Exploring mobile news reading interactions for news app

personalisation
Marios Constantinides1, John Dowell1, David Johnson1, Sylvain Malacria1, 2
1 2
University College London, Inria Lille
London, United Kingdom Lille, France
{m.constantinides, j.dowell}@cs.ucl.ac.uk [email protected]
[email protected]
ABSTRACT Mobile news access perfectly complements the
As news is increasingly accessed on smartphones and continuously updating, 24-hour nature of digital news
tablets, the need for personalising news app interactions is services. But if users are now never out of range of the
apparent. We report a series of three studies addressing key news, they need more than ever for that access to be
issues in the development of adaptive news app interfaces. adaptive and personalised. Personalised news services are
We first surveyed users’ news reading preferences and already able to help people find news that is relevant to
behaviours; analysis revealed three primary types of reader. them, to recommend the right news to the right users, and to
We then implemented and deployed an Android news app help users keep abreast of news by aggregation over
that logs users’ interactions with the app. We used the logs multiple sources. This adaptivity is achieved through
to train a classifier and showed that it is able to reliably several methods [5] including: news content personalisation
recognise a user according to their reader type. Finally we by pushing filtered articles predicted to match the user’s
evaluated alternative, adaptive user interfaces for each interests; adaptive news browsing by changing the order of
reader type. The evaluation demonstrates the differential news categories; contextual news access by offering users
benefit of the adaptation for different users of the news app access to additional information related to the news they are
and the feasibility of adaptive interfaces for news apps. reading; and news aggregation, by automatically identifying
main news topics emerging from multiple sources. This
Author Keywords previous work on adaptivity in digital news access has
Mobile news reading; Personalisation; Implicit sampling; focused on recommendation of news content. But,
Adaptive mobile User Interfaces; adaptation of the way people interact with news services
ACM Classification Keywords has not been investigated.
H.5.m. Information interfaces and presentation (e.g., HCI): Personalisation of news access clearly needs to extend
Miscellaneous. beyond ‘what’ content users access to ‘how’ they access it,
as evident in the abundance of mobile news apps offering
INTRODUCTION personalisation features. For example, Inside.com –
Mobile app ecosystems are transforming patterns of news Breaking News allows users to select news topics to follow
consumption. Until quite recently, reading the news was a and then provides 300-character summaries of relevant
niche use for smartphones [12], mostly for when users were stories along with links to the original sources. Another
‘on the go’; now however, two in every three users of example is Newsbeat, again an aggregator but one that
mobile devices in the US regularly access news and as creates ‘personalised radio news bulletins’. Users select
many as one in five read in-depth news articles daily [2]; a their preferred text news sources from which stories are
similar picture is found in the UK [1]. This growth in pulled each day, summaries created, then news podcasts
mobile news access continues the migration of news created using text-to-voice technology. A third example is
consumers to the Internet. Flipboard, which uses the metaphor of a ‘personal
magazine’ to present articles from conventional news
Permission to make digital or hard copies of all or part of this work for
providers as well as social media updates, and RSS feeds.
personal or classroom use is granted without fee provided that copies are Users curate and share their own mini-magazines within the
not made or distributed for profit or commercial advantage and that copies app, drawing in stories on their preferred topics.
bear this notice and the full citation on the first page. Copyrights for
components of this work owned by others than ACM must be honored. The personalisation of news app interaction in these
Abstracting with credit is permitted. To copy otherwise, or republish, to examples is achieved through making the interface
post on servers or to redistribute to lists, requires prior specific permission
adaptable. Adaptive news interfaces that ‘automatically’
and/or a fee. Request permissions from [email protected].
MobileHCI '15, August 25 - 28, 2015, Copenhagen, Denmark adapt to the way the user reads the news in particular
© 2015 ACM. ISBN 978-1-4503-3652-9/15/08...$15.00 contexts are not found, other than in re-ordering menus of
DOI: http://dx.doi.org/10.1145/2785830.2785860 headlines to take account of previous reading choices. This
adaptation could be far more extensive, for example, to take
account of users’ idiosyncratic patterns of browsing news
headlines or the different ways in which different users read
news articles [8, 12].
Adaptive personalisation relies on constructing and
exploiting an individual user profile to deliver a tailored
version of the user interface [9]. A user profile is a model of
the user that the system learns through interaction with the
user. The construction of a user profile can be based on
explicit or implicit information gathering approaches. The
former consists of information provided directly by the user
such as by forms and questionnaires; although data captured
this way will have a greater reliability, the disruption to the
user can be significant. Recent studies attempting to
compare both methods have found implicit data capture to
be preferred by users [7].
User profiling has been demonstrated with news service
Fig. 1 Our BBC-mimic mobile news app that logs users' news
applications. Billsus and Pazzani [4] developed the news
reading interactions
recommendation system NewsDude to recommend news
articles for desktop users. They used supervised machine of adaptive user interfaces for each of the three news reader
learning methods in the form of nearest-neighbor types was evaluated. Our results suggest that different
algorithms to model short-term interests, and a naive Bayes reader types would benefit from different user interfaces.
classifier for long-term interests. Carreira et al. [6] used IDENTIFICATION OF NEWS READER TYPES
interaction logs of mobile users of news services to We deployed an online questionnaire1 using CrowdFlower
implicitly capture user profiles as the basis for with the aim of identifying stereotypical patterns of
recommending articles of interest. Their prototype news behaviour and individual experiences on mobile news
application logged aspects of users’ reading behaviour such reading. Although other studies [2,3] reported interesting
as reading duration, estimated number of lines read, insights on news access and consumption, this survey was
estimated reading speed, etc. However, they logged designed to reveal reading and navigational behavior
interactions with news services on PDAs having much less especially on smartphones.
advanced capabilities (neither 3G nor high resolution
screen). Recently Tavakolifard et al. [11] proposed a news The questionnaire consisted of 24 questions probing
content recommendation system (including the mobile app) demographic information and news reading behaviour on
that efficiently delivers “tailored news in the palm of your mobile devices including the estimated time spent on news
hand”. No studies have attempted to log interactions for the reading each day, the frequency of reading, browsing
purpose of personalising the interface as opposed to strategies, reading styles and so on. The sample comprised
personalising the new content. 140 respondents (54 females, 72% aged between 19-35,
60% hold a higher education qualification). The only
Interaction data capture with smartphones has been requirement for participants was that they read news on
demonstrated in other studies but not in relation to news smartphones. Respondents received a token payment for
consumption. Oulasvirta et al. [10] used logs of smartphone participating.
interactions to examine users’ habits, in particular their
habitual checking of their smartphone state and Analysis revealed interesting tendencies in users’
notifications. Woerndl et al. [13] proposed a unified preferences. Respondents mainly reported that they read the
approach for collecting data about smartphone interactions news once a day for between 10 and 30 minutes, preferably
in an appropriate granularity for user modeling. during the mornings and at home. Regarding their
navigation and reading of news, no strategy dominates.
In this paper we report an investigation into implicit When they browse, users do it either through all sections or
profiling and adaptive user interfaces for mobile news apps. they jump to a particular section whereas when they read
First, a survey was conducted to examine news reading they might skim or read the whole article.
behaviour of users of mobile devices. A cluster analysis
revealed three main types of mobile news reader In addition to the descriptive analysis we performed a
characterized by five factors. Second, a study was hierarchical clustering analysis on the responses to all
conducted to investigate whether users of a news app could questions from all participants. The analysis revealed three
be identified in relation to the three types using a dedicated homogeneity clusters and was conducted in three steps:
news app (Fig. 1), logging user’s interactions during two
weeks. Five characteristic factors were extracted from these 1
logs and were used for training a classifier. Finally, a design Online questionnaire URL: http://goo.gl/HvoxBc
 Reader INTERACTION LOGGING AND CLASSIFICATION STUDY
types Having characterised mobile news readers as one of three
‘Trackers’ ‘Reviewers’ ‘Dippers’
Clustering types, the question that follows is whether a news app could
factors detect a user as being a particular reader type from their
Many times a Less than once interactions alone. We now report a study to examine this
Frequency Once a day question involving the development of a mobile news app
day a day
Total daily
capable of collecting and classifying users’ interaction data.
5-10 min 10+ min 0-5 min
reading
Habito News: the mobile news app
Browsing Through all Particular
strategy
Both
sections section We implemented an Android news app to log user’s
interactions whilst reading the news. To ensure its
Detailed
Reading style Skimming
reading
Scanning familiarity to users, our news app mimicked the BBC’s app
in terms of its visual presentation (Fig. 1). News stories are
Public
Location Home Home organised in rows of thumbnails and are retrieved from the
Transport
BBC API using an asynchronous background task. To log
Table 1 News reader types for mobile news reading the interactions we implemented a background service that
unobtrusively collects data associated with general
(1) We first run a hierarchical clustering technique using preferences and usage of the app, both navigational and
Ward’s cluster method to identify groups of users with the reading, based on the five factors found in the survey study.
same characteristics. Three clusters were identified within
the total data set; 31% in group A (Trackers), 36% in group Deployment and data collection
B (Reviewers) and 33% in group C (Dippers). The Habito News app was downloaded from Google Play
(2) We then used a Pearson’s correlation test to associate and used by 23 participants (11 females) aged 22 to 33
questions with clusters. Five significant clustering factors years old (M=27, SD=3.64) for two weeks. Participants
were identified frequency; daily reading time; browsing were instructed to read the news, as they would do with any
strategy; reading style, and; location. other available news app and to use it as their primary news
app. After installing the app, users completed a form
(3) We finally ran a cross tabulation statistical method to containing the same questions as contained in the survey.
examine the distribution of the five factors over the clusters. The app then recorded their interactions in relation to the
The three news reader types are summarized in Table 1 in five clustering factors in the survey (including precise scroll
relation to the five factors. Note that whilst the clusters are usage, reading time, frequency of reading, location, exact
independent of each other, two clusters may not be category scrolling event such as items browsed and swipe
discriminated on one factor, i.e., a user can belong to more direction, etc). Although, other researchers conducted
than one cluster on that factor alone. We labeled the three logging studies [5, 8], to the best of our knowledge none of
news reader types as ‘Trackers’, ‘Reviewers’ and ‘Dippers’. them logged data in such fine granularity.

Tracker: A person who likes to be informed about the latest User Modeling: Predicting News Reader Types
stories and any updates to stories he or she is following, The logged interaction data was used to identify each
usually reading the news for up to 10 minutes at a time and participant’s news reader type using a naïve Bayes
several times a day at intervals, for example, when classifier. The classifier was developed in three steps:
travelling. Due to her limited time she prefers to extract the (1) The five clustering factors were extracted from the raw
important bits of a story (i.e. reading by skimming). interaction data by abstraction. For example, reading style
Reviewer: A person who likes to catch up on the day’s was estimated using the formula2 in which the proportion of
news, preferably at home. He likes an in-depth analysis of the article exposed to the user by using scroll positions,
the stories he reads and will read at length to fully divided by the total size of the document and multiplied by
understand the story (i.e. a detailed reading). He usually the number of words to approximate the number of words
reads the news once a day, spending more than 10 minutes viewed. The number of words read per (wpm)3 was
to get through all the stories of interest and likes being estimated by dividing the approximate number of words
informed on a variety of topics. seen by the reading duration. Finally, the reading speed was
interpreted as one of the three reading styles: reading for
Dipper: A person with a casual interest in the news but comprehension (i.e. detailed reading) up to 230 wpm;
likes to read news on specific topics such as sport. She scanning, reading faster than 700 wpm, and; skimming, a
always knows what she is looking for so does not spend reading speed between normal reading and scanning.
more than 5 minutes accessing the news. She likes to
browse particular sections to find stories and looks for
specific facts or pieces of information without reading 2
everything (i.e. reading by scanning). words seen = scroll_offset/scroll_range*no_of_words
3
wpm = words seen/reading duration
 Fig. 2 The baseline interface with the adaptive variants for the three news reader type
Likewise, we extracted the frequency by keeping sessions For example, for the Tracker reader, the canonical profile
of how frequent the user accessed our app, and whether the says they read news many times a day, with a total reading
user accessed any other news app installed in their phones time of less than five minutes, using both browsing
Regarding to the total daily reading time we added up the strategies, a skimming reading style, and primarily a public
time spend on reading. Finally, we implemented another transport location.
background service to capture user’s location. To transform
The classifier was able to correctly classify the canonical
location (longitude, latitude) into public transport or home
profiles with up to 88% accuracy for Trackers, 91% for
we used a heuristic formula in which stable consecutive
Reviewers and 92% for Dippers. Future work will validate
location points were defined as home, and non-stable ones
the classifier over a new corpus of interaction logs to be
as public transport.
collected from users of the news app.
(2) Creating the training set: each user was classified by
ADAPTIVE UI AND EVALUATION STUDY
taking the transformed values of their interaction log for Having shown that we can recognise a user’s news reader
each of the five factors identified by the clustering of
type from a data log of their interactions, the next question
survey results (e.g. frequency, total daily reading, etc). We
that arises is whether the different types would benefit from
classified each user as a particular news reader type
different adaptive forms of the news app.
corresponding with the five factors and making the
simplifying assumption that the factors are equally To examine this question we devised adaptive user
weighted. So, each item in the training set was of the interfaces for each news reader type through a series of
following form: user id, F1, F2, F3, F4, F5, Class (where Fn: semi-structured interviews and contextual inquiries with 10
Factor, Class: news reader type). participants. The interviews probed participants’ about their
news reading and their opinions on a range of customizable
The training set therefore consisted of 23 users, each features of current mobile news apps and design
consisting of a profile of five factors derived from their suggestions of our own. Participants were also asked to
interaction logs and each identified as one of the three types experience reading the news with our news app and provide
of reader. A word of caution is appropriate here. As a comments on its design. For example, Trackers receive the
separate check of the classification of reader types, each
latest stories or updates in the top static area for quick
user had been asked to complete the survey. Only a 55%
access and we replace the horizontal organisation of stories
correspondence was found between participants’ answers to to a full-width layout because they like to get a quick
the survey (e.g., for how long they read the news each day snapshot of the news. For Reviewers we did not make
in total) and the collected interaction data. It was of course significant changes because it seemed that the baseline
such self-assessments in the survey that produced the interface almost meet their needs. For Dippers, we provide
original clustering factors. Hence a separate study is needed the search functionality to enable quick browsing of
to establish whether the same news reader types would specific facts and the jump-to category feature to allow
emerge from the cluster analysis if logged interaction data faster navigation. For Trackers and Dippers we also have
were added to the survey data. the extra feature of summaries to support their reading
(3) Implementing the classifier: We used a canonical profile behaviour. The way summary is presented in each of these
for each of the three types to validate the classifier. The types differs depending on their characteristics. For
canonical profiles represent the interaction log values for example, Trackers receive summaries as a paragraph
the five factors for a prototypical user of that reader type. because they are skimmers, whereas we display the
summary as bullet points for Dippers who read in a faster asked to find and read specific articles and then briefly
pace. The adaptive features of each type are listed below: described what the article was about. After completing the
tasks with both interfaces, participants completed the
(1) Adaptive UI for Trackers (Fig. 2b)
comparison questionnaire. A short debriefing at the end of
• Top static area for the tracked articles the experiment sought participants’ views on how easily
• Full-width categories layout they were able to find and read articles with each interface
• Option of a summarised version of the story and the features of the adaptive variant.
(2) Adaptive UI for Reviewers (Fig. 2c)
Design
• No visuals (articles’ thumbnails) in the menu structure, The experiment was a one way within-subject design on UI
only top stories kept open type (baseline, adaptive variant), conducted independently
• Accordion to ease access categories on each group of news reader type (Trackers, Reviewers,
(3) Adaptive UI for Dippers (Fig. 2d) Dippers). Dependent variables were the time taken to find
• Search functionality articles and the time taken to read them. Participants were
• Easy access to articles of a particular category (jump-to) not aware which interface was the baseline and which was
• Option of a summarised version of the story the adaptive design.
Evaluation Study of the Adaptive UI variants Results
We conducted a laboratory study to compare the baseline An independent-samples t-test was conducted to compare
and adaptive variant of the news app interface for each the time taken to find and read stories in the baseline and in
news reader type. The comparison was of time to find the adaptive variant interface for each news reader type.
(browsing) and read (reading) the news in benchmark tasks. Baseline vs. Adaptive for Trackers: There was a significant
The following hypotheses were formulated: difference in the scores for the time taken to find articles in
H1: The adaptive UI for Trackers improves their the baseline (M=21.42, SD=7.59) and the adaptive for
performance over the baseline UI. Trackers (M=9.85, SD=4.37) conditions; t(12), p=0.004.
H2: The adaptive UI for Reviewers improves their There was also a significant difference in the scores for the
performance over the baseline UI. time taken to read articles in the baseline (M=147.42,
H3: The adaptive UI for Dippers improves their SD=62.05) and the adaptive for Trackers (M=77,
performance over the baseline UI. SD=30.47) conditions; t(12) p=0.019. We therefore accept
Participants H1 – the adaptive interface for Trackers improved their
18 individuals (7 Trackers, 5 Reviewers and 6 Dippers) performance over the baseline.
aged 20 to 30 years old (M=23, SD=2.26) participated in
Baseline vs. Adaptive for Reviewers: There was not a
the study. All participants had considerable experience in
significant difference in the scores for the time taken to find
using mobile devices and particularly with Android OS.
articles in the baseline (M=21.40, SD=5.77) and the
Materials adaptive for Reviewers (M=25, SD=10.12) conditions; t(8),
We used Justinmind 4 to develop the interactive wireframes p=0.509. No significant difference found in the scores for
(Fig. 2) and deployed them on a Samsung Galaxy S3 (4,8- the time taken to read articles in the baseline (M=199,
inch screen, 1280x720 resolution). A video camera was SD=12.04) and the adaptive for Reviewers (M=221.60,
used to record participants’ interactions with the device. A SD=27.15) conditions; t(8) p=0.127. We therefore reject H2
comparison questionnaire was used to measure their – the adaptive interface for Reviewers did not improve their
subjective preferences between the two interfaces. We performance over the baseline.
asked participants to choose which interface they preferred
(using a scale ranging from “Mostly A” to “Mostly B”, Baseline vs. Adaptive for Dippers: There was not a
where Mostly A means strong preference for interface A - significant difference in the scores for the time taken to find
baseline- and Mostly B for interface B -adaptive-). articles in the baseline (M=18.67, SD=5.98) and the
adaptive for Dippers (M=25.50, SD=8.43) conditions; t(10),
Procedure p=0.137. There was also not a significant difference in the
At the beginning of each session participants completed a scores for the time taken to read articles in the baseline
questionnaire that would allow their news reader type to be (M=180.83, SD=79.17) and the adaptive for Dippers
determined. Demographic information was also collected. (M=117.33, SD=38.51) conditions; t(10) p=0.108. We
Once their type was identified, participants were instructed therefore reject H3 – the adaptive interface for Dippers did
to complete a set of predefined tasks on both interfaces. It not improve their performance over the baseline.
should be noted that the tasks necessarily varied between
news reader types, appropriate to their characteristic Satisfaction and User Comments
patterns of news reading. For example, Reviewers were 71% of Trackers preferred the adaptive interface to browse
with 29% of them stating a strong preference, whereas all
of them found the adaptive interface better for reading.
4
http://www.justinmind.com/ However, 80% of the Reviewers preferred the baseline for
navigation, and 60% easier for reading. Dippers stated a Dippers navigation, combining the jump-to category and
neutral preference on both interfaces with 33% for each. the return-to-top functionality.
In their post-experiment comments, Trackers praised the CONCLUSION AND FUTURE WORK
possibility to switch between the full article and its We explored the feasibility of recognising patterns of news
summary, but also suggested extra features to support reading interactions and evaluated three adaptive interface
tracked articles, e.g. visually differentiating read articles designs for different news reader types. We show that from
and new ones. They however strongly disliked the menu their interaction log, a specific user can be recognised as
structure; the vertical scrolling for categories, due to the one of three kinds. The reader types emerging from the
fact they want to be able to browse all categories with ease. online survey are well defined and distinct. The evaluation
of the three variant interfaces suggests that different news
Reviewers reported they did not find the adaptive menu reader types need different user interfaces. We have
beneficial and would prefer a snapshot of articles within demonstrated a method for monitoring users’ news reading
multiple categories as opposed to being restricted to one. behaviour and inferring news reader type from it.
None reported wanting article summaries despite reading
long articles and being probed in the post interviews. In the future we will further explore the design of adaptive
interfaces, in order to be in a position to demonstrate a
Dippers also preferred article summaries and the option to complete adaptive mobile news framework providing
switch between the full and the summarised version of an automatic personalisation of news apps.
article. They praised the menu structure with the jump-to
category but recommended a return-to-top functionality REFERENCES
(widely common technique at websites navigation). 1. Ofcom, News consumption in the UK, Public report
(2014).
DISCUSSION
2. Pew Research Centre, The Future of Mobile News,
The key findings of these three studies are: (1) mobile news
readers can be characterised within three types; (2) it is Public report (2012).
possible to detect a user’s news reader type from analysis of 3. Reuters Institute, Tracking the future of news, Public
their interactions with a news app, and; (3) different reader Report (2014)
types benefit from different forms of news app interface. 4. Billsus, D. & Pazzani, M. A hybrid user model for news
story classification. Springer Vienna (1999), 99-108.
We showed that mobile news readers can be distinguished
5. Billsus, D. & Pazzani, M. Adaptive news access. In The
as Trackers, Reviewers or Dippers according to their
adaptive web. Springer Berlin Heidelberg (2007).
frequency of news reading, their duration of news reading,
their browsing strategy, their reading style and their 6. Carreira, R., Crato, J. M., Gonçalves, D., & Jorge, J. A.
location. We showed that a Bayesian classifier can identify Evaluating adaptive user profiles for news classification.
the user of a news app according to their news reader type In ACM IUI 2004.
from a log of interactions with a news app from which the 7. Gauch, S., Speretta, M., Chandramouli, A., & Micarelli,
clustering factors could be extracted. The evaluation study A. User profiles for personalized information access.
demonstrated that Trackers performed better with the In The adaptive web. Springer Berlin Heidelberg (2007).
adapted interface variant, whereas Reviewers and Dippers 8. Grzeschik, K., Kruppa, Y., Marti, D., & Donner, P.
performed better with the baseline interface. Trackers also Reading in 2110-reading behavior and reading devices:
expressed a preference for the adapted variant created for a case study. The Electronic Library (2011).
them, whereas Reviewers and Dippers did not prefer the 9. Jameson, A. Adaptive interfaces and agents. Human-
variants created for them. Those two variants were clearly Computer Interaction: Design Issues, Solutions, and
not successful, however this does not preclude that Applications, (2009).
successful forms could be created and further investigation 10. Oulasvirta, A., Rattenbury, T., Ma, L., & Raita, E.
is needed. However the more significant finding is that their Habits make smartphone use more pervasive. Personal
needs and those of the Trackers are not met by the same
and Ubiquitous Computing (2012).
interface and that adaptation is desirable, which can reliably
been achieved by the system. 11. Tavakolifard, M., Gulla, J., Almeroth, K., Ingvaldesn, J.,
Nygreen, G. & Berg, E. Tailored news in the palm of
Based on evaluation findings, we modified the Reviewers your hand: a multi-perspective transparent approach to
interface and dropped the accordion style organisation to news recommendation. In ACM WWW 2013.
restore the horizontal rows organisation of the baseline 12. Westlund, O. From mobile phone to mobile device:
interface. We also added the feature of expanding News consumption on the go. Canadian Journal of
categories to provide the big picture of stories within a Communication (2008), 33(3).
category. Likewise, we suggest a simpler and more 13. Woerndl, W., Manhardt, A., & Prinz, V. A framework
straightforward accordion style organisation for improving
for mobile user activity logging. In MUSE 2010.