Leveraging Text Mining Techniques for Civil Aviation Service Improvement: Research on Key Topics and Association Rules of Passenger Complaints

Abstract

Airline customers will often complain to the relevant authorities if they encounter an unpleasant flight experience. The specific complaint information can directly reflect the various service problems encountered, so conducting in-depth research on public air transport passenger complaints can reveal important details for improving service. Therefore, by analyzing the passenger complaint data of relevant civil aviation departments in China, we propose a method for identifying key topics of passenger complaints based on text mining. We organically integrate sentiment analysis, topic modeling and association rule mining. A new complaint text analysis framework is constructed, which provides new perspectives and ideas for complaint text analysis and related application fields. First, we calculate the sentiment orientation of the complaint text based on the sentiment dictionary method and filter complaint texts with strong negative sentiment. Then, we compare the two topic modeling methods of LDA (Latent Dirichlet Allocation) and LSA (Latent Semantic Analysis). Finally, we select the better LDA method to extract the main topics hidden in the passenger complaint text with high negative emotional intensity. We use the Apriori algorithm to mine the association rules between the complaint topic words and the service problem classification labels on the complaint text. We use the FP-growth algorithm to mine the association rules between the complaint subject words and the service problem classification labels on the complaint text. By comparing the Apriori algorithm with the FP-growth algorithm, the results of mining the support, confidence and promotion of the association rules show that the Apriori algorithm is more efficient. Finally, we analyze the causes of specific service problems and suggest improvement strategies for airlines and airports.

1. Introduction

Complaints are voluntary acts of information sharing and feedback from customers, which can contain valuable information about service problems [1]. Complaint analysis (CA) can reveal urgent issues related to product attributes or service quality [2], and help businesses understand and respond to changes in consumer preferences, often by identifying needs to guide service improvement and prevent customer loss [3]. This is particularly important in the aviation industry, where airlines give importance to implementing complaint management and providing remedies for complaints [4]. The first step in complaint management is to understand the needs of passengers and the reasons for their complaints [5]. Based on this focus, we have posed the following questions to guide our research.

• What are the main topics of passenger complaints in public air transport?
• What are the causes or triggering mechanisms of service problems reflected in the topic of passenger complaints?
• How can we prevent or deal with identified service problems in public air transportation?

To address these research questions, we propose a unified framework of key topic identification methods for civil aviation passenger complaints based on text mining. The goal is to support effective operation of public air transportation, help civil aviation enterprises improve service weaknesses, and mitigate the negative impact of service complaints, thereby cultivating passengers’ confidence and loyalty, and developing their consumption potential. First, we use sentiment calculating methods and topic models to explore the most prominent topics in complaints. After collecting passenger complaint data, we calculate the sentiment orientation of complaint text using a sentiment dictionary, including sentiment polarity and sentiment intensity. It is then necessary to screen the negative complaint texts: the more negative the sentiment, the likely more prominent the service problems reflected. The LDA and LSA methods are used for topic modeling on complaint text with high negative sentiment value, and the topic consistency index is used to evaluate the interpretability and coherence of topics. This reduces the dimensionality of the complaint text based on topic characteristics. The results show that the LDA model is better than the LSA model on the whole. Second, we use association mining method to analyze the causes of service problems reflected in complaint topics, and match complaint texts after dimensionality reduction with the problem labels. Subsequently, we form an item set, mine the association rules between complaint topics and complaint issue labels using the Apriori algorithm and the FP-growth algorithm, obtain frequent item sets, and then sort them according to support, confidence, and lift. In this way, we obtain the Apriori algorithm to present more effective strong association rules and identify the corresponding service weaknesses and complaint reasons. Finally, the results of strong negative topics and strong association rules are comprehensively analyzed, and corresponding improvement suggestions and prevention strategies are proposed. This is performed to remediate public air transport service problems from the perspectives of different topics, such as airlines and airports.

Based on the above research questions, this paper proposes a text analysis framework for civil aviation passenger complaints that combines sentiment analysis, topic modeling and association rule mining. The follow-up structure of the article is as follows: Section 2 sorts out the relevant theoretical basis through a literature review; Section 3 describes the data and methods; Section 4 shows the experiments and the analysis of the results; Section 5 discusses conclusions and makes recommendations for improvement.

2. Literature Review

In view of the existing research, we will review the relevant research from four aspects: complaint application research, sentiment analysis, topic modeling and association rule mining.

2.1. Complaint Application Research

Customer complaints are an important way to measure customer satisfaction with the quality of products and services [6]. Paying less attention to customer complaints can lead to dissatisfaction, loss of loyalty, and customer churn. Complaint text usually encompasses a vast quantity of information. Most applications related to complaint information entail sentiment analysis, complaint text classification, complaint characterization, and complaint early warning, among other elements. For instance, Ref. [7] proposed a comprehensive method for emotion and statistical process control analysis. A review search of a game in the Apple App Store was conducted. Sentiment analysis was employed to determine customer satisfaction scores from the review data, while statistical process control analysis was utilized to detect major customer complaints at an early stage and avert service failures. By integrating these two types of analyses, it becomes possible to monitor customer complaints concerning mobile gaming services.

Ref. [8] conducted an analysis of online complaint behavior for six different hotels and showed that online complaint behavior is influenced by cultural context and varies between different levels of hotels. Ref. [9] applied data mining on complaints about the services of the North American transportation network to understand the problems behind the complaints, by constructing a classification model and performing correlation analysis. A meta-analytic investigation by Eshaghi et al. [10] systematically examined interconnections between critical determinants of passenger satisfaction and subsequent psychological responses within air travel contexts. Their findings reveal that traveler contentment significantly influences both immediate behavioral intentions and mediated pathways affecting airline recommendation likelihood, with gender emerging as a moderating variable in perceived value–satisfaction dynamics. Among service quality dimensions, cabin crew performance and in-flight amenities demonstrated paramount importance in shaping positive passenger evaluations. This research proposes an integrated framework elucidating how aviation enterprises can operationalize satisfaction metrics to cultivate customer loyalty, ultimately establishing actionable guidelines for sustainable competitive advantage in airline operations. A longitudinal investigation by Pereira et al. [11] delineated evolving patterns in air travel satisfaction through comparative analysis of pre-pandemic and pandemic-era service evaluation determinants. Their methodology leveraged 9745 user-generated evaluations from airlinequality.com, employing aviation-domain-specific sentiment classification systems to ensure contextual precision. Predictive modeling techniques employing gradient-boosted decision trees were subsequently applied to correlate sentiment polarity with carrier profiles, traveler demographics (including nationality and passenger segmentation), and temporal variables. The analysis uncovered a paradoxical pattern: baseline dissatisfaction levels observed pre-2020 intensified significantly post-pandemic onset, revealing systemic service quality vulnerabilities amplified by operational disruptions. This data-driven approach provides granular insights into crisis-era consumer behavior patterns, offering empirically grounded insights for service recovery strategies in aviation service design. The behavior of the staff is a major factor affecting passenger satisfaction. Therefore, there has been significant research on analysis and applications of complaint information, but usually, this research focuses on the causes of complaint occurrence, or analysis of complaint handling and prevention strategies [3]. Ref. [12] developed an automated text mining pipeline to derive operational insights from unstructured OCR-processed airline feedback. By deploying Latent Dirichlet Allocation (LDA) topic modeling, the study identifies latent service quality dimensions embedded in passenger narratives. A hybrid sentiment classifier combining lexicon-based polarity scoring and dependency parsing further dissects satisfaction drivers, while semantic pattern recognition techniques isolate causal phrases linked to passenger approval or grievances. Ref. [13] constructs a multidimensional benchmarking architecture through Large-Scale Text Analytics (LSTA), synthesizing 1.2 million reviews across 80 carriers into a hierarchical evaluation matrix. Their framework integrates semantic vector space modeling with eigenvector-based theme extraction, enabling cross-airline performance mapping through quantifiable sentiment indices and service excellence rankings. Both methodologies advance beyond conventional review analytics by transforming unstructured textual data into strategic decision-support tools for service gap diagnosis and competitive positioning. This study explores the mechanism of triggering passenger complaints through text mining technology, and finally proposes possible service improvement measures for airlines. Based on the research on the application of relevant complaints, the following is a review of the relevant literature on sentiment analysis.

2.2. Sentiment Analysis

Sentiment analysis attempts to measure feelings by extracting information from texts, videos, or emoticons [14]. Sentiment mainly consists of two aspects: sentiment polarity (positive or negative) and sentiment intensity (degree of positivity or negativity) [15]. Sentiment analysis allows us to measure a person’s inner evaluation of the topic presence of an object [16].

In existing research on airline customer sentiment, Ref. [17] put forward a technique for creating and choosing control variables for those possible psychological situational characteristics. By doing so, it effectively minimizes the risk of bias due to omitted variables. In a survey carried out at GRU Airport, which is the largest airport in Latin America and located in São Paulo, researchers formulated an econometric model concerning the determinants of passenger satisfaction. The aim was to investigate the influence of flight delays on the perception of service quality. The findings of this research validate the connection between flight delays and passengers’ overall satisfaction with airports. In addition, Ref. [18] devised a framework encompassing three operators: Assemble + Deft, Edify + Authenticate, and Forecast. This framework serves the purpose of categorizing opinion instances as either sarcastic or non-sarcastic. It was experimentally evaluated using a Twitter dataset and leveraging cutting-edge key technologies, namely a recurrent neural network (RNN) equipped with gated recursive units and support vector machines (SVMs). The dataset consists of opinions regarding the impact of the COVID-19 pandemic on air travel. The outcomes of this study will assist airlines in comprehending customer dissatisfaction and grievances, enabling them to make targeted decisions on enhancing their services. In the academic landscape, the research presented in Ref. [19] offers a significant enhancement to the existing body of sentiment analysis techniques. It broadens the application scope of data-driven and visual analytics methodologies, aiming to gain a more profound comprehension of customer satisfaction within the airline industry, especially during the COVID-19 pandemic. Moreover, Stefania [20] conducted an automated sentiment analysis on 639 comments. These comments, sourced from the website of the Italian National Consumer Union, pertained to the airline sector. The findings revealed that travelers’ anxieties predominantly centered around issues such as compensation, flight cancellations, and matters related to COVID-19. Additionally, their emotional responses were a complex blend, proving difficult to anticipate. Furthermore, the work described in Ref. [21] focused on a tailored dataset. This dataset comprised online reviews of four major airlines in Bangladesh. Based on this dataset, a multiclass sentiment analysis was carried out to explore the diverse sentiment expressions within the collected reviews.

Currently, the existing methods for sentiment analysis are mainly divided into two categories: one is the sentiment lexicon-based method, which uses a labeled sentiment lexicon. This lexicon is built from statistical and linguistic work by researchers on a rich corpus. The input text is matched to the lexicon to identify the sentiment of the words and calculate sentiment scores [22]. The other method is machine learning, which uses a large manually labeled corpus as a training set, and classifies sentiment by extracting text features and building models; typically, these models are support vector machines, decision trees, or plain Bayesian classifiers. Since most machine learning methods are supervised learning algorithms, manually labeling the text is necessary [23]. In contrast, sentiment lexicon-based methods are simple to operate, and users can also expand the sentiment lexicon themselves to improve accuracy [24]. This study uses the SnowNLP python library (https://pypi.org/project/snownlp (accessed on 25 February 2025)), whose main functions include lexical annotation, sentiment analysis, keyword extraction, and text summarization. SnowNLP is a Python-based natural language processing library that can process a wide variety of Chinese texts, including simplified Chinese and traditional Chinese. When processing traditional Chinese, SnowNLP can perform natural language processing tasks such as word segmentation, sentiment analysis, and part-of-speech annotation. SnowNLP is widely used for sentiment analysis of online comments, user sentiment recognition in Q&A communities, and text sentiment categorization in social software, and has achieved good performance [25]. Based on the research on the application of sentiment analysis, the following is a review of the relevant literature on topic modeling.

2.3. Topic Modeling

Topic modeling (TM) is a probability-based statistical model [26], used to find semantic topics that best describe the content of text documents, and the word information associated with each semantic topic [27]. Topic modeling technology is a powerful text mining method that has been applied in multiple fields, providing interpretable representations of document content. It can also be used to classify and reduce text dimensionality [28].

In the analysis of airline passenger sentiment, some studies have adopted a topic modeling approach. For example, Ref. [29] compared changes in air passengers’ perceptions of airline services before and during COVID-19. Two methods were used in the study, topic modeling and sentiment analysis. First, the Latent Dirichlet Assignment (LDA) model was used to extract impactful topics from online reviews that reflected passenger opinions. These topics are then assigned to the five dimensions of the SERVQUAL model to analyze the change in importance of each quality of service dimension. Based on this analysis, this paper puts forward some suggestions for improving the service quality of airlines according to the changes in air passengers’ cognition. Ref. [30] introduced a passenger satisfaction prediction model, which capitalized on a real-world dataset from Kaggle. Through the application of the Pearson correlation coefficient and the PCA-K-means clustering method, the study explored the interplay between individual in-flight services and passenger attributes. Based on these analyses, a new satisfaction prediction model was developed, utilizing the deep learning Wide & Deep algorithm. Moreover, the DeepLIFT algorithm was employed to interpret the deep learning model, and feature importance analysis was conducted to clarify the key features influencing passenger satisfaction. The results demonstrated that the proposed prediction model outperformed benchmark models like MLP, SVM, and Random Forest, achieving a higher level of accuracy. Ref. [31] harnessed online reviews to demonstrate the value of considering factors derived from topic modeling of unstructured data. This approach offered a flexible supplement to numerical scores, facilitating a better understanding of customer satisfaction and, in turn, service quality. Additionally, ref. [32] proposed a machine learning-based method for analyzing tweets with the goal of enhancing the customer experience. Ref. [33] uses topic modeling to identify passenger attitudes and evaluates the performance of three different topic modeling methods. This paper contributes to a deeper understanding of the multifaceted factors that affect passenger satisfaction after flight delays, and provides valuable insights and recommendations for improving the quality of service for airlines.

As a technology for extracting and discovering potential information or knowledge from a large amount of text data, text mining is of great significance for data analysis and decision-making in various fields. As two commonly used text mining models, LSA (Latent Semantic Analysis) and LDA (Latent Dirichlet Allocation) show their own characteristics and advantages in practical applications. LSA is mainly used in tasks such as text classification, information retrieval, and document clustering, and the model effect is improved through dimensionality reduction and semantic analysis. LDA, on the other hand, is widely used in topic modeling, sentiment analysis, and recommendation systems to help reveal the topic information behind the text. LSAs and LDAs have their own advantages and applicability in different fields. In principle, both LSA and LDA are probabilistic and statistical text mining methods, both of which attempt to mine hidden semantic information from text data. However, LSA focuses more on mining semantic information through dimensionality reduction, while LDA focuses more on inferring topic structure through probability distribution. In addition, in application scenarios, LSA is usually used for information retrieval and text similarity calculation, while LDA is more suitable for tasks such as topic modeling and sentiment analysis. Topic models can generally be classified into four categories: algebraic, fuzzy, probabilistic, and neural [34]. LDA is the most frequently used, and belongs to the category of Bayesian probabilistic models; it is simple, intuitive, scalable, and interpretable. LDA, as proposed by [35], is a classical unsupervised, probabilistic growth model for textual feature extraction [36]. It is widely used in various fields, particularly those involving in-depth text analysis, for example, online reviews [37,38,39], social media hot topic extraction and tracking [40], consumer complaint topic analysis [41], and multi-label categorization [42].

In general, the number of topics K needs to be determined before running LDA for topic modeling, and there are two main methods to select this number. The first is a statistical approach that relies on the calculation of specific metrics, such as topic coherence and perplexity [43]. The second method uses human expertise to select the number of topics. This expert method relies on human intuition to read the document and select the correct number of topics [44]. The quantity of topics has a significant impact on the semantic coherence and interpretability of the final outcome, making it a crucial factor that demands meticulous consideration. Furthermore, Rkia et al. [45] utilized natural language processing (NLP) techniques and various topic models, including Latent Dirichlet Allocation (LDA), Latent Semantic Analysis (LSA), and non-negative matrix factorization (NMF), to identify and analyze recurrent themes. Their research indicated that the LDA model with unary syntax achieved the most favorable results, attaining a coherence score of 0.59. In contrast, the NMF model demonstrated relatively poor performance. The experimental results were capable of detecting emerging issues and furnished valuable insights to support decision-making processes. In this paper, the method of LDA and LSA comparative experiments was used in topic modeling, and the topic consistency index was used to evaluate the model effect of LDA and LSA, and finally the appropriate model output optimal theme was selected.

2.4. Association Rule Mining

Association rule mining (ARM) falls under machine learning (ML), and its task is to discover correlations and interdependencies between many data items [46]. Association rules are derived from ARMs, which mainly represent the relationships between data; association rules also enable visualization of connected methods and reasons for data behavior [47]. The rule mining process consists of two phases: first, finding groups of high-frequency items from the original dataset based on minimum support, and second, seeking association and data relevance between high-frequency groups to generate association rules [48].

The Apriori algorithm and FP-growth algorithm are the two most common algorithms in association rule mining. The Apriori algorithm is a typical ARM method proposed by [49], which uses layer-by-layer iteration to obtain frequent items. It has shown great effectiveness on retail sales data, for example. The Apriori algorithm is commonly used for event analysis in the fields of transportation, healthcare, and others, aiming to analyze the causes of events that have occurred, or give early warnings of potential future events. It does this by mining the correlations between events and cases. For example, it has been used to mine causal features of ship traffic accidents [50], signal hazardous factors and unsafe behaviors of pilots [51], identify causes of gas explosions [52], and identify cancer-related genes [53]. Air passenger complaints can also be viewed as a set of events, and thus, association rule mining algorithms can be used to analyze their causal factors and facilitate management and prevention by airlines and airports.

In addition, FP-Growth, as an excellent association rule mining algorithm, can efficiently mine frequent item sets on large datasets by constructing compact data structures and efficient processing methods. The FP-Growth (Frequent Pattern Growth) algorithm is a frequent item set mining algorithm based on an FP-Tree structure, which can efficiently mine frequent item sets by constructing an FP-Tree, thus avoiding the process of generating candidate sets. Ref. [54] introduced Guided Frequent Pattern-Growth (GFP-Growth), a groundbreaking algorithm for multi-targeted mining. The GFP-Growth algorithm is engineered to efficiently mine a specified collection of item sets while consuming minimal memory resources. The study demonstrated that GFP-Growth accurately calculates the frequency counts for every item set under consideration. Moreover, it was shown that this algorithm significantly enhances the performance in addressing various problems that involve item set mining tasks. In summary, this study uses the Apriori algorithm and the FP-growth algorithm to mine the association rules between topic headings and complaint problem labels, respectively, and selects the best scheme through the experimental results.

At present, most of the studies on passenger sentiment analysis mainly focus on passenger satisfaction and travel experience, and there are few studies that start from passenger complaints and very few studies on civil aviation passenger complaints. In terms of research methods, the LDA model, K-means algorithm, and deep learning Wide & Deep algorithm are widely used in complaint application research, but few studies can combine multiple methods for text analysis and mining. Based on the existing mature methodology, this study will conduct an in-depth study on the problem of passenger complaints in public air transport, and creatively propose a method for identifying key topics of passenger complaints based on the sentiment dictionary method, LDA and LSA topic modeling, and the Apriori and FP-growth algorithms. The innovation of this research lies in the connection of the emotion dictionary method, LDA and LSA topic modeling, and the Apriori and FP-growth algorithms to more accurately mine the association rules between complaint topic words and complaint issue labels, and reveal the important details of service improvement.

2.5. Summary and Contribution

According to the summary of the relevant studies in Section 2.1, Section 2.2, Section 2.3 and Section 2.4 above, it can be found that in the field of complaint text analysis, traditional research methods often focus on single-dimensional analysis. Some traditional research methods simply conduct sentiment analysis without digging deep into topic association. Some directly model the topic of all texts without considering the influence of differences in emotional tendencies on topic mining. Some mine association rules for all texts without considering the guidance of their core themes. The method and process proposed in this study are significantly innovative in the research and application of complaint texts, which integrates sentiment analysis, topic modeling, and association rule methods for complaint texts, and uses the most classical methods to conduct experiments in each part, and comprehensively analyzes the results of strong negative topics and strong association rules, so as to solve the proposed research questions. It is the only study, to the best of the authors’ knowledge, that combines and applies all the listed techniques within the scope of a single project. From the perspective of different topics such as airlines and airports, the corresponding suggestions for dealing with and improving the problems and prevention strategies of public air transport services are proposed. Most of the existing studies focus on single-dimensional complaint analysis (such as emotional polarity or topic classification), and few attempts have been made to screen text based on emotional intensity. In addition, the complaint analysis in the field of civil aviation still lacks in-depth exploration of the rules of subject association. In this study, we innovatively integrate the sentiment dictionary method, LDA/LSA topic modeling, and the Apriori/FP-growth association rule algorithm to construct a multidimensional analysis framework to provide more accurate decision support for civil aviation service improvement.

3. Model Methodology and Data

3.1. Model Methodology

We propose a unified framework of key topic identification methods for civil aviation passenger complaints based on text mining technology. The goal is to analyze the complaints of air passengers, fully exploring their information and potential correlations. The proposed framework is shown in Figure 1.

The methodology of this study is divided into three stages: Firstly, in the sentiment analysis stage, we believe that complaint texts with strong negative emotions are more valuable in the field of complaint research, so we accurately locate the complaint texts with strong negative emotions and input them into the subsequent analysis to make them more targeted and in-depth, which can directly hit the core pain points of passenger complaints. Secondly, in the topic modeling process, compared with previous studies, we are concerned about whether the extracted topics are closely related to the key problem areas of passenger complaints, so by inputting concentrated and strong negative feedback into the topic modeling, the obtained topic words can better reflect the core topics that need to be solved and optimized, and we can avoid the situation that the topics are diluted or deviated from the core complaint concerns due to the mixing of a large number of neutral or positive emotional texts. Finally, the obtained topic words can be input into the association rule algorithm, and due to the accuracy and pertinence of the source of the topic headings, we can dig out more meaningful and closer association rules in the context of complaints with strong negative emotions. These association rules are no longer general associations based on generalized text data, but go deep into the internal connection between the core contradictions of passenger complaints.

Through the above unique method and process design, we organically combine sentiment analysis, topic modeling and association rule mining, which opens up a new research perspective and application approach for the field of complaint text analysis, provides accurate and deeply related information support for accurately locating the root cause of complaint problems and formulating effective complaint solutions, and carries out in-depth analysis and pragmatic suggestions according to the results of complaint topics and association rules in the context of real complaint data. The framework of this research is divided into three stages. The sentiment analysis stage screens the text of high-negative-emotion complaints through the sentiment dictionary method to ensure that the follow-up analysis focuses on the core pain points. In the topic modeling stage, the performance of LDA and LSA models was compared, and the optimal model was selected to extract the subject words of complaints. In the association rule mining stage, the association rules between subject headings and service problem tags are mined based on the Apriori and FP-growth algorithms.This framework avoids the interference of neutral or positive sentiment text on topic extraction through a layer-by-layer progressive analysis strategy, and at the same time enhances the actual explanatory power of association rules. This section details the dataset we used, and the specific methods involved in the experiments.

3.2. Data Collection

The data used in this study were recorded from May to June 2023, with a city in China as the starting point and destination, and consist of 5627 passenger complaints in total. Passenger complaint data were provided by the airline and include the complaint category, complaint and other fields, problem classification labels, complaint unit type, and complaint text content. Among them, passenger complaints are mainly categorized into ticket services, abnormal flight services, check-in and boarding, baggage services, in-air services, customer services, member services, information notifications, special passenger services, overselling, and other services. Each complaint has a corresponding complaint issue label. Among them, the complaint issue label refers to a way to classify, summarize and identify complaint issues, and use concise and clear labels to describe the key characteristics and essential content of complaints, so as to facilitate the management, analysis and processing of a large number of complaint information. Some examples from the dataset are shown in

Table 1. Examples of complaint data.

Complaint Category	Complaint Issue Label	Type of Complaint	Complaint Text Content (Translated from Chinese to English)
Baggage service	The amount and standard of compensation	Domestic airlines	The luggage was damaged during the transportation. The airline promised to contact me, but it has not taken the initiative to contact me and deal with it. I have been very dissatisfied.
Ticket service	Refund rules	Foreign airlines and airlines from Hong Kong, Macao and Taiwan	I applied for a change a month in advance, can’t change it, applied for a refund, and got no money back, isn’t this blatant cheating? I can’t agree, I must file a complaint.
Air service	Service attitude issues	Aviation sales network platform	I was on this flight and the cabin crew missed my in-flight meal. I called twice in a row before I came over to ask. I think the in-flight solution is perfunctory. I think the quality of service is poor.
Check-in and boarding	Problems with check-in procedures	Domestic airport	The person in charge of check-in for this flight left his post without authorization and did not answer his phone, which prevented all passengers from checking in for more than 40 min, delaying the flight queue and causing delayed boarding.
Abnormal flight service	Refund and change rules	Domestic airlines	On my friend’s flight, the arrival time was changed, the traveler did not recognize it, and the airline refused to refund the full amount of the flight change. As a consumer, do I have the right to ask for a refund for the change?

. The text preprocessing process will be detailed in Section 4.1.

3.3. Sentiment Calculation and Topic Modeling Methods

Based on the above datasets, we will perform the following sentiment analysis and topic modeling. In this study, we use the SnowNLP module in Python 3.7.3 to analyze the sentiment polarity and intensity of civil aviation passenger complaint texts. SnowNLP is a Chinese sentiment analysis library with a training set of positive and negative sentiments, mainly from an online UGC corpus, which uses the naive Bayes principle to train and predict the data [55]. SnowNLP is a Chinese emotion analysis database, which contains a set of positive and negative emotion training sets, mainly from the online user-generated content (UGC) corpus containing comments, articles and other content generated by Internet users and shared on the network. This corpus uses the naive Bayesian principle to train and predict data [56].

After calculating the sentiment score, complaint text with negative sentiment polarity and lower sentiment intensity than a certain threshold is filtered. Then, the filtered complaint text is entered into the topic model and used for topic extraction, exploring the topics reflected in text with higher negative intensity. In this way, the topic extraction results are more realistic.

This study uses LDA to find topics in complaint text. LDA first applies the Bag of Words (BOW) technique to reduce each complaint document to the form of a word vector distribution, and then creates a list of vocabulary terms based on word frequency, which assigns term weights using the Term Frequency–Inverse Document Frequency (TF-IDF) technique and optimizes the vocabulary [57]. Finally, given the expected number of topics K and top N keywords in the topics, the probabilities between the three layers of “topic-document-vocabulary” are computed. The topics are regarded as polynomial distributions over the vocabulary, and the documents are regarded as polynomial distributions over the topics. This allows us to infer the distribution of the hidden complaint topics based on the complaint documents, and to mine the implicit semantic information in the documents [58]. The LDA model is illustrated in Figure 2.

LDA involves the following steps [59]:

• Generate the topic distribution ($\theta$m) of document (m) by sampling from the prior Dirichlet distribution ($\alpha$).
• Generate the topic (Zm,n) of the (n)$th$ word of document (m) by sampling from the polynomial distribution of topics ($\theta$m).
• Generate the word distribution ($\Phi$$z_{m,n}$) of the topic (Zm,n) by sampling from the Dirichlet distribution ($\beta$) of the words.
• Sample from the polynomial distribution of the words ($\Phi z_{m,n}$) to finally generate the words (Wm,n).

3.4. Classical Algorithm for Association Rule Mining

Based on the above sentiment analysis and topic modeling results, the following experiments will be carried out on association rules. The Apriori algorithm and FP-growth algorithm are the two most common algorithms in association rule mining. Both are based on frequent item sets for data mining to find connections between item sets. The Apriori count method uses an iterative way to find the largest frequent item set, and then uses the obtained maximum frequent item set to achieve stronger association rule mining. The FP-growth algorithm uses the method of constructing frequent schema trees to perform database compression so that the entire algorithm only needs to scan the database twice, and no candidate sets are generated.

In this study, the Apriori and FP-growth algorithms were used to mine the association rules between topic headings and complaint problem labels, and the experimental results selected the best scheme. Taking the Apriori algorithm as an example, this method assumes that the set of transactions for association rule mining is T,$T=\left\{T1,T2,\dots ,Tm\right\}$, and the set of data items is $E=\left\{e1,e2,\dots ,em\right\}$. An association rule is generated if $X\Rightarrow Y$ under the condition $X\cap Y=\varnothing ,X\subset E,Y\subset E$ [51]. A flowchart of the Apriori algorithm is shown in Figure 3.

Figure 3 shows that the association rules are bounded by the minimum thresholds of support and confidence. Support denotes the frequency of the rule (i.e., the ratio between the number of transactions X and Y contained in the transaction set and the number of all transactions), denoted as:

$$Support(X\Rightarrow Y)=P(X\cup Y)={\displaystyle \frac{Count(X\cup Y)}{Count\left(N\right)}}$$

(1)

Confidence represents the strength of the rule (i.e., the ratio between the number of transactions containing X and Y and the number of transactions containing X in the transaction set), denoted as

$$Confidence(X\Rightarrow Y)=P(Y\mid X)={\displaystyle \frac{Count(X\cup Y)}{Count\left(X\right)}}$$

(2)

In addition to support and confidence, lift is a measure that describes the intrinsic value of a rule using correlation analysis, specifically the degree to which X is correlated with Y:

$$Lift(X\Rightarrow Y)={\displaystyle \frac{Support(X\cup Y)}{Support\left(X\right)\times Support\left(Y\right)}}$$

(3)

When the value of lift is 1, it indicates that X and Y are independent of each other; if the value is less than 1, it indicates that the two transactions are mutually exclusive. Generally, when the value of the lift is greater than 3, the mined association rule is valid and strong [60].

4. Experimental Results and Analysis

4.1. Data Preprocessing

In this section, we will conduct a complete experimental session, including data preprocessing, sentiment analysis based on sentiment dictionary, LDA and LSA topic modeling, and association rule mining using Apriori and FP-growth algorithms. Our data requires preprocessing before being put into the model. In this study, the following preprocessing operations are performed on the complaint text data:

• Remove punctuation: We remove useless punctuation, such as [ !"♯$\%{\&}^{\prime }\left(\right)*+,-./:\le >?@[\setminus ]\wedge {-}^{\prime }\{\mid \}\sim ].$)
• Remove numbers: Numbers mentioned in the text, such as phone numbers, ticket numbers, and mailbox numbers, are not needed; therefore, such elements are removed.
• Remove sparse words: Sometimes removing sparse words from the text data is necessary, which may include names of people, countries, cities, airports, etc.
• Remove stopwords: Stopwords are common intonational words used in language; we remove these words because they do not convey important semantic content.
• Cut words: The “jieba” python library based on Python third-party libraries is used for Chinese segmentation, and cuts the complaint text precisely.

Table 2. Examples of complaint texts before and after preprocessing.

Number	Preprocessing Complaint Text	Preprocessed Complaint Text
1	The passenger called and said the following: The passenger did not change successfully after paying the change fee, and the delay affected the passenger’s trip, so he complained.	Pay, change, fee, change, successful, delay, affect, trip
2	Passengers purchased tickets through the airline APP, paid for the meal beef stew, the beef was sour, the passengers checked the side dishes that did not have a sour taste in the ingredients, and complained to the flight attendant about mileage compensation, and the passengers were dissatisfied with the compensation and called to complain.	App, ticket, pay, buy, meal, beef, sour, view, ingredients, sour, side dish, flight attendant, give, miles, compensation, indemnity
3	Passengers called to say the following: Two passengers were planning to travel; due to a change in the itinerary, the friends buying tickets did not need to pay a handling fee to change, but the passengers needed to pay more than CNY 1000 to change. After the complaint platform’s customer service attitude was bad and did not give a good plan, the passengers were dissatisfied and called to complain.	Trip, itinerary, change, friend, purchase, ticket, don’t, pay, fee, change, payment, platform, customer service, bad attitude, give, program
4	The passenger called and said the following: There are four people in the group, two of whom are children, and they must be seated next to adults. The passengers were unable to check in through online channels and contacted the airline for feedback, and customer service could not assist with check-in, resulting in four passengers having to sit separately. Passengers were dissatisfied and called to complain.	Child, adult, seated, online, channel, check-in, procedure, contact, feedback, customer service, assist, handle, lead to, separate
5	The passenger called to say that after applying for a sick refund and submitting relevant information, the airline had not handled the review, and he was dissatisfied with this and called to complain.	Application, sick leave, submit, relevant, information, processing, review
…	……	……

shows an example of the complaint text before and after the preprocessing.

4.2. Sentiment Analysis of Text Based on a Sentiment Dictionary

Using a sentiment lexicon-based method, in this study, we extract texts with a high intensity of negative sentiment as the input to the topic model, making the results of the topic model more representative. The SnowNLP library primarily uses algorithms based on naive Bayesian classifiers in machine learning for sentiment scoring. The SnowNLP library was used to calculate the sentiment scores in 5627 complaint texts. The calculated sentiment scores represent the probability of semantic positivity between [0, 1], with 0.5 as the boundary to differentiate between positive and negative sentiments. The closer the sentiment is to 1, the more positive the sentiment performance, and the closer it is to 0, the more negative the sentiment performance; the sentiment of the answer is regarded as positive if the score is higher than 0.5, and otherwise is regarded as negative [61]. The overall distribution of the complaint text sentiment scores is shown in Figure 4.

The negative categories accounted for 99% of the total. Since the sentiment score is generally low, a threshold of ${10}^{-6}$ is set, resulting in 4097 complaint texts with high negative sentiment, as shown in Figure 5.

According to the results of the sentiment analysis, it is clear that the complaints, being a channel for travelers to express their dissatisfaction with the flight experience, have a generally negative sentiment polarity. Moreover, the sentiment intensity, or sentiment score, tends to be low.

4.3. Text Topic Modeling Based on the LDA and LSA Models

4.3.1. Experimental Comparison of LDA and LSA Models

Topic modeling is very important in the process of the method we propose, which directly affects the topic results of each complaint text on the one hand, and indirectly affects the rules generated by the association rule algorithm on the other hand. Therefore, in this study, a comparative experiment was set up, and the popular LSA and LDA methods were used to model the topic of complaint texts with sentiment scores below the threshold. After preprocessing the 4097 complaint texts, the BOW model was used to generate word vectors, and the TF-IDF method was used to assign weights to the words. Since LDA and LSA do not need to train the model in advance, and the implicit semantic information in the text can be mined through unsupervised learning methods, this study first determines the expected number of topics K and the top N keywords in the topic through statistical index method and expert experience.

In this study, we used the topic consistency index to evaluate the model effect of LDA and LSA, and selected the appropriate model to output the optimal theme, and the higher the agreement, the better the model [62]. Usually based on statistical analysis, the optimal number of topics is a number in the range of 50–200 [63]. Therefore, the number of topics in this study was set from 1 to 200, and the corresponding topic consistency of different models under different number of topics was calculated.

As shown in the figure below, Figure 6 shows the consistency scores of LDA and LSA under different number of topics, and

Table 3. The consistency score of LDA and LSA under the top 30 topic settings.

Number	LDA Score	LSA Score
1	0.4049	0.2985
2	0.4821	0.4200
3	0.4889	0.3781
4	0.5023	0.4267
5	0.5130	0.5100
6	0.5054	0.3521
7	0.4839	0.3786
8	0.4826	0.4391
9	0.4812	0.4527
10	0.4558	0.3973
11	0.4630	0.3514
12	0.4609	0.4018
13	0.4463	0.3922
14	0.4349	0.4183
15	0.4292	0.3613
16	0.4301	0.3491
17	0.4375	0.3612
18	0.4200	0.3760
19	0.4256	0.4077
20	0.4375	0.3572
21	0.4312	0.3524
22	0.4216	0.3720
23	0.4183	0.3649
24	0.4130	0.3518
25	0.4088	0.3426
26	0.4011	0.3807
27	0.3919	0.3826
28	0.4031	0.3439
29	0.4208	0.3588
30	0.4060	0.3392

lists the number of topics until the value of 30. According to the line chart, when the number of topics is 1–30, both LDA and LSA fluctuate greatly, and the fluctuation trend is about the same, and the LDA score in this range is slightly higher than that of LSA. When the number of topics is 30–78, LDA shows an obvious growth trend, but the fluctuation range is small in the growth process, LSA shows a downward trend, the fluctuation range decreases, the LDA score in this range is higher than that of LSA, and the gap between the two gradually increases. When the number of topics is 78–200, the LDA score reaches a peak when the number of topics is 78, and then decreases with an increase in the number of topics, and when the number of topics reaches 110, the score tends to stabilize after the number of topics reaches 78, and the LDA score in this range is much higher than the LSA score. On the whole, it can be seen that there are differences in the performance of LDA and LSA models in topic consistency, and the effect of LDA models is better than that of LSA models on the whole. Therefore, the number of topics with the highest consistency in the LDA model was 78, and the topic consistency score was 0.517, as shown by the red dot in Figure 6. In addition, based on business experience, the top N keywords in the theme are set to 3.

4.3.2. Discovered Topics

Figure 7 shows a visualization of the selected LDA model. We use the Gensim library (https://radimrehurek.com/gensim/ (accessed on 25 February 2025)) for LDA topic modeling. The left panel shows the distribution of each topic, each bubble represents a topic, and the size of the bubble indicates how often that topic appears in the document. The position of the bubbles indicates the similarity between the subjects, and the closer the distance indicates the more similar the topics. The vocabulary lists on the right panel show the terms corresponding to the selected topic, where the blue bar represents the overall term frequency in the dataset, and the red bar indicates the estimated term frequency in the selected topic. Only the three terms with the highest frequency were used to represent the selected topic in this experiment.

From Figure 7, it is clear that the distribution of the 78 topics is relatively even, with only a few topics overlapping. By counting the distribution of the 78 topics in the 4097 complaint texts, it is evident that the extracted topics of complaints have a wide range, which can comprehensively cover the complaint service categories.

In May and June 2023, the most prominent topic of passenger complaints about a certain city as the departure and destination was “submission, sick leave, materials”, which belongs to the category of ticketing services; this topic appears much more frequently than other topics. Passenger complaints under this topic were mostly refund disputes caused by passengers who were unable to take flights due to their own or their family members’ illnesses, and most of them mention that they could not take flights due to physical discomfort caused by COVID-19. Topics that appear frequently in the field of ticketing services include “free, charge, high”, “consumer, behavior, amounts”, and “name, change, returns”. It is clear that the many topics under ticketing services that passengers complained about are the most important types of services that need to be paid attention.

Table 4. Examples of selected topic terms.

Number	Topic Category	Main Topic Characterization Words
1	Ticket services	Submission, sick leave, materials
2	Abnormal flight service	Delays, take-offs, weather
3	Check-in and boarding	Staff, boarding, arrival
4	Baggage services	Baggage, compensation, check through
5	In-air service	Guests, flight attendants, meals
6	Customer service	Personnel, attitude, economy class
7	Member services	Membership, points, manager
8	Information notice	Display, aircraft type, adjustment
9	Special passenger service	Application, minutes, full refund
10	Overselling	Ticketing, mailing, overselling
11	Other services	Transit, department, accommodation

shows the frequently occurring topics under other complaint service categories. The topic categories in the table provide data directly to airlines.

4.4. Mining Association Rules Based on the Apriori and FP-Growth Algorithms

4.4.1. Association Rule Results

After obtaining the prominent topic words of passenger complaints through the LDA method, the association rules between the topic words and the complaint text service problem tags are further mined to analyze the possible causes of the complaint topics. This will allow us to put forth improvement suggestions for the corresponding service weaknesses.

In this study, we implemented the Apriori and FP-growth algorithms based on autonomous Python coding, and the minimum support and minimum confidence of two algorithms were 0.005 and 0.05, respectively. A total of 44 rules were generated, of which the 5 rules with the highest support, confidence, and lift are shown in

Table 5. Top five rules with the highest support of the Apriori algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Refund due to illness}	{Submission, sickness, material}	0.0447	0.7754	9.6270
2	{Delays, departures, weather}	{Dissatisfaction with compensation}	0.0217	0.5460	4.1503
3	{Price, difference, fare}	{Passenger ticket prices}	0.0203	0.7281	14.4102
4	{Fees, charges, high}	{Refund rules}	0.0178	0.4371	4.5803
5	{Free, booking, mistake}	{Wrong purchase}	0.0142	0.6374	8.1095

,

Table 6. Top five rules with the highest support of the FP-growth algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Submission, sick leave, material}	{Ticketing service}	0.0316	0.9576	0.0004
2	{Ticketing service}	{Submission, sickness, material}	0.0316	0.1433	0.0004
3	{Delays, take-offs, weather}	{Irregular flight service}	0.0156	0.9571	0.0009
4	{Irregular flight service}	{Delays, take-offs, weather}	0.0156	0.1545	0.0009
5	{Fees, charges, high amounts}	{Ticketing service}	0.0154	0.9222	0.0004

,

Table 7. Top five rules with the highest confidence of the Apriori algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Damaged baggage}	{Baggage, compensation, consignment}	0.0124	0.8095	23.8606
2	{Refund due to illness}	{Submission, sickness, materials}	0.0447	0.7754	9.6270
3	{Price, difference, fare}	{Passenger ticket price}	0.0203	0.7281	14.4102
4	{Compensation, arrangement, impact}	{Dissatisfaction with compensation}	0.0098	0.6452	4.9039
5	{Free, booking, mistake}	{Wrong purchase}	0.0142	0.6374	8.1095

,

Table 8. Top five rules with the highest confidence of the FP-growth algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Itinerary, there is a change, the second}	{Ticketing service}	0.0017	1.0000	0.0006
2	{Provide, antigen, swim}	{Ticketing service}	0.0015	1.0000	0.0006
3	{Ground transportation services}	{Staff, boarding, arrival}	0.0001	1.0000	0.0091
4	{Feedback, put, privately}	{Ticketing service}	0.0001	1.0000	0.0006
5	{Submission, sickness, material}	{Ticketing service}	0.0316	0.9576	0.0005

,

Table 9. The top five rules with the highest lift of the Apriori algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Display, model, adjustments}	{Ticketing session flight main service information}	0.0095	0.6094	26.0063
2	{Baggage damage}	{Baggage, compensation, consignment}	0.0124	0.8095	23.8606
3	{Missing a ride}	{Staffing, boarding, arrivals}	0.0073	0.5769	21.2942
4	{Seat, choice, self}	{Seat selection}	0.0056	0.5000	19.8883
5	{Price, difference, fare}	{Ticket prices}	0.0203	0.7281	14.4102

and

Table 10. The top five rules with the highest lift of the FP-growth algorithm.

Number	Lhs	Rhs	Support	Confidence	Lift
1	{Ticket purchase, mailing, overbooking}	{Overbooking}	0.0009	0.4737	0.0206
2	{Overbooking}	{Ticket purchase, mailing, overbooking}	0.0009	0.3913	0.0206
3	{Airport merchant services}	{Ride, on the day, pick up}	0.0002	0.2857	0.0159
4	{Ride, on the day, pick up}	{Airport merchant services}	0.0002	0.1111	0.0159
5	{Airport merchant services}	{Booking, knowing, channeling}	0.0001	0.1429	0.0110

. These represent the frequency of the rules, the strength of the rules, and the five rules with the highest correlation between the items before and after the rule, respectively. In the table, the left-hand side (Lhs) represents the rule’s front term and the right-hand side (Rhs) represents the rule’s back term. In the Apriori algorithm, the input data are the three topic words generated by the LDA model for each complaint and the corresponding category labels for each complaint. After calculation by the Apriori algorithm, the association rules between the topic words and category labels can be obtained. The expression of the association rules is “X→Y”, where X and Y are respectively referred to as the left-hand side (lhs) and the right-hand side (rhs) of the association rules.

According to the experimental results, we find that among the five rules with the highest support, confidence, and lift, the support and confidence of the association rules of the two algorithms are between [0, 1], and the improvement of the association rules of the FP-growth algorithm is much less than 1, while the improvement of the association rules of the Apriori algorithm is greater than 3, indicating that the association rules mined by the Apriori algorithm are effective and strong association rules. Therefore, the association rule results of the Apriori algorithm were selected as the final results of the experiment for subsequent analysis.

4.4.2. Analyzing the Results of the Association Rules

Based on the above experimental results, we know that the Apriori algorithm has more effective strong association rules, so the results of the association rules obtained according to the Apriori algorithm can help us analyze the main points of passenger complaints and the reasons for passenger complaints. Based on this, we will analyze the results of the Apriori algorithm in order to provide airlines with recommendations for improvement.

The rule “{Refund due to illness}⇒{Submission, sick-out, material}” had high support and confidence, and the main contradiction of this kind of complaint is that some passengers choose to voluntarily refund tickets because they did not know that there is a channel for a refund due to illness. This results in a large amount of handling fees: travelers fail to pass the authentication when submitting documents proving their illness, and their refund due to illness application is refused because they do not satisfy the requirements of the documents. Or after the review is passed, the corresponding handling fee will still be deducted. The requirements for materials stipulated by airlines are very strict. Therefore, before applying, airlines should provide more detailed information notifications, for instance in the ticket interface or the refund and change instructions. The channels of the sick refund and material requirements can be provided to the passenger after the successful purchase of the ticket, reminding them of the take-off time, refund and change instructions, and other important information. During the application process, airlines can provide travelers with templates for the corresponding materials, so that travelers can refer to them. After the passenger applies, the airline can speed up the review to prevent delays. At the same time, the airline can increase the input of digital review, so that after the passenger submits the materials, they can know whether the submitted materials are standardized or not.

The rule “{Damaged baggage}⇒{Baggage, compensation, consignment}” is at the top of both confidence and lift. This kind of dispute mostly happens when the passenger’s baggage is damaged during baggage checking, and the passenger is dissatisfied with the amount and standard of the compensation. The luggage may be damaged by force majeure during transportation and loading and unloading; the luggage might also be crushed when the plane is taking off, landing, or hitting bumps. Passengers should apply for compensation first, and the airline should inform the passengers of the amount of compensation, the standard, and the required time limit based on the degree of damage to the baggage. This can prevent ordinary incidents from escalating into a complaint. In the process of loading and unloading, the baggage may be damaged because the staff violently loads and unloads it, or because there is excessive friction on the conveyor belt. Therefore the training of staff should be strengthened to prohibit violent loading and unloading, and reduce the damage to the luggage.

The rule “{Price, difference, fare}⇒{Passenger ticket price}” is in the top five in terms of support, confidence, and lift. The frequency of such disputes is high because the price of air tickets is dynamically adjusted by the airlines according to supply and demand and other factors; passengers may therefore be dissatisfied with the monetary loss they perceive as a result of the fluctuation of airfares. Similar service issues include the strong rule of “{fees, charge, high} ⇒ {refund rules}”; this kind of dispute is mostly a problem of handling fee collection caused by passengers applying for a refund or change due to itinerary changes. Passengers may think that the handling fee is high, so they are dissatisfied and inquire about the refund rules. Both of these rules relate to fares and refunds.

The rule “{Free, booking, mistake}⇒{Wrong purchase}” also has a high level of support and confidence. Most of these disputes are about wrong purchases due to duplicate purchases, wrong document numbers, wrong names, wrong origin–destination, wrong flight time, etc., and travelers are dissatisfied with the lack of remedies provided. Airlines should provide free remedies to passengers who have purchased tickets by mistake, following the rules of the Chinese government. They should also strengthen the awareness training of staff on such rules, and increase the speed of identifying scenarios in which passengers have purchased tickets by mistake, to avoid delaying solutions.

There are also some rules on compensation standards, such as “{Compensation, arrangement, impact}⇒{Dissatisfaction with compensation}”, which has a high confidence level, and “{Delays, departures, weather}⇒{Dissatisfaction with compensation}”, which has a high support level. Such disputes are mostly due to weather conditions or machinery failures, resulting in delays in the take-off time of the aircraft. Passengers are often dissatisfied with the compensation standard and amount (sometimes being nothing). Airlines or airports should notify passengers in advance to protect passengers’ right to know, and should also provide water and meal services, humanized emotional comfort services, and if necessary, subsidies according to the delay time.

There are also some rules with a high lift, such as “{Display, model, adjustments}⇒{Tick-eting session flight main service information}”, which mostly involve complaints about the change of model, the inconsistency between the advertised model and the actual implementation, and failure to notify the passenger about the adjustment of the model among other things. With regards to “{Missing a ride}⇒{Staffing, boarding, arrivals}”, this rule is mostly due to passenger complaints about missed flights. Passengers often believe that they miss their flights due to no staff inquiries being made on-site, no radio broadcasts, no contact with passengers, and ferry or bus transportation time stops being too late, among other factors. The rule of “{Seat, choice, self}⇒{Seat selection}” is a complaint regarding check-in. To solve such complaints, airlines should strengthen their efforts to inform passengers in advance, and should also explain the reasons by text message, email, or phone call when there is a change in the flight plan. In addition, airports should provide multiple channels to ensure that passengers arrive at the boarding gate with time to spare. Therefore, these complaints are also consistent with the fairness theory mechanism, whereby people will compare their own input–output ratio with the input–output ratio of others to judge whether they have been treated fairly.

5. Conclusions and Suggestions for Improvement

We analyzed traveler complaints related to a certain city as the origin and destination in May and June 2023, and combined the results of sentiment values, topic results, and association rules between traveler complaints and problem labels. In the topic modeling process, the popular LSA and LDA methods were adopted, the model effect of LDA and LSA was evaluated by using the topic consistency index, and it was seen that there was a difference in the performance of the LDA and LSA models in topic consistency, and the effect of the LDA model was better than that of the LSA model as a whole. Therefore, the experimental results based on Section 4 show that the number of LDA models with the highest topic consistency score is 78 and the topic consistency score is 0.517, and the higher the consistency value, the better the LDA model. In addition, regarding the link of association rule mining, in association rule mining based on Apriori algorithm, the minimum support and minimum confidence scores are set to be 0.005 and 0.05, respectively. In the experiment, the five rules with the highest support, confidence and lift were selected, and on this basis, we find that the association rules mined by the Apriori algorithm are effective strong association rules, so the association rules of the Apriori algorithm are selected as the final result of the experiment. Here is a summary of our findings:

• We developed effective association rules with high frequencies and strong correlations.
• The topics with a high intensity of negative sentiments were mostly ticketing service problems or abnormal flight service problems.
• There were mainly refunding disputes due to sick refunds, price differences, handling fees, and wrong purchases.
• There were compensation disputes due to baggage damage, delays, missed flights, check-in issues, and other factors.

This study answers the three research questions raised in Section 1 through a three-stage analytical framework. The core topics of passenger complaints focus on ticketing services and abnormal flight services. A lack of transparency in information and employee service attitude problems are the main triggers in the trigger mechanism for service problems. It is recommended that airlines optimize information notification channels and strengthen staff training, while the government needs to strengthen closed-loop regulatory management.

The Chinese government department has attached great importance to passenger complaints, and has formulated numerous regulations to optimize service management in recent years. Informed by our research results on aviation complaints, we believe that the government should strengthen the supervision of major airlines and airline sales agents to ensure that they strictly implement the relevant requirements, and effectively protect the rights and interests of passengers. For airlines and airports, the closed-loop management of complaints should be perfected. Before the incident, contingency plans should be made; during the incident, service remedies should be provided; and after the incident, feedback should be taken and lessons should be learned. Based on the results of the experiment and the analysis of the results in the previous section, we can provide a valuable answer to question 3 in Section 1 The following are some suggestions for the closed-loop management of passenger complaints.

First, we saw that many complaints were caused by passengers not knowing the relevant regulations. This leads to a loss of money and great disappointment. Therefore, it is necessary to improve information notification services, broaden information channels, and protect the travelers’ right to know. This should include information on flight delays, cancellations, changes, refund rules, change rules, and other events. Second, many complaints are directly or indirectly related to the service attitude of staff. Therefore, better training of workers, such as airport service staff, flight service staff, or customer service staff, should be provided. This may include business knowledge training and service etiquette development; unified service standards and norms are imperative. Third, we recommend strengthening humanized management and construction. For example, airports can strengthen the construction of green channels and self-service check-in channels, and increase the presence and ability of intelligent question-and-answer robots. This could reduce the pressure of manual check-in procedures and manual inquiries, and reduce the occurrence of incorrectly booked flights and similar issues. Airlines should also improve the reward and punishment mechanism of abnormal flight management, develop new flight delay service measures, standardize ticket sales behavior, and optimize the environment in which passengers purchase tickets.

6. Limitations and Future Works

This study has important practical implications. First, through in-depth analysis of airline customer complaints and identification of key topics and areas of improvement, airlines and airports can optimize service processes and improve service quality in a targeted manner, thereby enhancing customer satisfaction and loyalty. Second, by mining the correlation rules in the complaint data, this method can help airlines and airports foresee potential service problems, take measures to prevent complaints in advance, and deal with complaints more effectively to reduce negative impacts. Third, through automated and intelligent text analysis methods, a large amount of complaint data can be quickly processed, processing efficiency can be improved, manual analysis costs can be reduced, and resources can be allocated more reasonably. In addition, this research has important theoretical implications. First, this study applied sentiment analysis, LDA and LSA thematic modeling methods. In addition, the two rule mining techniques of Apriori and FP-growth are linked to the analysis of aviation service complaints, which expands the application scope of text mining technology in the field of service management, and uses a variety of methods to compare and select the best in the topic modeling and association rule links, so as to increase the credibility of the article. Second, this study not only deepens the understanding of service quality management and customer feedback processing mechanisms, but also provides new perspectives and empirical support for related theoretical research.

However, there are some limitations to the research. First, the complaint data may only reflect the voices of some customers, especially those who choose to complain, while those dissatisfied customers who have endured silently or switched to other service channels are not included in the analysis, which may lead to some bias in the analysis results. Secondly, although the sentiment dictionary method is widely used in sentiment analysis, its accuracy depends on the perfection of the dictionary, and the recognition of complex or implicit emotional expressions may not be accurate enough. Finally, when extracting topics in LDA and LSA topic modeling, the granularity of the topic may need to be adjusted according to the actual application requirements, and too-coarse or too-fine topic division may affect the depth and practicability of the analysis. In future research, in addition to complaint texts, we can also consider integrating multi-channel data such as social media reviews, online surveys, and phone customer service records to obtain more comprehensive and multidimensional customer feedback. At the same time, we will further refine the topic classification, develop more accurate topic prediction models, and combine time series analysis and other methods to predict the trend and hot spots of service problems in advance. In addition, empirical studies were conducted in different airlines or airports to verify the universal applicability and effectiveness of the proposed method, and the model was continuously optimized based on practical application feedback.