Exploiting Reinforcement Learning to Profile Users and Personalize Web Pages

Ferretti, Stefano; Mirri, Silvia; Prandi, Catia; Salomoni, Paola

doi:10.1109/compsacw.2014.45

Cited by 11 publications

(7 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A first evaluation was conducted via simulation [1]. The rationale was to understand how the presented approach behaves when different types of users (i.e.…”

Section: Simulationsmentioning

confidence: 99%

See 1 more Smart Citation

Automatic web content personalization through reinforcement learning

Ferretti

Mirri

Prandi

et al. 2016

Journal of Systems and Software

Self Cite

View full text Add to dashboard Cite

This paper deals with the automatic adaptation of Web contents. It is recognized that quite often users need some personalized adaptations to access Web contents. This is more evident when we focus on people with some accessibility needs. Based on the user profile, it is possible to transcode or modify contents (e.g., adapt text fonts) so as to meet the user preferences. The problem is that applying such a kind of transformations to the whole content might significantly alter Web pages that might become unreadable, hence making matters worse. We present a system that employs Web intelligence to perform automatic adaptations on single elements composing a Web page. A reinforcement learning algorithm is utilized to manage user profiles. We evaluate our system through simulation and a real assessment where elderly users where asked to use for a time period our system prototype. Results confirm the feasibility of the proposal.who are equipped with devices with different capabilities, such as tablets, smart-phones, and smart TVs. In this context, both readability and legibility are affected by different issues, such as Web text characteristics [4,5,6] and users' abilities [7,8,9]. This work presents a system, called ExTraS (EXperiential TRAnscoding System [1,10,11]), and thought to improve Web content legibility. The approach adapts some characteristics, in particular the text formatting ones (e.g. font size, font face, luminance contrasts), according to users' preferences and needs. ExTraS tracks users' behavior to learn and model their preferences and to automatically provide the best adaptation, tailored for each user, predicting his/her needs. Such an adaptation is based on a machine learning mechanism (Reinforcement Learning), exploiting the idea of reward/punishment. We have built a prototype system and we have evaluated it by means of a simulation assessment and tests with real users. Both of them confirm the viability of our proposal.The remainder of the paper is organized as follows. Section 2 describes background and related work, while Section 3 presents the architecture of our system, detailing how it profiles users, how it applies the machine learning mechanism and how it adapts the content. Section 4 introduces the prototype we have developed, showing some screenshots. Section 5 reports the simulation assessment and the tests with users we have performed, detailed the obtained results. Finally, Section 6 concludes the paper illustrating main findings and further work. Background and related workOur work takes into account several issues and has been based on many related work. This section aims to briefly describe the most significant ones, which are related to improving Web pages legibility with font adaptation, adaptation and personalization of digital and Web contents and services, and the use of machine learning techniques and algorithms to track and understand users' experiences and behaviors.According to the definitions in the literature, legibility is related to perceiving text by distinguishing e...

show abstract

“…A first evaluation was conducted via simulation [1]. The rationale was to understand how the presented approach behaves when different types of users (i.e.…”

Section: Simulationsmentioning

confidence: 99%

“…Recent advances in Web technologies have revealed that adaptation of contents improves the access to the Web [1,2,3]. Among the different types of adaptation, self-adaptation is a way to organize contents based on the specific needs of a given user.…”

Section: Introductionmentioning

confidence: 99%

Automatic web content personalization through reinforcement learning

Ferretti

Mirri

Prandi

et al. 2016

Journal of Systems and Software

Self Cite

View full text Add to dashboard Cite

show abstract

“…In [20,21,33], authors make lo cal changes to the web page, which has the effect of keeping the designer preferences on the unaltered parts. To our knowledge, the other approaches do not consider this aspect, and transcode the web page only on the basis of the user preferences.…”

Section: Existing Approachesmentioning

confidence: 99%

“…This user profile can be entered by the user via an interface, giving values for the char acteristics [37,34] or from visual tests [22]. In a more innovative approach [20,21,33], the user profile is learned from user actions and Q-learning. In this paper, we do not address the user profile acquisition, which we will consider in a future work.…”

Section: Existing Approachesmentioning

confidence: 99%

See 1 more Smart Citation

Reconciling user and designer preferences in adapting web pages for people with low vision

Bonavero

Huchard

Meynard

2015

Proceedings of the 12th International Web for All Conference

View full text Add to dashboard Cite

The web has become a major tool for communication, services and an outstanding source of knowledge. It has also grown in com plexity, and end-users may experience difficulties in reading and acquiring good understanding of some overly complex or poorly designed web pages. This observation is even more valid for people with visual disabilities. In this paper, we focus on people with low or weakening vision, for whom we propose to adapt web pages to their needs, while preserving the spirit of the original design. In this context, obtaining a web page adaptation in a very short time may be a difficult problem, because user and designer needs and prefer ences may contradict each other, and because there may be a large number of adaptation possibilities. Finding a relevant adaptation in a large search space can hardly be done by an algorithm which com putes and assesses all possible solutions, which brings us to con sider evolutionary algorithms. A characteristic of our problem is to consider a set of preferences, each being implemented by an evalu ation function. This optimization problem can be dealt with multiobjective genetic algorithms, including the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and its next version (NSGA-III). NSGA-III has been recently introduced to address many-objective optimization problems (having more that four objectives). We com pare NSGA-II and NSGA-III performances in the context of adapt ing web pages in accordance to a set of preferences. The compar ison is based on running time, number of generations and quality of computed adaptation (number of satisfied objectives). We also show the importance of several parameters including population size, crossover/mutation probability, and the opportunity to aggre gate objective functions. From the obtained results, we conclude that the approach is feasible and effective on realistic web pages, especially with NSGA-III.

show abstract