Influence of Timeline and Named-entity Components on User Engagement
Download as PPTX, PDF2 likes862 views
Nowadays, successful applications are those which contain features that captivate and engage users. Using an interactive news retrieval system as a use case, in this paper we study the effect of timeline and named-entity components on user engagement. This is in contrast with previous studies where the importance of these components were studied from a retrieval effectiveness point of view. Our experimental results show significant improvements in user engagement when named-entity and timeline components were installed. Further, we investigate if we can predict user-centred metrics through user's interaction with the system. Results show that we can successfully learn a model that predicts all dimensions of user engagement and whether users will like the system or not. These findings might steer systems that apply a more personalised user experience, tailored to the user's preferences.
Ad
More Related Content
What's hot (20)
Viewers also liked (20)
Ad
Similar to Influence of Timeline and Named-entity Components on User Engagement (20)
Ad
More from Roi Blanco (12)
Recently uploaded (20)
Influence of Timeline and Named-entity Components on User Engagement
- 1. Influence of Timeline and Named-entity Components on User Engagement Yashar Moshfeghi1, Michael Matthews2, Roi Blanco2, Joemon M. Jose1 1 School of Computing Science, University of Glasgow, Glasgow, UK 2 Yahoo! Labs, Barcelona, Spain [email protected] ECIR 2013, Moscow, Russia
- 2. Outline • User Engagement • Prediction of User-centred metrics • Evaluation Methodology • Results • Conclusions
- 5. Time Named-Entity a Cranfield-style paradigm user engagement
- 6. Research Question • We aim to answer the following research question: – “can timeline and named-entity components improve user engagement in the context of a news retrieval system?”
- 7. Multi-faceted concept: emotional, cognitive and behavioural Subjective measures (O’Brien and Toms): focused attention, aesthetics, perceived usability, endurability, novelty, involvement Objective measures: Subjective Perception of Time
- 8. An increase of information-rich user experiences in the search realm (logged interaction data) Prediction of user preferences for web search results Prediction of user-centred metrics of an IIR system Build search applications in which the layout and elements displayed adapt to the needs of the user or context
- 9. Submit Query Retrieved Results The News System Anatomy Timeline Component Entity Component
- 10. Experimental Methodology • Design – A ‘within-subjects’ design was used in this study. • The independent variable – the system (with two levels: baseline, enriched), – controlled by the viewing timeline and named-entity components (enriched) or hiding them (baseline). • The dependent variables were: – (i) user engagement • (involvement, novelty, endurability, usability, aesthetics, attention) – (ii) system preference
- 11. Experimental Methodology - Task • We used a simulated information need situation. • The simulated task was defined as follow: – “Imagine you are reading today’s news events and one of them is very important or interesting to you, and you want to learn more. Find as much relevant news information as possible so that you can construct an overall (big) picture of the event and also cover the important parts of it.”
- 12. Experimental Methodology - Task • The search task was presented twice to each participant with different search topics.
- 13. • Advantages: • Reduced monetary cost • Ease of engaging a large number of users in the study. • Disadvantages • Low quality data and in turn, the challenge is to improve and assure data quality. • Need for techniques to minimise • spammers, • multiple account workers • Lazy worker
- 14. • Multiple response technique for our questionnaire • known to be very effective and cost efficient to improve the data quality • Browser cookies were used to guard against multiple account workers • To avoid spammers (as recommended in the literature), • Population screening based on location (United States) • HIT approval rate greater than 95% • To reduce attrition, demographic questions were put at the beginning of the experimental procedure.
- 15. Experimental Methodology - Procedure • Participants were instructed that the experiment would take approximately 60 minutes to complete • They were informed that they could only participate in this study once • Payment for study completion was $5 (The total cost of the evaluation was $510 ) • Each participant had to complete two search tasks, one for each level of independent variable (i.e. baseline and enriched system)
- 16. Experimental Methodology - Procedure
- 17. Experimental Methodology • We considered six dimensions introduced by O’Brien et al.: – focused attention, aesthetics, perceived usability, endurability, novelty, and involvement • The different dimensions were measured through a number of forced-choice type questions. • A 5-point scale respond (strong disagree to strong agree) – “Based on this news retrieval experience, please indicate whether you agree or disagree with each statement”. • In total, in each post-search questionnaire we have asked 31 questions related to user engagement – adapted from O’Brien et al. – randomised its assignment to participants
- 18. Experimental Methodology • Pilot Studies: – We run three pilot studies using 10 participants. – Other changes consisted of • modifications to the questionnaires to clarify questions, • modifications to the system to improve logging capabilities • improvements to the training video. – After the final pilot, it was determined that • the participants were able to complete the user study without problems • the system was correctly logging the interaction data.
- 19. Results Analysis – Data Preprocessing • To ensure the availability of relevant documents – two evaluators manually calculated • the Precision@1, 5, and 10 • for all the topics • a set of queries issued by the participants. – Precision@1, 5 and 10 were 0.85, 0.84, and 0.86 respectively, – Judges had a very high inter-annotator agreement with Kappa > 0.9. – This indicates that the queries the users issued into the system had good coverage and the ranking was accurate enough.
- 20. Results Analysis – Data Preprocessing • 63 out of 92 users successfully completed the study. • A relatively even split by condition, with 47% in the scenario where group 1, and 53% conversely. • We removed the – incomplete surveys – participants who repeated the study – participants who completed the survey incorrectly (based on task conditions) • they had to visit at least three relevant documents for a given topic, and • the issued queries should be related to the selected topic – identifying suspect attempts by checking • the extremely short task durations • comments that are repeated verbatim across multiple open-ended questions
- 21. Results Analysis – Demographic Info. • 126 search sessions that were successfully carried out by 63 participants. • The 63 participants – female=46%, male=54%, prefer not to say=0% – were mainly under the age of 41 (84%) • with the largest group between the ages of 24-29 (33.3%). • Participants had – a high school diploma or equivalent (11.11%), – associates degree (15.87%), – graduate degree (11.11%), – bachelor (31.7%) or – some college degree (30.15%). • They were – primarily employed for a company or organisation (39.68%), – though there were a number of self-employed (22.22%), – students (11.11%), and – not employed (26.98%).
- 22. Results Analysis Enriched ** Baseline Enriched * Baseline Enriched * Baseline Enriched Baseline Enriched ** Baseline Enriched Baseline 1 2 3 4 5 User Engagement Involvement Novelty Endurablility Usability Aesthetics Attention
- 23. Results Analysis • We did not find any statistically significant difference between the two systems for Subjective Perception of Time metric – with mean and standard deviation of 10.03, ± 5.22, and 10.12, ±4.95, for the baseline and enriched system respectively
- 24. Results Analysis - System Preference • the exit questionnaire posed the question – “Please select the system you preferred? (answer: 1: First System, 2: Second System)” – and overall, 76% of the participants preferred the enriched system better than the baseline system.
- 25. Prediction of User-centred Metrics: • The demographic features – participants’ age, gender, education, and occupation • The search habits features – the number of years they have used web search and online news systems, – the frequency they engaged in different news search intention such as browsing, navigating, searching, etc. – the news domain they are interested in • The interaction features (derived from log information) – the total time they spent on each component and to complete a task, – the number of clicks, retrieved documents, queries, – the number of times they used the previous/next button, and other functionality of the systems
- 26. Prediction of User-centred Metrics: • We chose – the System Preference question – all the user engagement dimensions. • For System Preference question, – we have a binary class of “−1” indicating the participant did not prefer the enriched system and “+1” otherwise. • For the user engagement dimensions, – we used the final value calculated by aggregating all the questions related to each dimension – We transformed the values for each dimension to binary by mapping 4-5 to “+1” and otherwise to “−1”
- 27. Prediction of User-centred Metrics: • We learned a model to discriminate between the two classes using – SVMs trained with a polynomial kernel, – based on our analysis in the majority of cases, outperformed other SVM kernels (linear, and radial-basis). • We also tried other models such as bayesian logistic regression and decision trees but they underperformed with respect to SVMs.
- 28. Prediction of User-centred Metrics: • classification performance – averaged over the 63 participants of the study – using 10-fold cross validation Results indicate that ◦ for all the user engagement dimensions (excluding focused attention), the combination of all features leads to the best prediction accuracy ◦ Regarding the system preference question, user-system interaction features determine with high accuracy the participants’ preference of a system (over 87%).
- 29. Summary • Given the competitiveness of the market on the web, applications nowadays are designed to be both efficient and engaging. • Thus, a new line of research is to identify system features that steer user engagement. • This work studies the interplay between user engagement and retrieval of named-entities and time, in an interactive search scenario. • We devised an experimental setup that exposed our participants on two news systems, one with a timeline and named-entity components and one without. • Two search tasks were performed by the participants and through questionnaires, user engagement was analysed.
- 30. Conclusions • Overall findings based on user questionnaires, show that substantial user engagement improvements can be achieved by integrating time and entity information into the system. • Further analysis of the results show that the majority of the participants preferred the enriched system over the baseline system. • We also investigated the hypothesis that user-centred metrics can be predicted in an IIR scenario given the participants’ demographics and search habits, and/or interaction with the system. • The results obtained across all the user engagement dimensions as well as System Preference question, supported our hypothesis. • As future work, we will continue to study how user interactions can be leveraged to predict satisfaction measures and possibly build interfaces that adapt based on user interaction patterns.
- 31. Acknowledgement: This work was partially supported by the EU FP7 LiMoSINe project (288024). This work was performed while intern at Yahoo! Research lab in
Editor's Notes
- #4: “how and why people develop a relationship with technology and integrate it into their lives.”
- #5: Thus, a new line of research is to identify system features that steer user engagement, which has become a key concept in designing user-centred web applications. Given the ubiquity of the choices on the web the competitiveness of the market, applications nowadays are designed to not only be efficient, effective, or satisfying but also engaging.
- #6: There has been great attention on retrieving named entities, and using the time dimension for retrieval. Those approaches are evaluated exclusively focusing on a Cranfield-style paradigm, with little or no attention on user input, context and interaction. However, it is difficult to correlate user engagement with traditional retrieval metrics such as MAP. This problem becomes exacerbated when the user has to cope with content-rich user interfaces that include different sources of evidence and information nuggets of a different nature. This work studies the interplay between user engagement and retrieval of named-entities and time, in an interactive search scenario.
- #8: User engagement is a multi-faceted concept associated with the emotional, cognitive and behavioural connection of user with a technological resource at any point of interaction period [1]. O’Brien and Toms defined a model characterising the key indicative dimensions of user engagement: focused attention, aesthetics, perceived usability, endurability, novelty, involvement. These factors elaborate the user engagement notion over the emotional, cognitive and behavioural aspects. Subjective and objective measures are proposed to evaluate user engagement [1], the former being considered to be the most for evaluation. We use the subjective measures proposed by O’Brien et al. [3]. Objective measures include subjective perception of time (SPT) and information retrieval metrics among others. SPT is calculated by asking participants to estimate the time taken to complete their searching task, which is compared with the actual time [1].
- #9: Given the increase of information-rich user experiences in the search realm, we leverage the amount of logged interaction data. Prediction of user preferences for web search results based on user interaction with the system has been studied previously. In this work, we try to predict user-centred metrics of an IIR system rather than user preferences for its search results. Our positive findings could steer research into building search applications in which the layout and elements displayed adapt to the needs of the user or context.
- #10: To provide a use case for our investigation, we experiment with a news search system, which encourages interaction due to the information overload problem associated with the news domain. One way to facilitate user interaction in such scenarios is to develop new methods of accessing such electronic resources. For this purpose, we carefully varied the components of a news retrieval system page. We experimented with a timeline and named-entity component (enriched) or hiding them (baseline), while keeping everything else fixed, and tested whether adding these components can help improve user engagement. To study the predictability of the user centred metrics, we repeat our interactive experiments at two different points in time, with a tightly controlled setting. As an outcome of those experiments, we conclude that the user-centred metrics can be predicted with high accuracy given their interaction with the system and their demographics and search habits are provided as an input.
- #12: We introduced a short cover story that helped us describe to our participants the source of their information need, the environment of the situation and the problem to be solved. 2) This facilitated a better understanding of the search objective and, in addition, introduced a layer of realism, while preserving well-defined relevance criteria.
- #13: We prepared a number of search topics that covered a variety of contexts, from entertainment and sport to crime and political issues, in order to capture participants’ interests as best as possible.
- #14: We make use of Amazon’s Mechanical Turk (M-Turk), as our crowdsourcing platform.
- #15: Particular attention was paid in our experimental design to help motivate participants to respond honestly to the self-report questions and take the tasks seriously.
- #16: - though they would be given 120 minutes between the time they accepted and submitted the HIT assignment. ------- - they would not be paid if they had participated in any of the previous pilot studies. ------- - Given the findings of Mason and Watts, we expect the increase in wage just to change the rate of incoming workers to accept the HITS, and not affect their performance. - The total cost of the evaluation was $510 including the cost of the pilot studies and some of the rejected participants, which we consider to be cost-effective. ------- - The order in which each participant was introduced to the systems was randomised to soften any bias, e.g. the effect of task and/or fatigue. - Subsequently, participants were assigned to one of two systems (baseline or enriched) by clicking the link to the external survey. -------
- #17: At the beginning of the experiment, the participants were introduced to an entry questionnaire, Demographic information, Previous experience with online news, in particular, browsing and search habits to estimate their familiarity with news retrieval systems and their related tasks. At the beginning of each task, the participants completed a Pre-search questionnaire, to understand the reason why a particular topic was selected At the end of each task, the participants completed a post-search questionnaire to elicit subject’s viewpoint on all user engagement dimensions. Finally, an exit questionnaire was introduced at the end of the study. In this questionnaire we gathered information about the user study in general: which system and task they preferred and why and their general comments.
- #18: For example, involvement was measured by adapting three questions from [3]: (1) I was really drawn into my news search task. (2) I felt involved in this news search task. (3) This news search experience was fun.
- #19: In each iteration, a number of changes were made to the system based on feedback from the pilot study. For example, for each dimension we computed Cronbach’s alpha to evaluate the reliability of the questions adopted for each dimension. We finalised the questions of each dimension by confirming their Cronbach’s alpha value (> 0.8). Cronbach’s alpha is used as a measure of the internal consistency of a psychometric test score for a sample of subjects.
- #20: This is further explained by the fact that the topics were timely and most news providers including in the index contained articles related to them.
- #21: after either abandoning it part-way through or had completed it once before.
- #23: Figures 2 shows the box plot for the user engagement analysis, for the two systems (baseline and enriched), based on the post-study questionnaire. The box plot reports, over the data gathered from 63 participants, five important pieces of information namely: the minimum, first, second (median), third, and maximum quartiles. We performed a paired Wilcoxon Mann-Whitney test between measures obtained for the enriched system for each user to check the significance of the difference with the baseline system. We use (*) and (**) to denote the fact that a dimension had results different from that of the baseline with the confidence levels (p < 0.05) and (p < 0.01) respectively. As shown in Figure 2, the enriched system has a better median and/or mean and lower variance than the baseline system across all dimensions. This shows that substantial user engagement improvements can be achieved by integrating time and entity information into the system. The findings also show that participants are significantly more engaged both from cognition (considering endurability and involvement) and emotion (considering the aesthetics and novelty) aspects when time and entity dimensions of the information space are provided (i.e. enriched system).
- #25: (we refer to as System Preference)
- #26: We investigate whether user engagement and in a more general sense user-centred metrics can be predicted, given the participants’ demographic and search habits information, and/or their interaction with the system.
- #27: taken from exit and post-search questionnaire respectively
- #29: Remarkably, the machine learned model is able to predict with a low error all of the user and system metrics. Given these positive findings, it is possible to move towards personalised search applications in which the layout and elements displayed adapt to the needs of the user or context which in turn results in increasing the users’ engagement as well as their preference of the system.