The integration of data elements scattered along di erent resources, with heterogeneous formats, can take advantage of an approach with progressive and lightweight steps, instead of pursuing costly upfront mappings. To support such approach, we de ned a multiscale-based dataspace architecture, called LinkedScales, which carries an integration process via graph-based transformations over a graph database. A series of scales in the dataspace systematizes an integration and enrichment chain of steps to leverage transformation processes, which incrementally go from raw representations towards ontology-like structures. However, how to record and keep track of the intermediary outcomes in the integration chain remains an open research challenge. This article proposes combining the concept of scales with trails { lightweight, scale-specialized semantic annotations to enable progressive integration towards a semantic representation. We conduct experiments involving organism-centric analysis in life science to show the bene ts of trails for transformation between scales.
Biologists often conduct organism-centric analysis in which organisms { i.e.,
species or taxonomic groups { are the central focus and data are collected and
integrated around them. In this context, biologists might compare organisms
in a systematic way and investigate conditions related to their hypotheses. In
this context, the construction of pro les [
This incremental process also produces three kinds of intermediary outcomes: semantic representations, knowledge discovery results and user feedback. They have operational purposes and drive transformation tasks in the production of content in the upper scales. However, there is no a systematic method to record and keep track of these intermediary outcomes. Operations built over them, like transformation and enrichment, can be better speci ed, managed and followed if they rely on a standard mechanism to document the outcomes.
In this article, we propose combining LinkedScales with the concept of trails.
Trails are \hints" represented as structured semantic annotations concerning
operational scale aspects { i.e., each scale emphasizes a particular step of the
integration chain, therefore each scale has distinct types of trails. Trails play
the role of metadata associated with portions of data [
We present a practical scenario of exploring trails with LinkedScales. We conducted an experimental analysis considering the integration and semantic enrichment of resources related to a particular organism pro le. In particular, trails are exploited to guide the process of linking content in the scales with external knowledge bases, like DBpedia, to better characterize the data conceptually. In order to show how trails can improve the linking process, in the rst step of our experimental procedure, we apply the transformation without the trails and we compare the results taking the trails into account afterward.
The remaining of this article is organized as follows: Section 2 presents foundations and related work. Section 3 describes the LinkedScales framework while Section 4 details the proposal of combining LinkedScales with trails. Section 5 presents the conclusion remarks. 2
Several data integration approaches have emerged, including federated databases,
schema integration and data warehouses [
Scenarios in which schemas often change and new data models must be
considered still lack an e cient solution. To this end, pay-as-you-go integration
approaches implement incremental integration based on progressive steps to
continuously re ne and improve the connections among sources. The proposal
of dataspaces aims at providing the bene ts of the classical data integration
approach but in a progressive fashion way [
This investigation explores the concept of trails in a pay-as-you-go integration
approach. Trails are keyword-based annotations that relate concepts to data
sources to be integrated. They are used for a gradual improvement of integration
among sources [
As an alternative for the one-step approach to de ne equivalences between
distinct data source elements, trails rely on services to support incremental
renements of mappings between schemas. Whenever the user feeds the system
with new \hints", it exploits them to improve the semantic equivalences
discovery. These \hints" are treated as a lightweight mechanism to de ne declarative
relationships between loosely integrated data sources [
LinkedScales [
The integration starts on the lowest scale, where all original data sources are ingested and transformed into graphs. Each subsequent scale from this point is a graph derived from the previous scale, taking advantage of the exibility of graphs to logically represent di erent structures along the scales. This model allows representing operations within and across the scales as transformation procedures in graphs. Fig. 2 presents the four scales aiming at going from the raw data sources (lower scales, containing more details about format and structure) to a conceptual scale (fewer details of format and structure, and focus on domain-speci c concepts). Scales are interconnected by an orthogonal graph, supporting traceability among them { i.e., it is possible to "track" sources/targets of transformations between scales.
The Physical Scale aims at representing the di erent data sources in their original physical format as a graph. The original raw data sources are transformed into a graph by an ingestion procedure. The Graph Translator reads several specialized formats { e.g., Excel, CSV, relational tables, XML { and converts them to an equivalent graph representation. The original structure, format and content of the underlying data sources are re ected in a graph.
The Logical Scale o ers a common view for data inside similar or equivalent logical models represented in the previous scale. Tables and hierarchical documents are examples of logical models present in the sources. In the previous scale, di erences might exist in the representation of a table within a PDF, a table from a spreadsheet and a table within an HTML le since they preserve speci cities of their formats.
The Description Scale emphasizes the content (e.g., labels of elements within an XML document or values in spreadsheet cells) and their relationships. Since models represent relations among data elements in di erent ways { e.g., a row in a table can represent data concerning the same entity while hierarchical relations in a document represent aggregations { the Description Scale reduces all logical models to a single uni ed one, to shift the focus towards the descriptive content. The uni ed model selected for this scale relies on the triple <resource, property, value>, which is usual in several meta-data standards as Resource Description Framework (RDF).
The highest scale refers to the Conceptual Scale . It integrates data from the lower scale at a semantic level by exploiting the content and relationships between nodes to discover and make explicit the semantics through ontologies. Entities are discovered, deduplicated and related to ontologies as instances of classes, or properties and their values. A \textual graph" of the previous scale becomes a graph containing interrelated entities and their properties/values, with explicit semantics supported by ontologies. 4
This work involves an enhancement of the LinkedScales framework to incorporate Trails as the driving component for transformations and provenance. It treats trails as scale-specialized operational semantic annotations, which indicates the role of data portions. Such hints are considered by scale transformation processes, incrementally conducting the re nement of the dataspace.
We conducted an experiment in the organism-centric scenario to investigate
how trails improve transformations between scales. We collected two
complementary sources coming from di erent scienti c publications { as illustrated in
Figure 3. The rst source is an XLS spreadsheet [
Both data sources are concerned with information about lung sh. While the rst data source contains morphological traits, behavioral aspects, habitat characteristics, etc. of several lung sh species, the second data source comprises a phylogenetic tree and a phenotypic description in a character/character state format. Even though both data sources are available for researchers, integrating such information conceptually by combining data of the same lung sh species remains a challenging laborious task.
In next sections we exploit the data sources as a running example to describe the use of di erent types of trails and their relationship with scales. Trails vary according to each scale, indicating relevant aspects of data that the transformation process takes into account during the production of an upper scale. We further describe roles of trails presenting the scale that they are inserted accompanied by the target transformation scale {e.g., a physical-logical trail refers to a trail to be inserted in the physical scale, impacting in the data production of the logical scale. 4.1
Lowest part of Figure 4 presents an excerpt of an XLS spreadsheet
containing information from a study of discrete characters change in the evolution of
lung sh (class Sarcopterygii ) [
Data is ingested into LinkedScales database as a graph. The middle part of Figure 4 shows partial representation of the ingestion result in the Physical Scale. Rows of nodes represent rows of the spreadsheet and their stream of cells. The graph focus on representing as much information as possible of the raw resource. Via such data, the logical organization can be inferred or derived { e.g., initial and boldly formatted cells usually are the table schema.
Physical-logical trails { pictured as colored hexagons in Figure 4 { are inserted to distinguish types of structures and their internal components. Figure 4 (middle part) illustrates how trails are used to conduct transformations from the physical to the logical scale. Trails associate structure-related roles to the nodes as: table (lst:table), row (lst:dataRow ) and the stream of cells corresponding to the schema (lst:schemaRow ). In the bootstrap phase of the dataspace, this type of trail is either automatically inferred by the ingestion module, according to the internal structures, or speci ed by the user. In short, the Physical-logical trails indicate how data is logically organized within the format-speci c graph representation of the resource.
Based on the associated physical-logical trails, a transformation process adopts a standard representation of structural elements of tables to logically represent the resource in the logical scale. Representing structures using a standard representation in the logical scale is particularly important, as it allows, for instance, reusing table-related algorithms to reach resources independently of formats. The Description Scale aims at shifting the focus to the content and their relationships, reducing logical models to an RDF-based structure. The bottom part of Figure 5 illustrates how logical-description trails are used to produce the description scale from the logical scale. At this point, trails indicate how structural elements should be organized as <resource, property, value> triples.
Figure 5 illustrates how trails (colored hexagons) are associated with structural elements on the Logical Scale, indicating, for instance, that rows (nodes r229 and r230) are resources, schema attributes (green nodes Genus, Species, Age) are properties and cells are values { e.g., < r230, Genus, N eoceratodus > and < r230, Species, f orsteri > are triples produced based on trails.
The transformation illustrated in Figure 5 can be represented by a rule which
matches a pattern (including speci c trails) as input and produces a transformed
output. Transformation patterns are already de ned in the LinkedScales model
[
Description-conceptual trails focus on reaching an expected perspective { e.g., organism pro les. Figure 5 (upper part) illustrates trails indicating the expected semantic interpretation of nodes in the description scale, making the semantic explicit by adopting speci c elements of ontologies. Such trails can be automatically discovered by the system in a semi-supervised process or be directly assigned.
The Conceptual Scale addresses fundamental semantic concerns by distinguishing entities and adopting controlled vocabularies to represent descriptive properties. Adjustments { removing or adding description-conceptual trails { made on previous scale are a way for handling the dynamicity of scenarios as organims-centric research in terms of testing di erent hypothesis.
Scales and trails play complementary roles in the progressive integration process. While a scale provides a homogeneous view of the lower layers, trails o er the proper clues for the transformation to the next scale. Consider, for example, the logical scale. It o ers a homogeneous view of data considering the logical model, i.e., all tables are represented in the same way, as well as, all trees. If on one hand, this is a powerful mechanism, as the heterogeneity of several table formats is hidden in a lower scale, enabling to reuse the same algorithms for several homogeneous tables, on the other hand, these algorithms need clues to interpret implicit di erences which will impact in the next scale.
Regarding the experiment of integrating both XLS and NeXML sources, at the bootstrap stage, after ingesting both data les and converting them to the Logical and Description scales in the graph, we used DBpedia (dbpedia.org) to automatically produce the trails that guided the production of the Conceptual scale. The experiment aimed at connecting portions of the data source with DBPedia resources (English release of October 2015), and therefore indirectly linking and enriching similar resources.
Our procedure searches in the DBpedia for the most similar resources of each node in the graph. The search method compares the input query against the DBpedia resource contents. This comparison uses the tf-idf measure and may return approximate/incorrect results like uncorrelated resources. To examine the bene ts brought by the trails in this transformation process, the next integration stage inserted trails associated to the nodes to give clues to our integration system about the nature of the nodes in the graph. In this experiment, two trails were considered associated with speci c procedures:
-Species related Trail: The user tags the nodes that represent Species, then the system can lter, via SPARQL queries, the resources returned by the bootstrap stage that are instances of taxonomy-related classes, according to the DBpedia ontology.
-Morphological related Trail: The user tags the nodes that represent
morphological characters. Such trails are used as input in an entity-quality
recognition algorithm [
Figure 6 depicts a portion of the conceptual scale with (right part) and without (left part) trails. Each node in the gure represents a speci c species from both data sources { the rst data source in green, and the second data source in red { and edges represent relationships concerning taxonomy and morphological traits (entity-quality pairs). When trails are associated to elements of the previous scale (description), the produced conceptual scale is semantically re ned according to the expected requirements in the organism pro les. Asigni cant part of the biological research work remains in an organism-centric perspective, which usually requires combining data regarding distinct aspects of organisms. In this article, we presented how our LinkedScales framework, based on the multiscale integration approach, can work aligned with trails as operational semantic annotations. Trails systematize intermediary outcomes, improving the transformation process and provenance records among the scales. Our experimental analysis demonstrated the overall potential bene ts of trails in LinkedScales to reach organism pro les. Future work involves conducting additional experimental evaluations to thoroughly examine the quality and scalability of data integration provided by the approach.
Acknowledgments Work partially nanced by FAPESP (#2014/14890-0), INCT in Web Science, CNPq (#141353/2015-5), FAPESP-PRONEX (eScience project).