=Paper=
{{Paper
|id=Vol-2068/exss5
|storemode=property
|title=Explainable Movie Recommendation Systems by using Story-based Similarity
|pdfUrl=https://ceur-ws.org/Vol-2068/exss5.pdf
|volume=Vol-2068
|authors=O-Joun Lee,Jason J. Jung
|dblpUrl=https://dblp.org/rec/conf/iui/LeeJ18
}}
==Explainable Movie Recommendation Systems by using Story-based Similarity==
https://ceur-ws.org/Vol-2068/exss5.pdf
Explainable Movie Recommendation Systems by using
Story-based Similarity
O-Joun Lee Jason J. Jung∗
Department of Computer Engineering, Department of Computer Engineering,
Chung-Ang University Chung-Ang University
Dongjak-gu, Seoul, Republic of Korea Dongjak-gu, Seoul, Republic of Korea
concerto34@cau.ac.kr j3ung@cau.ac.kr
ABSTRACT However, the previous studies could not identify what the
The goal of this paper is to provide a story-based explana- users will gain or feel from the recommended items. Most
tion for movie recommendation systems, achieved by a multi- of the studies only took into account the ‘scrutability’ for the
aspect explanation and narrative analysis methods. We explain recommended items. Moreover, content of the items were not
how and why particular movies are similar based on following considered.
two aspects: (i) composition of movie characters and (ii) in-
For recommending the narrative works, in this work, we as-
teractions among the characters. These aspects correspond to
sume that the content of the items directly affect the users’
story-based features of the movies that are extracted from char-
preference. We focus on analyzing and exploiting the major
acter networks (i.e., social networks among the characters).
characteristics of the content (such as drawing styles of comics
By using the story-based features, we can explain the reason
or stories of movies).
why two arbitrary movies are similar or not. We anticipate
that the proposed method could improve the explainability of We have conducted a simple user survey among 97 users of
the recommender systems for movies. ‘webtoon’, which is a novel media distributing comics through
the web. The survey simply consisted of one question, which
ACM Classification Keywords allowed plural responses: “What are criteria that affect your
H.5.m. Information Interfaces and Presentation (e.g. HCI): preferences for webtoons?”. Most of the users wrote two cri-
Miscellaneous; J.5. Computer Applications: Arts and Human- teria: stories (98.96%) and drawing styles (97.93%); 96.90%
ities; I.2.4. Artificial Intelligence: Knowledge Representation of the users selected both the criteria. Also, we interviewed
Formalisms and Methods Lehzin Comics1 , which is one of the major webtoon publishers
in Korea. For a question: “Why you do not use recommender
engines for your platform?”, they answered that users mostly
Author Keywords consume webtoons within a few limited genres and drawing
Explainable Recommender System; Movies; Character styles.
Network; Story Analysis; Computational Narrative.
Based on the results, we have found out that the following
INTRODUCTION
patterns can be used for explaining the recommendations.
Various online services have been employing recommender • stories contained in the narrative works
module to provide users with the most relevant items. How-
ever, with only a list of recommended items, it is difficult for • how the stories are physically described.
the users to understand why such items are selected. The users
should spend additional resources (mostly time or money) for The goal of this study is to improve the explainability of the
identifying whether the recommended items are really prefer- recommender systems by using the similarity among the sto-
able. The problem becomes even worse for recommending ries. Nevertheless, as an ongoing study, we limited our target
narrative works (e.g., movies, TV series, novels, and so on). domain into the movies. Also, using character networks (i.e.,
For example, we have experienced giving up watching TV social networks among the characters), we preserved the ex-
series after a few episodes. pandability of the proposed method for other types of narrative
works.
Hence, explaining the reason on the recommendation has been
regarded as an important research issue. There have been EXPLAINING STORY-BASED SIMILARITY
various studies [7, 2] on building ‘explainable’ recommender
Expected results of the proposed method are similar to what
systems.
Netflix2 is already providing, as displayed in Fig. 1. Netflix
* Corresponding author. recommends sets of movies or TV series with some reasons;
e.g., “Because you watched Madam Secretary.”
© 2018. Copyright for the individual papers remains with the authors. 1 https://www.lezhin.com/ko/
Copying permitted for private and academic purposes. ExSS ’18, March 2 https://www.netflix.com/browse
11, Tokyo, Japan.
In order to compare the compositions of characters, we clas-
sified the characters with two criteria: (i) importance of their
roles and (ii) proximity with a protagonist. In our former
studies, we proved that roles of the characters are easily dis-
tinguishable. We classified the characters’ roles into three
categories (i.e., main characters, minor characters, and ex-
Figure 1: Examples of explanations provided by Netflix. tras) by using their centralities on the character networks [6].
The centralities were estimated by the linear combination of
standard node centrality measurements: the degree centrality,
closeness centrality, and betweenness centrality.
However, we targeted on more detailed explanations than
Netflix’s. Our proposed method focused on providing causal In addition, we distinguished the protagonists who are the
evidence based on the movies’ stories. For example, when we most spotlighted and the antagonists who are secondly focused
recommend “Cinderella” for the users, we can say “Cinderella [6]. We considered the remaining ones as tritagonists. The
is similar with Snow White with focus on relationships among tritagonists were categorized into three sides: friendly, neutral,
characters appearing in the movies.” and hostile ones for the protagonists.
To achieve these goals, this study aimed for following ob- The categorization was conducted by measuring social ties
jectives: (i) discovering story-based features, (ii) estimating among the characters. If there are three characters (cP , cA , and
story-based similarity among the movies, (iii) providing ex- ci ), cP is a protagonist, and cA is an antagonist, we can identify
plainable recommendation on the basis of the story-based which character is closer with ci than the other by comparing
similarity, and (iv) making the proposed method expendable ci ’s social ties for cP and cA . It can be formulated as
to other media.
T (ci , cP ) > T (ci , cA ),
P, if T (ci , cP ) > median T (c j , cP ),
For the former two objectives, we attempted to discover fea-
∀c j
tures with two aspects: composition of the characters and
interactions among the characters. Based on the story-based ci ∈ T (ci , cA ) > T (ci , cP ), , (1)
A, if T (ci , cA ) > median T (c j , cA ),
features, we developed measurements to display the differ-
∀c j
ences between two arbitrary stories. Also, with focus on the
N, otherwise.
expandability, data sources of the features were limited within
the character network. where T (c j , ck ) indicates the degree of social tie between c j
and ck , P and A indicate sets of the characters who are friendly
Character Network with the protagonist and the antagonist, respectively, N denotes
Our previous studies [1, 5, 4, 3] have proposed SNA-based a set of the characters who take the neutral positions between
methods for computationally analyzing the movies’ stories. the protagonist and the antagonist, median∀c j T (c j , cP ) and
We modeled the stories by using the character network that median∀c j T (c j , cA ) refer to median values of the social ties
was defined as follows; of all the characters for cP and cA , respectively.
D EFINITION 1 (C HARACTER N ETWORK ). Suppose To measure the degree of social ties, we considered the fre-
that N is the number of characters that appeared in a movie, quency of interactions and the number of paths between target
Cα . When N(Cα ) indicates a character network of Cα , characters. It is formulated as:
N(Cα ) can be described as a matrix ∈ RN×N . It consists |pk |
of N × N components which are social affinities among the T (ci , c j ) = ∑ ∏ f (nl−1 , nl ), (2)
characters where, ai, j is the social affinity of ci for c j when ∀pk ∈Pi, j l=2
Cα is an universal set of characters that appeared in Cα and
ci is an i-th element of Cα . where Pi, j is a set of possible paths between ci and c j , pk
indicates a k-th path in Pi, j , nl denotes a l-th node (character)
In this study, we used frequency of the dialogues among the on pk , |pk | is the number of nodes included in the pk , and
characters as the social affinity among them. The dialogues f (nl−1 , nl ) means a weighting value (interaction frequency)
were extracted from the movies’ scripts that were collected between nl−1 and nl .
from the Internet Movie Script Database (IMSDb) 3 .
By combining the two classification criteria, we categorized
the characters into six groups, as displayed in Fig. 2.
Composition of the Characters
The directors have to compose characters with a focus on rep- The number of characters in each category was represented as
resentation of their stories. In other words, the users might a 2 × 3 matrix. We call this matrix a ‘character composition
recognize the movies’ stories from the composition of char- matrix’. As a naive approach, the difference between two
acters, whether it is intuitively or analytically. Hence, the movies can be estimated by the Frobenius distance among
similarity among movies’ stories is recognizable from the their character composition matrices as:
difference among compositions of characters. DC (Cα , Cβ ) =k Cα − Cβ kF , (3)
3 http://www.imsdb.com/
Importance Interactions among the Characters
Although the protagonists and antagonists interact with most
Main Minor of the characters, the others are mostly bounded in particular
communities. For example, acquaintances of the protagonists
usually interact and appear with the protagonists. If they start
Protagonist Friendly Friendly appearing frequently with the antagonists, there is a possible
Main Minor indication that a conflict or a crisis (e.g., betrayal, convert, kid-
Characters Characters napping, etc.) is likely to happen. In other words, interactions
among the characters’ groups reflect methods for developing
the stories.
Proximity
Neutral
Neutral Neutral Based on this intuition, we compared the stories by using two
Main Minor criteria: (i) frequency: proportion of inter-group interactions
Characters Characters and (ii) aggressiveness: external adjacency of the groups.
To utilize the two metrics, we had to discover the characters’
groups. The groups were built on the basis of P, A, and N that
Antagonist
Hostile Hostile were composed in Sect. 2.2. Procedures for discovering the
Main Minor groups can be summarized as follows.
Characters Characters 1. Subsume extras under P, A, and N with the same method
that is used for the main and minor characters.
Figure 2: The proposed category of the characters.
2. Calculate the internal compactness and the external adja-
cency of each group.
3. If the external adjacency of a particular group is too high,
compared with its internal compactness, partition the group.
where k · kF denotes the Frobenius norm. In here, DC (Cα , Cβ )
has smaller value, as Cα and Cβ have more similar number of 4. Iterate Step. 2 and Step. 3, until the groups have adequate
characters for all the categories. It means that DC is highly quality.
affected by the number of characters. The external adjacency is measured on the basis of the external
Therefore, we normalized the character composition matrix interactions’ frequency and the out-degree of the groups, as:
as:
Cα
∑ d( j, l)
cl ∈Gk
CαNorm = , (4) I (Gk ) = ∑ ai, j × , (7)
|Cα | |Gk |
∀ci ∈Gk ,
∀c j ∈Gk ,
where |Cα | is the number of characters that appeared in Cα . ci ,c j
By comparing the normalized composition matrices, we got
a scale-tolerant distance. Also, the number of characters was
directly compared with other movie’s. It can be formulated as: 1, if a j,l , 0,
d( j, l) = , (8)
0, otherwise.
DCNorm (Cα , Cβ ) =k CαNorm − CβNorm kF , (5)
where Gk denotes a k-th group, |Gk | indicates the number of
max(|Cα |, |Cβ |) characters included in Gk , and d( j, l) refers to an indicator
DCScale (Cα , Cβ ) = . (6) function that indicates existence of interactions between c j and
min(|Cα |, |Cβ |) cl . On the contrary, the internal compactness was estimated
from the internal interactions’ frequency and the in-degree of
Thus, the explanations to the users can be composed by con- the groups.
sidering the (i) proximity and (ii) importance. For example,
‘The Day After Tomorrow (2004)’ and ‘Gravity (2013)’ are ∑ d( j, l)
commonly disaster movies. In these two movies, nature takes cl