An Explainable Knowledge Graph-Based News Recommendation
System
Zühal Kurt
a
, Thomas Köllmer
b
and Patrick Aichroth
c
Fraunhofer IDMT, Ilmenau ILM 98693, Germany
{zuehal.kurt, patrick.aichroth, thomas.koellmer}@ idmt.fraunhofer.de
Keywords: Explainability, Recommendation System, Knowledge Graphs, News.
Abstract: The paper outlines an explainable knowledge graph-based recommendation system that aims to provide
personalized news recommendations and tries to explain why an item is recommended to a particular user.
The system leverages a knowledge graph (KG) that models the relationships between items and users’
preferences, as well as external knowledge sources such as item features and user profiles. The main
objectives of this study are to train a recommendation model that can predict whether a user will click on a
news article or not, and then obtain the explainable recommendations for the same purpose. This is achieved
with three steps: Firstly, KG of the MIND dataset are generated based on the history and, the clicked
information of the users, the category and subcategory of the news. Then, the path reasoning approaches are
utilized to reach explainable paths of recommended news/items. Thirdly, the proposed KG-based model is
evaluated using MIND News data sets. Experiments have been conducted using the MIND-demo and MIND-
small datasets, which are the open-source English news datasets for public research scope. Experimental
results indicate that the proposed approach performs better in terms of recommendation explainability, making
it a promising basis for developing transparent and interpretable recommendation systems.
1 INTRODUCTION
In recent years, knowledge graphs have emerged as a
powerful tool for representing and organizing large
amounts of structured and unstructured data. One of
the domains in which knowledge graphs can be used
are explainable recommendation systems, which aim
at providing personalized recommendations to users
based on their preferences and transactions in a more
transparent and interpretable way than traditional
recommendation systems. They allow users to
understand how recommendations are generated and
provide insights into why certain recommendations
are proposed, thereby improving transparency,
persuasiveness, effectiveness, trustworthiness, and
user satisfaction with recommendations (Zhang,
2020).
Traditional news recommendation systems have
relied on various techniques, such as collaborative
filtering (Das, 2007, Lian 2018), content-based
filtering (Joseph, 2019, Huang, 2013), and hybrid
a
https://orcid.org/0000-0003-1740-6982
b
https://orcid.org/0000-0002-2715-3766
c
https://orcid.org/0000-0003-4777-6335
recommendation methods (Meguebli, 2014).
However, these methods often face limitations in
capturing the complex relationships and nuances
present in news articles. Furthermore, these methods
disregard the sequential information in the user’s
browsing history, making it difficult to learn users’
changing interests, and ignoring the importance of
entities in the news article (Zeng, 2022).
To address these challenges, knowledge
embeddings can be considered as an effective tool in
the field of news recommendation. One of the key
advantages of knowledge embeddings is their ability
to capture both the explicit and implicit semantic
relationships between news articles. By incorporating
various contextual concepts, such as word co-
occurrence patterns, syntactic structures, and
semantic similarities, knowledge embeddings can
effectively connect the news articles, even in the
absence of explicit relations. Furthermore,
knowledge embeddings facilitate the integration of
external knowledge sources, such as knowledge
214
Kurt, Z., Köllmer, T. and Aichroth, P.
An Explainable Knowledge Graph-Based News Recommendation System.
DOI: 10.5220/0012161300003598
In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 1: KDIR, pages 214-221
ISBN: 978-989-758-671-2; ISSN: 2184-3228
Copyright © 2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
graphs and ontologies, into the recommendation
process. By leveraging the structured information
encoded in these knowledge sources, the embeddings
can enrich the semantic representation of news
articles, enabling more accurate and context-aware
recommendations (Liu, 2020, Sheu, 2021). This
integration opens up possibilities for incorporating
additional information, such as news article metadata,
user profiles, and social network connections, to
generate explainable recommendations.
To enhance the quality and accuracy of the
knowledge embedding based recommendations, path
reasoning models have emerged as a promising
approach. These models leverage the rich contextual
information embedded in users’ browsing histories to
infer their interests and preferences. However, these
path reasoning models encounter significant
challenges due to the dynamic nature of news
consumption patterns, the need to balance serendipity
and user preferences, and the ever-present risk of
filter bubbles (Tai, 2021, Zheng, 2018). In this study,
we will examine the complexities and obstacles
associated with path reasoning models in news
recommendation systems, exploring the limitations
they encounter and the potential solutions that can
enhance their performance in delivering diverse,
more precise, and reliable recommendations.
As traditional media platforms struggle to keep up
with the proliferation of online content, news
recommendation systems have emerged as essential
tools for personalized news delivery. These systems
aim to tailor news recommendations to users’
preferences and provide them with news articles of
interest. However, while news recommendation
systems offer convenience and customization, they
also face numerous challenges. From the filter bubble
effect to information overload and algorithmic bias,
these systems encounter complex obstacles that affect
the diversity, accuracy, and reliability of the news
individuals receive. In this context, understanding
and addressing the challenges of news
recommendation systems is crucial to promote a well-
informed society and responsible news consumption.
From filter bubbles and echo chambers to bias and
ethical considerations, navigating the complex
landscape of news recommendation presents
obstacles that require careful attention. Recent studies
have explored these challenges faced by news
recommendation systems and shed light on the
importance of balancing between personalization and
diversity in providing accurate and unbiased news
recommendations (Qi, 2022, Bernstein,
2020 and Chen
2023)
.
Knowledge embeddings are a means to that end,
providing rich information of users and items that can
be exploited to generate intuitive and more suitable
explanations for the recommended items. In this
study, we will explore the concept of explainable
recommendation systems based on knowledge
graphs, outlining some key terms, and benefits and
challenges regarding their development and
implementation for the domain of news
recommendations.
In Section 2, we delve into the existing
approaches that augment traditional models with
KGs for the explainable news recommendations.
In Section 3, we will provide an initial overview
of the recommendation pipeline, to characterize
how explanations can be enabled at several
stages, namely, data acquisition and storage, data
preprocessing, model training, model prediction,
model evaluation, and recommendation delivery.
Then, the details of the proposed model that
relies on path reasoning models are given to
access high-level relationships between users
and products, according to the KG structure, in
Section 3.
Finally, we show how different types of
explanations can emerge from a news
recommendation system via KGs.
2 KG FOR EXPLAINABLE NEWS
RECOMMENDATIONS
Knowledge embeddings are learned over the graph to
reach the embeddings from each user, item, relation,
and entity, then the recommendations are generated
for a particular user to find the most preferable item
under the main relation of the recommendation
system. The main relations can be used for that
purpose are ‘purchase’, ‘watch’ and ‘click’ etc., and
recommendations can be yielded using the shortest
path from the user to the respective item through the
knowledge graph (KG). Furthermore, to also address
the demand for explainability, several approaches can
be applied, including embedding-based, connection-
based, and propagation-based methods (Balloccu,
2022). That includes connection-based explainable
KG-based models, which utilize path embedding
approaches, and Policy Guided Path Reasoning
(PGPR) and CoArse-to-FinE neural symbolic
reasoning method (CAFÉ), which represent the
current state-of-the-art: PGPR can be considered an
adaptation of Reinforcement Learning (RL), in which
an agent starts from a given user, and learns to
An Explainable Knowledge Graph-Based News Recommendation System
215
navigate to the potential items of interest, such that
the path history can provide an explanation for why
the item is recommended to the user (Xian, 2019).
CAFÉ, on the other hand, represent a commonly used
path reasoning model that is based on finding a coarse
sketch of past user behaviour, and then conducting
path reasoning to derive recommendations based on
the user model for fine-grained modelling (Xian,
2020). The effectiveness of PGPR and CAFE
approaches have been evaluated for several Amazon
e-commerce domains, obtained explainable
reasoning paths in (Xian, 2019 & 2020). However,
utilizing the KG path reasoning algorithms for news
recommendation has not been fully examined, while
existing studies are utilizing KG try to improve news
recommendation accuracy.
Another path reasoning algorithm for news
recommendation via KG is Anchor-KG, which has
been proposed in (Liu, 2021). The main purpose of
this algorithm is to develop a policy network to
generate a subgraph from the KG for each news
article. Each subgraph needs to include key entities
from the news article and some necessary
neighbouring entities which connect the article
entities via multi-hop relational paths. The knowledge
subgraph is called “anchor graph” since it only uses
the most essential knowledge entities from the
exponential growth of multi-hop relations and
produces a footprint for each news article over the
KG. Hence, the knowledge-aware reasoning of any
two news articles can be conducted simply using the
interactions of their corresponding anchor graphs.
Anchor-KG utilizes an RL-based optimization
framework for training, and experiments are applied
on two real-world news datasets. However, while this
approach can be utilized for enhancing document
representation and providing knowledge reasoning
between news documents, it cannot generate
explainable recommendations.
Being inspired by the success of leveraging KGs,
deep knowledge-aware network (DKN) has also been
proposed for recommendation (Wang, 2018). DKN is
a content-based recommendation model for click-
through rate (CTR) predictions that utilizes users’
click history as input, and generates outputs based on
the probability of the user clicking the news. This
model is trying to take advantage of external
knowledge for news recommendation, especially
generating the KG by associating each word in the
news content with a relevant entity in (Wang, 2018).
The set of contextual entities of each entity is
searched and used to provide more complementary
and distinguishable information. After the user’s
embeddings and the candidate news’ embeddings are
generated based on the entities, a deep neural network
(DNN) is applied on these embeddings for CTR
prediction in the DKN model.
To provide meaningful news recommendations, it
is also necessary to incorporate additional KG
information (Wang, 2018). Several academic studies
are utilized such as NELL and DBpedia, as well as
commercial ones such as Google Knowledge Graph
and Microsoft Satori to generate KGs from a dataset
(Wang, 2018, Zhang, 2016). These KGs have been
successfully implemented in text classification,
machine reading, and word embedding research areas
but not yet for recommendation systems. To combine
user behaviour and news content information to
generate a KG, the News Graph (NG) model was
proposed (Liu, 2019), which include topic entities and
collaborative relations that are relevant for news
recommendations, while removing news-irrelevant
relations. User behaviours are incorporated, and three
types of collaborative relations are constructed
depending on the co-occurrence in the same news,
clicked by the same user, or clicked by the same user
in the same browsing session. The extraction of news-
relevant relations and enhancement of collaborative
information provides increased capabilities for
representing news articles and users’ reading
behaviours as a result. Hence, applications include
not only news recommendations, but also news
category classification, news popularity prediction
and local news detection.
Based on the above, an explainable news
recommendation system utilizing a KG graph model
is presented in this study. Firstly, conceptual
foundations are introduced, and the real-world
examples is described to understand how the KGs can
be integrated into the recommendation pipeline, also
for the aim of providing explanations. Using an
example algorithmic solution to model, integrate,
train, and evaluate a recommendation system with
KGs based on explainability perspective, we aim at
training a recommendation model that can predict
whether a user would click on a news article or not.
The implementation can be split in three domains: (1)
generation of metadata and KG of the MIND News
dataset based on (a) user history, (b) user clicks and
(c) news categories and subcategories. (2) PGPR and
CAFÉ approaches are utilized to reach the
explainable paths of recommended items/lists (3) the
proposed KG-based model is evaluated using two
real-world public MIND-News data sets.
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
216
3 THE PROPOSED KG-BASED
RECOMMENDATION MODEL
The following presents a new explainable KG-based
algorithm for personalized recommendations. With
its ability to incorporate a large amount of knowledge
from various sources and represent it in a
comprehensible and transparent manner, it can
provide users with recommendations that are not only
accurate but also easy to understand. It can be
especially useful in domains where trust and
transparency are crucial. We use the publicly
available MIND News datasets which vary in
domain, extensiveness, and sparsity. For scalability
purposes, the reduced version of this dataset is
utilized and obtained with the following steps:
In the first step, the raw news data is cleaned and
pre-processed to generate explainable
recommendations. The raw dataset is composed by
the following files and illustrated in Figure 1(a):
behaivours.tsv: List of users and some
demographical data.
news.tsv: the catalog of news and the entities.
A dataset reduced to its K-cores (i.e., dense
subsets) is a subset with removed items and users,
such that each of the remaining users and items have
k reviews each. Hence, these datasets are reduced to
its 5-cores and transformed to the standardized
format. Then, simple time-based data and knowledge
graph embeddings are obtained for training stage and
finally, the proposed Explainable Knowledge Graph-
Based Recommendation Model (EKG-RM) is
evaluated on this metadata. The standardized KG
model is composed by 4 different main files and
generated as in (Balloccu, 2022) and illustrated in
Figure 1(b):
(a)
(b)
Figure 1: a) The raw data and b) the standardized KG model
of the Mind News Dataset (Balloccu, 2022).
Item to KG (i2kg_map.tsv) includes the mapping
between the News dataset and the corresponding
entity in the KG.
Entity map (e_map.dat) includes the set of
entities and contains all the unique entities
present in the triplets including item entities.
Relation map (r_map.dat) includes set of
relations and contains all the unique relations
present in the triplets.
KG (kg.dat) includes set of triplets (the graph
itself) and contains all the triplets of the form
(entity_head, relation, entity_tail).
The relations are extracted from behaviors and the
news files of the MIND dataset. The history, the non-
clicked and clicked information from behaviors file
used to define history and clicked relations. The
category, and subcategory information from the news
file used to define same category relations. Hence
three different relations are defined in the KG model
of MIND dataset which are history, clicked and same
category. The detailed of the relations and KG model
is given in Section 3.1.
In the second step, the path reasoning models for
KG-based recommendation systems are applied on
the proposed model. A path-reasoning algorithm
starts from specific user and proceeds through the
graph to discover the preferable items in the graph for
the target user. The objective is that if the system
bases its results on an explicit reasoning path, it is
easy to interpret the reasoning process leading to each
recommendation, i.e., providing the relevant
reasoning paths in the graph as interpretable evidence
for why a particular recommendation is made. For
instance, considering the user A, the proposed model
is trying to find candidate News B and News C, along
with their explainable paths in the graph, as shown in
Figure 2, e.g.,
𝑈𝑠𝑒𝑟 𝐴 𝑁𝑒𝑤𝑠 𝐴 𝐶𝑙𝑖𝑐𝑘𝑒𝑑 𝑏𝑦
𝑈𝑠𝑒𝑟 𝐵 𝐴𝑙𝑠𝑜_𝐶𝑙𝑖𝑐𝑘𝑒𝑑 𝑁𝑒𝑤𝑠 𝐵 𝑎𝑛𝑑 𝑈𝑠𝑒𝑟 𝐴
𝑁𝑒𝑤𝑠 𝐴 𝐵𝑒𝑙𝑜𝑛𝑔 𝑡𝑜 𝑆𝑎𝑚𝑒 𝐶𝑎𝑡𝑒𝑔𝑜𝑟𝑦 𝑁𝑒𝑤𝑠 𝐷
.
The path-based approaches implement a path
selection algorithm or define a set of meta-path
patterns to constrain the path search space due to the
large number of nodes and edges in the KG,
(Balloccu, 2022). The path-based approaches rely on
pre-computed paths (tuples) that model high-order
relationships between users and items/news or
perform path reasoning to conduct recommendation
and path retrieval simultaneously according to the KG
structure. The path-based reasoning approaches can
extract one or multiple explainable paths between the
recommended item and already experienced items.
These paths can be translated into textual
explanations for the end users. For instance, in the
news domain, a path between the news already
An Explainable Knowledge Graph-Based News Recommendation System
217
Figure 2: A sample of KG-based structure for Mind News Dataset (Balloccu, 2022).
clicked by the user (news1) and a news recommended
to that user (news2), shaped in the form of user1
clicked news1 belongs to same category1 of the
news2, can be used to provide the explanation “news2
is recommended because you clicked another news1
belongs to same category1 of news2”.
Unfortunately, the path generation and selection
are not optimized for the recommendation purpose
and lead to sub-optimal recommendations and
explainable paths. Hence, path reasoning models,
more specifically connection-based explainable KG-
based approaches are proposed to overcome this issue
and to optimize the recommendation model while
searching for paths in the KG. The most recent path-
based reasoning algorithms that enable textual
explanations, such as PGPR and CAFÉ models, are
utilized for this purpose. These models depend on RL
framework to optimize recommendations by
navigating paths between users and recommended
items in the KG, (Balloccu, 2022).
The PGPR model relies on an RL agent that is
conditioned to the user and trained to navigate to
potentially relevant items. The RL agent can be
performed an explicit multi-step path reasoning over
the graph with starting from a given user node to find
out appropriate items in the graph for the target user.
Then, the path from the user 𝑢 to the item 𝑖 can be
used to explain the recommendation (Balloccu, 2022
& Xian, 2019).
The CAFÉ model (Xian, 2020) creates a
personalized user profile based on transactions of the
user in the KG and utilizes a neural symbolic
reasoning modules in the path reasoning stage. A
layout tree is generated with the modules based on the
user profile, then this tree is used by the path
reasoning algorithm to generate a set of
recommendation paths. The path inference is
structured using a layout tree which offers more
efficiency compared to PGPR. However, the CAFÉ
model is also needed for profile guided path
reasoning (Balloccu, 2022, Xian, 2020).
3.1 The Generated KG Model
In particular, the collaborative edges/relations are
extracted in four different ways: (1) in history of the
same user; (2) clicked by the same user; and (3)
clicked by the same user in the same browsing session
(4) the news in the same category or subcategory. In
this paper, the objective of the KG-based
recommendation system is to recommend news to
users and explain why the news are recommended. A
knowledge base as a set of triplets 𝑇 =
𝑓(𝑒𝑛𝑡𝑖𝑡𝑦_ℎ𝑒𝑎𝑑, 𝑒𝑛𝑡𝑖𝑡𝑦_𝑡𝑎𝑖𝑙, 𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛) or
recommendation are generated whether the relation
denotes the relationship between entity head and
entity tail. Then the proposed explainable
recommendation has two different purposes, the first
one is to find one or a set of items 𝑖 that are most
likely to be clicked by the user 𝑢, and the second one
is to generate an explanation based on 𝑇 to explain
why the user should click the item. Three types of
entities (i.e., entity_head or entity_tail) for
explainable recommendations are considered for
MIND News dataset:
user: the users of the recommendation system.
item: the news in the system to be recommended.
category: the categories that the news belongs to.
Furthermore, we consider 5 different types of
relationships between entities:
History: the relation from a user to an item/news,
which means that the user has already clicked
and read that news.
Clicked: the relation from a user or an item,
which means that the user has clicked the
item/news.
Belongs_to: the relation from an item to a
category, which means that the item/news
belongs to the same category.
Also_ History: the relation from an item to
another item, which means these news/items
have been already read by the same user.
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
218
Also_ Clicked: the relation from an item to
another item, which means this items/news have
been already clicked by the same user.
The relations are extracted from behaviours and
the news files of the MIND dataset. The history, the
non-clicked and clicked information are reached from
behaviours file to define history and clicked relations.
The category, and subcategory information are
reached from the news file to define belongs to same
category relations. The title entities, abstract entities
from news file can be also used to extract more
relations, but it is considered as a future work. The
KG standard format for MIND News dataset is
obtained as in (Balloccu, 2022).
To generate explainable news recommendation,
firstly the main relation clicked relationship between
user 𝑢 and item 𝑖 needs to be extracted, secondly
explanations for each retrieved user-item (𝑢, 𝑖) pair
based on the relations and entities related to them
needs to be provided. The following explainable
recommendations can be deduced from Figure 2:
The News B is recommended to User A, since
the News A is in the history of User A and
clicked by the User B.
The News D is recommended to User B, since
News A and News B belong to same category of
News D that are already clicked by the User B.
The News C is recommended to User C, since
News A is in the history of User C and clicked
by the User B.
Table 1: Detailed statistics of the MIND News dataset.
Dataset MIND-Demo MIND-Small
Users 5,000 65,238
News 28,603 94,057
Interactions 80,088 829,162
Relation T
yp
es 5 5
Tri
p
lets 634,630 4,066,384
4 EXPERIMENTAL ANALYSIS
AND EVALUATION METRICS
The proposed algorithm is applied on the open-source
English MIND News dataset (MIND website) that
was collected from Microsoft News website in 2019.
Users were randomly selected who had at least 5 news
clicks, and each user is hashed into an anonymized ID
to protect user privacy. The news click behaviours of
the users were formatted into impression logs and
valued as 1 for click and 0 for non-click. In addition,
small versions of MIND-Small and MIND-Demo
dataset were released by randomly sampling 50,000
and 5,000 users and their behaviour logs.
Table 2: The experimental results based on the evaluation
metrics of the MIND-Demo and MIND-Small datasets.
Datasets Metrics
Methods
Precision
(%)
Recall
(%)
Hit Ratio
(%)
NDCG
(%)
MIND-Demo PGPR 1.17 3.11 10.45 4.40
CAFE 2.85 8.02 5.17 6.28
MIND-Small PGPR 0.55 2.66 5.22 2.41
CAFE 1.28 5.88 2.33 3.97
The testing methodology adopted in this study is
the same as in a previous study (Balloccu, 2022). The
implicit ratings are split by two subsets that are named
by training and test sets for each dataset. Each split is
consisting of 80% for training and 20% for testing
sets. The detailed statistics of the utilized MIND
datasets are given in Table 1.
The four different top-N recommendation
measurements are utilized for evaluation as in
(Balloccu, 2022). The precision can be defined as the
percentage of correctly recommended items in the test
set that are also presented in a user’s the top-N
recommendation list. The precision of the
recommendation system can be indicated as:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
(
𝑁
)
=
#ℎ𝑖𝑡𝑠
|
𝑇
|
∙𝑁
where |𝑇| is the number of test ratings, 𝑁 is the length
of the recommendation list.
The recall can be defined as the percentage of
items in the test set that are also presented in a user’s
the top-N recommendation list. The recall of the
recommendation system can be represented as:
𝑅𝑒𝑐𝑎𝑙𝑙
(
𝑁
)
=
#ℎ𝑖𝑡𝑠
|
𝑇
|
where
|
𝑇
|
is the number of test ratings.
The NDCG takes into consideration the position
of correctly recommended items in the list of top-N
recommendations and is evaluated as the average of
all test users of the NDCG.
The hit-ratio is measured by looking at the number
of hits, i.e., the number of items in the test set that are
also presented in the top-N recommendation item list
returned for each user. Then, the hit-ratio of the
recommendation system can be represented as:
𝐻𝑖𝑡 − 𝑅𝑎𝑡𝑖𝑜
(
𝑁
)
=
#ℎ𝑖𝑡𝑠
𝑛
where 𝑛 is represented the total number of users,
#ℎ𝑖𝑡𝑠 is symbolized as the overall hit of the
recommendation system. The top-N recommendation
An Explainable Knowledge Graph-Based News Recommendation System
219
list is generated for each user in the testing set, where
𝑁 is selected 10.
The experimental results based on the evaluation
metrics of the MIND-Demo and MIND-Small
datasets are given in Table 2. The experimental
results demonstrate that the CAFÉ model performs
better than PGPR model on the MIND News datasets.
It can be concluded from the experiments results that
are encouraging; however, it needs to be further
improved in the future works. For example, these
datasets can be reduced to its 10-cores in the pre-
processing stage and transformed to the standardized
format as in (Ballocu, 2022).
4.1 The Case Study of the Explainable
Recommendation Generation
The generated predicted path file from the KG-based
recommendation system is constructed as in the
tutorial (Balloccu, 2022). Firstly, the explainable
recommendation is extracted for a random User #ID:
U8619 via path reasoning approaches. Then the meta-
path is extracted for the same user #ID: U8619 as
illustrated in Figure 3. Furthermore, the case study for
a random User #ID: 2683 is also given in Table 3.
Figure 3: The extracted meta-path for the random User #ID:
U8619.
To show the ability of the proposed model to generate
knowledge-enhanced explanations, a case study for a
random test user is conducted from MIND-Small
Dataset, for whom we have examined that the first
recommendation provided by the system is correct.
The top-5 explainable recommendation list for the
User #ID:2683 are listed along with their probabilities
computed based on PGPR models in Table 3.
Table 3: The top-5 explainable recommendation list for a
random User #ID: 2683.
Probability
Entity
Hea
d
Relation
T
yp
e
Entity
Tail
Relation
Type
Recommendation
0.572
User
2683
clicked
User
1336
clicked News 12289
0. 527
User
2683
clicked
User
3530
clicked News 6216
0.498
User
2683
clicked
User
2791
clicked News 10022
0.424
User
2683
clicked
User
2200
clicked News 21667
0.186
User
2683
belongs_to
News
2322
belongs_to News 9718
5 CONCLUSIONS
Explainable recommendation systems based on KGs
have emerged as a promising approach to address the
challenges of traditional recommendation systems.
By leveraging KGs, such systems can model complex
relationships between entities and provide
personalized recommendations that are more accurate
and diverse. Moreover, the explainability aspect of
these systems enables users to understand the
underlying reasoning behind the recommendations,
thereby improving their trustworthiness.
In this paper, an explainable KG-based news
recommendation model was proposed. The approach
applies KG graph generation combining extracted
news metadata and five different relationships, and
state-of-art PGPR and CAFE algorithms to find
explanation paths between a user and the
recommended items in the KG to explain the
recommendations. Experimental results on real-
world MIND-Small and MIND-Demo datasets
indicate flexibility of the model to incorporate
multiple relation types and show that the proposed
approach offers a promising solution for explainable
news recommendations to users. As for directions for
future work, the explainable recommendation quality
needs further improvement by extending the KG and
the selecting appropriate parameters for the MIND-
News dataset. Another objective as a future work is
to advance the development and implementation of
privacy-preserved news explainable recommendation
systems that prioritize user privacy and provide
enhanced user control over their personal data.
ACKNOWLEDGEMENTS
This work was supported by European Union´s
Horizon 2020 research and innovation programme
under grant number 951911.
REFERENCES
Zhang, Y., Chen, X. (2020). Explainable recommendation:
A survey and new perspectives. Foundations and
Trends® in Information Retrieval, 14(1), 1-101.
Balloccu, G., Boratto, L., Fenu, G., Marras, M. (2022).
Hands on explainable recommender systems with
knowledge graphs. In Proceedings of the 16th ACM
Conference on Recommender Systems, pp. 710-713.
MIND Dataset website: https://github.com/msnews, last
accessed 2023/04/25.
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
220
Wu, F., Qiao, Y., Chen, J. H., Wu, C., Qi, T., Lian, J., Zhou,
M. (2020). MIND: A large-scale dataset for news
recommendation. In Proceedings of the 58th Annual
Meeting of the Association for Computational
Linguistics, pp. 3597-3606.
Xian, Y., Fu, Z., Muthukrishnan, S., De Melo, G., Zhang,
Y. (2019). Reinforcement knowledge graph reasoning
for explainable recommendation. In Proceedings of the
42nd International ACM SIGIR Conference on
Research and Development in Information Retrieval,
pp. 285-294.
Xian, Y., Fu, Z., Zhao, H., Ge, Y., Chen, X., Huang, Q.,
Geng, S., Qin, Z., Melo, G., Muthukrishnan, S., Zhang,
Y. (2020). CAFE: Coarse-to-Fine neural symbolic
reasoning for explainable recommendation. In
Proceedings of the 29th ACM International Conference
on Information & Knowledge Management, pp.1645-
1654.
Liu, D., Lian, J., Liu, Z., Wang, X., Sun, G., Xie, X. (2021).
Reinforced anchor knowledge graph generation for
news recommendation reasoning. In Proceedings of the
27th ACM SIGKDD Conference on Knowledge
Discovery & Data Mining, pp. 1055-1065.
Wang, H., Zhang, F., Xie, X., Guo, M. (2018). DKN: Deep
knowledge-aware network for news recommendation.
In Proceedings of the 2018 world wide web conference,
pp. 1835-1844.
Liu, D., Bai, T., Lian, J., Zhao, X., Sun, G., Wen, J. R., Xie,
X. (2019). News Graph: An enhanced knowledge graph
for news recommendation. In Proceedings of the 28th
ACM International Conference on Information &
Knowledge Management KaRS@ CIKM, pp. 1-7.
Balloccu, G., Boratto, L., Fenu, G., Marras, M. (2022). Post
processing recommender systems with knowledge
graphs for recency, popularity, and diversity of
explanations. In Proceedings of the 45th International
ACM SIGIR Conference on Research and Development
in Information Retrieval, pp. 646-656.
Zhang, F., Yuan, N. J., Lian, D., Xie, X., Ma, W. Y. (2016).
Collaborative knowledge base embedding for
recommender systems. In Proceedings of the 22nd
ACM SIGKDD international conference on knowledge
discovery and data mining, pp. 353-362.
Das, A. S., Datar, M., Garg, A., & Rajaram, S. (2007).
Google news personalization: scalable online
collaborative filtering. In Proceedings of the 16th
international conference on World Wide Web, pp. 271-
280.
Joseph, K., & Jiang, H. (2019). Content based news
recommendation via shortest entity distance over
knowledge graphs. In Companion Proceedings of the
2019 World Wide Web Conference, pp. 690-699.
Huang, P. S., He, X., Gao, J., Deng, L., Acero, A., & Heck,
L. (2013). Learning deep structured semantic models
for web search using clickthrough data. In Proceedings
of the 22nd ACM international conference on
Information & Knowledge Management, pp. 2333-
2338.
Meguebli, Y., Kacimi, M., Doan, B. L., & Popineau, F.
(2014). Stories Around You- a Two-Stage Personalized
News Recommendation. In Proceedings of the
International Joint Conference on Knowledge
Discovery, Knowledge Engineering and Knowledge
Management, Volume 1, pp. 473-479.
Zeng, Y., Du, J., Xue, Z., & Li, A. (2022). Scientific and
Technological News Recommendation Based on
Knowledge Graph with User Perception. In 2022 IEEE
8th International Conference on Cloud Computing and
Intelligent Systems (CCIS), pp. 491-495. IEEE.
Lian, J., Zhang, F., Xie, X., & Sun, G. (2018). Towards
Better Representation Learning for Personalized News
Recommendation: A Multi-Channel Deep Fusion
Approach. In Proceedings of the Twenty-Seventh
International Joint Conference on Artificial
Intelligence (IJCAI), pp. 3805-3811.
Liu, D., Lian, J., Wang, S., Qiao, Y., Chen, J. H., Sun, G.,
& Xie, X. (2020). KRED: Knowledge-aware document
representation for news recommendations. In
Proceedings of the 14th ACM Conference on
Recommender Systems, pp. 200-209.
Sheu, H. S., Chu, Z., Qi, D., & Li, S. (2021). Knowledge-
guided article embedding refinement for session-based
news recommendation. IEEE Transactions on Neural
Networks and Learning Systems, 33(12), 7921-7927.
Tai, C. Y., Huang, L. Y., Huang, C. K., & Ku, L. W. (2021).
User-centric path reasoning towards explainable
recommendation. In Proceedings of the 44th
International ACM SIGIR Conference on Research and
Development in Information Retrieval, pp. 879-889.
Zheng, G., Zhang, F., Zheng, Z., Xiang, Y., Yuan, N. J.,
Xie, X., & Li, Z. (2018). DRN: A deep reinforcement
learning framework for news recommendation. In
Proceedings of the 2018 World Wide Web Conference,
pp. 167-176.
Qi, T., Wu, F., Wu, C., Sun, P., Wu, L., Wang, X., ... & Xie,
X. (2022, July). ProFairRec: Provider fairness-aware
news recommendation. In Proceedings of the 45th
International ACM SIGIR Conference on Research and
Development in Information Retrieval, pp. 1164-1173.
Bernstein, A., De Vreese, C., Helberger, N., Schulz, W.,
Zweig, K., Baden, C., ... & Zueger, T. (2020). Diversity
in news recommendations. arXiv preprint
arXiv:2005.09495.
Chen, J., Dong, H., Wang, X., Feng, F., Wang, M., & He,
X. (2023). Bias and debias in recommender system: A
survey and future directions. ACM Transactions on
Information Systems, 41(3), 1-39.
Balloccu, G., Boratto, L., Fenu, G., & Marras, M. (2022).
XRecSys: A framework for path reasoning quality in
explainable recommendation. Software Impacts, 14,
100404.
An Explainable Knowledge Graph-Based News Recommendation System
221