Tourism, as a system, is dependent on collaboration between stakeholders, in order to ensure the development of a destination as a whole and also benefit for each individual stakeholder. This fact triggered a growing interest in understanding tourism collaboration, which has led to a considerable diversification of conceptual and methodological approaches. One emerging and promising area of inquiry consist in approaching tourism collaboration from the perspective of multilayer networks (Baggio, 2017), a concept also known as “network of networks” in network science (Kivela et al., 2014).
However, very few papers have inquired the structure of tourism collaboration networks from a multilayer perspective and none of them systematically inquired about both structural variations from one collaboration layer to another and interdependencies that might exist between collaboration layers. The current research has aimed precisely at filling this gap by methodologically combining Social Network Analysis (SNA) with nonparametric statistical analysis, and by taking as case studies two Romanian destinations: an emerging one (Bran), and a stagnating one (Vatra Dornei).
Methodological approach and results
Data used in this paper has been collected through semi-structured interviews with 47 stakeholders from the two destinations. The focus of the interviews was on identifying all relationships of each interviewed stakeholder, along with important information related to the collaborator and the purpose of collaboration. The analysis has been conducted in two steps, with two corresponding sets of results.
First, SNA allowed for a thorough investigation of each sub-network identified based on 7 different purposes of collaboration. For this, the sociogram for each sub-network was created and several metrics were calculated. Important differences between the layers were identified in this stage: largest and densest sub-networks are those for marketing and for the supply of goods and services while accessing funds or sponsorships barely create network structures.
Second, the associations between the 7 sub-networks in each destination have been identified through multiple correspondence analyses. The most important result is that some interdependencies between layers of collaboration are common for both destinations, despite one being an emergent destination and the other a stagnating one. More precisely, results suggest significant interdependencies between the following pairs of activities: (1) policy design and accessing funds and (2) products creation and knowledge exchange and research.
This paper contributes to the literature by operationalizing the concept of multilayer networks in tourism studies and by delivering one of the first systematic analyses in tourism of interdependencies between different layers of collaboration. It also draws attention to the potential that the multilayer network approach has for providing important knowledge on stakeholders’ collaboration and calls for more analyses from such a perspective.
Baggio, R. (2017). Network science and tourism – the state of the art. Tourism Review, 72(1), 120-131.
Kivela, M., Arenas, A., Barthelemy, M., Gleeson, J.P., Moreno, Y., & Porter, M.A. (2014). Multilayer networks. Journal of Complex Networks, 2, 203-271.
Multimodal sentiment analysis aims to recognize people’s attitudes from multiple communication channels such as verbal content (i.e., text), voice, and facial expressions. It has become a vibrant and important research topic in natural language processing, which also benefits various applications, such as social robots, customer analysis, healthcare, and tourism applications.
What is M2Lens?
Currently, deep-learning-based models achieve superior performance over the traditional methods in multimodal sentiment analysis. Representative examples include transformers, CNNs, and RNNs. However, these models often work like black-boxes, hindering users from understanding the underlying model mechanism and fully trusting them in decision-makings. Post-hoc explainability techniques, such as LIME, SHAP, and IG, help identify important features (e.g., words or image patches) that influence model predictions. However, these methods often target providing local explanations on instances (e.g., sentences) in unimodal scenarios. They do not scale well to produce global explanations on how intra- and inter-modal in- teractions influence the model decisions, for example, how the models will behave when positive words and sad voices are presented.
We present M2Lens, a novel explanatory visual analytics tool to help both developers and users of multimodal machine learning models better understand and diagnose Multimodal Models for sentiment analysis.
How does it work?
Multi-modal sentiment analysis combines the heterogeneous data and captures two primary forms of interactions in different modalities: Intra-modal interactions refer to the dynamics of one modality. Inter-modal interactions consider the correspondence between different modalities across time.
We formulate a set of rules to characterize three typical types of interactions among modalities, including dominance, complement, and conflict. The dominance suggests that the influence of one modality dominates the polarity (i.e., positive sentiment or negative sentiment) of a prediction. The complement indicates that two or all three modalities affect a model prediction in the same direction. The conflict reveals that the influences of modalities differ from each other.
To facilitate a comprehensive understanding of multimodal models, M2Lens offered multi-level and multi-faceted explanations, including the influences of individual modalities and their interplay, and importance of multimodal features.
Case Study: Multimodal Transformer
In this case study, the expert explored and diagnosed a state-of-the-art model, Multimodal Transformer (MulT), for sentiment analysis using the CMU-MOSEI dataset. MulT fuses multimodal inputs with cross-modal transformers for all pairs of modalities, which learn the mappings between the source modality and target modality (e.g., vision → text). Then, the results are passed to sequence models (i.e., self-attention transformers) for final predictions. All the multi-modal features of the input data are aligned at the word level based on the word timestamps. Following the settings of previous work, we trained, validated, and evaluated MulT with the same data splits (training: 16,265, validation: 1,869, and testing: 4,643).
In our paper, we demonstrate how M2Lens helps users understand and diagnose multimodal models for sentiment analysis through two case studies. Below we describe two such examples.
Example I: Dominance of Language Modality
After loading the multimodal transformer in the system, an expert E1 referred to the Summary View to see how individual modalities and their interplay contribute to the model predictions.
By looking at the second layer, E1 found that the language modality has the largest influence among the three modalities since it has the longest bar to the left and widest range of dots in the bee swarm plot.
In the last layer, within the dominance group, he discovered that the longest bars attach to the language modality, and the color of the prediction barcode aligns well with that of the language barcode. Thus, E1 concluded that the language also plays a leading role in the dominance relationship. He noticed that there are a group of dense blue bars appearing at the end of the language barcode, where the errors are relatively large (as indicated by the yellow curve above the dashed line). He wondered what features or their combinations caused the high errors. Therefore, he brushed the corresponding area of the blue bars.
Then, the Template View lists the feature templates of the selected instance in the Summary View. By sorting them in descending order of error, E1 found that the “PRON + PART” appears at the top with a child feature. Then, he collapsed the row and found that 21 instances contain the word “not”, where it negatively influences the predictions.
Next, he clicked “not” to see the details about this feature in the Projection View. Zooming in on the word “not”, several similar negative words (e.g., “isn’t”, “wouldn’t”) were observed. E1 speculated that the model could not deal well with negations. Subsequently, he lassoed these words to closely examine the corresponding instances in the Instance View.
When exploring the examples with large errors, E1 noticed that when double negations appear in a sentence (e.g., “not…sin…” and “not…bad…”) , the model tends to treat them separately and regards both of them as indicators for negative sentiment. He thought that augmenting double negation examples or preprocessing them into positive forms may improve the model performance.
Example II: Dominance of Visual Modality
Afterward, E1 referred back to the “dominance” group in the Summary View, where a collection of red bars from the prediction barcode conform with the ones from the visual modality. It indicates that the visual modality dominates the predictions, and the error line chart above suggests a low error rate in contrast with the previous case. Motivated by this observation, E1 brushed the red bars to investigate the patterns in the visual features.
In the Template View, “Face Emotion” has the largest support. After unfolding the row, E1 found that “Sadness + Joy” is a frequent and important combination. This intrigued him to find out how a contrary emotion pair co-occurs. After clicking the template, the corresponding glyphs are highlighted in the Projection View. Most of them were found outside of the red area, which verifies that the instances with “Sadness + Joy” often have small prediction errors. Through browsing the instances and their videos in the Instance View, “Joy” and “Sadness” are often considered important to model predictions. Their co-occurrences may be due to the presence of intense and rich facial expressions in the videos. And the model seemed to capture these important visual facial expressions.
In summary, we characterize the intra- and inter-modal interactions learned by a multimodal model for sentiment analysis. Moreover, we provide multi-level and multi-faceted explanations on model behaviors regarding dominant, complementary, and conflicting relationships among modalities.
In the future, we can extend the system to other multimodal NLP tasks (e.g., emotion classification and visual question answering). And we can further conduct a comparative study of multimodal models, determing under what circumstances we should use multimodal models and when uni- or bi-modal models are sufficient for target applications.
Check out the following video for a quick look at M2Lens’s features.
How can tourism be reborn? How can we reach states of awe? How can events become experiences that change people? Transformation is no longer a buzzword. In fact, we can use experience design principles to intentionally design transformative experiences – from opening glimpses and triggers towards long-term integration. In this research project, we look into attendees’ experiences of Burning Man shared on Instagram. By adopting a data analytics approach, we identify the socio-physical factors of human transformative experiences within and beyond the festival environment.
A three-phase methodological procedure
Phase 1 Data collection: To determine Instagram posts related to Burning Man, the hashtag #burningman2019 was used. In 2019, the 29th edition of Burning Man took place from August 25th through September 2nd. Based on this timeframe, we extracted posts published six months before the event, during, and six months after, resulting in a total of 53,326 posts. The extracted data, including captions (i.e., texts, hashtags, emojis), posts’ dates, posts’ URLs, and types of user account (i.e., business/personal).
Phase 2 Data pre-processing: First, language identification was applied in Python using Spacy to eliminate non-English posts. Next, business accounts and their corresponding posts, duplicates, and posts without a description were removed. This resulted in 35,802 usable posts: 8953 thereof published before the event, 2840 published during the event, and 24,009 published after the event. Thereafter, a list of stopwords was prepared, and irrelevant signs and unknown characters, numbers, and references to usernames with @ were removed. Slang words were reformed and hashtags, as well as emojis, were extracted.
Phase 3 Deep topological data analysis: A deep topological data analysis approach is a combination of topological data analysis and deep generative models. Concerning the former, the topic list is used as an embedding space for dimension reduction and further clustering. With the latter, the purpose is to learn the true data distribution of the training set to generate new data points with some variation. Depending on the data type, common approaches thereof include variational autoencoders and generative adversarial networks. The Vietoris-Rips algorithm was used to connect nearby data points to build topological structures, and nested complexes were used to identify persistent elements of the data structure using Morse Theory. Finally, the manifolds of the original dimensions were simplified and visualized.
An overview of the results
The software DataRefiner was used to build thematic clusters and to visualize the topological structure of the data. The text was tokenized, and text parameters were identified by weighing the tokens. These parameters are the key terms that represent the different clusters and can be expressed via a correlation value between −1 and 1. By means of an iterative process, the number of clusters was adjusted until the number of noise points reached a suitable minimum, ultimately reaching 30 unique clusters.
From an epistemological viewpoint, the selection of the number of clusters and the interpretation thereof requires deep knowledge of a topic domain from a researcher. The naming of each cluster was based on extracted keywords, found parameters, and text summaries, following crosschecks and consensus among the research team. Finally, the graphical representation of the cluster map was visualized, presenting the topographic structure with similar clusters located close to each other. Clusters that are spatially opposite from each other show that their relevant parameters are fundamentally different, and the correlation of the parameters is close to −1. The detailed results can be found in our paper.
Our research outperforms prior studies that may suffer from the intrinsic drawbacks of traditional techniques. For instance, although LDA has been widely used in tourism research, it relies primarily on word co-appearance frequency. Relations between topics thus remain unknown. By showcasing the potential and usefulness of deep topological analysis in tourism research, in cases where the goal is to explore implicit messages based on short-text and unstructured Instagram data, our study supplies a robust and transparent compass to navigate through the epistemological and methodological questions and decision-making process – from theory-led research design and data collection to data analysis and theory generation.
How to cite: Neuhofer, B., Egger, R., Yu, J., & Celuch, K. (2021). Designing experiences in the age of human transformation: An analysis of Burning Man. Annals of Tourism Research, 91, 103310.
by: Egger, Roman, Oguzcan Gumus, Elza Kaiumova, Richard Mükisch, and Veronika Surkic In ENTER22 e-Tourism Conference, pp. 343-355. Springer, Cham, 2022.
Social media plays a key role in shaping the image of a destination. Although recent research has investigated factors influencing online users’ perception towards destination image, limited studies encompass and compare social media content shared by tourists and destination management organisations (DMOs) at the same time. This paper aims to determine whether the projected image of DMOs corresponds with the destination image perceived by tourists. By taking the Austrian Alpine resort Saalbach-Hinterglemm as a case, a netnographic approach was applied to analyse the visual and textual posts of DMO and user-generated content (UGC) on Instagram using machine learning. The findings reveal themes that are not covered in the posts published by marketers but do appear in UGC. This study adds to the existing literature by providing a deeper insight into destination image formation and uses a qualitative approach to assess destination brand image. It further highlights practical implications for the industry regarding DMOs’ social media marketing strategy.
The presentation below is from the ENTER 2022 conference
Pls. cite as: Egger, Roman, Oguzcan Gumus, Elza Kaiumova, Richard Mükisch, and Veronika Surkic. “Destination Image of DMO and UGC on Instagram: A Machine-Learning Approach.” In ENTER22 e-Tourism Conference, pp. 343-355. Springer, Cham, 2022.
by: Egger, R., Pagiri, A., Prodinger, B., Liu, R., & Wettinger, F. In ENTER22 e-Tourism Conference (pp. 356-368). Springer, Cham
The needs of travellers vary across cultures. When it comes to culinary aspects, there is a strong connection between gastronomy and culture. To optimise service offerings, investigation of the essential aspects of dining experiences in relation to cultural backgrounds is of great importance. In the age of digitalisation, tourists share their dining experiences throughout their multiphasic travel journey via online platforms. By considering nine distinct cultural backgrounds, this research aims to investigate tourist experiences based on TripAdvisor restaurant reviews through topic modelling, using the city of Salzburg as its study context. Depending on one’s cultural circumstances, the findings demonstrate that the most important aspects include staff, food-menu items, value for money, restaurant physical appearance, food authenticity, overall service, menu offers, food quality, atmosphere, and recommendations. This study advances the state-of-the-art knowledge of societal culture as a variable in the target market analysis of restaurant customers. Findings allow restaurant owners, other tourism service providers, and destination management organisations to analyse and adapt their service offerings and strategies accordingly.
The presentation below is from the ENTER 2022 conference
Pls. cite: Egger, R., Pagiri, A., Prodinger, B., Liu, R., & Wettinger, F. (2022, January). Topic Modelling of Tourist Dining Experiences Based on the GLOBE Model. In ENTER22 e-Tourism Conference (pp. 356-368). Springer, Cham.
As a result of travel activities, overtourism has become a global issue. Even after the COVID-19 pandemic, the topic of overtourism would benefit localized overcrowding as a new occurrence in the tourism industry. To investigate tourists’ feelings when visiting overcrowded attractions, the analysis of online reviews has been recognized as a reliable source given the rich data it provides. In the digital area, reviews posted by tourists become critical in influencing one’s decision-making process. One typical example is TripAdvisor which enables tourists to consult reviews on any hotel, restaurant or attractions shared by other users.
Due to the unstructured nature of online data, topic modeling and sentiment analysis has gained their popularity. Topic modeling identifies the main topics of the reviews and is particularly suitable for exploratory studies. Sentiment analysis quantifies subjective information by natural language processing and computational linguistics. By taking overtourism (using Paris) as the research context, this study aims to uncover the most common issues when tourists visit overcrowded attractions and to reveal their feelings through text analytic techniques.
Data Collection and Data Preprocessing
All available English posts of the top 10 cultural-related attractions in Paris in TripAdvisor were extracted, resulting in a total of 140,712 posts published by any user as of the end of 2019. The attractions include Notre-Dame de Paris, Basilica of the Sacred Heart of Paris, Louvre Museum, Tour Eiffel, Centre Pompidou, Musée d’Orsay, City of Science and Industry, Museum of Natural History, Arc de Triomphe, and Sainte-Chapelle.
An open-source visual programming software, Orange 3, was applied for the following procedures. First, online posts were pre-processed. A list of stopwords was prepared to eliminate non-informative text. The remaining corpus was transferred to lowercase, where diacritics were transformed to the basic format. Next, text data was tokenized. All words were converted into their basic form, using lemmatization (e.g., traveling to travel).
Latent Dirichlet Allocation (LDA) Topic Model
This study applied LDA topic models to identify the underlying topics in an unstructured corpus such as customer reviews. Specifically, LDA views a document of text as a mixture of topics that disclose words with certain probabilities. However, due to the restriction of Orange 3 on the number of data instances, 5,000 posts were randomly selected for each attraction in Paris using a random selection in excel. LDA topic modeling was conducted to generate term clusters from the extracted reviews, which yielded 10 topics for each attraction based on the default setting of Orange 3. The degree of how a token contributes to a given review was revealed based on TF-IDF representation (term frequency-inverse document frequency).
In the next step, based on the identified topics, a lexicon-based sentiment analysis using the Vader algorithm was adopted to extract online users’ feelings based on the posts. Sentiments refer to feelings based on attitudes, emotions, and opinions; it determines whether an expression is positive, negative, or neutral. The results are presented by a numerical spectrum where −1 is the most negative, +1 is the most positive, and 0 suggests the neutral point.
Summary of the Results
LDA identifies 14 topics relevant to the issue of overcrowding and 10 general tourism-related topics. The table below presents the 14 overtourism topics and their corresponding average sentiment scores. The naming of the topic was based on the top keywords with the highest TF-IDF scores detected by LDA. For instance, the findings suggest that visitors felt most negatively about “safety and security”, “service and staff”, and “queues of customers”. Yet, the sentiment scores were higher regarding “social interaction”, “reputation”, and “overall atmosphere”. The detailed results can be found in our paper.
Unlike earlier research built upon existing measurements, this study takes one step further by exploratorily discovering the critical dimensions in managing tourist experiences at overcrowded attractions. Thus, this study contributes from a methodological angle by incorporating topic modelling technique and sentiment analysis to reveal tourists’ subjective perceptions. The technique applied in this study is beneficial to marketers who want to examine tourists’ feelings based on UGC elsewhere.
How to cite: Yu, J., & Egger, R. (2021). Tourist Experiences at Overcrowded Attractions: A Text Analytics Approach. In Information and Communication Technologies in Tourism 2021 (pp. 231-243). Springer, Cham.
Introduction: Online complaints have become increasingly influential on the purchasing behavior of customers in recent years. In an effort to analyze large quantities of textual complaints and detail the various aspects of them, Aspect-Based Sentiment Analysis was looked to as an ideal framework to take on the task.
Purpose: This study set out to synthesize specific service failure items and categorize them into the groupings related to the hotel guest cycle and the corresponding operations, then compare the patterns of expression used by Asian and Non-Asian guests as they related to their hotel experiences.
Design/Methodology/Approach: A total of 390,236 online complaint terms posted about 353 hotels in the UK by hotel guests from 63 nations were manually derived from TripAdvisor for analysis. In line with previous studies by Sezgen, Mason, & Mayer (2019); and Xu & Li (2016), the textual data was processed according to the following steps (see Figure 1). Then, to detect the aspect category, the researchers made use of terms that would be indicators of the presence of the aspect—these are referred to as Aspect Terms. Before the aspect-based sentiment analysis could be processed, the aspect terms and categories were identified (c.f. Sann & Lai, 2020). Finally, a list was compiled in which the extracted terms and aspects were organized in order of importance. These results were based on the co-occurrence frequency-based method (Brun et al., 2014; Schouten et al., 2018; Sharma & Waghmare, 2019).
Findings: With consideration given to the homophily theory, we posited that Asian and non-Asian guests would exhibit similarities and differences with respect to the service failures they encountered when presented with various items of service. The results confirmed this as Asian guests were shown to encounter more service failures with respect to the engineering segment of operations (e.g. hotel room equipment issues), while non-Asian guests encountered more service failures on the housekeeping end of operations (e.g. toilets, public areas, cleanliness, and bedding). By organizing the failures according to the four stages of the guest cycle, it was observed that approximately 80% of the service failures occurred during the occupancy period.
Originality/Value: This study contributes to the existing literature on hotel guest satisfaction both with respect to the methodology it uses and the new findings it presents on differences in perceptions of service failures members among different cultures.
Brun, C., Popa, D. N., & Roux, C. (2014). XRCE: Hybrid Classification for Aspect-based Sentiment Analysis. Paper presented at the 8th International Workshop on Semantic Evaluation (SemEval 2014), Dublin, Ireland.
Sann, R., & Lai, P. C. (2020). Understanding homophily of service failure within the hotel guest cycle: Applying NLP-aspect-based sentiment analysis to the hospitality industry. International Journal of Hospitality Management, 91. 102678. https://doi.org/10.1016/j.ijhm.2020.102678
Schouten, K., van der Weijde, O., Frasincar, F., & Dekker, R. (2018). Supervised and Unsupervised Aspect Category Detection for Sentiment Analysis with Co-occurrence Data. IEEE Trans Cybern, 48(4), 1263-1275. doi:10.1109/TCYB.2017.2688801
Sezgen, E., Mason, K. J., & Mayer, R. (2019). Voice of airline passenger: A text mining approach to understand customer satisfaction. Journal of Air Transport Management, 77, 65-74. doi:10.1016/j.jairtraman.2019.04.001
Sharma, G., & Waghmare, M. (2019). Review Summarization and Aspect Category Detection with Co-occurrence data by refining Word Embeddings. International Journal of Scientific Research and Review, 07(05), 58-65.
Xu, X., & Li, Y. (2016). The antecedents of customer satisfaction and dissatisfaction toward various types of hotels: A text mining approach. International Journal of Hospitality Management, 55, 57-69. doi:10.1016/j.ijhm.2016.03.003
Recent tourism research has adopted social media analytics to examine tourism destination image (TDI) and gained timely insights for marketing purposes. Comparing the methodologies of social media analytics and intercept surveys would provide a more in-depth understanding of both methodologies and a more holistic understanding of TDI than each method on their own. This study aims to investigate the unique merits and biases of social media analytics and a traditional visitor intercept survey.
This study collected and compared data for the same tourism destination from two sources: responses from a visitor intercept survey (n=1,336) and photos and metadata from Flickr (n=11,775). Content analysis, machine learning, and text analysis were used to analyze and compare the destination image represented from both methods. In particular, Latent Dirichlet Allocation (LDA) was the main topic modeling approach.
The results indicated that the survey data and social media data shared major similarities in the identified key image phrases. Social media data revealed more diverse and more specific aspects of the destination, whereas survey data provided more insights in specific local landmarks. Survey data also included additional subjective judgment and attachment towards the destination. Together, the data suggested that social media data should serve as an additional and complementary source of information to traditional survey data.
This study fills a research gap by comparing two methodologies in obtaining TDI: social media analytics and a traditional visitor intercept survey. Furthermore, within social media analytics, photo and metadata are compared to offer additional awareness of social media data’s underlying complexity. The results demonstrated the limitations of text-based image questions in surveys. The findings provide meaningful insights for tourism marketers by having a more holistic understanding of TDI through multiple data sources.
by Leyla Atabay & Beykan Çizel Journal of Tourism and Services, Vol. 11 No. 21 (2020) Link to Paper: https://jots.cz/index.php/JoTS/article/view/163
With this paper, we aimed to obtain meaningful themes and emotions from traveler reviews of hotels in three mass tourism destinations (Antalya, Majorca, Sharm El Sheikh). For that, we asked three main questions.
1) What are the features of the service components of the mass tourism destination hotels?
2) What are the emotions that arise from the analysis of hotel reviews according to mass tourism destinations?
3) What are the similarities and differences between the tourists’ emotions about the service components of hotels operating with the same concept in different mass tourism destinations?
Figure 1: Hotel Selection – Source: own processing
In order to answer these questions and scale the data, we determined the most liked 9 hotels (3 in each) in 3 destinations with similar characteristics and collected a total of 3588 reviews. At this point, we preferred the “rvest” package (rvest.tidyverse.org) in the leading R program in data mining applications and coded a script data from a hotel review website. We also used dplyr, tidytext, readxl, tm, syuzhet, wordcloud, lubridate, ggplot2, reshape2, rlang, and purrr packages for the other analyzes process.
After collecting the data, we moved on to the “pre” data cleaning process. First, we fixed or deleted the corrupted characters in the corpus data. Then we combined the words that would be synonymous with each other and the plural forms of the words. Thus, we created a file ready for cleaning in the R program.
In the data cleaning process, R offers some auxiliary functions. We would like to present some examples of these below.
As can be observed in the codes, we first defined an excel file, namely “reviewdf”. Then we specified which column to read and moved on to the steps of cleaning the corpus data. First, we converted all letters to lowercase. In the second step, we defined it as plaintext. In the third, fourth, and fifth steps, respectively, we removed punctuation, numbers, stopwords. In step six, we deleted any other words we believed were unnecessary for this data. Thus, we also had to clean up the “whitespace” created in the previous steps.
Immediately after cleaning, we defined a term-document matrix in the clean dataset and created wordclouds to understand the main themes of reviews.
Figure 2: Wordclouds – Source: own processing
However, wordclouds were not enough to understand the importance of themes and which service components they are related to. For this reason, we thought it would be more informative to create a network analysis from the data and calculate the link strengths between nodes (namely words in reviews). So we created word networks for three destinations.
Figure 3. Review Networks – Source: own processing
Comparison of reviews with the help of previous analyzes was hardly possible. More precisely, it was clear that the comparison was based on the visual reading capability of the reader. To overcome this problem and answer the third research question, we applied “Correspondence Analysis”.
Figure 5: Correspondence Analysis
Our findings highlighted the most important service features and prevailing emotions for hotels in Mediterranean destinations. Furthermore, the results of the multiple correspondence analysis revealed how emotions towards hotel services differ in three different destinations.
Atabay, L., & Çizel, B. (2020). Comparative Content Analysis of Hotel Reviews by Mass Tourism Destinations, Journal of Tourism and Services, 21(11), 147-166. doi:10.29036/jots.v11i21.163
Plutchik, R., (1980). Emotion: A Psychoevolutionary Synthesis, New York: Harper & Row.
Have you ever thought about why you ‘liked’ and ‘commented’ on some of the photos on Instagram, and some not? Certainly, content matters. However, when you need to browse through hundreds of Instagram posts, one of the subtle factors that may unconsciously influence your behavior is the color of the image. Color, as one of the major components in tourism aesthetics, influences human physiological responses and leads to changes in our online behavioral reactions.
Yet, the interplay between pictorial content and user engagement remains unclear and difficult to investigate. Our study applies a machine learning approach in order to investigate the role of color in influencing user engagement on Instagram based on tourism pictures with different features.
A six-step methodological procedure
Step 1 Picture selection: To identify tourism photos as the data sources on Instagram, the most commonly mentioned typologies summarized by previous literature were treated as hashtags to facilitate the data crawling process. They include “#beach”, “#mountain”, “#heritage”, “#forest”, “#gastronomy”, “#temple”, “#lake”, “#museum”, and “#cityscape”.
Step 2 Data extraction: A total of 7,887 public posts published between 2017 and 2019 were crawled, including the date, page URLs, image URLs, username, and the number of likes and comments. Yet, because engagement rate changes logarithmically, data was re-collected from the page URLs after 14 days of the first data extraction. After excluding the posts removed by the users, the final dataset contained 4,757 pictures.
Step 3 Image annotation and clustering: Because hashtags do not necessarily reflect the pictorial content (e.g., one might post a selfie with #mountain when hiking), we re-assigned the extracted pictures. Specifically, the image labels (i.e., the entities of a picture such as general objects, locations, activities, and animal species) annotated by Google Cloud Vision API were transformed into vectors using tf-idf value, indicating to what extent a label contributes to a picture. Next, the Louvain algorithm was applied to convert the detected labels into several clusters based on highly interconnected nodes (entities in the data). The image labels were considered as the edges (relationships between those entities) that connect different pictures, forming an image-network-graph and leading to a clustering of highly-connected images.
Step 4 Calculation of engagement rate: The next step was to calculate the average engagement rate of each identified cluster by taking the total number of likes and comments of a post and dividing it by a given user’s follower numbers.
Step 5 Color conversion: Google Cloud Vision was applied to detect a picture’s dominant colors. Cloud Vision returned up to 10 RGB values and their representative scores for each image. To ensure that the color presented is in line with human visual perception, RGB color codes were converted to hue, resulting in 12 major colors: orange, orange-yellow, yellow, yellow-green, green, blue-green, blue, blue-violet, violet, violet-red, red, and red-orange. Finally, to attain the percentage of color across the entire image, an individual color’s score was divided by the sum of all the scores returned from Cloud Vision.
Step 6 Implementation of machine learning methods: To analyze the relationship between color and the engagement rate of each cluster, auto machine learning with SVM and random forest was conducted. The engagement rate was selected as the target variable, and the color attributes were treated as input variables for prediction. Note that only the attributes that would contribute most to the quality of the resulting model were selected.
SVM and random forest were implemented based on the automatic optimization feature, optimizing the number of trees for the random forest and the gamma and C hyperparameter for SVM. The contribution of the model’s selected color was ranked based on weight vectors, which were calculated using a local interpretable model explanations (LIME) method. Specifically, LIME generates random samples around neighboring inputs and finds correlation weights for each input in the dataset. By summing up the weights of the color attributes based on their extent of contribution to engagement rate, a final output is given.
An overview of the results
The Louvain algorithm generated 24 image clusters. The labels with the highest tf-idf value were included as keywords to facilitate the naming process of each cluster. Concerning the effects of color, since the accuracy score of SVM was better than the random forest’s in most of the cases (by evaluating MSE and RMSE), the results from SVM were presented.
The table below provides four image clusters as examples, followed by the results of SVM. Notably, LIME used in SVM returns local weights and only focuses on the most relevant ones. For instance, violet has the highest weight in “urban views”, while violet-red conquers in “seascape” and blue-violet in “water and natural impressions” and “high-end cuisine”. Take “high-end cuisine” as an example, to achieve the highest engagement rate possible within this cluster, a picture should be composed more of blue-violet shades while slightly minimizing yellow-green and blue touches. The detailed results can be found in our paper.
Overall, applying SVM to Instagram data offers a new form of analysis for tourism and digital marketing. Our study provides a hands-on guide for marketers to bring images presented on Instagram to light in order to optimize their marketing content based on consumers’ preferences and interests. Meanwhile, our research provides some indications as to where and how future studies could collect data in a more structured fashion.
How to cite: Yu, J., & Egger, R. (2021). Color and engagement in touristic Instagram pictures: A machine learning approach. Annals of Tourism Research, 103204.