THESIS
2023
1 online resource (x, 107 pages) : illustrations (some color)
Abstract
This study aims to understand the dynamic characteristics of recommender systems by
introducing users’ decision-making component in the systems and quantifying the dependence
on recommendations through stochastic models. Researchers constantly strive to
develop algorithms for recommender systems that can optimize the recommendations for
individual users. However, over-reliance on such algorithms might lead to sub-optimal
results, thereby diminishing the effectiveness of the recommender systems. While this
phenomenon is often observed, its mechanism and behaviors remain an open research
question. In this thesis, such a systematic study is performed on two classic recommendation
scenarios, namely the path recommendation in traffic networks and friend recommendation
in social networks. Fo...[
Read more ]
This study aims to understand the dynamic characteristics of recommender systems by
introducing users’ decision-making component in the systems and quantifying the dependence
on recommendations through stochastic models. Researchers constantly strive to
develop algorithms for recommender systems that can optimize the recommendations for
individual users. However, over-reliance on such algorithms might lead to sub-optimal
results, thereby diminishing the effectiveness of the recommender systems. While this
phenomenon is often observed, its mechanism and behaviors remain an open research
question. In this thesis, such a systematic study is performed on two classic recommendation
scenarios, namely the path recommendation in traffic networks and friend recommendation
in social networks. For each problem, a suitable agent-based model is selected
to effectively reflect its respective characteristics. In particular, a stochastic approach is
introduced to infuse the human decision-making component to the model. Using these
two cases, the impact that user’s degree of reliance on recommendation algorithm brings
to the system efficiency is studied and the findings generalized. In the traffic network case study, the classical cellular automata model is used to simulate urban traffic, where
the shortest path algorithm is used to find the optimum path. To add stochasticity into
the model, flexibility and generosity factors are introduced to quantify the dependence of
drivers on the shortest route recommended by the algorithm. They are also utilized to
observe the impact of driver’s choice on the efficiency of the overall transportation network
at different levels. In the social network case study, the combination of a classical
graph-based network model and the common neighbor algorithm are used to investigate
how these algorithms impact the choice of friends of individual users in social networks
and the overall structure of such networks. As before, a layer of stochasticity is added to
the model to evaluate how friend-selection mechanisms affect the final network structure.
This is achieved by introducing two probability parameters, namely algorithmic reliance
and attribute awareness. The results presented in this thesis show that the excessive reliance
on the recommendation algorithm can lead to observable degradation of system
performance. Furthermore, a trend akin to a phase change is also observed where at the
beginning, the effectiveness of the recommendation gradually increases with the increase
of user’s reliance on the algorithm. However, once it passes the critical point, a sharp decrease
is observed. Such behaviors can only be revealed by adding the human-factor dimension
into their classical explanatory models, which is demonstrated by the stochastic
approach developed in this thesis. The results provide insight into the inherent interaction
between recommendation algorithms and users’ decision-making processes, beyond
what existing models can reveal. In this thesis, the rationales behind this phenomenon are
discussed using the theoretical framework and recommended operational mechanisms to
avoid sub-optimal performances are presented.
Post a Comment