Students’ Research Works – Spring 2017: Recommenders (PeWe.Rec)

Personalized Recommendation of TV Program

Ondrej Čerman
master study, supervised by Mária Bieliková

Abstract. This thesis deals with personalized recommendation of live television broadcasting. Because of very high number of TV channels broadcasting high number of television programs, TV viewers have trouble with searching and choosing TV programs to watch. Personalized recommendation should help these people. We have developed application for Android smartphones that shows live television program. This application also collects data from users, which is used for personalized recommendation.

A recommendation of live TV broadcasting usually includes many challenges including: different user preferences in different times, a lot of brand new programs in TV broadcasting, noisy input data. Another challenge we have is to deal with fact, that we do not have any information that is automatically collected from TV devices. We are relying only on data that user share via our application. We are trying to solve these problems by changing recommendation techniques to count with these situations and presenting the results.

to the top | to the main

Personalized recommendation taking into account visual impacts

Martin Černák
master study, supervised by Michal Kompan

Abstract. Recommender system is important part of the web nowadays. Amount of information on the web is just unsearchable and to find useful informations and keep track about interesting topic is too time consuming. Well placed and selected recommendations can improve user experience which often transforms into the revenue.

Lists of results generated by recommenders are often accompanied with graphical elements such as images. These elements and their attributes in some domains significantly influence user preferences. Despite state-of-the-art approaches are focused on processing textual content of items, or observing and modeling behavioral relationships and interactions between users, they ignore graphical elements.

In our work we aim at proposing novel method for recommendation which takes into account the influence of graphical elements and analyze its viability.
We decided to use this method to recommend hotels. This domain have some specific problems like not enough data for individual users but also there are plenty of images for individual hotels. We believe that using our method we can ease problem with data.

to the top | to the main

User Modelling Using Visual Stimuli

Peter Gašpar
doctoral study, supervised by Mária Bieliková

Abstract. Information overload on the Web has been a well-known problem in many domains. Recommender systems have been trying to solve this issue by reducing the number of items presented to user, usually in a personalized way. However, a good recommender needs user model that can describe user preferences and use them to better understand user’s behaviour.

In our research we are exploring the domain of recommender systems. There have been many approaches based either on the users’ or items’ similarity. However, these is also a need to take into account many other inputs that can influence user’s preference, e.g. visual appearance of the content they consume.

In our work we incorporate visual features extracted from the images into the recommender and compare their performance to the basic content-based approaches.

to the top | to the main

Web services consumer behavior analysis for improving recommendation

Veronika Gondová
master study, supervised by Mária Bieliková

Abstract. Web provides the environment for various services. Visitors of these services (consumer of services) use them to find the most appropriate ones. A number of alternatives that are available together with other factors (e.g. time) cause that the customer can not choose the best one. Personalized recommendation based on data from user behavior can reduce this problem. However, the data itself do not have to carry the sufficient information to generate recommendation that follows the interests of user. User model has a major impact on the accuracy of recommendation. Model may include – finding similar users, segmentation of users and also prediction of user behavior in specific situations.

In our work we focus on the analysis of user behavior in services such as e-commerce or e-bank. Our aim is to reveal the key characteristics of user that are important for effective recommendation. It may be personality traits, interests of user, motivation, skills with web navigation and also short-term characteristics such as emotion. Our method will be verified by different types of recommendation.

to the top | to the main

Scalable personalized recommendation for one-to-one marketing

Mário Hunka
master study, supervised by Michal Kompan

Abstract. Every e-commerce has certain channels whereby they have to contact and somehow attract the customers. Common practice is to send daily/weekly/monthly set of deals that meets some criterion (e.g. the bestsellers, the cheapest etc.) and send out emails. This often ends up with lot of unread emails or unsubscribing from mailing by customers.

In our research, we are creating a model that fit user preferences and is able to recommend personalized deals to a customer. Based on user history and his activities combined with certain features of deals we can predict the most suitable deals for each user. We are also finding proper frequency of sending mails for each user to avoid making mailing routine.

We will evaluate our recommender by its scalability – ability to process bigger amount of data with minimal impact on its accuracy.

to the top | to the main

Supporting Online Student Communities by Utilization of Questions and Answers Archives

Adrián Huňa
master study, supervised by Ivan Srba

Abstract. Community Question Answering (CQA) sites (e.g. StackOverflow, Yahoo! Answers) provide an effective way to get answers for complex or too specific questions that cannot be understood well by search engines. Other users of CQA sites answer these questions by using their own knowledge and experience. During the recent years, CQA sites gained popularity and currently hold valuable knowledge in their archives of questions and answers.

Massive Open Online Courses (MOOC) emerged as an affordable way to access quality education. The courses are often provided by world-class universities and companies. Students are also provided with discussion forums, where they can, for example, get answers to their questions or can use them as a place to clear confusions about course’s material. We can observe similarities between knowledge search in MOOC discussion forums and CQA sites.

n our work, we focus on supporting online student communities in MOOCs by utilization of question retrieval for automatic question answering. We incorporate specifics of educational setting into our matching model between new question and archive of resolved questions. There is one big difference between questions in general CQA systems and MOOCs. CQA systems allow users to search in an archive of resolved questions, while in MOOCs that is not possible. Because of that, questions in MOOCs tend to repeat when a course is being rerun. In our model, we match a new question with questions from the archive by utilizing TF-IDF cosine similarity and also other features, and if a good match for the new question is found, we eliminate the time needed for manual question answering by automatically providing an answer from the archive.

to the top | to the main

Recommendation of New Questions in Online Student Communities

Jakub Mačina
master study, supervised by Ivan Srba

Abstract. Existing Community Question Answering(CQA) systems and discussion boards used by Massive Open Online Courses (MOOCs) have difficulties with an increasing proportion of unanswered questions. We propose an approach for a recommendation of new questions designed for CQA systems in educational settings to support collaboration among students. Majority of the existing methods tend to overload only few experts by recommending most of questions to them. We proposed two innovations in which makes our question routing method suitable for an educational domain. Firstly, our approach explicitly models user’s willingness to answer the question and combines it with an expertise of a user. Secondly, we utilized non-QA data from online course for question routing such as students’ grades, course activity and knowledge prerequisites for a new question.

Experimental results demonstrated that additional features specific to learning environments help in predictions of new question answerers. Moreover, online A/B experiment conducted on one MOOC course with more than 4500 students at EdX platform showed total impact on the online community. Question recommendation increased average number of contributions and portion of answered questions while preserving answer quality and time to answer.

to the top | to the main

User segmentation for personalization of newsletters in CQA systems

Matúš Salát
master study, supervised by Ivan Srba

Abstract. CQA systems are a common resource of knowledge sharing and obtaining information. Every day many questions and answers are created. One way how to inform about news in any CQA system is a newsletter. Nevertheless, in the most popular CQA systems (such as Stack Overflow), existing newsletters are not personalized yet and show random generated content for every user. As the result, newsletters are not very popular because their content has no information value for the most of users.

We propose to create user segments that group users with common interests and activity in a CQA system. For each group, we will generate a newsletter containing specific content dependent on group attributes. In addition, we want to modify one or several specific sections of newsletter by means of personalization based on user previous activity. Personalization of newsletter can help with providing the right content to new as well as stable users. We want to evaluate the proposed method on data from the selected existing CQA system.

to the top | to the main

Improving Diversity and Freshness of Newsletters in CQA Systems

Martin Šrank
master study, supervised by Ivan Srba

Abstract. Newsletters represent a standard way to inform users of online communities about new or interesting content. Their importance is even greater in online communities producing large amount of user-created data, such as CQA systems. Nevertheless, many popular CQA systems, such as Stack Overflow, only offer generic newsletters, which do not take into account users’ interests or diversity of the recommended content.

We plan to analyze existing approaches in personalized content recommendation in CQA systems and design a method for automatic creation of personalized newsletters for individual users. We want to focus on improving the diversity and freshness of the recommended content as a way to prevent filter bubbles. To avoid the problem of a cold start, we propose combining collaborative filtering and content-based filtering approaches when creating a newsletter for low-activity users.

to the top | to the main