An Analysis of Knowledge Representation for Anime Recommendation

Using Graph Neural Networks

Yuki Saito

, Shusaku Egami

2,1 a

, Yuichi Sei

1 b

, Yasuyuki Tahara

1 c

and Akihiko Ohsuga

1 d

Department of Informatics, University of Electro-Communications 1–5–1 Chofugaoka, Chofu, Tokyo, 182–8585, Japan

National Institute of Advanced Industrial Science and Technology (AIST), 2–4–7 Aomi, Koto-ku, Tokyo, 135–0064, Japan

Keywords:

Knowledge Graph, Graph Neural Networks, Recommender System.

Abstract:

In recent years, entertainment content, such as movies, music, and anime, has been gaining attention due

to the stay-at-home demand caused by the expansion of COVID-19. In the content domain, research in the

ﬁeld of knowledge representation is primarily concerned with accurately describing metadata. Therefore,

different knowledge representations are required for applications in downstream tasks. In this study, we aim

to clarify effective knowledge representation through a case study of recommending anime works. Thus, we

hypothesized how to represent anime works knowledge to improve recommendation performance from both

quantitative and qualitative aspects and veriﬁed the hypotheses by changing the knowledge representation

structure according to the hypothesis. Initially, we collected data about anime works from multiple data

sources and integrated them to construct a knowledge graph (KG). We also prepared several KGs by varying

the knowledge conﬁguration. Subsequently, we compared the recommendation performance of each KG as

an input to the graph neural networks. As a result, it was found that the amount of semantic relationships

was proportional to the recommendation performance and that the properties that can characterize the work

contributed to the recommendation.

1 INTRODUCTION

The COVID-19 pandemic has caused people to spend

more time at home, such as telecommuting and dis-

tance learning. Accordingly, entertainment content

including movies and music has gained signiﬁcant

attention due to the popularization of subscription-

based services. For instance, in 2020, Netﬂix ac-

quired approximately 37 million new subscribers

worldwide (Soldo, Lana and Schagerl, Christopher,

2023). Among much research in the content domain,

there is the ﬁeld of knowledge representation, which

covers the construction of systematic databases based

on content metadata. Knowledge Graph (KG) is a

core technology in the ﬁeld of knowledge represen-

tation and can represent the relationships among var-

ious types of knowledge in a graph structure. In

a KG, knowledge is represented as the structure of

triplets: subject, predicate, and object. For ex-

https://orcid.org/0000-0002-3821-6507

https://orcid.org/0000-0002-2552-6717

https://orcid.org/0000-0002-1939-4455

https://orcid.org/0000-0001-6717-7028

ample, the fact that Princess Mononoke is directed

by Hayao Miyazaki can be represented as a triplet

“Princess Mononoke, directed by, Hayao Miyazaki”

to provide the knowledge that humans understand

conceptually in a machine-readable format. The

standard format for sharing such representations on

the Web is called Resource Description Framework

(RDF); N-Triples and Turtle are used as notations for

the practical use of RDF. Linked Data (Bizer et al.,

2009) is a KG linked to other KGs using semantic

web technologies such as RDF.

Existing studies on KGs for content primarily fo-

cus on conceptually accurate descriptions of content

metadata. Thus, different knowledge representations

are often required when applying KGs to downstream

tasks. In this study, we focus on a recommendation

task, which is considered important among down-

stream tasks. Furthermore, we narrow our focus to

the domain of anime, which has gained attention as

one of the rapidly growing content industries. This

study aims to investigate effective knowledge repre-

sentation through a case study of anime recommen-

dation. The speciﬁc approach is to formulate the hy-

potheses on how to conﬁgure knowledge to improve

Saito, Y., Egami, S., Sei, Y., Tahara, Y. and Ohsuga, A.

An Analysis of Knowledge Representation for Anime Recommendation Using Graph Neural Networks.

DOI: 10.5220/0012359500003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 2, pages 243-252

ISBN: 978-989-758-680-4; ISSN: 2184-433X

243

recommendation performance and to verify them by

changing the structure of the KG according to the hy-

pothesis.

The remainder of this paper is organized as fol-

lows: Section 2 introduces prior studies on knowl-

edge representation and recommendation models in

the content domain; Section 3 formulates a hypothe-

sis and then explains our approach to verify the hy-

pothesis; Section 4 presents the construction process

of the knowledge graph and the results of the recom-

mendation experiment; Section 5 discusses the exper-

iment results; Section 6 summarizes this study includ-

ing prospects.

2 RELATED WORK

2.1 Knowledge Graph

2.1.1 Wikidata

Wikidata (Vrande

c and Kr

otzsch, 2014) is a typical

example of Linked Data. Launched by the Wikime-

dia Foundation in 2012, Wikidata is one of the Wiki-

media projects dedicated to structuring all knowledge

within the project, including Wikipedia. Each entity

on Wikidata is assigned a unique identiﬁer. The en-

tities encompass items that represent real-world ob-

jects, concepts, and events, along with properties that

deﬁne relationships and characteristics between items

and describe attributes related to the items. In ac-

cordance with the Linked Data principle (Heath and

Bizer, 2011), URIs are assigned based on identiﬁers,

enabling data retrieval from the URIs in JSON or RDF

format. Furthermore, SPARQL Protocol and RDF

Query Language (SPARQL) endpoint is available.

2.1.2 KG of Manga, Anime and Games

Oishi et al. (Oishi et al., 2019) focused on the

multimedia franchise and adaptation relationships of

manga, anime, and games (MAG). They constructed

a KG of MAG to develop technology to associate

bibliographic information with other works automat-

ically. The KG was constructed by collecting in-

formation from several Web resources specialized in

various adaptations and linking them to the Media

Arts Database (MADB) as a whole. MADB (Agency

for Cultural Affairs, 2015) is a database that holds

metadata related to manga, anime, games, and me-

dia arts. It is developed and managed by the Agency

for Cultural Affairs of Japan. To promote media arts

in Japan, it provides a digital archive of media arts

as an open database. The database contains approx-

Figure 1: Example of CKG (Wang, et al., pp.951).

imately 330,000 manga titles and 120,000 game ti-

tles as of March 2022 (Media Arts Consortium JV,

2022). Users can access SPARQL query services and

dataset dumps. In the study, they utilized the Super-

work model (Lee et al., 2018) to describe the adap-

tation relationships between works. The quantitative

quality of the constructed KGs was evaluated by ran-

dom sampling of triplets and visual conﬁrmation of

the validity of the superordinate-subordinate relation-

ships among entities.

2.1.3 Japanese Visual Media Graph

Japanese Visual Media Graph (JVMG) is a project to

build a KG for Japanese visual media such as manga,

anime, games, and visual novels (Pfeffer and Roth,

2020). The project aims to provide researchers who

focus on modern media, themes, and characters with

a systematic database. JVMG collected data from the

websites of several enthusiast communities. The en-

thusiast communities have the advantage of describ-

ing visual media at a more granular level of detail

than general-purpose data sources. In addition, they

tend to update information about visual media im-

mediately. The project is currently in the process

of integrating auxiliary data sources (e.g., Wikidata,

MADB) into three main data sources. The database is

expected to be further expanded in the future.

2.2 KG-Based Recommendation

2.2.1 Knowledge Graph Attention Network

Collaborative ﬁltering is a recommendation algorithm

for predicting a users’ latent preference based on the

interaction between users and items. However, it is

difﬁcult to capture the semantic relationships between

items because collaborative ﬁltering considers only

the users’ behavioral data.

Knowledge Graph Attention Network (KGAT)

is a graph neural network recommender model that

takes into account the feature and attribute of item

(Wang et al., 2019). KGAT takes a graph as its in-

put, which combines user behavioral (e.g., rating, pur-

chasing, and browsing) data and auxiliary information

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

244

about items. This graph is referred to as the Collabo-

rative Knowledge Graph (CKG) in the paper (Figure

1). In general, KGAT assumes that a CKG is con-

structed from a single data source. The model consists

of an embedding layer and attentive embedding prop-

agation layers. The embedding layer performs pre-

training embedding by applying TransR (Lin et al.,

2015) to the CKG. The attentive embedding propaga-

tion layers perform propagating embedding by mes-

sage passing (Gilmer et al., 2017) according to the

graph structure. The embedding of each node is up-

dated by aggregating information from neighboring

nodes based on the attention mechanism (Vaswani

et al., 2017). The output of the ﬁnal layer can be used

as an input for downstream tasks. In the paper, the

model is quantitatively evaluated by recommending

movies and music.

2.2.2 KG-Boosted Reinforcement Learning for

Recommendation

Sakurai et al. proposed an artist recommendation

method (Sakurai et al., 2022) that combines a KG of

music with reinforcement learning by agents. The KG

consists of four types of nodes: users, music, music

genres, and artists. They deﬁned the following three

types of triplets.



, e

(

)

, m



: user u

listened to music m



, e

(

)

, m



: artist a

created music m



, e

(

)

, g



: music m

belongs to genre g

Applying TransE (Bordes et al., 2013) to the KG

constructed from the triplets described above results

in obtaining embeddings attached to each node. In

the KG with embeddings, agents are positioned at

the target user nodes, allowing them to freely explore

the graph. It is noteworthy that reinforcement learn-

ing for the agent to reach the artist’s node from the

user’s node can ensure the explainability of the rec-

ommendation based on the arrival path. To evalu-

ate the proposed method, they conducted experiments

for artist recommendation using the Spotify Million

Dataset (Chen et al., 2018), which contains 57,880

music, 1,006 users, 14,973 artists, and 2,517 genres.

As a result, the proposed method improves the recom-

mendation performance by about 1–2 % against the

baseline.

Figure 2: Overview of this study.

3 APPROACH

3.1 Overview

As mentioned in Section 1, this study aims to clarify

effective knowledge representation for anime recom-

mendation tasks. We hypothesize how to conﬁgure

knowledge to improve recommendation performance

and verify the hypotheses by changing the structure

of the KG correspondingly. Speciﬁcally, we have de-

ﬁned the following two hypotheses based on the quan-

titative and qualitative aspects of knowledge consid-

ered in the KG construction.

(A) Quantitative aspect: Can we improve recommen-

dation performance by enriching the semantics of

the graph?

(B) Qualitative aspect: Can information that charac-

terizes the work improve recommendation perfor-

mance among metadata?

Regarding hypothesis (A), we construct KGs by

collecting anime data from multiple sources, and

then we evaluate the recommendation performance by

varying the conﬁguration of data sources. Regarding

hypothesis (B), we construct KGs from triplets except

for a single arbitrary property and compare the recom-

mendation performance among them. In the manner

of ablation study, we quantify what information con-

tributes to a recommendation performance. Figure 2

shows the entire ﬂow of approach from KG construc-

tion to recommendation.

3.2 Knowledge Graph Construction

3.2.1 Collecting Data

We considered several data sources with different

characteristics to construct a KG of anime. First, we

selected MyAnimeList

as a data source speciﬁc to

the domain of anime. MyAnimeList is an interna-

tional website dedicated to anime and manga. It is

also a kind of enthusiast community as deﬁned by

https://myanimelist.net

An Analysis of Knowledge Representation for Anime Recommendation Using Graph Neural Networks

245

Pfeffer et al. in Section 2.1.3. MyAnimeList was

selected because it provides not only metadata but

also a wealth of rating data from registered users.

In this study, we use the constructed KG to recom-

mend anime content; therefore, MyAnimeList is suit-

able for our study. Second, we selected Wikidata as a

data source for general-purpose KG. Because it func-

tions as a hub among many datasets and is suitable

for entity matching due to its pre-deﬁned properties

for identifying the entities in external datasets. We

expect to improve recommendation performance not

only by simply increasing the amount of information

but also by using complementary data sources with

different characteristics.

Each data source assigns a unique identiﬁer to an

anime work. In MyAnimeList, the anime work is

identiﬁed by an integer value under the namespace

of “https://myanimelist.net/anime/”. In Wikidata, the

anime is described as an item identiﬁed by an in-

teger value preﬁxed by Q under the namespace of

“http://www.wikidata.org/entity/”.

3.2.2 Entity Matching

In order to link metadata about an anime entity de-

rived from two different data sources with a single

anime entity, we applied entity matching and inte-

grated the data sources. In this process, the inte-

gration means creating a dictionary of MyAnimeList

identiﬁers and Wikidata identiﬁers for the anime men-

tioned in Section 3.2.1. Entity matching consists of

two main steps. The ﬁrst step is direct entity match-

ing. When a Wikidata anime item has MyAnimeList

anime ID

as one of its properties, we got the pair

of the item’s Wikidata identiﬁer and MyAnimeList

anime ID (MyAnimeList identiﬁer). This allowed

us to identify anime entities described in two sep-

arate data sources. The second step is indirect en-

tity matching using a third-party data source as an

intermediate dictionary. MyAnimeList entities often

have third-party website links, such as Wikipedia or

other streaming platforms. Similarly, Wikidata enti-

ties have identiﬁers of third-party data sources in ad-

dition to the MyAnimeList anime ID. We identiﬁed

the anime entities by using third-party data sources

that connect to both MyAnimeList and Wikidata as

intermediate dictionaries.

Note that we had to deal with differences in the

granularity of anime work descriptions between data

sources. For instance, there is a case where a work W

is deﬁned as a single entity in Wikidata. In contrast,

it is divided into two entities in the third-party, and

in MyAnimeList, it is split into three entities. This

https://www.wikidata.org/wiki/Property:P4086

Figure 3: Differences in anime work description granularity

and data source priority.

discrepancy is caused by different data sources often

having different editing policies, such as whether to

regard an anime series as a single work or as several

different works. We handled the problem by assigning

priorities to each data source. Speciﬁcally, priority is

assigned based on the following order relation.

wikidata

≻ W

third-party

≻ W

myanimelist

(1)

Therefore, in the aforementioned case, multiple

work entities described in MyAnimeList are aggre-

gated into a single entity described in Wikidata as

shown in Figure 3. Since Wikidata is a general-

purpose dataset in a wide range of domains, it is dif-

ﬁcult to capture detailed adaptation relationships spe-

ciﬁc to anime works. In contrast, MyAnimeList is a

data source specialized in anime works; thus, it can

describe works in detail. Consequently, we assigned

priorities following the order relationship as shown in

Equation 1.

3.2.3 Property Selection

Entity matching allowed us to link knowledge de-

rived from different data sources to a single work en-

tity. Among the knowledge, resources (such as cate-

gories or classes) that have unique identiﬁers enrich

the semantics of KG. Therefore, properties that re-

fer to them are required. For this reason, we selected

properties by imposing constraints. We listed prop-

erties and their values that meet all of the following

conditions.

• Conditions (a) on Wikidata properties:

– A property whose subject belongs to the class

“anime (Q1107)” in a chain

– A property whose number of appearances in

Wikidata is 1000 or more

– A property whose data type is “Item”

• Conditions (b) on MyAnimeList properties:

– A property whose value functions as a resource

or category

– A property whose value does not contain a URL

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

246

Figure 4: Example of entity matching and peripheral struc-

ture across “Made in Abyss” and “Tokyo Magnitude 8.0”.

For example, the partial structure of the KG

around “Made in Abyss” and “Tokyo Magnitude 8.0”

is shown in Figure 4. “Made in Abyss” is as-

signed the identiﬁer “Q61853333” (Made in Abyss:

Japanese anime television series)

in Wikidata, and

the identiﬁer “34599” in MyAnimeList. We ap-

plied entity matching by creating a key-value pair

of Q61853333 and 34599 according to the above

methodology, and properties were selected according

to Section 3.2.3. “Tokyo Magnitude 8.0” also has

Wikidata and MyAnimeList identiﬁers respectively,

and entity matching is applied similarly. On the sub-

graph, these two works are linked to a common entity,

because they share common elements in production

company, voice actor, and genre. Such co-references

are repeated in many other works, allowing KG to

capture the higher-order relationships between works.

3.3 Recommendation

We performed an experiment for the anime recom-

mendation task using the constructed KGs as a part

of the model’s input. The recommendation model is

KGAT introduced in Section 2.2.1. As mentioned,

KGAT inputs a CKG that combines the User-Item bi-

partite graph used in collaborative ﬁltering and the

Item KG based on the auxiliary information. The

User-Item bipartite graph corresponds to an existing

dataset by MyAnimeList users. The Item KG means

the KG based on the metadata constructed in Section

3.2. By linking the KGs and the rating dataset with

MyAnimeList identiﬁers, we constructed a CKG as

shown in Figure 1, and KGAT learned its structure to

recommend the anime works to the users.

Table 1: Results of entity matching (EM) per step.

Step 0 Step 1 Step 2

Number of EM work 0 2,989 5,740

Percentage of EM (%) 0 18 35

http://www.wikidata.org/entity/Q61853333

4 EXPERIMENTS

4.1 Evaluation of KG Construction

4.1.1 Dataset

In this study, since we are planning to use the KG for

the recommendation, we have selected Anime Dataset

with Reviews - MyAnimeList (Marlesson, 2020).

This dataset consists of three tables of data: “ani-

mes.csv”, “proﬁles.csv”, and “reviews.csv”. Since

“animes.csv” contains metadata about anime works,

we constructed KG based on this data. However, the

metadata is outdated because it covers data up to 2020

and the MyAnimeList API changed in speciﬁcations.

To update the data, from June 8 to June 9, 2023, we

crawled and supplemented data for 16,216 works in-

cluded in the dataset. As a result, we have obtained

16,082 works with detailed data. The 134 missing

works are thought to be due to differences between

the present and when the dataset was created, such as

page deletion.

4.1.2 Evaluation of Entity Matching

According to Section 3.2.2, we applied entity match-

ing in two steps. In the ﬁrst step, we obtained a Wiki-

data item that has the property P4086 (MyAnimeList

anime ID). In the second step, three third-party

websites were selected as intermediate dictionaries:

“AniDB”

, “Anime News Network”

and “Ofﬁcial

Website”. These websites were considered to be sufﬁ-

ciently active as intermediate dictionaries, because in

the existing dataset (Marlesson, 2020), 12,348 exter-

nal links were assigned to AniDB, 8,864 to the Anime

News Network, and 10,927 to the Ofﬁcial Website.

In both steps, we collected the data in Wikidata on

June 12, 2023. The number and percentage of entity

matching per step are shown in Table 1.

4.1.3 Selected Property

As for Wikidata, we used Wikidata Query Service to

list the properties that meet the conditions (a) as of

September 27, 2023. As for MyAnimeList, we listed

the properties that meet conditions (b) based on the

added detailed data. Table 5 and Table 6 show the

properties listed by data source in the Appendix.

When changing the conﬁguration of data sources,

the prerequisites of the experiments should be stan-

dardized for all the KGs. Thereby, we ﬁltered the

works that have both MyAnimeList and Wikidata

https://anidb.net

https://www.animenewsnetwork.com

An Analysis of Knowledge Representation for Anime Recommendation Using Graph Neural Networks

247

Table 2: Statistics of constructed KGs

# triplet # entity # property

KG-all 191,699 28,666 50

KG-mal 106,049 17,626 25

KG-wikidata 85,489 16,577 18

properties. Consequently, 5,621 works out of 5,740

ended up being included in the KG.

Eventually, we constructed a KG consisting of

191,699 triplets by integrating MyAnimeList and

Wikidata. Including it, all the KGs were constructed

with RDFLib, a Python library. The RDF exported in

Turtle format was stored in GraphDB

. All datasets

of this study are available at https://anonymous.

4open.science/r/kgat-input-DB18 in a format that can

be inputted for KGAT.

4.2 Quantitative Aspect:

Recommendation Experiment I

4.2.1 Experimental Setup

Firstly, we evaluated the constructed KGs through the

anime recommendation task. We constructed multiple

KGs with different conﬁgurations of the data sources

used, in order to evaluate the effect of the amount

of knowledge on the recommendation performance.

Speciﬁcally, we constructed the following 3 types of

KGs as follows.

• KG-All: based on MyAnimeList and Wikidata

• KG-Mal: based only on MyAnimeList

• KG-Wikidata: based only on Wikidata

We compared their recommendation perfor-

mances, as the inputs to the KGAT model. The statis-

tics by knowledge considered in the KG construction

are shown in Table 2.

As already mentioned in Section 2.2.1, the KGAT

model takes a CKG as its input. In order to con-

struct the CKG, we selected a dataset “reviews.csv”

from the aforementioned dataset. We implemented

a model based on KGAT-pytorch

, which is designed

for a problem known as Top-K recommendation. This

problem predicts the top K items in each user’s list

of works to be rated. Recall@K and nDCG@K are

ranking metrics to evaluate the quality of Top-K rec-

ommendation.

These metrics for the recommendation task are

calculated as follows:

https://graphdb.ontotext.com

There are 7 other properties that are added automati-

cally in the process of loading RDF into GraphDB.

https://github.com/LunaBlack/KGAT-pytorch

Table 3: Typical hyperparameters of KGAT model.

# epoch LR Dimensions @K value

100 1e-4 [64, 32, 16] [20, 40, 60, 80, 100]

Recall@K =

Number of correct items in the top K

Number of possible items in total

(2)

nDCG@K =

DCG@K

IDCG@K

(3)

, where DCG@K can be calculated as follows if the

i-th prediction is in fact the evaluated value rel

DCG@K =

∑

i=1

rel

− 1

log

(i + 1)

(4)

, where IDCG@K denotes the DCG for the ideal rank-

ing order. That is, it expresses how much DCG can be

obtained for the best possible ranking. nDCG@K is

equivalent to normalized DCG@K by IDCG@K and

it ranges from 0 to 1. We evaluated the quantitative

quality of recommendation by means of these met-

rics.

In addition, we converted reviews.csv to the fol-

lowing dictionary format:

: [i

, i

, ·· ·], u

: [i

, i

, ·· ·], ···} (5)

, where u

denotes a unique user id, and i

denotes a

unique item id that user rated. We split it into the train

and test data by each user’s item list. In the original

KGAT, the split ratio is train:test = 8:2. Following

this, the number of works rated by each user should be

at least 5, so we extracted users who have ﬁve ratings

or more as a threshold. As a result, 53,000 rating data

by 4,330 users remained.

We experimented with three different KGs and

preprocessed rating data as a common input for the

recommendation. Table 3 shows the typical hyperpa-

rameters used in training the model. Hyperparameters

are kept at the default values in the implementation.

We used the same conﬁguration for all experiments in

this paper.

4.2.2 Experimental Result

Table 4 shows the results of the evaluation experiment

by recommendation. We compared the recommenda-

tion performance for each KG conﬁguration. As a

baseline input, the recommendation performance of

collaborative ﬁltering without KG was set to “None”.

The @K values of metrics are listed only for K = 20,

60, and 100 as representative.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

248

Table 4: Recommendation performance by each KG conﬁguration.

Recall nDCG

@K = 20 @K = 60 @K = 100 @K = 20 @K = 60 @K = 100

KG-all 0.0959 0.2052 0.2828 0.0487 0.0752 0.0910

KG-mal 0.0916 0.2007 0.2717 0.0435 0.0702 0.0845

KG-wikidata 0.0911 0.2054 0.2869 0.0450 0.0730 0.0897

None 0.0779 0.1650 0.2323 0.0376 0.0588 0.0722

4.3 Qualitative Aspect:

Recommendation Experiment II

Secondly, we constructed new KGs that exclude an

arbitrary property in the manner of ablation study. To

accurately measure the effect of the excluded proper-

ties in the recommendation, a KG constructed from a

single data source is set as a baseline. When build-

ing a KG from multiple data sources, the information

about a work described by properties may not neces-

sarily be unique. We can quantify the effect of unique

information described excluded an arbitrary property

by setting a single data source as the baseline.

In particular, we set KG-wikidata as the baseline.

Speciﬁcally, all 18 properties of Table 5 were ex-

cluded one by one, and the KGs constructed in each

case were used as input for the recommendation ex-

periment. Calculating the recommendation perfor-

mance ratio of the KG for each excluded property

to the baseline KG, we quantiﬁed how much the ex-

cluded properties contributed to the recommendation.

The results are shown in Figure 5. In more details,

results for all the metrics are shown in Figure 6 in the

Appendix.

5 DISCUSSION

5.1 Effects of Data Source Variation

First, in the evaluation experiment regarding rec-

ommendation, we varied the data source conﬁgura-

tion considered in KG construction. Table 4 reveals

that when “None” is used, the recommendation per-

formance is at its lowest. This indicates that tak-

ing some knowledge into account would lead to im-

proved recommendation performance. When com-

paring “KG-all” with “KG-mal” and “KG-wikidata”,

KG-all demonstrated the best recommendation per-

formance across four metrics. Additionally, Table

2 shows that the recommendation performance is

roughly proportional to the number of triplets in the

KG. Despite KGAT typically assuming knowledge

from a single data source, “KG-all” achieved the best

performance. This result can be attributed not only

to the increased volume of data but also to the mutu-

ally complementary role of the two data sources with

different characteristics. Speciﬁcally, it appears that

Section 3.2.2 dealt with the description granularity

gap, and Section 3.2.3 selected properties with rich

semantics, both of which played an effective role.

In summary, Hypothesis A, as outlined in Section

3.1, is substantiated. This is because the experiments

conﬁrm that recommendation performance improves

as the number of triplets increases. In other words, the

experiments highlight the positive effect of the quan-

titative aspect of knowledge considered during KG

construction on recommendation performance.

5.2 Effects of Property Exclusion

Second, we excluded properties from KG-wikidata in

the manner of ablation study. Figure 5 shows the Re-

call@20 ratio concerning the baseline for each ex-

cluded property. To interpret the graph, red bars sig-

nify properties with a negative effect on the recom-

mendation, as their exclusion led to performance im-

provements. Conversely, blue bars indicate properties

with a positive effect on the recommendation, as their

exclusion resulted in performance declines.

According to Figure 5, the properties that had a

negative effect by being excluded are “genre (P136)”,

“has part(s) (P527)” and “country of origin (P495)”.

The ﬁrst property describes the work’s genre, the sec-

ond describes the hierarchical relationships between

the works, and the third describes the country where

the work was produced. Compared within the meta-

data, these properties are information that can charac-

terize the work. In particular, “genre” most directly

characterizes the work. The other two properties do

not necessarily directly characterize the work. How-

ever, works from the same series or produced in the

same country are likely to have similar characteristics.

This result seems to support Hypothesis B in Section

3.1.

On the other hand, the property that had a

large positive effect by being excluded is “characters

(P674)”. This is a property that describes the charac-

ters appearing in the work. Since it describes a char-

acter’s name speciﬁc to an anime work, the associated

An Analysis of Knowledge Representation for Anime Recommendation Using Graph Neural Networks

249

Figure 5: The effect of excluded property in recommendation (Recall@20 ratio to baseline).

triplets frequently contain named entities. Almost all

named entities rarely link to entities of other works,

thus they often exist as leaf nodes in the KG. When

the graph neural networks propagate the embeddings

according to the structure of KG, the leaf nodes are

dead ends, thus preventing effective propagation of

the embeddings. Therefore, properties of named en-

tities, such as “characters”, are thought to have im-

paired recommendation performance.

The “voice actor (P725)” property also has a pos-

itive effect due to its exclusion. Referring to Table 5

in the Appendix, “voice actor” appears the most fre-

quently among the Wikidata properties. Nevertheless,

the exclusion of “voice actor” does not signiﬁcantly

affect the recommendation performance in Figure 5.

This result emphasizes that qualitative factors such as

KG’s topology and property semantics should also be

considered, as quantitative factors do not solely deter-

mine recommendation performance.

6 CONCLUSION

In this study, we investigated effective knowledge rep-

resentation through a case study of recommending

anime works. Our approach is to ﬁrst hypothesize

how to conﬁgure knowledge in order to improve rec-

ommendation performance, and then verify the hy-

potheses by changing the KG’s structure according to

the hypothesis. We aimed to identify which aspects of

knowledge about anime contributed to the recommen-

dation by inputting multiple KGs into a graph neural

network model and comparing their recommendation

performance. Our ﬁndings revealed that: in terms

of quantity, incorporating more data sources into the

KG led to improved recommendation performance, in

terms of quality, the knowledge that characterized the

works more effectively contributed to the better rec-

ommendation. This study offers a novel methodol-

ogy for constructing downstream-task-aware knowl-

edge representation by evaluating the effects of struc-

tural changes in KGs on the performance of recom-

mendation. Our method is generalizable to other con-

tent domains: inductive studies that investigate differ-

ences for every domain, and the development of sys-

tems that automatically discover important properties

for KG-based recomendation.

However, there is still room for discussion regard-

ing property exclusion results. It is important to stan-

dardize the prerequisites for all properties, such as

equalizing the number of triplets for each property.

Additionally, it is advisable to introduce new metrics

or utilize existing ones to evaluate the contribution of

each property to KG-based recommendation.

As an alternative methodology, text-based KG

construction is worth considering. For example, the

automatic KG construction from a work’s synopsis

text is feasible. As such texts provide detailed de-

scriptions of a work’s storyline or worldview, they

have the potential to enrich properties that character-

ize the work. In recent years, there have been numer-

ous studies on the construction of KGs from a natural

language using Large Language Models (Mihinduku-

lasooriya et al., 2023) (Giglou et al., 2023); thus, we

consider enriching our KG by using such advanced

technologies.

ACKNOWLEDGEMENTS

This research was supported by JSPS Grants-in-Aid

for Scientiﬁc Research JP21H03496, JP22K12157,

JP23H03688, JP22K18008 and Sumitomo Electric

Group Social Contribution Fund.

REFERENCES

Agency for Cultural Affairs (2015). Media arts database.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

250

Bizer, C., Heath, T., and Berners-Lee, T. (2009). Linked

data - the story so far. Linking the World’s Informa-

tion.

Bordes, A., Usunier, N., Garcia-Dur

an, A., Weston, J., and

Yakhnenko, O. (2013). Translating embeddings for

modeling multi-relational data. In Proceedings of the

26th International Conference on Neural Information

Processing Systems - Volume 2, NIPS’13, pages 2787–

2795, Red Hook, NY, USA. Curran Associates Inc.

Chen, W., Lamere, P., Schedl, M., and Zamani, H. (2018).

Recsys challenge 2018: Automatic music playlist

continuation. In Proceedings of the 12th ACM Con-

ference on Recommender Systems, pages 527–528.

Giglou, H. B., D’Souza, J., and Auer, S. (2023). Llms4ol:

Large language models for ontology learning.

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and

Dahl, G. E. (2017). Neural message passing for quan-

tum chemistry. In Precup, D. and Teh, Y. W., editors,

Proceedings of the 34th International Conference on

Machine Learning, volume 70 of Proceedings of Ma-

chine Learning Research, pages 1263–1272. PMLR.

Heath, T. and Bizer, C. (2011). Principles of Linked Data,

pages 7–27. Springer International Publishing, Cham.

Lee, J. H., Jett, J., Cho, H., Windleharth, T., Disher, T.,

Kiryakos, S., and Sugimoto, S. (2018). Reconcep-

tualizing superwork for improved access to popular

cultural objects. Proceedings of the Association for

Information Science and Technology, 55(1):274–281.

Lin, Y., Liu, Z., Sun, M., Liu, Y., and Zhu, X. (2015). Learn-

ing entity and relation embeddings for knowledge

graph completion. In Proceedings of the Twenty-Ninth

AAAI Conference on Artiﬁcial Intelligence, AAAI’15,

pages 2181–2187. AAAI Press.

Marlesson (2020). Kaggle: Anime Dataset with Reviews -

MyAnimeList.

Media Arts Consortium JV (2022). Implementation Report

on the Project to Promote the Development of Me-

dia Arts Collaboration Infrastructure, etc. Number

FY2022. Media Arts Consortium JV.

Mihindukulasooriya, N., Tiwari, S., Enguix, C. F., and Lata,

K. (2023). Text2kgbench: A benchmark for ontology-

driven knowledge graph generation from text.

Oishi, K., Mihara, T., Nagamori, M., and Sugimoto, S.

(2019). Identifying and linking entities of multime-

dia franchise on manga, anime and video game from

wikipedia. In Jatowt, A., Maeda, A., and Syn, S. Y.,

editors, Digital Libraries at the Crossroads of Digi-

tal Information for the Future, pages 95–101, Cham.

Springer International Publishing.

Pfeffer, M. and Roth, M. (2020). Japanese visual media

graph: Providing researchers with data from enthusi-

ast communities. International Conference on Dublin

Core and Metadata Applications, pages 136–141.

Sakurai, K., Togo, R., Ogawa, T., and Haseyama, M.

(2022). Explainable artist recommendation based on

reinforcement knowledge graph exploration. In Inter-

national Workshop on Advanced Imaging Technology

(IWAIT) 2022, volume 12177, pages 91–96. SPIE.

Soldo, Lana and Schagerl, Christopher (2023). Impact of

the Covid-19 Pandemic on Netﬂix. MAP Education

and Humanities, 3(1):75–82.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., and Polosukhin, I.

(2017). Attention is all you need. In Guyon, I.,

Luxburg, U. V., Bengio, S., Wallach, H., Fergus,

R., Vishwanathan, S., and Garnett, R., editors, Ad-

vances in Neural Information Processing Systems,

volume 30. Curran Associates, Inc.

Vrande

c, D. and Kr

otzsch, M. (2014). Wikidata: a free

collaborative knowledgebase. Communications of the

ACM, 57(10):78–85.

Wang, X., He, X., Cao, Y., Liu, M., and Chua, T.-S. (2019).

Kgat: Knowledge graph attention network for recom-

mendation. In Proceedings of the 25th ACM SIGKDD

International Conference on Knowledge Discovery &

Data Mining, KDD ’19, pages 950–958, New York,

NY, USA. Association for Computing Machinery.

APPENDIX

Table 5: Wikidata properties in KG (18 in total).

Property label Identiﬁer # triplet

voice actor P725 28681

genre P136 9386

instance of P31 7485

country of origin P495 5560

original language of ﬁlm ··· P364 5350

distributed by P750 3417

production company P272 3112

has part(s) P527 3110

distribution format P437 3013

director P57 2680

character designer P8670 2146

based on P144 2063

screenwriter P58 1928

characters P674 1842

language of work or name P407 1724

composer P86 1670

broadcast by P3301 1220

original broadcaster P449 1102

An Analysis of Knowledge Representation for Anime Recommendation Using Graph Neural Networks

251

Figure 6: All the effect of excluded property in recommendation in heatmap.

Table 6: MyAnimeList properties in KG (25 in total).

Property label # triplet

type 15941

genres 13451

producers 12179

themes 6245

studios 5767

status 5621

source 5621

title 5621

year 5621

rating 5621

season 5621

licensors 3911

demographics 2019

relations:Adaptation 3229

relations:Other 1660

relations:Sequel 1613

relations:Side story 1324

relations:Prequel 1194

relations:Alternative version 858

relations:Parent story 846

relations:Alternative setting 633

relations:Summary 510

relations:Spin-off 408

relations:Character 294

relations:Full story 241

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

252