Resumen de Supporting decentralized collaborative processes in the digital transformation
David Sánchez Charles
Crowdsourcing, the art of involving several individuals in the decentralized execution of business activities, is being positioned as the replacement of outsourcing, as it allows organization to reach a capable workforce whenever it is necessary for the business. Nevertheless, adoption among industry is still low, as the technology is yet not mature and, in particular, it is difficult to monitor the execution of the business activities in a crowdsourcing platform. In this thesis, we advance towards creating better monitoring tools for crowdsourcing processes and a mechanism for modelling the worker's behavior.
Formalizing the work to be done in a process is the first step for improving the overall efficiency and quality of problem resolution. Still, there is a lack of mechanisms for defining business processes capable of adapting to the needs of the crowd. Therefore, we start this thesis by introducing a graphical modelling language for describing decentralized collaborative processes. The focus of this work is to allow the definition of complex worker requirements, as well as provide a quick overview and assessment of the implemented quality assurance mechanisms. In a longer-term vision, having well-defined processes will help in making more predictable the performance of any crowdsourcing project.
For those cases in which there is no formal process and the crowd can self-organize how they execute the business activities, we have also made the first steps for designing a method capable of discovering processes by analyzing the factual work done in the platform. Assuming that all steps recorded by the platform have some textual description of the work done, we propose to use novel natural language processing tools for generating groups of similar activities and, hence, enabling later analytics and insights, such as a process discovery for understanding, monitoring, or simply formalize the underlying crowd-process.
As for modelling the worker's behavior, we started by studying a particular crowdsourced process pattern that enables the platform to rank users based on their performance. The novelty of such prototype relies on the role of the reviewer, played by skilled individuals on the platform, that acts as reviewers of the translations done by in-training translators. The feedback provided by the reviewers is later reused for deciding if an in-training translator should be promoted to the reviewer role.
Unfortunately, there is no clear way of extrapolating the previous user evaluation to other processes. In this thesis, we propose to let the platform monitor the actions performed by individuals in order to create a profile of their behavior. We assume that those actions can be thougth as events that can be later processed by a discovery method, summarizing such actions in the form of a process model. Apart from the fitness of the resulting process models, precision is a key quality metric of these behavioral profiles. Low-precision models are more likely to describe the behavior of several users, reducing the insights obtained by analyzing or comparing process models. In particular, repetition of activities -- very often due to the human nature -- is one of the key trace characteristic that reduces precision of models discovered with most process mining techniques as we highlight, and palliate, during this chapter.
We also propose a new similarity metric between process models, enabling platforms to compare users based on the similarity of the user profiles. In particular, we have applied this similarity metric with an industrial dataset compromising several workers with access to a source code repository, and it turns out that their role in the organization is partially seen in how they access such source code repository.
