This paper presents an original approach based on a recent metric called feature maximization for developing accurate diachronic analysis tools. In such process, querying of bibliographic databases is firstly exploited to provide a thematic corpus of scientific publications covering a large time period. In a second step, two strategies based on contrast graphs generated by the use of feature maximization metric are proposed. The first one is based on the direct use contrast graphs who relates time periods and publication contents. The second strategy combines a preliminary step of clustering with the use of contrast graph generated by feature maximization applied on cluster contents to highlight the relation between topics represented in clusters as well as to embed them in a temporal path. Both techniques are parameter-free and knowledge agnostic. We illustrate the efficiency and the complementarity of the proposed technique by experimenting then on a dataset related to gerontology research extracted from the data collected by the ISTEX project, a project whose aims is to construct a general purpose and open access database of scientific documents.
The ISTEX1 project’s main objective is to provide the whole French higher education and research community with on-line access to retrospective collections of scientific literature in all disciplines.
On the basis of the initial platform services, we are currently working towards proposing new added-value services. One of our central concern is then to develop tools for highlighting the dynamics of the collection. Hence, the development of dynamic information analysis methods, like incremental clustering and novelty detection techniques, is becoming a central concern in a bunch of applications whose main goal is to deal with large volume of textual information whose content is varying over time, such as ISTEX. The purpose of the analysis and diachronic mapping is to track, for a given domain, changes in contexts (sub-themes) and the evolution of vocabularies and actors http://www.istex.fr/ that materialize these changes in terms of appearances, disappearances, divergence or convergence. The applications relate to very various and highly strategic domains, including web mining, technological and scientific survey.
In order to identify and analyze the emergence, or to detect changes in the data, we have previously proposed two different and complementary approaches: 1. Performing static classifications at different periods of time and analyzing changes between these periods (time-step approach or diachronic analysis); 2. Developing methods of classification that can directly track the changes: incremental clustering methods (incremental clustering) and novelty detection methods (incremental supervised classification).
The development of direct methods being still an ongoing research, we present
hereafter two original word-based methods relying on the first approach and using a metric
called feature maximization we have recently developed
We first present our feature maximization metrics and related contrast graph exploited throughout our approach. In a next step, we describe our experimental data and associated preprocessing. Lastly, we highlight our results with the two proposed approaches and our conclusion. 2
Feature maximization (F-max) is an unbiased cluster quality metrics that exploits the
properties of the data associated to each cluster without prior consideration of clusters
profiles. This metrics has been initially proposed in
Consider a partition which results from a clustering method applied to a dataset represented by a group of features . The feature F-measure ( ) of a feature associated with a cluster is defined as the harmonic mean of the feature recall ( ) and of the feature predominance ( ), which are themselves defined as follows: with ( ) = ( ) =
∈ ∈ ∈ ∈
′∈
,∈ ′ ( ) = 2 ( ( )× ( )) ( )+ ( ) (1) (2) (3) where represents the weight of the feature for the data and represents all the features present in the dataset associated with the cluster . Feature Predominance measures the ability of to describe cluster . In a complementary way, Feature Recall allows to characterize according to its ability to discriminate from other clusters.
Feature recall is a scale independent measure but feature predominance is not. We
have however shown experimentally in
In supervised context, feature maximization measure can be exploited to generate a
powerful feature selection process
A feature having some negative values can be separated in 2 different positive sub-features, the first one representing the positive part of original feature and the second one, its negative (5) (6) (7) =∪∈
() = ′∈ ′() | / | = ∈ () || where represents the subset of in which the feature occurs.
Finally, the set of all selected features is the subset of defined by: In other words, the features judged relevant for a given cluster are those whose representations are better than average in this cluster, and better than the average representation of all the features in the partition, in terms of Feature F-measure. Features which never respect the second condition in any cluster are discarded. 2.3
A specific concept of contrast ( ) can be defined to calculate the performance of a retained feature for a given cluster . It is an indicator value which is proportional to the ratio between the F-measure ( ) of a feature in the cluster and the average Fmeasure
of this feature for the whole partition.Contrast of a feature for a cluster is expressed as: ( ) = ( )⁄
( ) The active features of a cluster are those for which the contrast is greater than 1. Moreover, the higher the contrast of a feature for one cluster, the better its performance in describing the cluster content. 2.4
In the mathematical field of graph theory, a bipartite graph (or bigraph) is a graph whose
vertices can be divided into two disjoint and independent sets U and V such that every
edge connects a vertex in U to one in V. Contrast graphs are bipartite graphs based on
the relations between a set of features S and a set of labels L
Such kind of graphs have many interesting properties. First, they reduce the cognitive overload produced with classical graphs representation because of the associated feature selection process that reduces the number of potential connections. Second, they can be used to indirectly highlight relationships between labels, whenever features have contrasted interaction with several labels. Third, the combination of this approach with weighted force-directed model (Kobourov 2012) for graph representation permits altogether to highlight central or most influent labels of the L set and to easily identify the labels that are the most densely connected through associated features, these latter appearing in close neighborhood position in the graph.
We have proposed a first original use of contrast graph in the case of the analysis of
the transdisciplinarity between different research domains and time periods in
Our experimental data is a collection of 9801 scientific papers in English language related to gerontology domain published between 1995 and 2010, extracted from ISTEX database by INIST documentary engineers specialized in the medical domain. After a tokenization step, the keywords are extracted from the abstracts by a part-of-speech method developed in Python. However, the NLP treatments are minimalist (just morpho-lexical and syntactic) and we thus don’t use any vocabulary resource except a stop word dictionary.
We present on the following experimental section the two different approaches we have applied on the extracted metadata of our dataset, that are, the GRAFSEL approach which is a supervised approach based on a direct exploitation of the relations between document content extracted form keywords and document publication year to build a contrast graph and the CLUSTSEL approach that exploit a clustering process on the extracted document content and build up a contrast graph highlighting relation between cluster content with a further use of document publication years to highlight diachronic changes. 4 4.1
To clarify the principle of our supervised approach, we named GRAFSEL, we follow three steps that are schematically presented in figure 1: 1. The papers of the experimental dataset are assigned to a class that represents their publication year; 2. The papers being represented by their extracted keywords, we select keywords related to each year and compute the strength of the relations (i.e. the contrast) between selected keywords and years exploiting the feature maximization metric shortly described in section 3; 3. The last step is to build the graph highlighting the relationships between the years and the selected keywords by weighting the links of the graph with the formerly obtained contrast values.
Figure 2a shows the global result obtained by taking into account altogether the whole corpus and the fifteen considered years. All of the following graphs are obtained with a force-directed algorithm (Spring algorithm) taking into account as weight of a yearword edge the contrast of the word for the considered year.
We perform hereafter an attempt of interpretation of our results, mostly to illustrate the potential of the method. Such attempt is not substituting an in-deep expert validation that is planned in a near future.
As show in the figure 2, which is a zoom on the 2000s, we can observe in the years 2003-2006 the emergence of terms like "nurse”, “nursing”, “home care”, “medicare”, “family caregivers”, “home”, ”satisfaction”, ... that denote the development of home help services to maintain autonomy of elderly people.
In a complementary way, we can also exploit radar chart representation for each year in order to facilitate the interpretation of the results. In such representation each extracted word is a radius of a circle whose length depends on its contrast value whilst the description space is the same for all considered years allowing in such a way to detect changes.
Considering all the above-mentioned years, it is possible to detect invariant directions during each related period (figure 3). Furthermore, sudden changes of said directions suggest new scientific domains but also changes in professionals’ practices. As we have formerly observed in figures 2, the direction corresponding to “nursing”, “home”, “care” appears in the year 2002-2003, and terms as “risk”, “cancer”, ”mortality” and thereafter “exposure”, “stress” emerge in the years 2006. Additionally, if the first years were solely marked by the term “women”, the use of “people”, “personality” in years 2003-2006 indicates that a humanization of care might appear. On its own side, the term “mice” is often used until 2001 and disappear after: it might figure out an indicator of changes in experimentation protocols.
This short discussion shows that the use complementary modes of representation obviously enables a quick and simple view of the evolution of a thematic corpus through time. 4.2
The overall principle of our unsupervised approach, we named CLUSTSEL is presented in figure 4: 1. The papers being represented their extracted keywords, we cluster their descriptions using a clustering algorithm. Several experiments are achieved by varying the number of expected clusters and consequently obtaining clustering models of various sizes. 2. A clustering quality measurement based on feature maximization is exploited to find out the optimal model among all the ones that have been generated. 3. The clusters of the optimal model being represented by the keywords extracted from their associated papers, we select keywords related to each cluster and compute the strength of the relations (i.e. the contrast) between selected keywords and clusters exploiting the feature maximization metric shortly described in section 3; 4. The graph highlighting the relationships between the clusters (i.e. the topics) and the selected keywords is built by weighting its links with the formerly obtained contrast values; 5. Papers publication years are used to find out dominant period of the topics as well as to build up a diachronic chart figuring out the comparative influence of topics during each year.
For clustering, we exploit 2 different usual clustering methods, namely k-means
In the specific case of your experiment we propose to build up a contrast graph between a set of clusters representing the main research topics of the domain that have been extracted by the clustering process and the most contrasted features issued from the clusters’ descriptions. This approach that combines clustering and contrast graph in an original way highlighting the most connected topics.
In the case of our experiment we focus on one type of external labels that are papers’ publication years. Papers’ publication years are exploited to perform a diachronic analysis of the topics’ activity, highlighting the importance of each topic in each time period, either this activity is considered individually (see figure 7) or relatively to the other topics (see figure 6). As it is shown in the next part related to the analysis of the results, this approach helps to precisely understand the chronology of the research activity of a global research domain, like in our specific case.
In the context of our dataset we obtained an optimal model comprising 12 clusters (i.e. topics). The spatial distribution of 12 topics presented on the graph of figure 6 highlights clearly interpretable structure of the domain. Such graph provides generic although detailed representation of the domain-related research topics whilst highlighting the main relationships between the said topics whenever those topics appears as close neighbors on the graph. As an example, the ‘Homecare” topic is directly related to logically connected topics like “Physical performance” and “Health condition”. Similarly, “Menopause related problems”, an early topic, appears to be accurately related to “Cancer studies” and “Gene senescence” that figure out more recent and more general research topics.
On its own side, diachronic representations that are presented in figure 6-7 can then be used to get a better understanding of the gerontology development from early research (“Menopause related problems”, Hearing loss, “Age change” general studies) to more up to date research (“Home care”, “Risks factors”, “Physical performance”, “Health condition”, “Sociology of health”) that fits well with the global changes regarding health politics. In that context research on “Physical performance” becomes the most prominent in the recent years and seems thus clearly represent a central focus because of its obvious influence on the other recent research areas.
Last but not least, research on “Neurodegenerative diseases” (Alzheimer, Parkinson, …) seems to have split into two parts by generating a new specialized area related to “Memory performance”.
In such a way, results provide by CLUSTSEL approach appears clearly complementary to the ones obtained by the GRAFSEL approach. Hence, the two methods provide similar results although they highlight those ones with different levels of generality.
We have presented an original overall methodology for the diachronic analysis of large and heterogeneous text collections based on feature maximization and associated contrast graphs. The originality of that approach comes from the fact that the nodes of the obtained graphs result from the combination of a feature selection processes and a classification or a clustering process, depending on the chosen option. Thus, one main advantage of the approach is to avoid cognitive overload in the current case of management of high dimensional data. Another of its main advantage is to be altogether parameter-free and knowledge/language-agnostic. Our first experimental results obtained from the analysis of a realistic dataset extracted from the ISTEX bibliographic database are promising. Hence, they prove to be easily interpretable by an expert of the analyzed domain. Moreover, the supervised and unsupervised options of our approach provide similar results that can be considered of different levels of generality.
One further and encouraging domain of investigation would concern to check the scalability of our approach to the context of massive data analysis. 6
Acknowledgments
ISTEX receives assistance from the French state managed by the National Research Agency under the program "Future Investments" bearing the reference ANR-10-IDEX0004-12.