The goal of this paper is to provide a story-based explanation for movie recommendation systems, achieved by a multiaspect explanation and narrative analysis methods. We explain how and why particular movies are similar based on following two aspects: (i) composition of movie characters and (ii) interactions among the characters. These aspects correspond to story-based features of the movies that are extracted from character networks (i.e., social networks among the characters). By using the story-based features, we can explain the reason why two arbitrary movies are similar or not. We anticipate that the proposed method could improve the explainability of the recommender systems for movies.
Various online services have been employing recommender module to provide users with the most relevant items. However, with only a list of recommended items, it is difficult for the users to understand why such items are selected. The users should spend additional resources (mostly time or money) for identifying whether the recommended items are really preferable. The problem becomes even worse for recommending narrative works (e.g., movies, TV series, novels, and so on). For example, we have experienced giving up watching TV series after a few episodes.
Hence, explaining the reason on the recommendation has been
regarded as an important research issue. There have been
various studies [
For recommending the narrative works, in this work, we assume that the content of the items directly affect the users’ preference. We focus on analyzing and exploiting the major characteristics of the content (such as drawing styles of comics or stories of movies).
We have conducted a simple user survey among 97 users of ‘webtoon’, which is a novel media distributing comics through the web. The survey simply consisted of one question, which allowed plural responses: “What are criteria that affect your preferences for webtoons?”. Most of the users wrote two criteria: stories (98.96%) and drawing styles (97.93%); 96.90% of the users selected both the criteria. Also, we interviewed 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 consume webtoons within a few limited genres and drawing styles.
Based on the results, we have found out that the following patterns can be used for explaining the recommendations. stories contained in the narrative works how the stories are physically described.
The goal of this study is to improve the explainability of the recommender systems by using the similarity among the stories. Nevertheless, as an ongoing study, we limited our target domain into the movies. Also, using character networks (i.e., social networks among the characters), we preserved the expandability of the proposed method for other types of narrative works.
EXPLAINING STORY-BASED SIMILARITY Expected results of the proposed method are similar to what Netflix2 is already providing, as displayed in Fig. 1. Netflix 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. Copying permitted for private and academic purposes. ExSS ’18, March 11, Tokyo, Japan.
However, we targeted on more detailed explanations than Netflix’s. Our proposed method focused on providing causal evidence based on the movies’ stories. For example, when we recommend “Cinderella” for the users, we can say “Cinderella is similar with Snow White with focus on relationships among characters appearing in the movies.” To achieve these goals, this study aimed for following objectives: (i) discovering story-based features, (ii) estimating story-based similarity among the movies, (iii) providing explainable recommendation on the basis of the story-based similarity, and (iv) making the proposed method expendable to other media.
For the former two objectives, we attempted to discover features with two aspects: composition of the characters and interactions among the characters. Based on the story-based features, we developed measurements to display the differences between two arbitrary stories. Also, with focus on the expandability, data sources of the features were limited within the character network.
Our previous studies [
DEFINITION 1 (CHARACTER NETWORK). Suppose that N is the number of characters that appeared in a movie, Ca . When N(Ca ) indicates a character network of Ca , N(Ca ) can be described as a matrix 2 RN N . It consists of N N components which are social affinities among the characters where, ai; j is the social affinity of ci for c j when Ca is an universal set of characters that appeared in Ca and ci is an i-th element of Ca .
In this study, we used frequency of the dialogues among the characters as the social affinity among them. The dialogues were extracted from the movies’ scripts that were collected from the Internet Movie Script Database (IMSDb) 3.
The directors have to compose characters with a focus on
representation of their stories. In other words, the users might
recognize the movies’ stories from the composition of
characters, whether it is intuitively or analytically. Hence, the
similarity among movies’ stories is recognizable from the
difference among compositions of characters.
In order to compare the compositions of characters, we
classified 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
distinguishable. We classified the characters’ roles into three
categories (i.e., main characters, minor characters, and
extras) by using their centralities on the character networks [
In addition, we distinguished the protagonists who are the
most spotlighted and the antagonists who are secondly focused
[
The categorization was conducted by measuring social ties among the characters. If there are three characters (cP, cA, and ci), cP is a protagonist, and cA is an antagonist, we can identify which character is closer with ci than the other by comparing ci’s social ties for cP and cA. It can be formulated as ci 2 8 >>>P; > > > <
T (ci; cP) > T (ci; cA); if T (ci; cP) > median T (c j; cP);
8c j
T (ci; cA) > T (ci; cP); >A; if T (ci; cA) > median T (c j; cA); >>> 8c j > >:N; otherwise: ; (1) 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 with the protagonist and the antagonist, respectively, N denotes a set of the characters who take the neutral positions between the protagonist and the antagonist, median8c j T (c j; cP) and median8c j T (c j; cA) refer to median values of the social ties of all the characters for cP and cA, respectively.
To measure the degree of social ties, we considered the frequency of interactions and the number of paths between target characters. It is formulated as:
T (ci; c j) =
jpkj å Õ f (nl 1; nl ); 8pk2Pi; j l=2 (2) 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) on pk, jpkj is the number of nodes included in the pk, and f (nl 1; nl ) means a weighting value (interaction frequency) between nl 1 and nl .
By combining the two classification criteria, we categorized the characters into six groups, as displayed in Fig. 2. The number of characters in each category was represented as a 2 3 matrix. We call this matrix a ‘character composition matrix’. As a naive approach, the difference between two movies can be estimated by the Frobenius distance among their character composition matrices as:
DC (Ca ; Cb ) =k Ca
Cb kF ; (3)
Importance t s i n o g a t o r P t s i n o g a t n A
Main
Main Characters y t i m i x o r P la Neutral tru Main eN Characters Minor
Minor Characters where k kF denotes the Frobenius norm. In here, DC (Ca ; Cb ) has smaller value, as Ca and Cb have more similar number of characters for all the categories. It means that DC is highly affected by the number of characters.
Therefore, we normalized the character composition matrix as:
CaNorm =
Ca ; jCa j where jCa j is the number of characters that appeared in Ca . 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: DCNorm(Ca ; Cb ) =k CaNorm
CbNorm kF ; DCScale(Ca ; Cb ) = max(jCa j; jCb j) min(jCa j; jCb j) : Thus, the explanations to the users can be composed by considering the (i) proximity and (ii) importance. For example, ‘The Day After Tomorrow (2004)’ and ‘Gravity (2013)’ are commonly disaster movies. In these two movies, nature takes the roles of antagonists. However, tritagonists within ‘The Day After Tomorrow (2004)’ are mostly main characters, while a protagonist of ‘Gravity (2013)’ almost solely appeared. Also, the number of character of ‘The Day After Tomorrow (2004)’ is bigger than that of ‘Gravity (2013)’. because ‘The Day After Tomorrow (2004)’ is closer to a drama movie dealing with family affection and ‘Gravity (2013)’ is also a sort of thriller movie. (4) (5) (6)
Interactions among the Characters Although the protagonists and antagonists interact with most 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 appearing frequently with the antagonists, there is a possible indication that a conflict or a crisis (e.g., betrayal, convert, kidnapping, etc.) is likely to happen. In other words, interactions among the characters’ groups reflect methods for developing the stories.
Based on this intuition, we compared the stories by using two criteria: (i) frequency: proportion of inter-group interactions 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 were composed in Sect. 2.2. Procedures for discovering the groups can be summarized as follows. 1. Subsume extras under P, A, and N with the same method that is used for the main and minor characters. 2. Calculate the internal compactness and the external adjacency of each group. 3. If the external adjacency of a particular group is too high, compared with its internal compactness, partition the group. 4. Iterate Step. 2 and Step. 3, until the groups have adequate quality.
The external adjacency is measured on the basis of the external interactions’ frequency and the out-degree of the groups, as: I (Gk) =
å 8ci2Gk; 8c j2Gk; ci,c j ai; j å d( j; l) cl2Gk jGkj ; d( j; l) = 1; 0; if a j;l , 0; otherwise.
; where Gk denotes a k-th group, jGkj indicates the number of characters included in Gk, and d( j; l) refers to an indicator function that indicates existence of interactions between c j and cl . On the contrary, the internal compactness was estimated from the internal interactions’ frequency and the in-degree of the groups.
E (Gk) =
å 8ci2Gk; 8c j<Gk ai; j å d( j; l) cl<Gk å d( j; l) 8cl : In order to compare I (Gk) with E (Gk), we defined a unified metric by the linear combination of I (Gk) and E (Gk). It can be formulated as:
Q(Gk) = a (I (Gk)) (1 a) (E (Gk)) ; where a 2 [0; 1] is a weighting factor; in this study, we set a = 0:80. Based on Q(Gk), we determined whether Gk had to be reconstructed or not.
Therefore, if Q(Gk) was lower than a user-defined minimum threshold, q , we partitioned Gk into several groups. The method for partitioning the group is the same with Eq. 1. It can be summarized as follows: 1. Choose top-two characters as centers for new groups based on the characters’ centralities. 2. Partition the target group into three new groups by using the same method with Eq. 1. 3. Calculate Q(Gk) of novel groups. If the average quality of the novel model is worse than previous model’s, restore the model and designate next order as the center.
From the characters’ groups, we measure the proposed metrics. The frequency can be measured by the proportion of inter-group interactions for the total interactions. It can be formulated as:
F (Ca ) = å å 8Gk 8ci2Gk; 8c j<Gk ai; j ai; j ; (11) (12) where Gk indicates a group of the characters.
In order to formulate the aggressiveness, we considered both of the external adjacency and the internal compactness of the groups. It is formulated as:
A (Ca ) = å E (Gk) 8Gk å I (Gk): 8Gk Therefore, the explanations to the users can be generated by considering the (i) frequency and (ii) aggressiveness. With the same example in Sect. 2.2, stories of ‘The Day After Tomorrow (2004)’ and ‘Gravity (2013)’ are commonly led by two groups. However, groups in ‘The Day After Tomorrow (2004)’ are highly separated, while groups of ‘Gravity (2013)’ frequently interact with each other, relatively. In addition, groups within ‘The Day After Tomorrow (2004)’ commonly have high internal compactness and low external adjacency, since this movie describes an ‘adventure’, where a father rescues his son. On the other hand, groups of ‘Gravity (2013)’ show opposite characteristics, because ‘Gravity (2013)’ draws a person who has to save herself, relying on tritagonists’ advice.
We propose a method for explaining the reason behind existence or non-existence of similarity between the two arbitrary movies. This method can improve explainability of the recommender systems for movies. Also, it is expandable to other media and formats, since the proposed method is built on the character network.
Nevertheless, the proposed method is not yet verified with a large scale dataset. Our future research will be focused on implementing appropriate datasets and conducting an evaluation for the proposed method. Also, we are planning to consider multiple features of the movies (e.g., genre and tempo) for generating explanation.
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (NRF-2017R1A41015675).