Understanding Customers’ Insights Using Attribution Theory: A Text Mining and Rule-Based Machine Learning Two-Step Multifaceted Method

Featured Application

Text Mining, Apriori Algorithm, Association Rules, Rule-based Machine Learning Method, Web Graph Analysis, Content Analysis, Qualitative Projective Technique, Textual Complaints.

Abstract

By looking at complaints made by guests of different star-rated hotels, this study attempts to detect associations between complaint attributions and specific consequences. A multifaceted approach is applied. First, a content analysis is conducted to transform textual complaints into categorically structured data. Furthermore, a web graph analysis and rule-based machine learning method are applied to discover potential relationships among complaint antecedents and consequences. These are validated using a qualitative projective technique. Using an Apriori rule-based machine learning algorithm, optimal priority rules for this study were determined for the respective complaining attributions for both the antecedents and consequences. Based on attribution theory, we found that Customer Service, Room Space, and Miscellaneous Issues received more attention from guests staying at higher star-rated hotels. Conversely, cleanliness was a consideration more prevalent amongst guests staying at lower star-rated hotels. Qualitative research was conducted to corroborate the findings. Other machine learning techniques (i.e., Decision Tree) build rules based on only a single conclusion, while association rules attempt to determine many rules, each of which may lead to a different conclusion. The main contributions of this study lie in the fact that this is one of the first attempts to detect correlations within the online complaining behaviors of guests of different star-rated hotels by utilizing rule-based machine learning.

1. Introduction

Comprised of numerous subsectors, the hospitality and tourism industry in England generates revenue in various ways: accommodations, F&B services, transportation, cultural attractions, and other recreational activities [1]. Travel agencies, tour operators, and other reservation services alone have generated revenue of over 40 billion pounds a year since 2013 [1], and in 2016, travel agents and tour operators generated over 33 billion British Pounds [2]. More pertinent to the current study, online travel booking revenue in 2017 amounted to roughly 18 billion USD, with expectations of reaching 24.1 billion in 2022 [3].

Travellers are becoming increasingly reliant on online booking platforms when it comes to making accommodation arrangements. And similar to what is seen with other online purchasing decisions, customers are relying heavily on peer reviews. This trend can be partially attributed to the fact that potential customers regard the information provided by peers to be independent and thus more trustworthy than information provided by business entities [4]. Within the tourism industry, the hotel sector is experiencing some of the most substantial effects of online reviews and, thus, has much to gain by understanding the situation and responding appropriately [5].

With respect to the data found in online booking reviews, existing research on hotel guests’ preferences, experiences, and expectations has mainly investigated behaviour using big data and textual analysis models. As such, these studies have not drawn out the connections between attribution and consequence. For instance, Liu, et al. [6] applied big data to explore language-specific factors of hotel satisfaction and found that some lower star-rated hotels offer better perceived service quality than high star-rated hotels based on guest ratings. Hu, et al. [7] studied hotel guest complaints in different hotel categories using text-based analytics. Their study revealed that cleanliness and problems with facilities are major sources of guest dissatisfaction at lower star-rated lodgings, while service quality problems and issues related to high pricing are major targets of guest complaints for higher star-rated hotels. This highlights the importance of discriminating between guests’ preferences for individual hotel attributes depending on the hotel category. Important insights to be gathered here are that hotel guests’ preferences, experience, and expectations are inclined to vary depending on the star-rating. Inasmuch as it is important for hoteliers to clearly understand the ways in which different standard hotel systems impact guests’ expectations so that they can give due attention to the most relevant hotel attributes. To add greater clarity to the picture, this study explores relationships that exist between complaint attributions and the consequences of guest behaviour by applying attribution theory as its main pre-determined framework. For these purposes, attributes that can typically be extracted from internet complaints (e.g., negative online reviews) were the focus of consideration. Differing from prior research on online reviews, this study adopted Apriori rule-based machine learning to identify associations between attributions and consequences within the online complaints of customers staying at different star-rated hotels.

In order to develop a better understanding of the customer review landscape for the hotel industry, this study follows an emerging style of research that makes use of user-generated data in the form of complaint reviews. Focus is given to the differences in the patterns of online complaining behaviour (OCB) exhibited by guests who stayed at higher star-rated and lower star-rated hotels. By applying newly-developed association rule (AR) algorithms to the analysis of online complaint reviews, the researchers have the opportunity to conduct a far-reaching empirical and quantitative exploration. And also, since few studies have made use of machine learning analyses to examine potential factors that may lead to OCB in the hospitality industry, the main purposes of this study are:

i.

To broaden the body of literature related to machine learning applications by applying it to the field of hospitality.

ii.

To examine complaint attributions of guests staying at different star-rated hotels to determine the relationships that exist between those attributions and their consequences.

iii.

To use attribution theory to develop a better understanding of the complaining behavior of hotel guests.

A total of more than 2000 online complaints from more than 350 hotels that are considered representative of their categories were collected manually. First, content analysis was conducted to structure the textual data. This process created ten analytical variables for the new qualification algorithms. Furthermore, web graph analysis and Apriori algorithms were linked together to perform the analysis. The approach was adopted in an attempt to enrich the literature related to machine learning by applying it to the hospitality industry. In this respect, Apriori was used to detect relationships that may exist between complaint antecedents and consequences by looking at complaints of travellers who stayed at different star-rated hotels. Finally, to ensure the credibility and trustworthiness of the findings produced by the algorithm, a qualitative projective technique was performed by hotel and academic experts who have more than 20 years’ experience in the hospitality industry. With the combination of the quantitative algorithms and the qualitative projective technique, the results of the study were reinforced via triangulation. The hope is that the results of this study will help hotel managers address the potential consequences of the concerns that guests have while visiting various star-rated hotels.

The main contributions of this study lie in the fact that it is one of the first attempts to use machine learning applications to detect correlations between complaint antecedents and consequences by exploring the online complaining behavior of guests of different star-rated hotels. Since previous research has not tested the links between attributions and consequences, this study synthesizes and expands existing literature by performing this task. In addition, a ruled-based machine learning method (i.e., Apriori Algorithm) has been adopted to various topics and areas including glasses sales [8], education [9], engineering [10], transaction reduction [11], minimizing candidate generation [12], sales pattern (association of items) at the O! Fish restaurant [13], computer science and communication [14], cyberspace security [15], and finally, in medical industry (e.g., heard disease prediction [16], identify frequent diseases [17], and medical billing [18]). However, regarding detecting the relationships between complaint attributions and specific consequences of guests’ online complaining behavior at various star-rated hotels, no previous studies have been so far conducted by utilizing Apriori algorithm; therefore, this study attempts to fill gaps and amplify the hospitality body of literature.

The following paragraph provides the details of this study, which is organized into five sections: a literature review, research methodology, summary of results, discussions and conclusions, and limitations and recommendations for future studies.

2. Literature Review

2.1. Rule-Based Machine Learning Applications: Association Rule Algorithms

In machine learning, a commonly applied technique that is used to discover interesting relations in the data is known as Association Rules (ARs) [19]. ARs is also referred to as Shopping Cart (or Market Basket) Analytics, and can be used to detect relationships or associations that exist between the specific values of categorical variables in large data sets [20]. ARs have been used to uncover hidden patterns. For instance, customers who order product A often also order product B and/or C; or employees who said positive things about initiative X also frequently complain about issue Y but are happy with issue Z. For retailers, ARs can be used to predict which products consumers will potentially purchase or make suggestions for future purchases. For instance, Amazon.com and other retailers use ARs to make recommendations for similar products or items that are frequently purchased together [21]. In this respect, the new knowledge that is produced via mining ARs is considered as informative as the knowledge derived from classical ones [22].

In this study, the researchers apply the ARs to the work of discovering frequent associated antecedents and consequences of complaints made online by guests of different star-rated hotels. The study also attempts to determine the associations that exist among multiple hotel attributes and use that combined information to make further predictions. For instance, a customer who complains about attribute A also complains about attribute B and will belong to hotel category X. The following paragraph briefly introduces two of the most popular AR models: Apriori and Carma.

Apriori Algorithm: This algorithm is used for extracting association rules from data [23] and pulling out rules with the highest information content [24]. It was proposed by Agrawal and Srikant [25]. Apriori is well-known to be the representative algorithm for Data Mining [26] and is one of the most popular and widely used algorithms for mining frequent item sets [27]. The technique was designed to operate on databases containing transactions [28]. According to the threshold value (or minimum support value), the algorithm is able to identify the frequent item sets which are subsets of transactions in the database [29]. Apriori deals with items and item sets that make up transactions. Items are flag-type conditions that indicate the presence or absence of a particular feature in a specific transaction. An item set is a group of items that may or may not tend to co-occur within transactions [30]. Apriori offers five different methods of selecting rules and uses a sophisticated indexing scheme to process large data sets efficiently. For large problems, Apriori is generally faster to use, has no arbitrary limit on the number of rules that can be retained, and can handle rules with up to 32 preconditions. While Apriori requires all input and output fields to be categorical, its optimization for this type of data allows it to deliver better performance [24].

Carma Algorithm: The Continuous Association Rule Mining Algorithm (Carma) is an alternative to Apriori that reduces I/O costs, time, and space requirements [23]. Using only two data passes, Carma is able to deliver results at much lower support levels than Apriori. It also allows changes to be made to the support level during execution. Similar to Apriori, Carma also deals with items and item sets that make up transactions. Items are flag-type conditions that indicate the presence or absence of a particular feature in a specific transaction. An item set is a group of items that may or may not tend to co-occur within transactions. Carma extracts a set of rules from the data without requiring the user to specify input or target fields [24]. In contrast to Apriori, the Carma node offers ‘build settings’ for rule support (support for both antecedent and consequent) rather than just antecedent support. This means that the generated rules can be used in a wider variety of applications—for example, to find a list of products or services (antecedents) whose consequent is the item that one wants to promote this holiday season [24].

2.2. Hotel Star-Rating System

The terms “hotel grading,” “hotel classification,” “hotel segment,” and “hotel rating” are used interchangeably to classify hotels according to service, facility level, and price [31]. The star-rating system is used to categorize hotels by level of operating performance [32], and send signals to guests about the hotel’s intrinsic quality [33]. However, as Tsao [34] points out, there is no universal standard for hotel ratings. According to the World Tourism Organization (UNWTO) and International Hotel and Restaurant Association (IHRA), “accommodation establishment of the same type (e.g., hotels, motels, and inns) have been conventionally broken down into classes, categories, or grades according to their common physical and service characteristics and established at government, industry or other private levels” [35]. Hotel rating systems are generally classified based on basic registration standards, the physical standards hotels need to follow, grading standards, and qualitative service-related aspects [35,36]. Hotel star-ratings can be expressed in various forms; for example, stars (e.g., Chinese National Tourism Administration), diamonds (e.g., American Automobile Association), or crowns (e.g., English Tourism Boards) [31,35]. The different star-ratings may be based on different objective criteria, such as infrastructure, services, amenities, and/or the sizes of the rooms and common spaces [37]. Higher star-ratings are also usually considered to be a good indicators of higher levels of quality [37] and greater likelihoods that guest expectations will be met during service delivery [35]. In Spain, for instance, higher star-rated hotels have been shown to offer the best facilities and services for the most expensive room rates, while lower star-rated hotels are more likely to have low-quality equipment, services, and levels of computerization—as well as lower prices [32]. Hensens [38] argued that a hotel star-rating has a major impact on hotel guests’ experiences and expectations; therefore, studies connecting rating systems with customer (dis)satisfaction and service quality are inevitable. Adopted and modified from Hensens [38], Figure 1 depicts organizations involved in rating systems.

2.3. Guest Expectations and Preferences Based on Different Hotel Star-Rating

Discussing the differences between star-rated hotel guest preferences, Liu, Teichert, Rossi, Li, and Hu [6] discovered that sometimes hotels of lower star designation are able to outperform hotels of higher designation in terms of the ratings that the guests assign. In terms of customer complaints, Ekiz, et al. [39] found Rooms received the highest number of complaints from luxury hotel guests, followed by Service Failures caused by inexperienced, unprofessional, or misbehaving staff. Fernandes and Fernandes [40] analysed 1191 guest reviews for hotels listed on TripAdvisor. In this study, it was shown that the categories of Cleanliness, Rooms, Sleep Quality, Bathrooms, Breakfast, and Facilities received the most negative statements for 2- to 3-star hotels. Comparatively, 4- to 5-star hotels received more negative statements directed at Customer Care, Location, and Value. The results of the study indicate that core activities are central to service providers at budget hotels, while peripheral activities help luxury hotels meet the expectations of guests. The findings also point out that facility problems or cleanliness are the biggest issues for guest dissatisfaction at lower-end hotels, while service-related problems and overpricing are the major sources of discontent among guests at high-end hotels. Ultimately, this reaffirms the importance of identifying the preferences and expectations of the guests and recognizing how they can vary across different categories of hotels [6,7,41,42]. As discussed, while extant literature has identified causes of guest dissatisfaction by applying traditional approaches (e.g., critical incident technique “CIT” or multivariate technique), this study may be the first attempt to explore the causes of online customer complaints by applying a rule-based machine learning method. More specifically, this study attempts to explore whether different attributes of online complaints made by guests of different star-rated hotels can be explained by differences in attribution and consequence. The hope is that the findings will be of considerable value to those in the industry who are seeking to improve service standards and quality.

2.4. Negative Experiences and Attribution Theory

Service quality and customer satisfaction are the two most important factors for service organizations. Service quality occurs during the service encounter, or “moment of truth”, where the customer has an interaction with the service provider. If the customer encounters any incident during this moment, it may result in the customer leaving the situation dissatisfied [43]. Chua, et al. [44] demonstrated that intangible and tangible elements were the two most important factors that determined customers’ perceptions of service quality. If the service expectations are not met, service failures occur. This, in turn, gives potential rise to complaints and/or negative experiences [45,46,47,48]. The complaints themselves may also be regarded as negative experiences, even though they may be determined by factors outside the control of service providers. Therefore, whether or not the customers attribute the causes of the negative experience to the service provider will depend on their perception of the surrounding details [32].

Attribution theory is a framework that explains the causal relationship between two incidents, assuming that people tend to uncover and evaluate the causes of their (dis)satisfaction [49], p. 549). The word attribution refers to the perceived or inferred cause of the result. The common principle is that individuals interpret behaviors in terms of their causes and that these interpretations play an important role in determining the reactions to those behaviors [50], p. 458). The theory also focuses on the interactions that take place between the attributions one makes about one’s own and others’ performances (on the “antecedents side”) and the effect they have on one’s subsequent choices and actions (on the “consequences side”) [51,52]. The antecedent side includes certain information about the behavior or circumstances leading up to the complaint behavior. This information is applied by the subject to infer the cause of the outcome [50]. For a guest staying at a high-end hotel, for instance, one can assume “high expectations” would be part of the antecedents. In this situation, compliance would be regarded as “inside” the expectations due to the fact that the guest is spending more money and, perhaps, has past staying experiences to judge by. If another guest were to spend less, he or she might have lower expectations, and the cause for compliance would be “outside” the expectation. Meanwhile, the consequences side deals with one particular result of the complaint behavior [50]. In situations where compliance is considered to fall within expectations, the subject credits the responsible party with positive or negative attitudes and/or traits based on the outcome of the event. Based on the attribution-consequence relationship, this study attempts to explore factors that might result in a subject attributing a particular cause to an event and examine how one cause might lead to another cause due to that particular attribution. In general, attribution theory has been widely discussed in literature (e.g., Weiner [51,52], Graham [53], LaBelle and Martin [54], and Orth, et al. [51,52,53,54,55]); however, little research has been pursued to investigate the causes for guests’ complaint levels, and even less work has been completed on psychological explanations [56,57]. Therefore, this study utilizes attribution theory to explain how to understand guests’ complaining behavior by investigating which variables may be critical in understanding these phenomena. Understanding the causes behind dissatisfying service encounters is essential for hoteliers who hope to reduce dissonance levels among customers. To help develop this understanding, this study uses a conceptual framework adopted from [50] Kelley and Michela’s (1980) comprehensive attribution model, while exploring the role that reliance on negative reviews or online complaints plays in the initial formation of opinions. Figure 2 illustrates the research framework adopted from [50] Kelley and Michela’s general models of attribution.

3. Methodology

3.1. Preparation and Processing Data

Step 1: collection and samples

In total, 2020 usable online complaints were collected from the TripAdvisor site for the London market in the United Kingdom [58]. These complaints were directed at 353 hotels, which rank from two- to five-star based on the British hotel rating system. A maximum of 20 of the most recently posted complaint reviews with details of the complaint for each hotel were manually extracted for analysis to ensure efficiency and proper representation of complaint data. All the reviews used had an Overall ratings of one- or two-star, as these ratings are recognized as “complaints” within the TripAdvisor system [59,60]. These complaints were organized according to higher star-rated and lower star-rated hotel groupings (

Table 1. Dataset of each hotel category.

Hotel Category	Complaints	(%)
Hotel Star-Rating
2-Star	315	15.6
3-Star	557	27.6
4-Star	603	29.9
5-Star	545	27.0
Hotel Class
Higher Star-Rated	1148	43.2
Lower Star-Rated	872	56.8
Total	2020	100

).

Step 2: Coding procedures

To begin, each online complaint was obtained manually from each hotel webpage. Furthermore, qualitative content analysis was utilized to transform unstructured textual content into structured data [61]. Thirdly, coding categories were developed, and manual coding was performed. In order to avoid overlapping themes and reduce ambiguity in the content, coding items and variables were added, removed, or merged based on the deliberations of two independent coders [62]. Coding variables and their descriptions were identified and then verified with previous literature [6,62,63]. Following this, another categorical data file was created for the purpose of testing the algorithm models with the coding attributes and items described above. In total, this study identified 10 complaint attributes.

Table 2. Coding descriptions and complaint variables.

Source	Data Type	Variables	Coding	Coding Descriptions
Hotel	Categorical	Hotel Star-Rating	H/L	Higher Star-Rated (4- to 5-star) and Lower Star-Rated (2- to 3-star)
Coding	Categorical	Customer Service	Yes/No	Not friendly or warm, unhappy at work, rude, slow, unprofessional
Coding	Categorical	Room Facility	Yes/No	No toiletries, no complementary tea/coffee bags, no drinking water
Coding	Categorical	Hotel Facility	Yes/No	5* charges internet, slow Internet, No concierge service, no parking on site
Coding	Categorical	Cleanliness	Yes/No	Dirty waiter uniform, dirty sheets/duvet, blood on bed/duvet, stained carpet
Coding	Categorical	Location Accessibility	Yes/No	Out of center, far from bus station, away from tourist spots/shopping center
Coding	Categorical	Value for Money	Yes/No	High price tag for average hotel, tiny yet expensive room, poor value for money
Coding	Categorical	Safety	Yes/No	No floor-plan indicating fire exit, stolen money, unlocked my safe, burglarized
Coding	Categorical	Room Space	Yes/No	Tiny, no space to open your suitcase! Cramped shower room, very small scale
Coding	Categorical	F & B Issue	Yes/No	Noisy dinning, insufficient staff to serve, poor atmosphere, food very ordinary
Coding	Categorical	Miscellaneous Issue	Yes/No	bed bug bites, full of dead flies, too soft mattress, hard pillow, very rough sheet

presents the coding descriptions and complaint variables.

Step 3: Coding reliability

Lombard, et al. [64] and Gerdt, et al. [65] suggested that for the coding results to be accepted, two independent coders should be used. In order to test the inter-coder reliability, a percentage of agreement test was performed. This study followed the Sann, Lai, and Chang [62] and Cenni and Goethals [66] two-step inter-code reliability test. A preliminary inter-code reliability test was conducted after coding 5% of both of the full sets of coded reviews, and a follow-up reliability test was performed after coding another 10% of the total samples. The results for both coding grids were >90%, which was judged acceptable.

Step 4: Testing correlation among variables

Due to the fact that our variables are categorical data with two levels (i.e., yes or no); therefore, Phi Coefficient and Cramer’s V correlation should be tested [67,68,69,70]. Phi is “a measure for the strength of an association between two categorical variables in a 2 × 2 contingency table” [71] (p. 92). It is “calculated by taking the chi-square value, dividing it by the sample size, and then taking the square root of this value” [71] (p. 92). It ranges between 0 and 1 without any negative values [72,73]. Cramer’s V is an “alternative to phi in tables bigger than 2 × 2 tabulation” [71] (p. 92). Cramer’s V ranges between 0 and 1 without any negative values. Similar to Pearson’s r, “a value close to 0 means no association. However, a value bigger than 0.25 is named as a very strong relationship for Cramer’s V” [71] (p. 92) [74,75,76]. From the analysis, all the variables have a p-value of a pearson Chi-Square test higher than 0.05; and Phi and Cramer’s V values vary between 0.015—0.036 with a p-value higher than 0.05; thus, the Phi Coefficient and Cramer’s V Correlation are met.

Table 3. Phi and Cramer’s V interpretation.

Phi and Cramer’s V	Interpretation
>0.25	Very strong
>0.15	Strong
>0.10	Moderate
>0.05	Weak
>0	No or very weak

Source: Akoglu [71].

depicts the Phi and Cramer’s V interpretation, which was adopted from Akoglu [71].

3.2. Apriori Algorithm Knowledge Modellingble

An Apriori algorithm was applied in this study in order to identify the best rules for the model. This type of algorithm is among the most popular for conducting rule-based machine learning association rule concepts [77]. A total of ten online complaint attributes (see Table 2) were set to input, meaning they supplied the antecedent side, while the Hotel Star-Ratings were set to target, meaning they provided support for the consequent. In this study, ARs are applied to guest OCB in hotel contexts, working to develop an in-depth understanding of online complaining based on a cross-section of varied hotel attributes. Ultimately, the researchers are interested in learning how hotel guests complain about different types of hotel service attributes. A simple example of the complaining pattern sequence can be represented in the form of A → B, where A (called the antecedent or first attribute) and B (called the consequent or second attribute) are two disjoined item sets. This can be stated in the following form:

ifantecedent (s) then consequent (s)

ARs are used in this study to help determine how certain requirements should be related to each other so as to get an appreciable output. The minimum rule confidence threshold for those taken into consideration should be at least 70%, with a minimum antecedent support of 10% and a maximum number of four antecedents. Those outside the threshold should be rejected. The confidence threshold value represents the percentage by which the main objective of the rule is satisfied [28]. The confidence of the rule is defined as:

$$\mathrm{Confidence}\left(\mathrm{A}\to \mathrm{B}\right)=\frac{\mathrm{Support}\left(\mathrm{A}\cup \left.\mathrm{B}\right)\right.}{\mathrm{Support}\left(\mathrm{A}\right)}$$

where,

• Support (A ∪ B) = number of transactions including A ∪ B;
• Support (A) = number of transactions including A.

The Apriori algorithm proceeds in two stages. First, it identifies frequent itemsets in the data, and then it generates rules from the table/lattice of the frequent itemsets [23,30]. The top rules are generated with the assistance of SPSS Modeler 18. The full dataset can be freely accessed at http://dx.doi.org/10.17632/zn66tdcbh2.1 (accessed on 20 January 2023) or you can freely obtain from the MDPI Supplementary File. Figure 3 shows the process of AR algorithm analysis.

3.3. Algorithm Credibility: Qualitative Projective Technique

In order to ensure the credibility of the algorithm analyses, a qualitative projective technique was used to verify and validate them. The validation is important considering that this is a qualitative study dealing with “the why” behind guest behavior, with hopes of revealing underlying motives, values, and attitudes—including feelings and emotions—toward particular services or products [78]. The projective technique is suitable for elucidating unconscious feelings or thoughts by applying indirect questioning and stimuli (e.g., pictures, scenarios, or words) [78,79]. Data collection and related procedures were employed according to a non-probability, purposive sampling method. This allowed the researchers to choose participants appropriate to the purpose of the study. To achieve this aim, a social network sampling approach was adopted where participants were initially selected from within the researchers’ social network and added to via the networks of those initially inducted [80].

Data was gathered using face-to-face, semi-structured interviews, as this approach allowed the researchers to gain a better understanding of the subjective experiences and perceptions of the participants [78]. Before the interview commenced, interviewees were provided with explanations and asked to read a study information sheet which included the study purpose, procedures, and findings from the machine learning application. The purpose of this study was described to the participants in a general manner, and an example question was: “we are interested in learning what you think and feel regarding the study’s findings with respect to guest complaining behavior”. Since the intentions of this study are not to construct a comprehensive theory but instead to look into the thoughts and feelings that were evoked by the findings of the machine learning application, we only detail the subjective experiences and perceptions of interviewees with respect to the guest complaining behavior. In total, four interviewees were interviewed for the study. The interviews took place at various locations (e.g., offices, hotels). On average, each of the interviews lasted 30 min. The interviews were conducted in person and tape-recorded. The recordings were later fully transcribed. The participants comprised hotel industry and academic experts who have more than 20 years of experience in the fields of hotel and hospitality. By collecting the comments of multiple experts, the researchers were able to cross-check for individual bias by performing subsequent comparisons [81]. The hotel experts included individuals from each of the different star-rated establishments. By interviewing a mixture of hotel and academic experts, data triangulation could be performed to reinforce the validity of the findings.

Table 4. Participants’ demographic information.

Participant No.	ID.	Gender	Age	Occupation	Education Level	Interview Location
P01	Acad. 01	Male	56	Professor	Ph.D.	University Office
P02	Acad. 02	Female	47	Professor	Ph.D.	University Office
P03	Mgr. 01	Female	45	Hotel Manager (Higher star-rated hotel)	Ph.D.	Office
P04	Mgr. 02	Female	43	Hotel Manager (Lower star-rated hotel)	Master	Hotel lobby

Note: Acad.: Academic; Mgr.: Manager.

demonstrates the participants’ demographic information. Figure 4 describes the summary of the research process, including both the quantitative algorithm analyses and the qualitative projective technique.

4. Findings and Discussion

4.1. Web Graph Analysis

Web graphs show the strength of a relationship that exists between the values of two or more symbolic (categorical) fields. The graph utilizes lines of various widths to indicate the strength of the connection. These, for instance, have been used to explore the relationship that exists between the purchases of multiple different items at an e-commerce site or a traditional retail outlet. In the medical industry, web graph analyses have been used to explore relationships between cholesterol levels, blood pressure, and drugs that were effective in treating the patient’s illnesses [23,30]. Strong connections are presented with heavier lines. These indicate that the two values are strongly related and deserve further exploration. Medium connections are illustrated with lines of a lighter thickness, and weak connections are represented with dotted lines. If no line is shown between two values, it means either that the two values never occur in the same record or that their rate of occurrence does not meet the threshold specified in the Web node dialog box [82]. In this study, a maximum of 80 links may be displayed; weak links are those below 15, and strong links are those above 35.

4.1.1. Web Graph Analysis for Higher Star-Rated Hotels

Figure 5 is a web graph analysis of online complaint attributes for higher star-rated hotels. A clear difference was observed between the networks that formed. For instance, Location Accessibility and F & B Issue exhibited a strong direct connection (98.73%), meaning that these two attributes have a strong relationship. From this, it could be inferred that if hotel guests staying at higher star-rated hotels do not complain about Location Accessibility, they might complain about F & B Issue. Clearly, this is a relationship that could benefit from further research. The graph also shows several medium connections between various pairs of complaint attributes; for example, Location Accessibility and Customer Service (34.62%). This connection could suggest that if hotel guests complain about Location Accessibility, they might also complain about Customer Service or vice versa. Room Space and Room Facility (33.15%) and Miscellaneous Issue and F&B Issue (31.21%) also were among the top three medium connections. Some examples of weak connections include Safety and Cleanliness (14.75%) and Hotel Facility and F&B Issue (14.65%) (see

Table 5. Top 10 links for higher star-rated hotel.

Strong links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	98.73%	Location Accessibility = “No”	F & B Issue = “Yes”
2	98.73%	Safety = “No”	F & B Issue = “Yes”
3	98.59%	Location Accessibility = “No”	Customer Service = “Yes”
4	98.36%	Location Accessibility = “No”	Safety = “Yes”
5	98.31%	Location Accessibility = “No”	Room Facility = “Yes”
6	98.10%	Location Accessibility = “No”	Cleanliness = “Yes”
7	98.09%	Location Accessibility = “No”	Hotel Facility = “Yes”
8	98.02%	Safety = “No”	Miscellaneous Issue = “Yes”
9	97.98%	Location Accessibility = “No”	Miscellaneous Issue = “No”
10	97.95%	Location Accessibility = “No”	Room Space = “No”
Medium links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	34.62%	Location Accessibility = “Yes”	Customer Service = “Yes”
2	33.15%	Room Space = “Yes”	Room Facility = “Yes”
3	31.21%	Miscellaneous Issue = “Yes”	F & B Issue = “Yes”
4	30.77%	Location Accessibility = “Yes”	Room Space = “Yes”
5	29.51%	Safety = “Yes”	Customer Service = “Yes”
6	28.66%	Hotel Facility = “Yes”	Room Space = “Yes”
7	27.55%	Value for Money = “Yes”	Miscellaneous Issue = “Yes”
8	26.11%	Hotel Facility = “Yes”	Cleanliness = “Yes”
9	23.08%	Location Accessibility = “Yes”	Value for Money = “Yes”
10	22.93%	Room Space = “Yes”	F & B Issue = “Yes”
Low Links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	14.75%	Safety = “Yes”	Cleanliness = “Yes”
2	14.65%	Hotel Facility = “Yes”	F & B Issue = “Yes”
3	14.29%	Value for Money = “Yes”	Cleanliness = “Yes”
4	13.11%	Safety = “Yes”	Miscellaneous Issue = “Yes”
5	12.10%	F & B Issue = “Yes”	Room Facility = “Yes”
6	11.54%	Location Accessibility = “Yes”	Hotel Facility = “Yes”
7	11.54%	Location Accessibility = “Yes”	Room Facility = “Yes”
8	11.48%	Safety = “Yes”	Room Space = “Yes”
9	8.20%	Safety = “Yes”	Hotel Facility = “Yes”
10	8.20%	Value for Money = “Yes”	Safety = “Yes”

Note: We selected only top 10 links due to the space limitation; Field: means attribute.

).

4.1.2. Web Graph Analysis for Lower Star-Rated Hotels

Figure 6 is a web graph analysis of online complaint attributes for lower star-rated hotels. Here, a clear difference was observed between the networks that formed. For instance, Value for Money and Safety have a strong and direct path of connection (99.08%), meaning that these two attributes have a strong relationship. Based on the connection, it could be inferred that if the hotel guests at lower star-rated hotels complain about Value for Money, they might neglect to voice concerns about Safety. This question could benefit from further research. The graph also shows medium connections between several pairs of complaint attributes, for example, Value for Money and Miscellaneous Issue (33.94%). This may indicate that hotel guests who complain about Value for Money, may also complain about Miscellaneous Issues and vice versa. Location Accessibility and Miscellaneous Issue (33.33%) and Safety and Miscellaneous Issue (33.33%) were also among the top three medium connections. Some examples of weak connections include Value for Money and Hotel Facility (14.68%) and Location Accessibility and Cleanliness (14.29%) (see

Table 6. Top 10 links for lower star-rated hotel.

Strong links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	99.08%	Value for Money = “Yes”	Safety = “No”
2	99.07%	Location Accessibility = “No”	Room Facility = “Yes”
3	99.07%	Safety = “No”	Room Facility = “Yes”
4	99.03%	Safety = “No”	F & B Issue = “Yes”
5	98.90%	Location Accessibility = “No”	Cleanliness = “Yes”
6	98.48%	Location Accessibility = “No”	Room Space = “Yes”
7	98.48%	Safety = “No”	Room Space = “Yes”
8	98.44%	Location Accessibility = “No”	Hotel Facility = “Yes”
9	98.22%	Safety = “No”	Customer Service = “Yes”
10	98.01%	Location Accessibility = “No”	Miscellaneous Issue = “Yes”
Medium links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	33.94%	Value for Money = “Yes”	Miscellaneous Issue = “Yes”
2	33.33%	Location Accessibility = “Yes”	Miscellaneous Issue = “Yes”
3	33.33%	Safety = “Yes”	Miscellaneous Issue = “Yes”
4	32.41%	Room Facility = “Yes”	Cleanliness = “Yes”
5	32.04%	Miscellaneous Issue = “Yes”	F & B Issue = “Yes”
6	28.79%	Room Space = “Yes”	Cleanliness = “Yes”
7	23.81%	Location Accessibility = “Yes”	Value for Money = “Yes”
8	23.49%	Miscellaneous Issue = “Yes”	Customer Service = “Yes”
9	23.33%	Safety = “Yes”	Cleanliness = “Yes”
10	23.30%	F & B Issue = “Yes”	Customer Service = “Yes”
Low links	Links	Web of 10 fields: Overall %
ID		Field 1	Field 2
1	14.68%	Value for Money = “Yes”	Hotel Facility = “Yes”
2	14.29%	Location Accessibility = “Yes”	Cleanliness = “Yes”
3	12.04%	Value for Money = “Yes”	Room Facility = “Yes”
4	11.65%	Value for Money = “Yes”	F & B Issue = “Yes”
5	10.68%	Hotel Facility = “Yes”	F & B Issue = “Yes”
6	10.68%	Room Space = “Yes”	F & B Issue = “Yes”
7	9.52%	Location Accessibility = “Yes”	Hotel Facility = “Yes”
8	9.52%	Location Accessibility = “Yes”	Room Space = “Yes”
9	6.80%	F & B Issue = “Yes”	Room Facility = “Yes”
10	6.67%	Safety = “Yes”	Room Space = “Yes”

Note: We selected only top 10 links due to the space limitation; Field: means attribute.

).

The Web graph depicts information related to the pair-wise associations [21]. However, it is important to decompose the network for the purpose of interpreting the sets of frequent associations. The processes can provide greater insight on the associations that exist between the online complaining attributes. To develop a more in-depth understanding of the sets of frequent associations, this study employed Apriori rule-based machine learning.

4.2. Apriori Algorithm Findings for Online Complaining Behavior

The generated rules are presented based on a minimum confidence of 70% and a maximum of four antecedents. Based on the analysis, the top seven priority rules for the different star-rated hotels were compiled from the transactional data for the online complaint attributes (

Table 7. The top seven priority association rules for online complaining behavior related to different hotel-star ratings.

Rule Ranking	Rule ID	Consequent	Antecedent	Support (%)	Confidence (%)	Instances	Rule Support (%)	Lift	Deployability	Rule
1	15	Hotel Star-Rating = Higher Star-Rating	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No and Location Accessibility = No	10.446	71.090	211	7.426	1.251	3.02	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No and Location Accessibility = No ==> Higher Star-Rating
2	6	Hotel Star-Rating = Higher Star-Rating	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No	10.644	70.698	215	7.525	1.244	3.119	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No ==> Higher Star-Rating
3	1	Hotel Star-Rating = Higher Star-Rating	Room Space = Yes and Cleanliness = No	15.842	70.625	320	11.188	1.243	4.653	Room Space = Yes and Cleanliness = No ==> Higher Star-Rating
4	10	Hotel Star-Rating = Higher Star-Rating	Customer Service = Yes and Cleanliness = No and Location Accessibility = No	37.228	70.612	752	26.287	1.242	10.941	Customer Service = Yes and Cleanliness = No and Location Accessibility = No ==> Higher Star-Rating
5	16	Hotel Star-Rating = Higher Star-Rating	Customer Service = Yes and Miscellaneous Issue = No and Cleanliness = No and Location Accessibility = No	28.960	70.598	585	20.446	1.242	8.515	Customer Service = Yes and Miscellaneous Issue = No and Cleanliness = No and Location Accessibility = No ==> Higher Star-Rating
6	13	Hotel Star-Rating = Lower Star-Rating	Cleanliness = Yes and Customer Service = No and Room Space = No and Hotel Facility = No	10.941	70.588	221	7.723	1.635	3.218	Cleanliness = Yes and Customer Service = No and Room Space = No and Hotel Facility = No ==> Lower Star-Rating
7	14	Hotel Star-Rating = Lower Star-Rating	Cleanliness = Yes and Customer Service = No and Value for Money = No and Hotel Facility = No	11.436	70.563	231	8.069	1.635	3.366	Cleanliness = Yes and Customer Service = No and Value for Money = No and Hotel Facility = No ==> Lower Star-Rating

Note: We selected only top 7 best priority rules due to the space limitation. Note: Rule ID, Instances, Support, Confidence, Rule Support, Lift and Deployability are defined by IBM SPSS Modeler 18.0 Modeling Nodes [30], page 251. Rule ID: displays the rule ID assigned during model building. A rule ID enables you to identify which rules are being applied for a given prediction. Instances: displays information about the number of unique IDs to which the rule applies—that is, for which antecedents are true. Support: displays the proportion of IDs in the training data for which the antecedents are true. Confidence: displays the percentage of the IDs where a correct prediction is made, out of all the IDs for which the rule makes a prediction. It is calculated as the number of IDs for which the entire sequence is found divided by the number of IDs for which the antecedents are found, based on the training data. Rule Support: for Sequence models is based on instances and displays the proportion of training records for which the entire rule, antecedents, and consequent(s), are true. Lift: displays the ratio of confidence for the rule to the prior probability of having the consequent. Deployability: is a measure of what percentage of the training data satisfies the conditions of the antecedent but does not satisfy the consequent.

). The first rule listed in Table 6 posits that if a guest makes an online complaint about Miscellaneous Issues and Customer Service but doesn’t complain about Safety and Location Accessibility, then the guest stayed at a higher star-rated hotel. This rule has a confidence level of 71.09%. This means that guests making complaints about Miscellaneous Issues and Customer Service stayed at higher star-rated hotels 71.09% of the time. Furthermore, the support of the rule is 10.45%, meaning that in the entire transactional database, if a guest made online complaints about Miscellaneous Issues and Customer Service, it was associated with higher star-rated hotels 10.45% of the time. In another example, according to the 6th rule, if Cleanliness = Yes and Customer Service = No and Room Space = No and Hotel Facility = No, then the confidence is 70.59% that the complaint was made by a customer of a lower star-rated hotel. Therefore, if a customer makes an online complaint about Cleanliness, but not Customer Service, Room Space, and Hotel Facility, then in 70.59% of the time, the complaint applies to a lower star-rated hotel. Furthermore, the support for this rule is 10.94%, meaning that in the entire transactional database, if a guest makes an online complaint about Cleanliness, but not about Customer Service, Room Space, and Hotel Facility, it will be associated with a lower-star-rated hotel 10.94% of the time. Table 6 depicts the top seven priority association rules for online complaining behavior related to different hotel-star ratings.

Using the Apriori algorithm, the top seven priority rules were established for the antecedents and consequences of online complaint attributions. The results of these tests reveal that if guests make online complaints about Customer Service, Room Space, and Miscellaneous Issue but do not complain about Safety, Location Accessibility, Hotel Facility, and Cleanliness, then the guests are likely coming from higher star-rated hotels. First, with respect to Customer Service, the source of complaints is often related to interactions that take place between service providers and guests. The service failures may be related to the process or the outcome of the service delivery [83,84]. It is not necessary to assume that guests always have high expectations for high-profile services; in fact, often times it is enough to simply provide basic greetings, demonstrate welcoming attitudes, and maintain harmonious atmospheres [85,86,87]. One reason guests can end up dissatisfied may be related to their preparation processes. Guests who develop high expectations based on their personal preparations and planning might more easily find themselves in situations where those expectations are not met, and perceived service failures will occur—this in turn leads to complaining behavior as per the theory of expectancy disconfirmation [88,89]. Based on this theory, once a service has been experienced, the outcomes are evaluated in comparison with expectations [90,91,92]. When an outcome meets expectations, confirmation occurs. Disconfirmations (positive or negative) take place when there is a discrepancy between expectations and outcomes. A negative disconfirmation occurs when the outcome is worse than expected or when the service performance is less than anticipated [93,94]. Negative disconfirmations are particularly prone to producing dissatisfaction or complaining behavior. Mok and Armstrong [88] also pointed out that guest satisfaction is largely based on meeting or exceeding expectations. In order to explain the results of the current study, guests may have developed expectations for high quality services from past experiences staying at higher-star-rated hotels. On subsequent travels if there is a gap between expectations and actual services, a negative disconfirmation will occur. When it comes to Room Space, expectations may vary from one individual to another—they may also be influenced by locational factors. Guests who travel to London, for instance, might be accustomed to living in less-crowded environments, and when confronted with the spatial limitations, they may find it an unhappy experience [62]. Finally, with respect to Miscellaneous Issues, negative results could be related to a wide variety of things, including insects in the room, mattresses that are too soft, pillows that are too hard, or sheets that are too rough. For these issues, hotel managers can only try to ensure that their tangible and intangible in-house facilities are always fully operational before welcoming new arrivals.

When it came to online complaints made about on Cleanliness, but not about Customer Service, Room Space, or complaints made about Hotel Facility and Value for Money, guests were found to have stayed at lower star-rated hotels. This is comparable with results from Dutta and Das [45], Dutta, Jauhari, Venkatesh and Parsa [46] and Xu and Li [95], who asserted that cleanliness and dirty rooms caused the most dissatisfaction for guests staying in suite hotels. Since hotels are mainly “accommodations-led”, the main reason for guests to visit or return is the guest rooms and their cleanliness. Simply stated, clean and comfortable rooms help satisfy customers, while dirty ones lead to guest dissatisfaction [95]. Alhelalat, et al. [96], Ju, et al. [97], Radojevic, et al. [98,99], and Radojevic, et al. [100] all point out that a comfortable bed, a clean room and bathroom, and a certain amount of aesthetic appeal are considered important and necessary features to achieve customer satisfaction—these are also factors that will influence customer evaluations of their staying experiences. Other studies on public transportation also concluded that cleanliness and comfort are the most important factors when it comes to (dis)satisfaction, although the impact of these factors does differ depending on the socio-economic characteristics of the users [89,101,102,103,104,105,106]. Appendix A 

Table A1. Total of 17 association rules.

Rule Ranking	Consequent	Antecedent	Rule ID	Instances	Support %	Confidence %	Rule Support %	Lift	Deployability
1	Hotel Star-Rating = Higher Star-Rated	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No and Location Accessibility = No	15	211	10.446	71.090	7.426	1.251	3.020
2	Hotel Star-Rating = Higher Star-Rated	Miscellaneous Issue = Yes and Customer Service = Yes and Safety = No	6	215	10.644	70.698	7.525	1.244	3.119
3	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No	1	320	15.842	70.625	11.188	1.243	4.653
4	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Cleanliness = No and Location Accessibility = No	10	752	37.228	70.612	26.287	1.242	10.941
5	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Miscellaneous Issue = No and Cleanliness = No and Location Accessibility = No	16	585	28.960	70.598	20.446	1.242	8.515
6	Hotel Star-Rating = Lower Star-Rated	Cleanliness = Yes and Customer Service = No and Room Space = No and Hotel Facility = No	13	221	10.941	70.588	7.723	1.635	3.218
7	Hotel Star-Rating = Lower Star-Rated	Cleanliness = Yes and Customer Service = No and Value for Money = No and Hotel Facility = No	14	231	11.436	70.563	8.069	1.635	3.366
8	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Cleanliness = No and Safety = No and Location Accessibility = No	17	733	36.287	70.532	25.594	1.241	10.693
9	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No and Location Accessibility = No	5	312	15.446	70.513	10.891	1.241	4.554
10	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No and Hotel Facility = No	3	271	13.416	70.480	9.455	1.240	3.960
11	Hotel Star-Rating = Higher Star-Rated	Miscellaneous Issue = Yes and Customer Service = Yes and Location Accessibility = No	7	215	10.644	70.233	7.475	1.236	3.168
12	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No and Safety = No	4	312	15.446	70.192	10.842	1.235	4.604
13	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No and Hotel Facility = No and Location Accessibility = No	11	265	13.119	70.189	9.208	1.235	3.911
14	Hotel Star-Rating = Higher Star-Rated	Room Space = Yes and Cleanliness = No and Safety = No and Location Accessibility = No	12	305	15.099	70.164	10.594	1.235	4.505
15	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Cleanliness = No	2	767	37.970	70.143	26.634	1.234	11.337
16	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Miscellaneous Issue = No and Cleanliness = No	8	596	29.505	70.134	20.693	1.234	8.812
17	Hotel Star-Rating = Higher Star-Rated	Customer Service = Yes and Cleanliness = No and Safety = No	9	747	36.980	70.013	25.891	1.232	11.089

demonstrates a total of 17 rules.

4.3. Qualitative Study Results

As mentioned earlier, the intention of the qualitative portion of this study is to elucidate the subjective experiences and perceptions of the participants with respect to the results of the machine learning applied in the quantitative portion of the study. During the interviews, the participants expressed their understanding that guests place their focus on different features when staying at different categories of hotels. Based on their subjective assessments, the participants also conjectured that the guests staying at higher star-rated hotel categories stressed customer service and value for money, while those staying at lower star-rated hotel categories were more inclined to focus their concern on cleanliness and facility-related problems. Often times, the guests will place their main focus on one attribute and build on it by connecting it with other attributes. This tendency is expressed in an example given by one of the participating hotel managers:

★ “Normally, when the guest complains about services, what she/he would like to focus on is the importance of this issue, and they will connect it to other attributes to complain together…for example, ‘it might impact on my safety, comfort, and my whatever’…” (Mgr. 01)

The academic experts expressed their views that the results obtained from the algorithm analysis were both practical and useful, not only for hoteliers but also for those working in related literature. For those reasons, more research applying machine learning to the field of hospitality is encouraged. In sum, the qualitative portion of the study corroborates the results of the algorithm work and affirms its practical application to the industry. Below is one of the academic professor’s comments:

★ “This kind of algorithm technique is practical and useful both for industry and academics. I hope to see more research applying this approach to the hospitality industry…” (Acad. 01, Acad. 02)

5. Conclusions

This study aimed to detect relationships between complaint attributions and consequences for travellers who stayed at different star-rated hotels by analyzing their online complaining behavior. The study achieved this goal by applying machine learning techniques to TripAdvisor complaint reviews of hotels in the United Kingdom. Benefiting from a fourfold approach that applied content analysis, web graph analysis, Apriori algorithms, and a qualitative projective technique, this study was not only able to corroborate but also go beyond the conclusions reached in previous studies by revealing significant differences in online complaining behavior for different star-rated hotels.

5.1. Practical Implications

The visual representations produced through the analysis in this study offer an important contribution, as they allow industry managers to easily understand and utilize the data from the models and rule-mining for the purpose of improving their business strategies. For instance, based on the AR model, this study presented a web graph analysis that shows the strength of pair-wise relationships between different complaint attributes. Additionally, when more than two items are employed for rule generation, the ARs are able to depict even more insights and discover frequent item sets [21]. For these reasons, the Apriori algorithm was applied in this study in order to find the most frequent attributes of complaints and uncover hidden predictive information. After joining and pruning, the top 17 Apriori rules for different star-rated hotels were identified. These rules provide segment-specific information about the online complaint behavior of guests, and can be utilized by management to conduct hotel service quality policy analyses. By applying rule mining techniques, researchers and industry managers can develop a broad view of the relationships that exist between different parameters and simplify the processes of interpretation using models that are easier to understand [107].

Additionally, based on the results of the qualitative projective technique, we suggest the need for more in-depth investigations into applications of machine learning for the purpose of shedding light on online complaining behavior related to hospitality. More information is needed to understand the way in which algorithms can be used to predict customer complaining behavior. These new techniques will be able to provide new insight by overcoming some of the barriers that traditional methods are hindered by.

In sum, AR models are able to find patterns in the data where one or more entities (such as events, purchases, or attributes) are associated with one or more other entities [24]. The models construct rule sets that define these relationships, and the fields within the data can act as both inputs and targets. While traditional ARs allow for the manual discovery of patterns, AR algorithms (e.g., Apriori) harness technology to execute the work rapidly while allowing for greater insight into more complex patterns. ARs are also incredibly useful for predicting multiple outcomes—for example, if customers complain about attribute X, they will likely complain about Y and then Z as well. They can be used to associate a particular conclusion (such as the decision to complain about something) with a certain set of conditions, and they have an advantage over the more standard decision tree because of their ability to find connections between any of the attributes. The reason for this is that decision trees build rules with only a single conclusion, whereas ARs attempt to find many rules, each of which may have a different conclusion [24]. Therefore, if hoteliers don’t know which specific issue(s) to improve on, we recommend using ARs to help identify issues while also establishing their interconnected relationships. Depending on time, priorities, and company goals, hoteliers may want to place their focus on anywhere between 5 and 10 of the top rules to gain a better understanding of multiple issues at once. However, if they have very specific problems that need to be solved, decision trees are still recommended as they are able to go into greater depth and detail.

5.2. Theoretical Implications

This study applies attribution theory to examine the connections between online complaint attributions and their consequences. Attribution-consequence links are worth exploring because of the important role they play in determining guest dissatisfaction and electronic word-of-mouth behavior. The theory claims that an individual’s own perceptions of the success or failure of any activity will determine the amount of effort he or she will willingly commit to the same activity in the future [56]. The current study’s findings show that guests who had positive past experiences have high expectations for future stays. Consequently, failures to meet expectations, lead to negative outcomes and result in negative word-of-mouth behavior on online platforms. Using guest complaints from online platforms, the current study found links between the perceived causes of previous experiences (or attributions) and the affects these have on the behaviors of guests (consequences) who stayed at different star-rated hotels. These findings were consistent with Heider [108], who, by applying attribution theory, found that the way in which tourists attempt to comprehend their travelling experiences can influence their future travelling decisions or their avoidance behavior, and that word-of-mouth can also influence the behavior of others. Thus, it is evident that the link between guests’ perceived or understood attributions of previous experiences influences their future behavior. In response, this study calls for more attribution theory-based research that gives focus to guest complaints on online platforms (e.g., TripAdvisor, Agoda) in order to find practical solutions to issues within the hospitality industry.

5.3. Contributions and Future Research

The main contribution of this study is its innovative application of machine learning to detect correlations between antecedents and consequences in online complaining behavior related to the hospitality industry rather than relying on traditional methods used in past research (e.g., the critical incident technique or the multivariate approach). By analyzing real-world data (i.e., online complaints), the researchers were able to successfully conduct empirical and quantitative studies using ARs algorithms and the data they produced. Based on the results, we contend that there is a need for further, in-depth investigations into the details of online complaining behavior in hospitality contexts. The goals of these are not only to understand how algorithms may be useful in predicting customers’ complaining behavior but also to provide new insights that traditional methods have not been able to uncover due to their limitations. The machine learning approach has the potential to contribute greatly to the hospitality discipline and will certainly continue to gain popularity and importance in the industry. Another contribution of this study is that it extends previous research by applying attribution theory to online complaining behavior for the purpose of finding links between certain various parameters and predicting patterns for future activity. The results of this effort confirmed that specific online complaint attributes related to different star-rated hotels can be explained according to predictable attributions and consequences. Consequently, the findings of this study are practical and applicable to the industry.

5.4. Study’s Limitations

The study’s limitations are inevitable. First, this study collected data before the COVID-19 pandemic hit the hospitality industry. Further research might consider retrieving data during the post-crisis period and then conducting the comparison to see the differences in guest complaint behavior before and after the pandemic. Second, this study applied one of the most popular rule-based techniques (i.e., the Apriori algorithm); however, advanced algorithms (e.g., Generalized Sequential Pattern—GSP [Cf. Verma and Mehta [109]], An Efficient Algorithm for Mining Frequent Sequences—SPADE [Cf. Verma and Mehta [109]], or Pattern Growth-Based Approach [Cf. Cheng and Han [110], Hamdi, et al. [111]]) should be replicated in order to validate and advance the study’s outputs. In addition, given the contact-heavy nature of the hospitality service industry, COVID-19 has dramatically threatened employees’ safety and health. Prior research has shown that COVID-19 directly influences employees’ anxiety [112], work attitudes [113], turnover intention [114], psychological distress [115], work engagement [116], and cycling leisure activity [117]. However, with respect to the impacts of this pandemic, there are some gaps that need further research; for example, employees’ reactions toward working from home in different cultural contexts, employees’ commitment and trust, various psychological characteristics (e.g., ambition, risk aversion, family dedication), and family motivation for enhancing job performance during the crisis.