Content uploaded by Thiago H. Silva
Author content
All content in this area was uploaded by Thiago H. Silva on Aug 23, 2024
Content may be subject to copyright.
Culture Fingerprint: Identification of Culturally
Similar Urban Areas Using Google Places Data
Fernanda R. Gubert1, Gustavo H. Santos1, Myriam Delgado1, Daniel Silver2,
and Thiago H Silva1
1Universidade Tecnológica Federal do Paraná, Curitiba, Brazil
2University of Toronto, Toronto, Canada
fernandagubert,gustavohenriquesantos}@alunos.utfpr.edu.br,
{myriamdelg,thiagoh}@utfpr.edu.br, dan.silver@utoronto.ca
Abstract. This study investigates methods using a global data source,
Google Places, to identify culturally similar urban areas without relying
on difficult-to-access data like user preferences shown through check-
ins. We propose and assess a simple method requiring only information
about place types and their frequency in the studied areas, and a more
advanced method that enhances venue categories using Scenes Theory -
it helps us understand the cultural significance of everyday urban life. We
tested our methods in 14 cities worldwide and all US states. The results
suggest that a straightforward approach based on category frequencies
can highlight major cultural differences. However, the Scenes Theory-
based method provides a better understanding of cultural nuances, as
the ones supported by survey data.
Keywords: Cultural signature ·large scale assessment ·Google Places
1 Introduction
Traditional methods like surveys and interviews are important data sources for
studying culture in its complexity. However, these methods have drawbacks (e.g.
high costs and time-consuming), which limit their scalability. To remedy this
situation, some works evaluate alternative geolocalized data sources from the web
to study culture. These sources exist on a global scale and are faster to obtain.
Studies have shown the usefulness of these data sources in several domains [6,
16, 19, 21], including the cultural ones [2, 3, 8, 15, 17, 18].
Bancilhon et al. [2] explore an approach to quantifying a society’s culture
through city street names, revealing that these names reflect cultural values.
Using Foursquare data, Senefonte et al. [15] examine how regional and cultural
characteristics affect the mobility patterns of both tourists and residents. The
results indicate that the tourist’s origin significantly influences their behavior,
especially in large cultural differences between the origin and destination. Silva
and Silver [18] introduce a graph neural network method for predicting local
culture. They evaluate their approach on Yelp data, showing that it could help
predict local culture even when traditional local information is unavailable.
2 Gubert et al.
When aiming to provide methods based on geolocalized web data to describe
local culture, some research indicates that eating and drinking habits can be a
valuable option [3,8, 17]. These studies illustrate promising approaches to iden-
tifying cultural boundaries and similarities between different societies at differ-
ent scales. However, they rely on user preferences, typically manifested through
check-in data, which is challenging to obtain in practice for many users or with
global coverage. Another perspective follows the argument presented in [11],
which suggests that the availability of resources and services that meet the pop-
ulation’s needs contributes to forming a local identity. What is notable about
this approach is the opportunity to consider multiple aspects of culture, as the
resources of a region can be associated with various categories like religion, cui-
sine, and arts, providing a format that is still little explored. Our approach aligns
with this direction by exploring Scenes Theory [20], which captures local public
cultural dimensions embodied in venues such as cafes, churches, restaurants, and
nightclubs. This enables the creation of a cultural description of local areas, al-
lowing comparison with other areas—a step we perform in this study to identify
cultural similarities. This differs from previous studies [1,4, 9, 14], which tend to
disregard the cultural component in their analyses.
Extending previous studies, the approaches proposed here to describing local
culture rely on simple data from the Google Places API. One can provide an
expressive cultural abstraction of any covered urban area, thanks to the mapping
to Scenes Theory – see Section 2. Unlike studies that explored the cultural
characteristics of regions using eating habits and user mobility, this study aims
to derive such characteristics from the categories of venues present in a city.
This allows us to evaluate whether our proposed approaches can adequately
express key cultural aspects without relying on user actions, such as check-ins
and evaluations.
We evaluate the approaches using data from 14 cities on different continents
and all states of the United States. The results indicate that a simple approach,
Frequency, can capture significant cultural differences satisfactorily. However, a
more sophisticated approach, Scenes, can add extra semantic expressiveness in
capturing cultural characteristics. This added expressiveness is evident in our
outcomes and survey data comparison, indicating that Scenes better captures
cultural nuances.
2 Cultural Signatures obtained from Google Places (GP)
2.1 Data From GP
GP is a location-based social network that allows users to discover and share in-
formation about local venues, geographic locations or points of interest, such as
universities, cafes, bus stations, and parks. No type of location was disregarded.
GP API provides geolocated venue data, resulting in one of the world’s most
accurate, up-to-date, and comprehensive venue models. In addition to latitude
and longitude coordinates, venues are associated with at least one category de-
Title Suppressed Due to Excessive Length 3
signed to describe the venue type. In this study, we consider two datasets from
GP, States and Cities, as described next.
The Dataset States, presented in [10, 22], includes business metadata (geo-
graphic info, category information, etc.) from GP up to September 2021 for all
U.S. states. The dataset is composed of 4,963,111 unique venues and has 4,501
unique categories. The District of Columbia has the lowest number of distinct
venues, totaling 11,003, while California has the highest count at 513,134 unique
venues. We explore this dataset to study states focusing on geographic and cat-
egory information.
For Dataset Cities we have collected data from a set of cities. GP API pro-
vides, by default, 141 unique categories. However, these categories do not provide
the level of specificity necessary in this study. For example, the API assigns the
category “restaurant” to all venues of this type, but it does not offer more spe-
cific categories related to cuisine, such as Italian or Japanese, which is necessary
for this work. The optional “keyword” parameter is used in requests to the GP
API aiming to overcome the limitation. The API documentation3guarantees
valid results when inputs to this parameter are categories of venues, making it a
convenient option for the desired purposes. The categories chosen to use in this
parameter are those from the Yelp database due to the higher specificity, e.g.,
Yelp offers specific types of restaurants, such as Italian Restaurants. Yelp cate-
gories have a four-level hierarchical structure, making it suitable for our work to
adopt only those at the last level. Some of them were excluded because they were
not relevant to the purpose of the study, such as Provencal and Northeastern
Brazilian, resulting in a total of 888 categories.
Using the proposed strategy, we have collected data from 14 cities, namely:
Curitiba and Rio de Janeiro in Brazil; Toronto and Vancouver in Canada;
Chicago and Los Angeles in the USA; Berlin and Frankfurt in Germany; Paris
and Lyon in France; Seoul and Busan in South Korea; and Nairobi and Mom-
basa in Kenya. These cities are important in their respective countries and cover
regions with different cultural characteristics. A publicly available tool4details
the data acquisition process and clarifies the need for a balance between costs
and data volume, which leads us to have a summarized set of venues. This tool
aids in reproducing our study [5].
2.2 Urban Areas’ Cultural Dimensions
Following research on local “scenescapes,” we measure local scenes for the urban
areas by aggregating the set of available venue categories in terms of qualitative
meanings they express. To translate these concepts into measurements, for each
venue category (e.g., restaurant, university, or bar), a team of trained coders has
assigned a score of 1-5 on a set of 15 cultural dimensions si∈S={s1, s2, ..., s15},
such as transgression, tradition, local, authenticity, or glamour. Each area then
receives a score for each of the 15 dimensions, calculated as a weighted average.
3https://developers.google.com/maps/documentation/places/web-service/overview.
4https://github.com/FerGubert/google_places_enricher.
4 Gubert et al.
Detailed descriptions of the theoretical meaning of each dimension can be found
in [20].
2.3 Transfer Knowledge Procedure
The categories retrieved from GP need to be mapped to the appropriate set
of 15 dimensions scores of the Scenes Theory. Without trained coders for our
particular areas (States and Cities) we examine the Scenes’ dimension scores of
the Yelp categories presented in [19]. This knowledge is then adapted for use
with GP/ categories, a transferring knowledge outlined in Figure 1. It illustrates
an example for two different venues, each provided by a different dataset, venue
A from Dataset States and venue B from Dataset Cities .
Fig. 1. Overview of mapping GP categories to the local cultural dimensions.
As depicted in Figure 1, for a better description of the venues, both the
selected Yelp categories used in the requests and the broader categories made
available by GP are used. To increase semantic capacity and mapping accuracy,
sentences are created for each venue, following the procedures for each dataset.
In Dataset States, one sentence is created per venue, combining all associated
categories. For example, if the venue has the categories “Italian”, “Restaurant”
and “Food”, the sentence is: “Italian Restaurant Food”. Dataset Cities on the
other hand, lacks specific categories by default. Therefore, sentences include a
requested Yelp category and all GP categories associated with that venue. For
Title Suppressed Due to Excessive Length 5
example, if a venue has “Amusement Parks” and “Water Parks” due to Yelp
requests and “Tourist Attraction” as a default GP category, the sentences are:
“Amusement Parks Tourist Attraction” and “Water Parks Tourist Attraction”.
Yelp categories are organized in a 4-level hierarchical structure. To expand
semantic capacity, Yelp sentences are created using all hierarchical levels. In
other words, for each category at the last level, the associated sentence returns
to the first level. This is why “Active Life” was added to the Yelp sentences in
Figure 1; these Yelp categories are immediately below its root category.
In possession of sentences describing the venue, the mapping process is car-
ried out with SBERT, using the Sentence Transformers framework, in which
several pre-trained models with a large and diverse dataset of more than 1 bil-
lion training pairs are made available and can be used to calculate embeddings
from sentences and texts to more than 100 languages [12]. The cosine similarity
compares the generated embeddings, and for each sentence related to the venues,
the Yelp sentence with the highest score is retrieved. With this mapping, each
venue is associated with one or more vectors (depending on the number of related
sentences) containing the 15 dimensions of the Scenes Theory.
2.4 Cultural Signatures
We propose two approaches for creating cultural signatures, Scenes-based ap-
proachand Frequency-based approach.
For a particular urban area, the Scenes-based approachconsiders a vector
Sarea ={sarea
1, sarea
2, ..., sarea
15 }, where sarea
i=1
ωPω
v=1 1
mPm
ϕ=1 Sv,ϕ
i, with ω
representing the number of unique venues in an urban area, mis the number
of categories a venue has, and Sv,ϕ
iis the i-th element of the vector of cultural
dimensions for a certain venue vand one of its category ϕ; thus, sarea
irepresents
the average score of all venues in the urban area for a specific cultural dimension,
considering the average scores of all categories for each venue.
We also present an alternative approach, Frequency, aimed at creating cul-
tural signatures that disregard Scenes information, using only location cate-
gories. This approach considers the frequency of the category in the area, i.e., for
a particular urban area, a vector describes it by all the unique categories found
in that area. For example, an area could be described by the categories [Uni-
versity, Restaurant, Coffee Shop, American Restaurant] and another by [Italian
Restaurant, Wine Shop]. The frequency values are normalized per category.
Frequency helps answer the question: Are the existence and the number of
certain types of venues in two different urban areas enough to explain their
cultural differences?
3 Cultural Signatures Identify Culturally Similar Areas
3.1 Cities Worldwide
Scenes for Dataset Cities First, we evaluate the results of the cultural signa-
tures generated by the Scenes-based approach. We perform hierarchical clustering
6 Gubert et al.
using Ward’s linkage method and Euclidean distance, with the 15 dimensions of
Scenes Theory as features. The results are represented in the dendrogram de-
picted at the top of Figure 2, where a division into six clusters is identified.
Fig. 2. Hierarchical clustering dendrogram of cities resulting from Scenes (top) and
Frequency (bottom).
The result aligns with what is expected concerning the cultural characteristics
of the areas studied. Most of the clusters coherently grouped cities from the
same country - in general, countries have distinct cultural characteristics; the
exceptions in this sense are clusters 1 and 4. In cluster 1, Toronto was grouped
with Chicago and Los Angeles; note also that Los Angeles is the most dissimilar
city in the grouping. The result of Chicago and Toronto being together and more
similar makes sense, in that they are often considered to be culturally similar to
one another, even compared to Los Angeles. Regarding cluster 4, Vancouver was
grouped with Paris and Lyon. We found significant similarities between the most
recurrent categories of French cities and Vancouver, such as “Art galleries,” which
could help explain this result. Although German cities (Berlin and Frankfurt)
and French cities (Paris and Lyon) are on the same continent, they are quite
distinct culturally, and so their location in separate clusters seems reasonable.
To facilitate a comparative analysis by contrasting the values of each cluster
dimension with its corresponding overall average, we calculate the Z-Score, as
shown in Figure 3. The Z-Score is the number of standard deviations concerning
the average of what is being observed. This facilitates comparing clusters by
extracting the characteristics that stand out in each, compared with a general
overview, i.e., the centroid of clusters’ centroids. For example, cluster 3, repre-
senting Kenya, has one of the lowest values for Tradition. In contrast, for cluster
4 with the cities Vancouver, Paris, and Lyon, this dimension represents one of
the most important characteristics. Looking at cluster 1, composed of Chicago,
Los Angeles and Toronto, we see that Tradition is not as predominant as in clus-
Title Suppressed Due to Excessive Length 7
ter 4. This highlights the potential to identify cultural signatures and provide
an overview of geographic areas by extracting their key dimensions.
Fig. 3. Z-Score values of Scenes dimensions per cluster.
Frequency for Dataset Cities For Frequency, we perform hierarchical clus-
tering using the Complete linkage criteria and Cosine distance – the best combi-
nation tested. As depicted at the bottom of Figure 2, the results for Frequency,
as with Scenes, align with what is expected when grouping cities of the same
country. However, using Frequency differently, Chicago is more similar to Los
Angeles, and Vancouver is more related to Toronto than to the French cities.
The results obtained demand reflection because although Toronto and Van-
couver are in the same country, they are not necessarily similar in terms of im-
migration patterns, governance, geography, ecology, and cultural style. Toronto
and Chicago, on the other hand, have much in common: they are both Great
Lakes cities, with strong industrial heritages and are now in the midst of a post-
industrial transformation. Hence, they are often compared as similar cases [7,13].
We can reveal specific characteristics of each cluster by extracting the five
most distinct categories for each of them – we do that by calculating the distance
of the category from its cluster centroid. After that, we calculate the Z-Score
for the selected categories against the overall average. The result of this process
is illustrated in Figure 4. Certain categories in some clusters stand out so no-
tably that they not only significantly deviate from their overall average, but also
emerge as the sole positive value compared to others. For example, in French
cities, “municipality”, in Brazilian cities, “hang gliding”, and in Korean cities,
“face painting” exhibits this distinct characteristic. Making a comparison with
the Z-Score values illustrated in Figure 3, we can relate these specific findings
depicted in Figure 4 to the aspects highlighted in Tradition for cluster 4 (pre-
dominantly French), Transgression for cluster 5 (Brazil) and Self-Expression and
Charisma for cluster 6 (South Korea).
To analyze the clusters that differ between the Scenes and Frequency ap-
proaches, we examine the most evident characteristics in each. For Scenes, we
focus on clusters 1 and 4, selecting the three most prominent dimensions in
each and retrieving the most important sentences for those dimensions. For Fre-
quency, we look at cluster 3 and identify the 50 most frequent categories. For
example, Los Angeles, Chicago, and Toronto have “Business Consulting”, “Li-
braries” and “Gastropubs” in common, whereas Vancouver, Paris, and Lyon are
8 Gubert et al.
Fig. 4. Z-Score values for the most distinct categories per cluster (Frequency).
marked by “Antiques Book Store”, “Art Gallery”, “Comedy and Night Club” and
gastronomic diversity, such as “Portuguese Bakery”, “Spanish Meal Delivery”,
“Sushi Bars” and “Tapas Bars”. In Frequency, many categories can be found
that summarize these characteristics, such as “Gastropubs”, “Art Installation”,
“Imported Food”, “Meal Takeaway” and “Souvenir Shops”. The result indicates
that, unlike Frequency, through human knowledge in its dimensions, Scenes can
detect subtle differences among categories with similar meanings.
3.2 All States in the USA
Using Dataset States, we apply the transfer knowledge methodology (Section 2.3)
and create cultural signatures for all states in the country.
Evaluating Scenes-based approach for Dataset States To analyze cultural
signatures in this dataset using Scenes, we also perform hierarchical clustering
with 15 dimensions of the Scenes Theory as features, Ward linkage criteria, and
Euclidean distance. By inspecting the dendrogram, we observe a tendency to
group regions by geographic proximity. By mapping one of the clearest cuts in the
dendrogram, we obtain Figure 5 (right). It shows that culturally similar regions,
such as the US South, are grouped. These results reinforce the effectiveness of
the proposed method in identifying culturally similar regions.
Evaluating Frequency-based approach for Dataset States For this case,
we perform hierarchical clustering using the Ward linkage criterion and Eu-
clidean distance. Other combinations were experimented with, but none proved
Title Suppressed Due to Excessive Length 9
superior. We observe difference between this approach and the results obtained
with Scenes. Figure 5 (left) illustrates the mapped clusters provided by Frequency-
based approach.
Fig. 5. Results of hierarchical clustering considering all states in the USA represented
by Frequency (left) and Scenes (right).
It is not possible to detect clear patterns in the Frequency results, at least
as clear as identified by Scenes, regardless of the number of clusters adopted.
Surprisingly, Alaska and Maine are positioned within clusters larger than with
Scenes. Alaska is situated among states such as Washington, Oregon, North
Dakota, Minnesota, and Michigan. Maine is part of the largest cluster, which
includes most of the remaining states. Thus, Scenes provides extra semantic
expressiveness in smaller dimensions.
4 Comparing with Survey Data
There is no clear way to access the ground truth of our results. However, we
explore in this work a source where we expect some correlation: the American
Value Survey (AVS, access https://www.prri.org). The survey was conducted
among a representative sample of 5,031 adults (age 18 and up) living in all 50
states in the United States, having a statistically valid representation of the USA
population, including many minorities or hard-to-reach populations. Interviews
were conducted online between September 16-29, 2021 and September 1-11, 2022.
Additional details about the methodology can be found on the Ipsos website5.
The survey questions include political aspects and basic beliefs. We represent
these questions as features to describe states, where the values are the mean
answers of all participants for each state. We exclude political questions and
focus solely on basic beliefs6.
To assess the relationship between the results of the AVS and our propos-
als (Scenes and Frequency), we use the Pearson correlation for the Euclidean
distance between all pairs of states when describing them by AVS and our ap-
proaches. By doing that, we got a moderate correlation of 0.51 (p < 10−4) for
5https://www.ipsos.com/en-us/solutions/public-affairs/knowledgepanel.
6The complete list of questions used can be found at:
https://sites.google.com/view/neighbourhood-change.
10 Gubert et al.
Scenes. Using Frequency, on the other hand, resulted in a Pearson correlation
of −0.06 (p < 10−1) for the Euclidean distance between all pairs of states.
To better understand the correlation results individually we calculated the
Euclidean distance of each state in comparison to all others, considering its
descriptions using AVS and each of our proposals. Then, we calculate the Pearson
correlation (ρ) of these values. For Scenes,ρ∈[−0.221,0.709] and approximately
75% of all states exhibit either a moderate or high correlation. Alaska is the only
state with a negative correlation. By looking at the results for Frequency, with
ρ∈[−0.257,0.149], it is clear that it shows a worse association with another
source (AVS) regarding cultural beliefs.
5 Conclusion
In the present work, we examined data from Google Places (GP) and developed
two methods to establish cultural signatures of urban areas. The proposals (Fre-
quency and Scenes) were then assessed for their effectiveness in cities worldwide
and all states in the United States. We obtained evidence that the proposed
approaches, even a simple one based on frequency, could capture the cultural
character of geographic areas. We gathered evidence based on a comparison with
survey data that one of the approaches, based on the Scenes Theory, could cap-
ture better cultural nuances. Unlike other approaches that demand proxy data
for users’ preferences, e.g., user check-ins, our approach only demands simple
data, i.e., categories of venues, which are easily obtainable in GP for almost any
urban area. Hence, there is significant potential to utilize the proposed method-
ology for identifying cultural similarities between different locations. This could
facilitate the development of numerous new services and applications, such as
innovative location recommendation systems based on cultural criteria.
There are several ways to expand this work, such as expanding the dissimilar-
ity analysis to both approaches, Frequency and Scenes, or testing the proposed
methodology with other data sources. Since GP data is not free and acquiring
a considerable amount can be costly, this could also allow for expanding the set
of venues. Another possibility is to evaluate different levels of granularity, such
as neighborhoods and countries.
Acknowledgment
SocialNet project (process 2023/00148-0 of FAPESP) and CNPq (processes 313122/2023-
7, 314603/2023-9 and 441444/2023-7).
References
1. Arribas-Bel, D., Fleischmann, M.: Spatial signatures-understanding (urban) spaces
through form and function. Habitat Int 128, 102641 (2022)
2. Bancilhon, M., Constantinides, M., Bogucka, E.P., Aiello, L.M., Quercia, D.:
Streetonomics: Quantifying culture using street names. Plos one 16(6), e0252869
(2021)
Title Suppressed Due to Excessive Length 11
3. de Brito, S.A., Baldykowski, A.L., Miczevski, S.A., Silva, T.H.: Cheers to untappd!
preferences for beer reflect cultural differences around the world. In: Proc. of AM-
CIS’18. New Orleans, USA (2018)
4. Çelikten, E., Le Falher, G., Mathioudakis, M.: Modeling urban behavior by mining
geotagged social data. IEEE Trans on Big Data 3(2), 220–233 (2016)
5. Gubert, F., Silva, T.: Google places enricher: A tool that makes it easy to get and
enrich google places api data. In: Proc. of WebMedia’22, Extended Proceedings.
pp. 91–94. SBC, Curitiba, PR, Brasil (2022)
6. Hu, L., Li, Z., Ye, X.: Delineating and modeling activity space using geotagged
social media data. Cartogr Geogr Inf Sci 47(3), 277–288 (2020)
7. Kolpak, P., Wang, L.: Exploring the social and neighbourhood predictors of di-
abetes: a comparison between toronto and chicago. Prim. health care resear. &
devel. 18(3), 291–299 (2017)
8. Laufer, P., Wagner, C., Flöck, F., Strohmaier, M.: Mining cross-cultural relations
from wikipedia: a study of 31 european food cultures. In: Proc. of the ACM Web-
Sci’15. pp. 1–10. Oxford, UK (2015)
9. Le Falher, G., Gionis, A., Mathioudakis, M.: Where is the soho of rome? mea-
sures and algorithms for finding similar neighborhoods in cities. In: Proc. of the
ICWSM’15. Oxford, UK (2015)
10. Li, J., Shang, J., McAuley, J.: UCTopic: Unsupervised Contrastive Learning for
Phrase Representations and Topic Mining. In: Proc. of the ACL’22. pp. 6159–6169.
ACL, Dublin, Ireland (2022)
11. Mehta, V., Mahato, B.: Measuring the robustness of neighbourhood business dis-
tricts. J. of Urban Design 24(1), 99–118 (2019)
12. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-
networks. In: Proc. of the EMNLP’19. ACL, Hong Kong, China (11 2019)
13. Robson, K., Anisef, P., Brown, R.S., Nagaoka, J.: A comparison of factors deter-
mining the transition to postsecondary education in toronto and chicago. Res. in
Comp. Inter. Educ. 14, 338–356 (2019)
14. Sen, R., Quercia, D.: World wide spatial capital. PloS one 13(2), e0190346 (2018)
15. Senefonte, H., Frizzo, G., Delgado, M., Luders, R., Silver, D., Silva, T.: Regional
Influences on Tourists Mobility Through the Lens of Social Sensing. In: Proc. of
SocInfo’20. Pisa, Italy (2020)
16. Senefonte, H.C.M., Delgado, M.R., Lüders, R., Silva, T.H.: Predictour: Predicting
mobility patterns of tourists based on social media user’s profiles. IEEE Access 10,
9257–9270 (2022)
17. Silva, T.H., de Melo, P.O.V., Almeida, J.M., Musolesi, M., Loureiro, A.A.: A large-
scale study of cultural differences using urban data about eating and drinking
preferences. Information Systems 72, 95–116 (2017)
18. Silva, T.H., Silver, D.: Using graph neural networks to predict local culture. Envi-
ronment and Planning B: Urban Analytics and City Science 0(0), 12 (0)
19. Silver, D., Silva, T.H.: Complex causal structures of neighbourhood change: Evi-
dence from a functionalist model and yelp data. Cities 133, 104130 (2023)
20. Silver, D.A., Clark, T.N.: Scenescapes: How qualities of place shape social life. The
University of Chicago (2016)
21. Skora, L.E., Senefonte, H.C., Delgado, M.R., Lüders, R., Silva, T.H.: Comparing
global tourism flows measured by official census and social sensing. Online Soc
Netw Media 29, 100204 (2022)
22. Yan, A., He, Z., Li, J., Zhang, T., McAuley, J.: Personalized showcases: Generat-
ing multi-modal explanations for recommendations. In: Proc. of the SIGIR’23. p.
2251–2255. ACM, Taipei, Taiwan (2023)
