Personal agents and APIs can enable personalized applications that facilitate the daily processes that a person uses to gather, classify, persist and retrieve information in daily activities. Generalpurpose APIs will cover part of the functionality, but others need higher degrees of customization. This thesis focus on systems where to count with common models is difficult, and investigate a bottom-up approach where personal agents expose hypermedia APIs and copy, transform and combine API features from others, as a mechanism to improve their means of interaction.
We refer to a personal dataspace as the grouping of all kind of digital information
that is gathered and stored by an individual over long periods of time, under
that person’s control (but not necessarily exclusively so)[
In the last years, several decentralized social networks have been proposed [
Furthermore, the agents might interact with other agents through APIs, requiring (i) syntactical interoperability, i.e., establish the basic symbols to exchange a representation without losing structure or information, and (ii) semantic interoperability, i.e., reducing interpretation gaps to trigger acceptable behaviors through interactions.
Interoperability is traditionally achieved when peers can refer to a common reference model or ontology but become increasingly difficult in decentralized scenarios. The representations agents exchange, depends on how their schemes are aligned, but in some cases, direct mappings to common models are not available. In these setups, an agent would prefer to gather, infer or compose ad-hoc mappings during interaction with their peers, to interpret, transform or negotiate structured content. Also, in a growing ecosystem of applications that are separated from storage, we can expect a significant number of models changing over time. In this context, provenance is desired to help to interpret the original intended meaning in the long term.
This work focuses on cases where to reach global interoperability is difficult, and investigate an approach where APIs are decomposed in features to be exchanged in a network utilizing clone, mutation, divergence and convergence mechanisms while persisting provenance trails of the API evolution. The utility of such approach is three-fold: (i) to promote API evolution through agents that collect API functionality from others, (ii) to generate useful artifacts for service discovery (iii) and to incrementally build mappings between the underlying models of each agent, to improve their means of interaction. 2
There is a growing interest in APIs that adapt to the data of a user, as they show
significant improvements over fixed designs according to a specific user profile.
One use case is health-care applications, where care plans need to be tailored
to medical patient profiles and coordinated across several care providers. A
centralized approach has drawbacks since is difficult to acquire and continuously
update the user models required to provide sophisticated tailoring [
To create the proposed system I will use existing research from two fields, which I categorize as Hypermedia-driven APIs and Bottom-up Interoperability Elicitation. 3.1
Extensive research has been done regarding the discovery, composition, and
orchestration of Semantic Web Services [
Since this study focuses on cases where to reach global interoperability is difficult,
of particular interest are the approaches that do not aim to reach it. One is the
Frisco report [
The contrast of classical data integration techniques such as ontology matching
[
The following questions (Q) and related hypothesis (H) investigate how hypermedia APIs can be applied in the context of personal agents interacting in the Web, and how can provenance trails be used for service discovery in a decentralized setup. This work refers to provenance trails as a record of how API features are copied, transformed and exchanged between agents and corresponds to sequences of operators that transform API features into others and can be collected by following links.
– (Q1) Can agents expose resources and service descriptions from their underlying dataspaces as Linked Data to be discoverable by others? – (Q2) How can trails be represented, so they are suitable to characterize API feature change? – (Q3) To what extent can information derived from provenance trails influence the relevance and precision of API feature search and discovery? – (H1) An agent can discover and clone services from other agents by relying solely on Linked Data principles. – (H2) It is possible to gather provenance trails and discover Semantic Web
Services of high relevance without the need of a central service registry. 5
The model is used in the GPS4IC agent, which is part of the GPS4IntegratedCare2 project to provide RESTful APIs to be used by integrated healthcare applications. The agent is a nodeJS application that interprets JSON-LD3 files that contain 2 https://www.imec-int.com/en/what-we-offer/research-portfolio/gps4integratedcare 3 https://json-ld.org/spec/latest/json-ld sets of API features that use patient data, Domain-specific knowledge rules, and other external data sources to build at APIs at runtime for the health-care domain.
The GPS4IC agent organizes data in private and public dataspaces, namely (1) patient dataspaces, (2) service definition dataspaces and (3) knowledge dataspaces, which are logical constructs that group resources in a directory structure for simple storage, navigation, and access. Within each directory node, one can define inference operations O with links to input data and queries. Data can be N3 files, responses from a SPARQL endpoint, invocations to other operations and so on. An API is generated at runtime and exposed at a path P using a set of O that resides in a path equivalent to P of the directory. All the core functionality such as generating dynamic workflows for the patient, detecting medication conflicts and domain-specific knowledge like the definition of diseases, etc. do not reside in the Agent.
Functionality exists as resources that can be linked, cloned and shared using Change Operators such as: (i) add/remove inference feature (ii) import feature (iii) add/remove a resource. Operations often need definitions of resources sets which in this case, are matches of URI template patterns4 applied to the corresponding directory structure. Agents do not maintain explicit internal schemas but commit to (believe in) a set of health-care knowledge resources. 6
6.1
This approach is framed in Web environments, specifically in the domain of Web APIs elicitation and intelligent agent support.
From an API perspective, this work refers to agents that can play client and
server roles indistinguishably, consuming and also exposing APIs to others.
Each API consists of an interface implemented by components called features,
a logical construct that can afford interaction or functionality to be re-used
[
This approach requires features to be transferred between agents, although this
can be done in several ways, I already had results with an agent design where
service descriptors are interpreted to expose APIs by using only N3 rules, RDF
data, and queries. In this way, the API functionality can be directly exchanged
through the Web using simple dereferenciation or linking. API responses also
expose links to functional descriptions [
Each time a feature changes within an agent or is cloned from another agent, new data is registered into what I call provenance trails, that refer to the history of feature change and exchange between agents.
To characterize and keep track of feature change, I adopt an approach used in
ontology engineering based on layered change operators [
In the network, agents might have joined semantic communities and peers
providing similar contents (semantic neighborhoods) [
Data about how API features change can be useful to learn about the dynamics
of agents and their underlying models. For example, operator sequences allow
generating quantitative indicators of feature articulation efforts, the speed of
change or reusability. The data can then be analyzed by tools and techniques
for mining software repositories, such as dependency network analysis [
The hypothesis will be tested through two research aspects: 1. To evaluate the effects of provenance trails via simulations of agents that exchange features through hypermedia in a decentralized environment. To be measured through API evolution metrics and benchmarked against current semantic web service discovery approaches. 2. The impact that keeping such trails has for users while customizing their agents assisted by service discovery mechanisms without a central service registry.
H1 will be validated if instruments of (1) demonstrate automatic feature discovery and exchange using hypermedia APIs and collecting trails. H2 will be validated if (2) demonstrate fewer API articulation efforts by part of the users. Several tasks still need to be done to perform the studies and evaluate results, answering the research questions.
– To develop vocabularies or ontologies to represent (i) feature exposition, to expose and transfer functional descriptions and their dependencies, (iii) feature change operators, with representative instances organized in layers, (ii) and feature exchange. (Q1),(Q2).
– An extensive trade-off analysis of the approach (Q3). 8
In this work, we present an approach where a personal agent collects functionality
from other agents, improving their means of interaction while keeping model
autonomy. It also introduces provenance trails to support automated API
evolution when possible, and if not, diminish the costs that API composition
has for humans. I still have considerable work regarding the implementation
of ontologies and experiments, which I believe will contribute to state of the
art. Being this approach specific, I understand it as complementary to existing
approaches such as automatic schema mapping algorithms [
I would like to thank Miel Vander Sande for his useful advice.