Users searching for scienti c information are confronted by a \hidden face" of searchable space: their own selection of items, which could help map their navigation on recorded search \routes", is not open to consultation, and remains either concealed or even unavailable. At best, users bene t from recommendations but, as data on their own choices are not shared, networked information for global navigation remains nebulous. This position paper tackles the following research question: How could users searching for scienti c information bene t from each other's search sessions? Our answer comprises two steps. First, using examples from our own data, we look at the characteristics of user behavior, approached here mainly via structures of collaborative personalized search. The second step proposes a mapping of recorded search sessions: for similar queries, search sessions are modeled by sets of typical user/item pairs in networks we call \eco-systems" of queries. These eco-systems connect search sessions and are open to navigation for users. Final output is threefold: 1) visualization of interconnection between search sessions, 2) localization of the search session that best suits a user's needs, 3) navigation on alternative \routes" between search sessions selecting alternative paths to answers. The conclusion raises questions about typical uses of this new kind of a documentary object, exploiting functionalities originally adapted from a bipartite graph.
Scienti c and technical information (STI) has been recently quali ed as a \complex system" that has to improve its accuracy and its digital organization: these opinions, which appear in the research agenda (The National Academies of Science and Medicine 2017) of the National Academies of Sciences of the United States, underpin this position paper which raises questions about \searchability" of STI.
\Searchable" means \capable of being computationally searched", within an
independent community of search, open to serendipity (Conrad and Moeller,
2017), and navigation in an unknown space, even if we could now consider that:
\Web search is governed by a uni ed hidden space, and each involved element
such as query and URL has its inborn position" ([
The main issue tackled by this position paper is to approach search sessions for any query as \communities" of information retrieval, which are together confronted by the \hidden face" of their common searchable space. Users of search sessions never receive a record of other users' generated contents. There is no mapping of the various results of search sessions, showing clusters of user selections of items: these data, even anonymized, are not open to consultation, and remain either concealed or even unavailable. The lack of these kinds of data signi cantly a ects STI searches, where a single keyword could give access to a very wide range of structured discussions and interpretations, each delivered with their own speci c \version" and \vocabulary" or eld of experiment. Because of this limitation on alternative documentary selections, it remains di cult to identify roots of controversies: accurate interpretation of views, downloads, citations, is uncertain, and searching can sometimes appear as a lonely walk in a forest of hazy homonyms.
However, current research barely addresses the issue, as only \few approaches
take advantage of searches performed previously by users" ([
Our answer is twofold. First, we look at the variety of search behaviors and the need for a clearer approach to networks of search sessions: for similar queries, search sessions could di er signi cantly in their selections of items according to users' variable needs, which are still hardly modeled. Second, we propose a mapping of search sessions with the help of a bipartite graph that we developed for navigation inside what we call the \eco-system" of answers to a query. The nal point of this article is to review issues of e ciency relating to this new documentary object, modeling alternative searches, and helping to select, among recorded options, the answer that best ts a query. 2
Part 1. Search session and users' behaviors
STI search strategies are revealed by behaviors of the searches performed by
users. These behaviors could be identi ed from structured \relations between
the topics and the use of documents" ([
We will look at the characterization of user behavior in two areas: rst is the community behavior of searches, which appears to di er among disciplines; second, we review elements of individual behaviors which appear to vary with wide-ranging reference chasing purposes. Both seem to interact. 2.1
To start with examples, di erences in STI use behaviors among disciplines could
be considered via the large-scale survey of research data management carried
out at the CNRS (Centre National de la Recherche Scienti que) in 2014 ([
Another type of community behavior appears in a further recent national survey (COPIST): also at the CNRS, we carried out a survey of STI management at the national scale for all research organizations. Results include answers from 106 research and higher education institutions and events, overall showing a strong desire to share STI digital practices and fairly large di erences between current uses and management of groups of institutions. Detailed results on bibliometric uses and strategies are available:2 they reveal signi cant di erences in users' practices at the various institutions, together with their desire to share uses and data.3 2.2
It is established that emerging results create the risk of a \cold start" biased
answer to a query: new content is di cult to retrieve as it has, by de nition,
not yet been produced elsewhere. Conversely, a versatile answer could be seen
as fruitful when trying to nd out items which could be identi ed together
as \typical" and \original". It can also be the case that the attempted search
1
http://www.cnrs.fr/dist/z-outils/documents/Enqu%C3%AAte%20DU%20%20DIST%20mars%202015.pdf
2 http://www.cnrs.fr/dist/z-outils/documents/copist-premiers-resultats.pdf, p. 55
and further
3 Ibid, pp. 11 and 19
result is bound to unpredictable instrumental bases. [
Search sessions can thus have numerous goals, be grounded in variable
motivations, and di er along with variable discovery methods ([
Part 2: Networking search sessions: a \Query
Eco-System" 3.1
Interactive IR proposes networking of search sessions in various type of
frameworks like a \collaborative query management system " for \search and browse
interaction" ([
In this context, the \query ecosystem" presented here takes an innovative
approach to personalization techniques, aiming at an \interactive information
retrieval process" ([
In a \query ecosystem", interaction of a set of search sessions corresponds
to speci c IR needs in given STI contexts: an example is when research
infrastructures using the same kind of analysis source (X Rays, Neutrons, etc.) for a
large variety of experiments, each with their own practices in the same science,
produce comparable search sessions as \versions" of the same scienti c object. In
these contexts, the main goal of a query eco-system is to \re-rank results based
on the user's intention" by formalizing an analyzed interaction between similar
queries and retrieved results, and thus \lower the cognitive burden" of search by
mapping search sessions on a documentary topic of common interest ([
To build up representation of search sessions, we use graphs which have
long been used for a large variety of tasks related to knowledge representation.
Bipartite graph characteristics ([
Let us consider that our goal is to record \query routes" for a given keyword, and
that these routes, which of course could di er, are designed to be compared ([
Let us then write:
Qn = f (Nn; Kn) where Qn is a number of queries produced by users of a browser. Let us also consider that for each query Q, we can record a number of users N and a number of items (URL, documents, articles) K recording the search for a keyword at the same time among the same corpus of items: we will call this group a community of users of the same query \route" in a search session .
For positioning any distinct route, we could express the limits of its system
of \typical" search sessions ([
Q1 = f (Nmax;Kmin or
Q1 = f (Nmin;Kmax Or, in a general form: [max; min Qn = fN; K [max; min 3.3
Let us now note the coe cient of increase of N and the coe cient of increase of K when Q varies by one unit, that is to say, when a new additional query on the same keywords is recorded (for instance: check users interested in items of a domain before and after publication of a famous article). The above-mentioned coe cients of variation will allow measurement of the \stability" or instability of the search session's content, according to variations in coe cients of the increase or decrease of quantities n and k between two queries of the same \route". We will then record whether or not users and/or items have varied in a correlative way, between queries Q1 and Q2.
We assume here that the data structure of any query could vary from one to the other, and that this variation could be expressed by an observable and recordable index of change in the relation between number of users and number of items.
We could then write:
Qn = f [nmax; min kmax; min
Let us note that value \min" or \max" of n and k provide data on the measured quantity of these variables to \produce" Q, but let us also note that, from one unit Q to the other (from one query to the other), the coe cient of increase or of decrease between value \min" or \max" of n and k provides additional data on the quantity of these variables in the production of Q. These additional data provide a value of the \stability" or \instability" of the process of querying an additional search session, with a dynamic index of behavior. For any unit n and k in production of Q, the variation of quantities measured by and is critical information about Q. 3.4
We know that with regard to the slope of a set of queries, with the added mix of user and items that it measures, the slope tends to be stable when + 1 or ( + ) ! 1: in the latter case, transition between Qn 1 and Qn is steady because, as measured by + , variation of (n + k) is still of the same order as the variation of Q. There is no signi cant increase or decrease in the number of units (users and items) recorded between these two or more successive queries and corresponding search session content.
Conversely, there could be also an alternative situation in which ( + ) 1, i.e. that for each variation of one unit of Q, we have an unstable variation of characteristics of Q, indicated by an unstable variation of ( + ), which will increase or decrease abruptly: relations between users and items will change strongly. In the latter case, between units Qn 1 and Qn, the variation of the quantities x and y would be unstable, which means that the query's user and item pair will change signi cantly. This situation could indicate that a threshold has been reached in the e ectiveness of the query.
In this case, why continue to allocate (n; k) to Q if ( + ) tends toward minus in nity 1? This would mean that the combination (n; k) becomes about \performing", which is questionable when the time comes to decide on a new query Qn+1 with the same characteristics of n and k. With such characteristics for a query, it could be more interesting to change n or k than to reproduce the same quantities in the next query. More generally, for transition between Qn 1 and Qn, the choice of stability in n and k could be \justi ed" if we observe that stability prevails with ( + ) tending to 1 and \questionable" when instability prevails with + tending to 1.
On a long series of queries, there will be \learning" phenomena which will give meaning to the expression of limits in variations of the purposed investigation on a long set of queries covering large elds. In that perspective, it is possible to x arbitrary limits of variations to a given set of queries, with the \best" and the \worst", for instance: And
If we have selected these two opposite limits to queries belonging to an STI data search, we could write our limits with the following framework: Best = Nmax; ( + ) ! 1; Kmin Worst = Nmin; ( + )
1; Kmax \Best"
\Worst" Nmax Nmin ( + ) ! 1 ( + ) Kmin
Kmax 1 3.5
On the basis of the preceding expression, the two triplets could give shape to a formal graph-based presentation of paths existing between these two limits and possessing the characteristic of positioning all the non-contradictory solutions existing between these two limits with three parameters. The following bipartite Hamiltonian graph with connected nodes o ers that representation.
Nmαax
β
Kmin (α + β ) → 1 a • b • c • • d • e • f (α + β ) → ±∞
Nαmin
β Kmax keyword in a database. This has consequences, de ned by three characteristics, on GRAPHYP operating rules.
NODE (Query) TYPICAL DIRECTION OF SEARCH (Users, Items) a b c d e f
N max, K max, ( + ) !
1 Characteristic 1: GRAPHYP functions like a compass as it supplies all possible directions, and \locates" the recorded direction of a query Q. This set of possible \positions" during a search gives an overview of the proposed modeling of the searchable space for STI queries.
Characteristic 2: the second feature of GRAPHYP search modeling is that
it can be used to record and compare search \behaviors" in querying. It allows
users to learn from their past recorded behavior as well as from the recording
of other users of the same base, if data is made accessible to all users. The
anonymization solution is of course an important optional condition. [
In these two steps, the \compass" has the function of \orienting" and \locating" search attitudes as being recorded as individual dynamic behaviors in a search session.
Characteristic 3: GRAPHYP also allows the expression, by summation of individual attitudes, of the global trend of a community of users according to the characteristics of their uses in their own search sessions.
It could be possible to arbitrarily select one node, between a and e, to x an optimum point or optimum search session there, and measure the distance of a set of observed routes toward the selected node. Mapping of STI search routes becomes possible, like in air or sea routes. Cooperative or con icting routes could be identi ed by the GRAPHYP data structure.
GRAPHYP creates opportunity to reach output (STI data search session best results) while identifying alternative ways to reach it for a given input (query). In this way, users' learning pro les express the way they access items and which items could be thus compared and located locally and globally in their search sessions. GRAPHYP could be used as a kind of a browser for mapping uses of search sessions for an STI community.
Like any compass, GRAPHYP secures navigational parameters at any scale (see infra), owing to its built-in characteristics of information processing and transfer, its functions for positioning previously recorded routes, and its capacity to record and recommend navigation paths. GRAPHYP thus belongs to a new generation of learning systems combining intuitive choices and automatic path de nition at any operational scale. 3.7
Let us take all nodes (a; b; c; d; e; f ) as possible starting points for de nitions of \goals" of routes of search sessions at time T . Any node could be designed as a \goal" to be reached on a route built from the starting point of a given other node. We could then record \search session pro les" corresponding to circulation from one node to the other on a recorded route between search sessions. a b c 1 From the design of Figure 2 above, we can observe that a continuous circulation from node to node shows a common edge in all the pairs corresponding to any node: node a has a common edge with node e, b has a common edge with f , etc. From this standpoint, we could remark that it is consistently possible to have circular \travel" inside GRAPHYP, from edge to edge, while coming back to the starting node by the way of the complementary edge of that node. In this way, search sessions could be interconnected and explored. 3.8
As shown on Figure 2, node a could allow an exploration of GRAPHYP's other nodes which will represent six steps on one route, as designed here. Another sixstep route could be practiced from the same node a starting from its other edge. There are at least 12 possible steps which could be explored from any node in the Eulerian circuit, and GRAPHYP thus o ers 72 possibilities for identifying \search pro le" characteristics which could be modeled and recorded in an STI query.
Numerous characterized positions in exploration of possible search sessions could be identi ed in this way: it o ers users and their community detailed records of explored and unexplored ways of discovering information.
We could remark that these explorations of \possible" routes could be managed overall with any conventional probabilistic approach, with which the possibilistic approach could combine for any purposed uses. 3.9
We could also apply the design of Figure 2 to propose another signi cant feature of traceability for searchable STI space, namely the recording of search pro les, used to \read" characteristic steps of a navigation route in a database. This result could be produced when projecting on x; y axes as shown in Figure 3. b
Based on the design shown in Figure 3, we could, for any query applied to this grid of structured data (in our example: nx + 3 and nx + 5, ny2), identify the origin and thus the \kinship" of any search session located by GRAPHYP. We could then evaluate all diverging or converging solutions surrounding any observed node located on the grid. For any observed network, there exists an \induced network" which applies the property of \mutual reachability" of connected nodes in a network.
Mapping of search session routes using these ideas should help build strategies for discovering information.
Exploring neighbor routes between search sessions is an important feature of the searchable space, as it provides information on its depth and treewidth and, as far as we know, there are hardly any existing developments in mapping of this aspect of searchable STI space.
Reducing uncertainty on the size and direction of future routes to discovery could bene t from representations of courses which have been followed and/or abandoned (Figure 3). Based on our example, exploration of neighbor arriving routes could bene t from tree couples of edges which shape this node: a = (e; f ) b = (d; f ) c = (d; e) d = (b; c) c = (a; c) f = (a; b) e d e
Here we have the grid of nodes which are linked by a speci ed common edge, resulting from the position of each node on the grid. We could then propose the mapping of a grid of correlated neighbor routes. The grid observed on Figure 4, structured with the same nodes as the grid in Figure 3, o ers the option of projection that could be studied in a further work.
A nal property of the bipartite graph that gave birth to GRAPHYP is selfsimilarity, which is designed on the Figure 5 below.
The graph with nodes A, B, C, etc. is built on a larger dimension than the preceding one, and the addition of these \compasses" on a self-similar basis enables the building of information architectures of the same type for information processing at any scale, and the application to the GRAPHYP operating frame of basic addition, subtraction, multiplication and division operations. Scalability of the system could thus be established as a nal feature of descriptive functionalities of the system. It could then be used as a tool connecting approaches at various steps or in di erent scienti c domains.
Conclusions
The STI search community asks for accuracy and dynamic adaptation of
delivered items, as highlighted by reviews of \observational" approaches to
information retrieval ([
We propose mapping of recordable categories of typical search sessions, clustered in user-item groups, which, together, shape what we called the \query ecosystem". Navigation in this system is organized for complete \searchability" of a query, modeling alternative paths to answers. It raises several issues: rst, this new kind of a documentary object could help map \versioning" of search sessions, and thus add a new piece of semantic analysis for identi cation of alternative paths to answers, thus providing a \chronicle" of new incoming ideas through the corresponding modeling of searches. The status of sharing and privacy relating to this modeling system is a highly important issue.
Further work will seek to test accuracy and fast identi cation of versioning of search sessions; we will also check clustering at various scopes of fractal development of search session mapping, which would be useful for interactions between multipurpose searches (interdisciplinary, etc.) and the speci c geometry of items networking inside this type of data structure.