the topographical classification process have to be highlighted to the analyst in order to provide him an overview (i.e. a global summary) of the analysis results. This task is difficult because the profiles of the obtained classes are mostly complex weighted combination of indexes extracted from the data. We have previously observed that single extraction strategy like the one proposed by [17] could cause shortcomings or mistakes in the interpretation of the database contents. The first set of solutions we proposed for solving this problem, like class labeling and zoning strategies or generalization mechanisms, are presented in [14]. Figure 3 of section 4 presents a map resulting from these processes.</Paragraph>
    <Paragraph position="18"> In all the following sections, we will consider that the classification process deals with electronic documents associated with their description in the form of index vectors. Classes will be represented by node vectors or class profile; each component of the vectors being the coordinate of a document index element (keyword). The list of the input data, which are the documents affected to the node, will represent the &amp;quot;class members&amp;quot; profile. The conceptual mean of the classes will be below called a topic. This semantic information is supplied by the classified keywords and documents.</Paragraph>
  </Section>
class="xml-element"></Paper>