Learning user preferences for participatory sensing tasks
Moz Preza, Carlos Josue
School of Science and Technology, MSc in Information & Communication Technology Systems
Nowadays, we experience the rapid proliferation of smartphones and wearable technol-ogies, with several embedded sensors that are capable to sense humidity, temperature, light, and proximity among others. This leads to a new pervasive service paradigm called participatory sensing (or, often interchangeably, mobile crowdsensing). Participatory sensing leverages the power of the crowds in that, end users can collect sensing data through their smart mobile devices, and contribute them to a central platform that pro-cesses them to build services out of them. Therefore, new applications and services can be generated out of the collective effort of many users, some of which might be totally infeasible or far more resource demanding otherwise. However, to ensure the sustained participation of end users in these services, it is important to provide them with proper incentives for their contributions. These can be either monetary or non-monetary; in any case, it is mandatory to tailor these incentives to the users’ individual preferences.
This dissertation’s objective is to explore new ways of inferring and subsequently mod-eling the individual user preferences in the context of participatory sensing applications. To this end, it draws on historical data from the interaction of the end users with the application and applies machine learning techniques to efficiently profile users. The ul-timate aim is to take advantage of these user profiles in the process of targeting incen-tives to them.
The project research work is divided in two phases. The first phase is concerned with the data collection process. In the absence of data from real crowdsensing applications, the relevant data for the proof-of-concept experimentation is collected through an online questionnaire. The questionnaire is addressed to students at IHU and also uploaded to special-purpose websites that crowdsource responses to such research efforts. A total of 132 user responses was collected during this first phase of the project.
Then, in the second phase of the research work we apply machine learning models (spe-cifically: logistic regression models), to infer the individual end user preferences and an-alyze the similarity/ diversity characterizing them. More specifically, the choice problem users face when presented with multiple offers for participatory sensing tasks is cap-tured as an instance of multi-attribute decision making problems with multiple alterna-tives and modeled through probabilistic multi-class logistic regression models. Cluster-ing and community detection techniques are used to identify similarity and diversity trends in these preferences, with the aim to specify “classes” of users with distinct pref-erence features.
The outcome of this work, the user models and the related accuracy scores together with the classes that segregate the individual user preferences, form a major part of a paper that will be submitted to IEEE Transaction on Mobile Computing. The paper builds on the models derived in this Dissertation to analytically optimize the offered incentives to participatory sensing users.