Skip to Main content Skip to Navigation

Vers une modélisation plus réaliste de la diffusion d'innovations à l'aide de la simulation multi-agents

Abstract : Diffusion of innovations is defined as the communication process by which a new practice, idea or product spreads within a social system. For example, an innovation could be the adoption of a new contraceptive solution, practicing water boiling or buying new ergonomic handphones. Institutions and firms search to understand and predict whether their innovations will diffuse, in what extend, or fail. They are also interested in decreasing communication costs and attach an increasing interest in viral marketing. Agent-based modeling is a powerfull approach for studying such a collective process. However, existing models oversimplify adopters' beliefs and information content, despite the evidence of their central role in diffusion. Indeed, beliefs and information content explain diffusion failure due to misunderstanding of institutional messages, and they constitute decisionnal variables of institutions. In this thesis, we challenge the feasability and utility of a more realistic representation of beliefs. Such a model should nevertheless remain simple enough to be parametered, manipulated and validated against field data. We propose to represent adopters' beliefs as associative networks and define in what ways these beliefs are manipulated. We propose and test an interview protocol, which enables to parameterize and validate the model against real diffusion cases. We also define a communication protocol describing proactive communication and active information search. Associative networks improve the model's descriptivity, as illustrated by the simulations of event marketing, or of diffusion failure caused by information-beliefs inadequacy. The exploration of the model's parameter space highlights the importance of active search in information dynamics, and leads to question several expectations from viral marketing. Our agent-based model, like most others, reveals a strong sensitivity to the interaction network (also called ``social network''). As a consequence, the lack of descriptive large-scale networks forbids rigourous validation of agent-based models and limits their prediction potential. The wide use of trivial networks generators is justified by the untractability of data collection for large-scale networks. Hence, we highlight the existence of scattered statistics and qualitative knowledge about these networks, which may be used to constraint interaction graphs. We propose to formalize these statistics using Bayesian networks, and define an algorithm inspired by social choice theories. To illustrate this generator functionnality, we create a large interaction network for rural Kenya in which family links, colleagues and friendships links are generated in a spatialized environement. This generator opens the way to an intuitive and low-cost generation of interaction networks, probably leading to more descriptive simulations of innovation diffusion and other social phenomena.
Complete list of metadatas
Contributor : Samuel Thiriot <>
Submitted on : Friday, September 17, 2010 - 6:05:35 PM
Last modification on : Friday, January 8, 2021 - 5:50:03 PM
Long-term archiving on: : Tuesday, October 23, 2012 - 4:20:19 PM



  • HAL Id : tel-00519016, version 1


Samuel Thiriot. Vers une modélisation plus réaliste de la diffusion d'innovations à l'aide de la simulation multi-agents. Modélisation et simulation. Université Pierre et Marie Curie - Paris VI, 2009. Français. ⟨tel-00519016⟩



Record views


Files downloads