logo SBA


Digital archive of theses discussed at the University of Pisa


Thesis etd-05112019-180434

Thesis type
Tesi di dottorato di ricerca
Thesis title
Context-Aware Recommender Systems for Opportunistic Environments
Academic discipline
Course of study
tutor Gregori, Enrico
correlatore Delmastro, Franca
  • context-aware
  • context-aware recommender systems
  • internet of people
  • internet-of-people
  • iop
  • opportunistic environments
  • opportunistic networks
  • recommender systems
  • reti opportunistiche
  • sistemi di raccomandazione
Graduation session start date
Connected smart devices are now pervasive and they represent an important part of our daily lives.
The widespread diffusion of these devices greatly contributes to the rapid expansion and evolution of the Internet at its edges, where personal mobile devices follow the behaviour of their human users and can interact with other smart objects located in the surroundings.
This phenomenon paves the way towards the emergence of a new and more human-centric Internet paradigm called Internet of People (IoP), in which humans and their personal mobile devices are not seen merely as end users of applications, but they become active elements of the whole network.

According to IoP, data transmission leverages on both the Internet core and direct communications among devices in proximity.
Indeed, mobile devices can opportunistically exploit both the mobility of their human users and their physical contacts to discover and share information among each other by using self-organising wireless networks, without relying on any centralised infrastructure.
Moreover, humans and their personal devices can leverage direct communications also to share their computational resources and to establish new physical and virtual relationships with other people and devices.
In this thesis, we generally refer to Opportunistic Environments as the combination of all the aforementioned characteristics of the envisioned future Internet, where mobile devices opportunistically exploit their smart features and the surrounding environment to provide personalised services to their human users.

One of the most critical challenges in the opportunistic environment is the identification of the relevant information for the user.
In fact, in this scenario, users can potentially access a massive quantity of data coming from either remote servers or devices in proximity.
Moreover, the actual utility of the discovered contents typically depends on the current situation in which the user is involved.
Traditional approaches for data dissemination in self-organising networks are based on publish/subscribe mechanisms, assuming that user's preferences are well-defined a priori in content categories and her device should collect all the contents related to these categories.
However, in the real world, users' interests are not static, but they change over time and they often depend on the situation in which the user is currently involved and the other people in the nearby.
In this scenario, mobile devices must be able to act as avatars of their respective human users by exploring congested cyber landscapes and filtering the available contents according to their preferences and contexts.

In this dissertation, we design and analyse a new human-centric approach to identify relevant contents for the users in opportunistic environments, based on the use of Context-Aware Recommender Systems (CARS).
CARS represent a specific type of Recommender Systems that exploits information related to the user context to provide more accurate and personalised recommendations, based on the situation in which the user is currently involved.
CARS for centralised architectures has been widely studied in the literature, and several solutions have been proposed during the last years to address different target domains by exploiting various context information.
On the other hand, their use in highly distributed environments is still in its infancy, and ad-hoc solutions that fit the unique characteristics of the opportunistic scenario are required.
The main difference between CARS for centralised and distributed environments relies on the knowledge they can use to provide recommendations to the user.
While centralised CARS are able to access a complete knowledge about both users and items to suggest, distributed CARS can rely only on the local knowledge of each device, represented by the available contents generated by the local user and those advertised by other devices in proximity through direct communications.
Moreover, in this scenario, CARS must be lightweight and efficient in terms of response time.
This is due to the fact that they must be executed on devices with limited resources and, more important, opportunistic contacts have limited and unpredictable duration, during which mobile devices can exchange their knowledge and the recommended contents.

Therefore, the development of CARS in opportunistic environments must be supported by additional components able to establish opportunistic communications among mobile devices in order to discover new contents in the environment, and to model and recognise the user context to further personalise the provided recommendations.
To this aim, we propose a middleware solution that implements all the necessary features to provide personalised recommendations and services to mobile users, where CARS represent the main component of the whole architecture.

This thesis provides several contributions in the research area. As a first step, we propose a novel CARS called PopuLarity ItEm-based Recommender System (PLIERS) that exploits user-defined tags as a practical way to characterise both the user's context and her items in order to provide contextual recommendations in a centralised scenario.
Then, we leverage on PLIERS's principles to define a new solution specifically designed for opportunistic environments, called Pervasive PLIERS (p-PLIERS).
It represents a general framework for identifying useful and interesting contents in highly distributed scenarios, relying only on the exchange of single devices knowledge during proximity contacts and through direct wireless communications.

To this aim, we propose Wi-Fi Direct Group Manager (WFD-GM), a novel context-aware networking protocol based on the Wi-Fi Direct standard to establish self-forming direct communications among mobile devices.
The main goal of WFD-GM is the autonomous configuration and management of Wi-Fi Direct communication groups, without relying on the manual intervention of the human user.
In this way, our protocol enables the creation of self-forming wireless networks on commercial mobile devices and allows p-PLIERS to evaluate contents discovered in the opportunistic environment.

Finally, we define a lightweight approach to model and infer the user context directly on mobile devices, without the need of relying on remote computational resources (e.g., cloud architecture).
Our approach leverages on machine learning techniques to recognise the user’s context by combining the sensing capabilities of mobile devices with information related to the user's activities and social interactions, both in the physical and virtual worlds.
In addition, a sensing framework called ContextKit has been developed to monitor context data from mobile devices and collect real-world datasets to define and evaluate new CARS for opportunistic environments, and to support the development of new context-aware applications.