Artificial and crowd intelligence based recommender system framework

  • Andreas Theodoros Lianos

    Student thesis: Doctoral Thesis


    Tackling information overloading in online shopping is a significant challenge for e-shops. The current solutions demand that consumers become educated before they are able to distinguish which products are good for them. This research suggests a novel recommender system, namely ACIBa, which attempts to do part of the market research on behalf of the consumer and offer only a handful of products as recommendations.
    The novelty of ACIBa is not just in its approach of limiting rather than widening the product alternatives, but also in that it bases its reasoning partially on artificial intelligence and partially on the collective intelligence of an arbitrary crowd of people. This work aimed to create a framework which provides the guidelines and basic tools to produce an ACIBa-like recommender system, as well as to create a live proof-of-concept system to showcase its use.
    ACIBa made contributions in various areas. First and foremost it introduced a new
    recommender system that tackles the issue of information overloading from a very different perspective. The individual subcomponents of ACIBa have also extended the knowledge of their respective fields. The classifier used to cover the need for artificial intelligence, led to the creation of a novel ensemble classification methodology that allows training from typically unusable training sources. The software developed to interface between ACIBa and the crowd has been published as open-source software, effectively allowing other developers to create their own crowdsourcing systems. Finally, a new methodology has been introduced (namely ANA), that allows for an easy way to reduce the number of answers required from the crowd before conclusions can be reached.
    Both these parts, i.e. the ensemble classifier and ANA, have undergone separate testing as standalone components. The ensemble classifier was shown to drastically outperform individual member classifiers. ANA has been found to consistently outperform the fixed number of answers distribution. An ACIBa based recommender system has been developed and demonstrated live to consumers. We used online questionnaires to gather feedback on the quality of results. Despite its infant stage, the response was generally positive and the system well received.
    Date of AwardJun 2016
    Original languageEnglish
    Awarding Institution
    • University of Portsmouth
    SupervisorLinda Yang (Supervisor), David Ndzi (Supervisor) & Branislav Vuksanovic (Supervisor)

    Cite this