Motivation and Goals

User modelling is nowadays a highly important concept in the context of Web. As a process covering acquisition, processing and maintenance of a data about the user, his behaviour and preferences, it has a crucial influence on the task applied: a personalization, recommendation or various predictions on the site. Moreover, indirectly it increases the user experience within a website. Traditionally, the process of user modelling aims on long-term user activity and is focused mainly on the visited content. For this reason, there are modelled mainly preferences and long-term interests of the user.

Traditional modelling approaches capture quality information about the modelled user, they, however, need a sufficient amount of data about the user’s activity. For new users, they often suffer by a cold start problem. This is a problem mainly in domains with dynamically changing content and/or minority of loyal users, who visit the site regularly. Additionally, traditional modelling approaches capture mostly the stable long-term user behaviour and thus the updates and new directions in user preferences appear only slowly.

Modelling of user preferences offers also information only for selected types of tasks as mentioned personalised recommendation etc. To be able to capture additional traits about the user and his browsing habits in a website, there is need to look for new approaches and also types of a data.

In our thesis, we address a task of modelling changes in user behaviour. We focus primarily on the short-term changes, based on which we observe how the user behaviour evolve within a single session. By a session we understand a set of site visits (e.g., page views), realised by the same user, in which couples of consequent page visit time-stamps are distant no more than 30 minutes. Identification of behavioural changes between individual session actions is important to understand actual user behaviour. In this way we are able to react to user intent in an online time and to improve his experience almost intermediately.

Results

To evaluate characteristics of proposed user model, we chose the session end intent prediction task. This task depends mostly on short-term behavioural changes, as it aims at predicting if the user will end his session in the next step or not. By experimenting with a data from two domains with highly different characteristics, we were able to prove that the model is able to capture changes in the user behaviour in general.

The results for both domains showed that such a short-term behaviour is very difficult to recognise exactly. As a rule, the user behaviour before session end however changes. For this reason, the prediction of multiple actions in advance brought a significant improvement in the prediction precision. The high improvement was observed even when considering the only one action in advance. The reason is that based on proposed user model, there is possible to predict that the session end is near.

The slightly better results were reached for e-learning domain with more loyal and stable users who typically visit multiple pages per session. The reason is that in this case, the model data was fully used, the global part (attributes based on comparisons to sessions of all users) as well as the personal part (attributes based on comparisons to sessions of modelled user). In news domain, for high number of users, the personal model part did not contain enough data, so the prediction relied more on global part. Despite that, the prediction based on our proposed user model was able to find and learn important valuable attributes.

The user model was evaluated only by a one task, however, we believe that its potential is higher. It could be used for identification of session author (by looking for the most similar sessions in the past) or fraud detection (by identification of highly unusual user activity in comparison to user’s history). Another area for future improvements lies in usage of additional data sources allowing to model new types of data describing different behavioural traits. These could be data from eye trackers, which become nowadays very popular, or another biometric data describing user movements, or gestures.

Conclusions

In our thesis we introduced the task of modelling changes in user short-term behaviour. We proposed the model able to capture changes in user behaviour on the level of actions within a session. The main model idea lies in the description of actual user session and its comparison to previous sessions, realised by various users within multiple time periods. In this way, the user model contains a comparison to stable long-term behaviour as well as actual short-term one. Comparison to previous session of modelled user ensures personalised modelling. Potential lack of user’s browsing history is compensated by comparison to average behaviour of other users.

In addition, the model is proposed to be domain and language independent, as it uses only generally available usage data describing user activity and indirect descriptors of the structure and content of visited pages. For this reason the model is applicable to almost any site without a need to change its way of user activity logging or content transformation.

Selected publications

O. Kassak, M. Kompan, M. Bielikova

Students’ Behavior in a Web-Based Educational System: Exit Intent Prediction. Engineering Applications of Artificial Intelligence Journal, Mining the Humanities: Technologies and Applications, volume 51, pages 136–149. Elsevier, 2016. Audioslides  pdf

O. Kassak, M. Kompan, M. Bielikova

Personalized Hybrid Recommendation for Group of Users: Top-N Multimedia Recommender. Information Processing and Management, volume 52(3), pages 459–477. Elsevier, 2016.  pdf

O. Kassak, M. Kompan, M. Bielikova

User Preference Modeling by Global and Individual Weights for Recommendation. Acta Polytechnica Hungarica, volume 12(8), pages 27–41. IEEE, 2015.

M. Kompan, O. Kassak, M. Bielikova

Beyond User Preferences: The Short-Term Behaviour Modelling. In Proceedings of the Temporal Reasoning in Recommender Systems Workshop (RecTemp) collocated with RecSys’17, pages 3. Ceur, 2017.

T. Chovanak, O. Kassak, M. Bielikova

Behavioral Patterns Mining for Online Time Personalization. ACM User Modelling, Adaptation and Personalization conference, (UMAP’17), pages 2. ACM, 2017.

O. Kassak, M. Kompan, M. Bielikova

Short-term User Behaviour Changes Modelling. New Trends in Databases and Information Systems: Short Papers and Workshops, BigDap, DCSA, DC, pages 223–228. Springer, 2016.

O. Kassak, M. Kompan, M. Bielikova

Web User Short-term Behavior Prediction: The Attrition Rate in User Sessions. ECML PKDD 2015 - European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases - Doctoral consortium, pages 115–124. Library of Aalto University, 2015.

J. Visnovsky, O. Kassak, M. Kompan, M. Bielikova

The cold-start problem: minimal user’s activity estimation. Proceedings of the 1st workshop on recommender systems for television and online video (RecSysTV’14), pages 4, 2014.  pdf

to Homepage

to Teaching

to the Top