NewsRecs: A Mobile App Framework for Conducting and Evaluating
Online Experiments for News Recommender Systems
Noah Janzen
a
and Fatih Gedikli
b
Institute of Computer Science, University of Applied Sciences Ruhr West, M
¨
ulheim an der Ruhr, Germany
Keywords:
News Recommender Systems, Online Evaluation, News App, Mobile.
Abstract:
News recommender systems are ubiquitous on the web. Intensive research has been conducted over the last
decades, resulting in the continuous proposal of new recommendation techniques based on Machine Learning
models. To evaluate the performance of recommendation algorithms, offline experiments, user studies, and
online experiments should ideally be carried out one after the other so that the candidates move through a
quality funnel. However, our literature review of multiple academic papers shows that new models have
generally been evaluated using offline experiments only. Presumably, this is because researchers rarely have
access to a production system. This work attempts to alleviate this problem by presenting a framework that can
be used to evaluate recommendation models for news articles in an online scenario. The framework consists
of a mobile app in which users can receive recommendations from different algorithms depending on their
assigned group and rate them in multiple ways. The backend collects log data and makes it available for the
final evaluation. The specific contributions our article will make are as follows: (1) A thematic review of
27 academic experiments from the news recommendation domain focusing on the evaluation design. (2) An
open-source mobile app framework for conducting and evaluating online experiments.
1 INTRODUCTION
In the last decade, a lot of research has been done
in the field of news recommender systems, focus-
ing on the development of new recommender models.
Hence, a variety of different models are available to-
day for different use cases. In a real-world scenario,
an application designer must make a decision about
the most appropriate model for the specific applica-
tion in which the model is to be deployed.
The selection of a recommendation model is usu-
ally based on experiments comparing the perfor-
mance of a set of models. Offline experiments utilize
existing data sets and a predefined protocol simulat-
ing the user behaviour to determine the performance
of a model (e.g. accuracy). A more elaborate option
is a user study: A small group of potential users will
be asked to perform a series of tasks with the system.
Subsequently, the test persons answer questions about
their experience. Lastly, experiments can also be con-
ducted with a real system, which are referred to as
online experiments. The performance of the recom-
a
https://orcid.org/0000-0001-7066-7861
b
https://orcid.org/0000-0001-6190-0449
mendation algorithms is evaluated on the basis of real
users, who are typically unaware of an experiment,
which is why online experiments are seen as the most
trustworthy experiments. Typically, the selection of a
suitable recommendation model is a gradual process
involving both offline and online experiments, as well
as user studies where appropriate. First, an extensive
online experiment is typically conducted to filter out
unsuitable approaches (screening). A relatively small
set of potentially suitable models remains, which are
tested with the help of more expensive user studies or
online experiments to confirm the usefulness in prac-
tice (Ricci et al., 2022).
Offline experiments are appealing because they do
not require interaction with real users and allow com-
parison of a wide range of possible algorithms at rel-
atively low cost. Online experiments, on the other
hand, offer the advantage that their results provide the
strongest evidence because the system is used by real
users performing real tasks. In the real world, a rec-
ommender system typically influences the behavior of
its users. This change in behavior cannot be measured
using offline experiments alone because they work
with synthetic or historical user interaction data, i.e.,
the historical user behavior will remain static. A seri-
Janzen, N. and Gedikli, F.
NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems.
DOI: 10.5220/0011658000003393
In Proceedings of the 15th International Conference on Agents and Artificial Intelligence (ICAART 2023) - Volume 3, pages 267-275
ISBN: 978-989-758-623-1; ISSN: 2184-433X
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
267
ous problem for academic research is that it rarely has
access to a production system that allows online ex-
periments, and therefore often relies on offline exper-
iments (Ricci et al., 2022; Castells and Moffat, 2022).
The following sections are structured as follows.
Section 2 examines evaluation practices in academic
research, analyzing 27 papers that propose novel news
recommendation models. Section 3 presents the re-
quirements and the system design, Section 4 the im-
plementation and Section 5 the major features of the
proposed framework. Finally, the main findings are
concluded in Section 6.
2 RELATED WORK
News analytics plays an important role in news rec-
ommender systems (Gedikli et al., 2021), and also
in communication sciences in general (Nicholls and
Bright, 2019). One of the main challenges is the
implementation of a long-term evaluation of a news
recommender system. We analyzed various papers
proposing novel news recommendation models in or-
der to answer the following questions: How often are
models evaluated using offline or online experiments
in academic research? What metrics are used to eval-
uate the models?
2.1 Methodology
A semi-systematic approach was used to identify rele-
vant papers on the topic of news recommenders. First,
the Google Scholar search engine and a number of
digital libraries
1
were queried using the search term
“news recommendations”. Results prior to 2017 were
filtered out, as only current evaluation practice is of
interest. The papers received were then examined
with regard to their relevance in terms of title and ab-
stract. To answer the above research questions, only
papers that present and evaluate a model for generat-
ing recommendations for news are of interest. The re-
maining papers were then used as the starting point of
a snowball process. Additionally, a meta-study could
be found that also analyzes research for news recom-
mendation systems (Amir et al., 2022). Finally, this
served to validate our own findings.
2.2 Survey of Existing Approaches
A total of 27 research papers were examined that fo-
cus on the development and evaluation of novel news
1
Springer Link, ACM Digital Library, arXiv.org and
ResearchGate were reviewed.
recommendation models. The eldest paper dates from
2017, the most recent from 2022, and the majority
of research papers were published in 2019 or later
(85%).
Only 5 papers evaluated the developed model in
an online experiment. In 22 research papers, the
model was evaluated using only offline experiments.
All 5 research papers that examined their model with
online experiments also set up offline experiments in
advance. This confirms our assumption from the be-
ginning that academic research often relies only on
offline experiments (Castells and Moffat, 2022). The
reason for this is assumed to be that they rarely have
access to a productive system that allows online ex-
periments.
2.3 Evaluation Metrics
On average, the proposed models were evaluated us-
ing 3.7 evaluation metrics (median: 3). All papers
used at least two metrics. 70% of all models were
evaluated with 2 or 3 metrics. Research papers that
evaluated both offline and online tended to use more
metrics (5, 5, 7, 7, 12 metrics).
Table 1 shows all research papers examined with
the evaluation metrics used in each case. The last row
indicates in how many papers the respective metric
was used. Metrics that were used less than twice are
not shown for clarity. It can be seen that the Nor-
malized Discounted Cumulative Gain (nDCG), Area
Under the ROC Curve (AUC), and Mean Reciprocal
Rank (MRR) metrics are by far the most frequently
used for offline evaluation of news recommendation
systems (18, 17, and 15 times, respectively). This is
likely due to the fact that news recommendations are
typically displayed in the form of a list and these met-
rics are used to evaluate recommendation lists. This
is followed far behind by F1 score (F1), hit rate (also
hit ratio), click through rate (CTR) and precision (6,
5 and 4 uses).
Since metrics for online evaluation are particu-
larly relevant to the app framework, Table 2 includes
only the research papers that conducted online exper-
iments. The table shows all metrics that were used in
the online experiments of the papers.
By far the most popular is the Click Through Rate
(CTR) used for online evaluation. The CTR divides
the number of clicks on a recommendation by the
number of views of that recommendation. The second
most used metric is the number of clicks per session.
Session is defined here as the use of the recommenda-
tion service by a user. If no action is performed within
a defined period of time (e.g. 15 minutes), the session
ends. Except for the metrics CTR and Clicks, there
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
268
Table 1: Evaluation metrics used in the experiments.
Experiment
nDCG
AUC
MRR
F1
Hit rate
CTR
Precision
Recall
ILD
Clicks
1 DRN (Zheng et al., 2018)
2 RAKGG (Liu et al., 2021)
3 OF-SMR (Mulder et al., 2021)
4 PLM-NR (Wu et al., 2021)
5 RWTHIN (Symeonidis et al., 2020)
6 NPA (Wu et al., 2019b)
7 D-HAN (Zhao et al., 2021)
8 PTG (Koo et al., 2020)
9 GERL (Ge et al., 2020)
10 UPDNN-TNR (Hu et al., 2020a)
11 CSRN (Bai et al., 2020)
12 ENR-MU (Okura et al., 2017)
13 DNN-NR (Park et al., 2017)
14 NRMS (Wu et al., 2019c)
15 NRAML (Wu et al., 2019a)
16 NRLSUR (An et al., 2019)
17 PNRML (Qi et al., 2020)
18 AMM (Zhang et al., 2021)
19 JKP-RGC (Tian et al., 2021)
20 DKN (Wang et al., 2018)
21 FIM (Wang et al., 2020)
22 GNNR-UPD (Hu et al., 2020b)
23 KRED (Liu et al., 2020)
24 ASA-IPNR (Yoneda et al., 2019)
25 GNNR-UEPIM (Qiu et al., 2022)
26 EHUP-KG (Wang et al., 2019)
27 TEKG (Lee et al., 2020)
Frequency 18 17 15 6 5 4 4 3 2 2
Metric used in an offline experiment Metric used in an online experiment
are no common metrics that have been used for eval-
uation in several research papers (see Table 2). This
could be due to differences in the production systems
with which the experiments were conducted.
In one paper, an online experiment was conducted
on the Dutch news platform Blendle (Mulder et al.,
2021). On the platform, after reading a news arti-
cle, users can mark the article as a favorite. The re-
searchers defined the favorite ratio as the number of
users per user group (A or B) who marked an article
as a favorite, divided by the number of users in the
same group who finished reading the article. The re-
searchers assume that the ratio is an indicator of user
satisfaction with the article. Such a metric is highly
dependent on the productive system under which the
experiments are performed. Consequently, if the Fa-
vorites feature does not exist in the productive system,
the Favourite Ratio cannot be determined either.
Despite the differences in the use of the various
metrics, many metrics have in common that they are
based on clicks on a recommendation (CTR, Click-
s/Session, Precision@k, nDCG, CTR per article, CTR
per recommendation set, Clicks, Click Users/Ses-
sion). An essential requirement of the app framework
can be derived from this: Clicks on a recommendation
must be logged.
3 REQUIREMENTS AND
SYSTEM DESIGN
Next, we introduce our NewsRecs framework, which
allows researchers to compare several recommenders
NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems
269
Table 2: Evaluation metrics used in the online experiments.
Experiment
CTR
Clicks/Session
Precision@k
nDCG
ILS
CTR per article
CTR per rec. set
Compl. rate of rec.
Favourite ratio
Presentation charac.
Source diversity
Clicks
Pageviews
Sessions
Duration
Click Users/Session
DRN (Zheng et al., 2018)
OF-SMR (Mulder et al., 2021)
PLM-NR (Wu et al., 2021)
ENR-MU (Okura et al., 2017)
IPNR (Yoneda et al., 2019)
Frequency 4 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1
in an online experimental setting. Since the global
web traffic on mobile devices has surpassed desktop
long ago, we developed a mobile app for mobile de-
vices and therefore focus on mobile users. However,
the frontend can easily be converted into a web appli-
cation because the UI-framework we use is based on
web technologies.
3.1 Framework Requirements
The target users can be divided into two groups: On
the one hand, there are app users who use the mobile
app to receive news recommendations. On the other
hand, researchers use the software to conduct online
experiments, to create user surveys, and to evaluate
new algorithms or potentially user interfaces.
In order to be able to associate data with users, a
registration step is required. Therefore, an app user
can create an account and log in. He can also edit
his user profile within the app. The core feature of
the app is to receive news recommendations in a list.
A user can press on a news recommendation so that
the entire article opens within an in-app browser for
reading. After reading the article, the user can rate
the recommendation. On the other side, researchers
are able to design A/B-tests and create surveys for app
users. Researchers can also download all log data to
evaluate the alternatives tested.
3.2 Software Architecture
The NewsRecs framework implements a client-server
architecture. Users consume news recommendations
from a central server on their mobile client (see Fig-
ure 1). An external news server must provide meta-
data of news articles via a GraphQL or REST inter-
face, such as the titles, publication dates and the links
of the news articles. The framework then takes care
of the rest. The news server and news database are
not part of the NewsRecs framework, but can be eas-
ily implemented by using, e.g., one of many Python-
libraries for scraping news articles.
The NewsRecs server accesses the news server to
load news items. The recommendation algorithms
to be evaluated must be implemented in the News-
Recs server to generate recommendations to users.
The NewsRecs server is connected to a database that
stores user, usage, and survey data. Users access
the NewsRecs server via a REST interface, for ex-
ample, to register, receive news recommendations, or
rate recommendations on a Likert scale designed by
the researcher. Researchers can directly access the
NewsRecs database to download user usage data and
survey responses. Researchers can also create new
surveys on the NewsRecs server. Online experiments
can run over several days, weeks or months, making
long-term studies possible, which is one of the biggest
challenges in news recommender systems.
3.3 Technology Stack
NewsRecs uses the established MERN stack, con-
sisting of MongoDB as a document-oriented NoSQL
database, Express.js as a server-side web framework,
ReactNative as a UI framework for cross-platform
development of native apps for the Android and
iOS operating systems, and Node.js as a server-side
JavaScript runtime environment. NestJS has been
added to this stack. NestJS is a framework to develop
scalable server applications and provides an abstrac-
tion layer above the common Node.js frameworks.
Both the frontend and the backend are written in the
TypeScript programming language.
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
270
News
Database
News
Server
Mongoose
NewsRecs
Database
NewsRecs
Server
Mongoose
NewsRecs
External
GraphQL/REST
Researcher
Client
App User
Client
REST
Figure 1: Architecture of the NewsRecs Framework.
frontend/
api/
components/
constants/
i18n/
model/
screens/
store/
util/
.env
App.tsx
backend/
src/
auth/
common/
interactions/
mail/
middleware/
newsArticles/
surveys/
users/
app.module.ts
main.ts
.env
Figure 2: File structure of the NewsRecs framework.
4 IMPLEMENTATION
NewsRecs code and examples are MIT-licensed to
minimize legal impediments to adoption. The code
is available on GitHub
2
.
4.1 Frontend
The mobile app forms the front end, which was im-
plemented using the React Native framework. Figure
2 shows the relevant folders and files in the root direc-
tory of the frontend, each of which is briefly discussed
below:
In the api folder there are TypeScript modules,
that provide functions for communicating with the
backend.
All reusable React components are stored in the
components directory.
2
https://github.com/noah-janzen/news-recs
The constants folder contains modules with
general JavaScript objects like the app’s color
scheme.
The mobile app currently supports the languages
German and English. Depending on the system
language of the smartphone, the language of the
app is automatically adjusted. Localization files
are stored in the i18n directory.
The different screens of the app are located in the
screens folder.
The widely used Redux library is used to manage
state information within the mobile app. The Re-
dux configuration is performed in the store di-
rectory.
The util folder contains modules with helper
functions that are used in various React compo-
nents.
The environment variables of the app are stored in
an .env text file.
The starting point of the app is the App.tsx file.
This is where the navigation logic is located, im-
plemented using the React Navigation library.
4.2 Backend
The file structure of the backend is described below
(see Figure 2).
The auth directory contains the application’s au-
thentication logic. For authentication, the OAuth
2.0 protocol was implemented, which is an indus-
try standard for authentication.
The interactions folder contains the module,
which manages interactions between an user and
a news article.
The mail directory contains the module responsi-
ble for sending e-mails which is, e.g., required for
user registration or password reset.
The newsArticles folder contains the module
which provides an interface to load news recom-
mendations. News metadata is first loaded from
the news server. Depending on his group member-
ship, a user receives news recommendations pro-
vided by different algorithms. The membership of
a group depends on the ID of the user; thus, it is
de facto randomized. At this point, when used in
research, the recommendation model to be tested
and a baseline model must be implemented for
comparison. In this way, A/B-tests can be carried
out.
NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems
271
The surveys directory contains a module, which
provides an interface to create surveys and to
query and store survey responses.
The users directory contains a module, which
provides an interface for retrieving and editing the
user data of the logged-in user.
The app.module.ts file contains the root module
of the Nest application.
The main.ts file is the initial file of the applica-
tion that creates the nest application instance.
As in the frontend, there is an .env file in the
backend that contains all environment variables.
5 MAJOR FEATURES
This section briefly explains the main use cases of the
framework.
5.1 Create User Surveys
A core feature of the application is the creation of
surveys that users can answer in the app. This al-
lows quantitative and qualitative questions to be an-
swered. Surveys are defined by text files in JSON
format in the backend. A researcher can create
a new JSON file in the surveys/data/ path for
this purpose. The basic structure of this file can
be seen in Listing 1. The surveyId must be an
unique positive integer number. The startDate
property specifies the time when the survey is en-
abled for answering. Similarly, the endDate property
determines until when the survey can be answered.
The startsAfterDaysSinceRegistration prop-
erty specifies after how many days since registration
an user can view and answer the survey, since users
should use the app for a certain period of time be-
fore being surveyed. The durationInDays property
contains the time period in days that the user has to
answer the questions. The questions property is an
array that contains the questions.
Listing 1: Definition of a survey.
1 {
2 " sur ve yI d " : 0,
3 " s ta rtDat e " : " 2 0 22 - 07 -2 0T1 0 :0 0 : 00 .
000 + 00: 0 0" ,
4 " e nd Da te ": "2 0 22 -0 8 -2 7 T10 : 00 : 0 0. 0 0
0+0 0 :00 " ,
5 " st art sAf ter D ay s Si n ce R eg i st r at i on "
: 3 ,
6 " du ra ti o nI nD ays ": 7 ,
7 " q ue stion s " : [
8 // Q ue st io n Ob je cts
9 ]
10 }
Listing 2 shows the definition of a single ques-
tion object in the context of a concrete survey. It
contains the question in the supported languages of
the app. Currently, three different question type op-
tions are implemented (see Figure 3). Depending on
the question type, the property answerOptions must
be defined. This property is an array of multiple
answerOption objects. Such an object is identified
by an integer, non-negative value and the translations
for the answer.
Listing 2: Definition of a question object.
1 {
2 " q ue st io n " : {
3 "en " : " Ho w s at is fi ed are you
wit h the r ec o mm en d at io n
al go ri th m ?" ,
4 "de " : " Wi e z uf ri ed en bist Du mi t
dem E mp f eh l un gsa lg o ri t hm u s
?"
5 },
6 " q u es ti onT yp e " : " RA DIO ",
7 " an sw er Opt io ns " : [
8 {
9 " v alu e " : 4 ,
10 " a ns wer " : {
11 " en " : " Ver y s at is fied " ,
12 " de " : " Seh r z uf ri eden "
13 }
14 },
15 // Mo re an swe r op ti ons
16 ]
17 }
In the backend database, survey responses are
stored as individual documents. Users are able to
change previous answers while conducting a survey.
Once the user has completed the survey, no more an-
swers can be edited. Each document includes a times-
tamp indicating when the question was first answered,
the user id, the survey id, the question id, the ques-
tion type, the given answer and an optional timestamp
recording when the answer was last modified.
5.2 Evaluate Interaction Data
An interaction is a link between an user and a news
item that has been suggested to him. A distinction
is made between the following three types of interac-
tion:
(a) In the newsfeed, a news article is displayed to an
user in the visible area (viewport).
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
272
(a) Newsfeed. (b) Question type RADIO. (c) Question type RANGE. (d) Question type TEXT.
Figure 3: Screenshots of the mobile app. Subfigures (b), (c) and (d) show the different question types available for surveys.
(b) As in (a), but here the user presses on the news
article to read it.
(c) After reading a news article, a modal dialog opens
where the user can rate the recommendation. Rat-
ing a news recommendation is the third interac-
tion type.
A separate document is created in the database for
each step. For example, if a user sees a recommenda-
tion in the newsfeed, reads the corresponding article
and then rates it, three interactions are stored in the
database.
Three different input elements are available for
rating a recommendation (a binary input element, a
range input element and a free-text field). Before de-
ploying the app, a researcher has to decide between
one of these three input element types, which must
be specified in the .env file in the frontend. The
structure of a document depends on the interaction
type. All interaction documents contain the user ID,
the message article ID, and a timestamp. Interac-
tions of type (a) and (b) additionally store the flag
clicked, which indicates whether the user pressed
on the news article to read it. Interactions of type (c),
on the other hand, additionally store the input element
type (binary, range, or text) and the actual rating
(a string or a numeric value). The interactions can be
loaded via direct database access and analyzed locally
on a researcher’s computer, for example.
6 CONCLUSIONS
Newly proposed recommendation models need to be
evaluated to assess their effectiveness from the user’s
perspective. For this purpose, offline experiments,
user studies and online experiments are conducted
gradually. Especially in academic research, there
is rarely an opportunity to implement online exper-
iments because it usually does not have access to a
production system. To solve this problem, we devel-
oped a software framework that allows researchers to
conduct online experiments by using a mobile news
app with a recommendation engine.
First, 27 recent and related papers proposing new
recommendation techniques for news articles were
examined. In the majority of the research (22 pa-
pers, 81%), only offline experiments were conducted,
and no further user studies or online experiments were
performed. The most popular metrics for online eval-
uation are click-based metrics.
In general, the metrics used are highly dependent
on the production system used to implement the ex-
periment.
In a second step, a mobile app was developed that
displays news recommendations to users. Users are
divided randomly into two groups, each of which re-
ceives news recommendations from a different algo-
rithm. User interactions are recorded and stored in a
database. In addition, researchers can create surveys
that are displayed to users in the app.
We also conducted a usability study with three test
users to obtain feedback for future usability improve-
ments. All users had a positive impression of the app
NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems
273
and made some suggestions for improvement. For ex-
ample, it is important for the users that they receive
relevant recommendations in their native language.
With NewsRecs we want to contribute to the dis-
semination of online experiments for news recom-
mender in order to obtain more trustworthy results
in future work. We welcome the community to con-
tribute on GitHub and provide feedback in order to
jointly set the direction for future developments.
REFERENCES
Amir, N., Jabeen, F., Ali, Z., Ullah, I., Jan, A., and Kefalas,
P. (2022). On the current state of deep learning for
news recommendation. Artificial Intelligence Review.
An, M., Wu, F., Wu, C., Zhang, K., Liu, Z., and Xie, X.
(2019). Neural news recommendation with long- and
short-term user representations. In ACL 2019, pages
336–345. ACL.
Bai, B., Zhang, G., Lin, Y., Li, H., Bai, K., and Luo,
B. (2020). CSRN: Collaborative sequential rec-
ommendation networks for news retrieval. CoRR,
abs/2004.04816.
Castells, P. and Moffat, A. (2022). Offline recommender
system evaluation: Challenges and new directions. AI
Magazine, 43(2):225–238.
Ge, S., Wu, C., Wu, F., Qi, T., and Huang, Y. (2020). Graph
enhanced representation learning for news recommen-
dation. In WWW 2020, pages 2863–2869. ACM.
Gedikli, F., Stockem Novo, A., and Jannach, D. (2021).
Semi-automated identification of news story chains:
A new dataset and entity-based labeling method. In
INRA 2021, pages 29–42. CEUR-WS.org.
Hu, L., Li, C., Shi, C., Yang, C., and Shao, C. (2020a).
Graph neural news recommendation with long-term
and short-term interest modeling. Inf. Process. Man-
age., 57(2).
Hu, L., Xu, S., Li, C., Yang, C., Shi, C., Duan, N., Xie, X.,
and Zhou, M. (2020b). Graph neural news recommen-
dation with unsupervised preference disentanglement.
In ACL 2020, pages 4255–4264. ACL.
Koo, B., Jeon, H., and Kang, U. (2020). Accurate news
recommendation coalescing personal and global tem-
poral preferences. In PAKDD 2020, pages 78–90.
Springer.
Lee, D., Oh, B., Seo, S., and Lee, K.-H. (2020). News rec-
ommendation with topic-enriched knowledge graphs.
In CIKM 2020, pages 695–704. ACM.
Liu, D., Lian, J., Liu, Z., Wang, X., Sun, G., and Xie, X.
(2021). Reinforced anchor knowledge graph gener-
ation for news recommendation reasoning. In KDD
2021, pages 1055–1065. ACM.
Liu, D., Lian, J., Wang, S., Qiao, Y., Chen, J.-H., Sun, G.,
and Xie, X. (2020). KRED: Knowledge-aware doc-
ument representation for news recommendations. In
RecSys 2020, pages 200–209. ACM.
Mulder, M., Inel, O., Oosterman, J., and Tintarev, N.
(2021). Operationalizing framing to support multi-
perspective recommendations of opinion pieces. In
FAccT 2021, pages 478–488. ACM.
Nicholls, T. and Bright, J. (2019). Understanding news
story chains using information retrieval and network
clustering techniques. Communication Methods and
Measures, 13(1):43–59.
Okura, S., Tagami, Y., Ono, S., and Tajima, A. (2017).
Embedding-based news recommendation for millions
of users. In KDD 2017, pages 1933–1942. ACM.
Park, K., Lee, J., and Choi, J. (2017). Deep neural net-
works for news recommendations. In CIKM 2017,
pages 2255–2258. ACM.
Qi, T., Wu, F., Wu, C., Huang, Y., and Xie, X.
(2020). Privacy-preserving news recommendation
model learning. In EMNLP 2020, pages 1423–1432.
ACL.
Qiu, Z., Hu, Y., and Wu, X. (2022). Graph neural news rec-
ommendation with user existing and potential interest
modeling. ACM Trans. Knowl. Discov. Data, 16(5).
Ricci, F., Rokach, L., and Shapira, B., editors (2022). Rec-
ommender systems handbook. Springer, third edition.
Symeonidis, P., Kirjackaja, L., and Zanker, M. (2020).
Session-aware news recommendations using random
walks on time-evolving heterogeneous information
networks. User Modeling and User-Adapted Interac-
tion, 30:1–29.
Tian, Y., Yang, Y., Ren, X., Wang, P., Wu, F., Wang, Q.,
and Li, C. (2021). Joint knowledge pruning and recur-
rent graph convolution for news recommendation. In
SIGIR 2021, pages 51–60. ACM.
Wang, H., Wu, F., Liu, Z., and Xie, X. (2020). Fine-grained
interest matching for neural news recommendation. In
ACL 2020, pages 836–845. ACL.
Wang, H., Zhang, F., Wang, J., Zhao, M., Li, W., Xie, X.,
and Guo, M. (2019). Exploring high-order user pref-
erence on the knowledge graph for recommender sys-
tems. ACM Trans. Inf. Syst., 37(3).
Wang, H., Zhang, F., Xie, X., and Guo, M. (2018). DKN:
Deep knowledge-aware network for news recommen-
dation. In WWW 2018, pages 1835–1844. IW3C2.
Wu, C., Wu, F., An, M., Huang, J., Huang, Y., and Xie,
X. (2019a). Neural news recommendation with atten-
tive multi-view learning. In IJCAI 2019, pages 3863–
3869. AAAI Press.
Wu, C., Wu, F., An, M., Huang, J., Huang, Y., and Xie,
X. (2019b). NPA: Neural news recommendation with
personalized attention. In KDD 2019, pages 2576–
2584. ACM.
Wu, C., Wu, F., Ge, S., Qi, T., Huang, Y., and Xie, X.
(2019c). Neural news recommendation with multi-
head self-attention. In EMNLP-IJCNLP 2019, pages
6389–6394. ACL.
Wu, C., Wu, F., Qi, T., and Huang, Y. (2021). Empower-
ing news recommendation with pre-trained language
models. In SIGIR 2021, pages 1652–1656. ACM.
Yoneda, T., Kozawa, S., Osone, K., Koide, Y., Abe, Y., and
Seki, Y. (2019). Algorithms and system architecture
ICAART 2023 - 15th International Conference on Agents and Artificial Intelligence
274
for immediate personalized news recommendations.
In WI 2019, pages 124–131. ACM.
Zhang, Q., Jia, Q., Wang, C., Li, J., Wang, Z., and He,
X. (2021). AMM: Attentive multi-field matching for
news recommendation. In SIGIR 2021, pages 1588–
1592. ACM.
Zhao, Q., Chen, X., Zhang, H., and Ma, S. (2021). D-HAN:
Dynamic news recommendation with hierarchical at-
tention network. CoRR, abs/2112.10085.
Zheng, G., Zhang, F., Zheng, Z., Xiang, Y., Yuan, N. J.,
Xie, X., and Li, Z. (2018). DRN: A deep reinforce-
ment learning framework for news recommendation.
In WWW 2018, pages 167–176. ACM.
NewsRecs: A Mobile App Framework for Conducting and Evaluating Online Experiments for News Recommender Systems
275