Detecting digital voice of customer anomalies to improve product quality tracking

In today’s digital age, consumers increasingly share their experiences, opinions, and feedback on products and services across various online platforms (Chen et al., 2022). This user-generated content, collectively known as digital voice of the customer (VoC) or electronic word of mouth (eWOM), includes a wide range of reviews, comments and discussions on social media (Mastrogiacomo et al., 2021; Verma and Yadav, 2021).

Digital VoC provides organizations with valuable insights into customer perceptions, expectations and satisfaction (Colicev et al., 2019; Sykora et al., 2022). In this consideration, effectively leveraging Digital VoC has become a critical component of maintaining and improving product quality, particularly in the context of Quality 4.0, where advanced technologies such as big data analytics and artificial intelligence (AI) are driving the next generation of quality management systems (Escobar et al., 2021; Watson, 2019; Zonnenshain and Kenett, 2020).

Quality 4.0 has shifted traditional approaches to quality control and improvement towards more dynamic, real-time systems capable of self-monitoring and predictive maintenance (Agrawal et al., 2024; Antony et al., 2022). By leveraging data-driven methods, such as digital VoC analysis, organizations can better understand and respond to customer needs, thereby enhancing customer satisfaction (Shafiq, 2024). This paradigm shift reflects a broader movement in quality management from reactive problem-solving to proactive and continuous quality enhancement, facilitated by real-time customer feedback (Broday, 2022).

However, the vast amount of digital VoC presents significant challenges for organizations aiming to extract actionable insights. The volume, velocity, and variety of this data require new digital tools and novel analytical approaches, as traditional quality management tools are insufficient to handle such complexity (Mastrogiacomo et al., 2021). One widely used method to address this challenge is topic modeling, a technique designed to uncover latent topics discussed within large collections of unstructured textual data (Blei et al., 2003; Roberts et al., 2016). By applying topic modeling to Digital VoC data, organizations can identify recurring topics in customer feedback, which can be interpreted as key factors influencing customer satisfaction (i.e. quality determinants). Once the topics are identified, their evolution over time can be tracked, offering valuable insights into how customer feeling change. Detecting digital VoC anomalies, i.e. unexpected shifts or deviations in these views, becomes critical for identifying emerging quality issues or changes in customer satisfaction (Abrahams et al., 2015; Choi et al., 2020; Song et al., 2016).

While prior research has explored the potential of digital VoC in improving customer satisfaction (Barravecchia et al., 2022; 2023a, b; Özdağoğlu et al., 2018), there is a lack of focused studies on how anomalies can be systematically detected to inform quality management (Xu et al., 2022; Zaman et al., 2022). Current approaches for analyzing Digital VoC data have predominantly focused on static analysis of customer feedback. These methods do not account for the evolution of feedback over time. Consequently, they often overlook the anomalous patterns that may develop in consumer feelings. Addressing this gap, this study aims to introduce a dynamic approach to VoC analysis. In detail, this study targets the following two Research Questions (RQ):

The article is organized as follows. Section 2 provides a review of the literature with an overview of the connection between Quality 4.0 and Digital VoC. Section 3 outlines the methodology used in the study, including the exploration of dynamic trends in Digital VoC and the definition of categories of anomalies. In Section 4, the proposed taxonomy of Digital VoC anomalies for quality tracking is presented. Section 5 examines the ways in which these anomalies can be linked to quality improvement initiatives. The discussion in Section 6 compares the findings with other studies on the topic, highlights both practical and theoretical implications, and addresses the study’s limitations and future research directions. Finally, Section 7 presents the overall conclusions of the study.

The advent of Quality 4.0 represents a transformation in how organizations manage and enhance the quality of their products and services (Verma and Yadav, 2021; Antony et al., 2022). This novel quality paradigm allows organizations to move beyond traditional quality control methods toward more dynamic and data-driven systems. In this context, real-time monitoring and decision-making are essential for maintaining high-quality standards and quickly addressing emerging issues (Zonnenshain and Kenett, 2020).

A critical aspect of Quality 4.0 is the growing role of the Digital VoC (Barravecchia et al., 2022; Özdağoğlu et al., 2018). These unstructured data provide companies with real-time insights into customer experiences and expectations, offering a detailed understanding of product performance from the customer’s perspective (Pun and Chin, 2005). The ability to capture and analyze this feedback is critical for organizations seeking to align their offerings with evolving market demands, especially as customer needs become more dynamic and complex (Barravecchia et al., 2023a, b).

In the era of Quality 4.0, the integration of Digital VoC analysis into quality management practices allows for a more agile and responsive approach to identifying and addressing quality concerns (Sony et al., 2020). Organizations can leverage advanced text mining and analytics tools to process vast amounts of customer feedbacks, enabling the detection of emerging trends, patterns, and potential anomalies (Sony et al., 2020). This proactive approach, powered by digital tools, facilitates continuous monitoring and improvement.

Digital VoC data poses a challenge for organizations that need to extract actionable insights. To address this challenge, a variety of text mining and machine learning techniques have been proposed (see Table 1).

One of the most widely used approaches in Digital VoC analysis is sentiment analysis, which aims to classify customer feedback based on the emotional tone expressed in the text, whether positive, negative, or neutral (Kim, 2021). Sentiment analysis techniques have been applied across various contexts to monitor customer satisfaction, detect dissatisfaction, and anticipate potential quality issues (Jain et al., 2021). For example, Pang and Lee (2008) explored how sentiment classification could be used to evaluate online reviews, while Huang et al. (2023) provided an extensive review of sentiment analysis methods and their applications. By understanding the sentiment behind customer feedback, companies can identify areas where product performance meets or falls short of expectations (Liang and Wang, 2019).

A second significant category of technique for Digital VoC analysis is topic modeling (Mastrogiacomo et al., 2021; Özdağoğlu et al., 2018). Topic modeling techniques, such as Latent Dirichlet Allocation (LDA) (Blei et al., 2003), Latent Semantic Analysis (LSA) (Dumais, 2004), Non-negative Matrix Factorization (NMF) (Lee and Seung, 2000), or Structural Topic Model (STM) (Roberts et al., 2014), help uncover hidden topics within large text corpora by grouping words that frequently occur together into coherent themes. Topic modeling has been extensively used to analyze Digital VoC across multiple sectors, such as the hotel industry (Nguyen and Ho, 2023), food services (Park et al., 2020), shared mobility (Barravecchia et al., 2023a, b), and healthcare (Goto et al., 2022; Jia and Wu, 2021).

A further category of techniques frequently employed in Digital VoC analysis is clustering, which groups similar digital VoC records based on their textual features (Tabianan et al., 2022). K-means clustering (Nainggolan and Purba, 2020) and hierarchical clustering (Nguyen et al., 2020) are commonly used to segment digital VoC records into different categories, enabling organizations to identify clusters of complaints, suggestions, or praises.

More recent AI advancements introduced deep learning techniques, such as Bidirectional Encoder Representations from Transformers (BERT), which provide an understanding of language in context (Bilal and Almazroi, 2023). BERT models were successfully applied to analyze the sentiment and intent behind customer feedback, offering businesses deeper insights into the reasons behind customer satisfaction or dissatisfaction (Durairaj and Chinnalagu, 2021). This category of models, trained on large corpora of text, can capture the relationships between words and their surrounding context, making them highly effective in tasks like topic extraction (Mutinda et al., 2023).

Among the various text mining techniques presented in the previous section, Topic Modeling has emerged as a particularly effective method (Blei, 2012). The application of topic modeling algorithms yields two primary outputs (Roberts et al., 2016):

• (1)

  Topical content (⁠$TC$⁠) offering a structured representation of the topics within the text corpus. Each topic is characterized by a set of keywords that collectively define its focus. $TC$ quantitatively describes the weight or importance of each keyword within a topic.

• (2)

  Topic prevalence (⁠$TP$⁠) quantifies the distribution of topics across the textual documents. For each document, the algorithm assigns a multinomial distribution of probabilities, where each topic’s prevalence is indicated. This measure reflects the probability that a given document discusses a particular topic.

When applied to digital VoC related to a specific product or service, this approach enables the identification of latent quality determinants (Barravecchia et al., 2022; Mastrogiacomo et al., 2021; Özdağoğlu et al., 2018). As a general rule, if a topic is frequently discussed, it is likely to be of significant importance to the customer and thus crucial to perception of quality (Mastrogiacomo et al., 2021).

Topic Modeling has been selected for this study due to its robustness and proven efficacy, as evidenced by its widespread application (Mustak et al., 2021). However, the insights on anomalies drawn from Topic Modeling can be applicable to outcomes from other text mining methods.

Based on literature review on Digital VoC and time series analysis, this section outlines the methodology employed to address the first research question, namely the development of a taxonomy of anomalies specific to digital VoC analysis.

Previous studies introduced methodologies for analyzing the results of topic modeling applied to Digital VoC, with the aim of tracking the dynamic behavior of quality determinants for products and services over time (Barravecchia et al., 2023a, b).

The results of topic modeling are commonly expressed by the frequency with which topics are discussed by customers. For each topic, the Mean Topic Prevalence (⁠$MTP$⁠) quantifies the average prominence of a particular topic within the dataset and is computed as the mean of the Topic Prevalences (⁠$TP$⁠) across all VoC records, as shown in the formula (Mastrogiacomo et al., 2021):

Where, $N$ represents the total number of Digital VoC records analyzed, and $TPj,d$ is the topical prevalence corresponding to the $d$-th topic in the $j$-th Digital VoC record.

The cumulative $MTP$ s for all the $D$ identified topic is equal to one:

Table 2 provides some examples of the calculation of the $MTP$ indicator.

To move beyond a static analysis of customer feedback, the Interval Mean Topical Prevalence (⁠$IMTP$⁠) metric was introduced (Barravecchia et al., 2023a, b). Unlike static metrics, which provide a snapshot of topic prevalence at a single point in time, the $IMTPd,t$ captures how the prevalence of a given topic (the $d$-th topic) evolves over different time intervals (the $t$-th sampling period). Specifically, the $IMTP$ measures the average topical prevalence during a specified time period, allowing organizations to track and analyze change in the prevalence of topics across various timeframes. This approach facilitates the detection of dynamic trends in customer discussions. $IMTP$ can be calculated using the following formula (Barravecchia et al., 2023a, b):

In this equation, $Rt$ denotes the set of Digital VoC records from the $t$-th sampling period, with $|Rt|$ being its size.

For each sampling period, the sum of the $IMTP$ s for all quality determinants equals 1.

Table 3 provides some examples of the calculation of the $IMTP$ indicator.

While analyzing $IMTP$ provides valuable insights into the evolution of topical prevalence over time, it doesn’t offer information about the sentiment associated with these topics. As previously reported, various approaches are available for conducting sentiment analysis in textual documents (Kumar et al., 2016; Sun et al., 2019). However, sentiment analysis also has its limitations, primarily due to the subjective nature of interpreting textual emotions and the complexity of linguistic nuances (Yue et al., 2019). To enhance objectivity in assessing customer sentiment, the Mean Rating Proportion (⁠$MRP$⁠) indicator can be employed (Barravecchia et al., 2022). $MRP$ utilizes the objective rating value typically assigned by customers to their reviews, often on a scale of 1–5 (Barravecchia et al., 2022). The $MRP$ quantifies the prevalence of a topic within reviews of a specific rating, calculated using the formula:

where $t$ is the topic; $k$ is the level of the rating scale; $Rk$ is the subset of reviews associated to a rating level equal to $k$⁠; $TPi,t$ is the topical prevalence of the $t$-th topic in the $i$-th review; $|Rk|$ is the cardinality of $Rk$⁠.

Table 4 provides examples of the calculation of the $MRP$ indicator.

By analyzing the $MRP$ profile associated with each topic a link between product or service attributes (topic) and customer satisfaction or dissatisfaction can be established. Different attributes exhibit distinct $MRP$ profiles, which can be classified into positive, negative, or neutral categories based on their shape (see Figure 1).

Given the temporal nature of Digital VoC, particularly the $IMTP$ indicator, time series analysis can offer a robust framework for identifying and categorizing deviations in customer feedback. Time series data, which consist of observations recorded at regular intervals (Hamilton, 2020; Esling and Agon, 2012). These patterns may follow predictable trends, but can also be disrupted by random variations, commonly referred to as noise (Kirchgässner et al., 2013). Anomalies, also referred to as outliers, are significant deviations from the expected patterns and can signal important changes in the underlying system (Hawkins, 1984). Anomalies in $IMTP$ trends are critical for understanding shifts in customer sentiment or behavior, where deviations in feedback may highlight emerging issues or new trends that require attention.

In time series analysis, anomalies typically fall into two main categories (Box-Steffensmeier et al., 2014):

• (1)

  Noise or erroneous data: data points that result from errors in collection or processing, and do not reflect meaningful trends.

• (2)

  Indicators of unusual phenomena: data points that signal significant deviations due to changes in the underlying system, such as customer dissatisfaction or product failures.

Detecting these anomalies requires sophisticated analysis techniques to differentiate between random variations and meaningful deviations. Drawing on literature from time series analysis (Aggarwal, 2017), anomalies can be classified into four key categories:

• (1)

  Point anomaly: An isolated data point or short sequence that deviates sharply from the expected range.

• (2)

  Contextual anomaly: A data point that is anomalous in its specific context, even if it falls within the general range of expected values.

• (3)

  Collective anomaly: A group of data points that together exhibit a pattern deviating from the norm, requiring longitudinal analysis.

• (4)

  Other anomaly types: Further specialized categories that capture more granular variations, based on specific patterns within the dataset.

Building upon the foundational concepts of anomaly detection in time series analysis, this study introduces a novel taxonomy specifically designed for Digital VoC data. The established categories of anomalies are adapted to accommodate the characteristics of Digital VoC, where customer feedback evolves over time and is influenced by various external factors, such as product updates, market trends, and changes in customer expectations. By interpreting the $IMTP$ as a time series, where the prevalence of specific topics change over time, it becomes possible to detect meaningful deviations that signal shifts in customer sentiment or emerging quality issues.

The authors recognizes that Digital VoC, due to its nature, requires specific interpretation of anomalies compared to traditional time series data. While time series anomalies typically involve deviations in purely quantitative measurements (e.g. sales, temperatures, or stock prices), Digital VoC anomalies involve deviations in collective customer opinions. As a result, the proposed taxonomy of anomalies was developed to classify these shifts by focusing on how the discussion of topics evolve over time. The integration of time series concepts into the analysis of customer feedback data forms the foundation for this new anomaly taxonomy, specifically tailored to the dynamic of Digital VoC.

The proposed taxonomy delineates four types of anomalies, each with distinct implications for quality tracking. The taxonomy is structured as follows:

• (1)

  Spike anomalies: Sudden deviations in $IMTP$ values, either as upward or downward spikes. These anomalies represent sharp changes in customer discussion around specific topics. For instance, an upward spike could indicate increased dissatisfaction with a particular product feature, while a downward spike may suggest a resolution of previous concerns.

• (2)

  Level anomalies: Sustained changes in the baseline level of $IMTP$ over time. These anomalies reflect long-term shifts in customer focus, either increasing or decreasing attention on a specific topic. Level anomalies suggest persistent changes in customer behavior that may warrant strategic adjustments in quality management.

• (3)

  Trend anomalies: Gradual changes in $IMTP$ that occur over extended periods. Trend anomalies indicate evolving customer preferences or emerging issues that could gradually impact product quality or satisfaction. Identifying these trends enables companies to proactively address potential concerns.

• (4)

  Seasonal anomalies: Deviations from expected periodic patterns in $IMTP$⁠. Seasonal anomalies occur when the normal periodicity of a topic’s discussion is disrupted. For example, an unexpected surge in customer complaints during an off-peak season might signal a product defect that is sensitive to environmental factors.

This section provides a detailed explanation of the characteristics of the four types of Digital VoC Anomalies: Spike Anomaly, Level Anomaly, Trend Anomaly and Seasonal Anomaly.

Spike Anomalies are observed as acute deviations from the expected trend within the $IMTP$ data (see Figure 2). Characteristically, these anomalies present themselves as single data points or short sequences of points that exhibit a significant divergence from a previously trend.

Unlike gradual shifts or recurring patterns, spike anomalies are marked by their abrupt nature. They are detectable where the $IMTP$ for a given topic suddenly spikes (upward spike) or dips (downward spike), standing out against the trend line that has been established by preceding data points. These spikes may appear as isolated incidents in a stable trend, where they are easily discernible due to their contrast with the baseline level. However, their detection might be more nuanced within upward or downward trending data, where they represent an unexpected acceleration or deceleration in topic prevalence.

Level Anomalies within Digital VoC are identified when there’s a significant and sustained change in the baseline level of the IMTP. This alteration manifests as a distinct jump or drop in the data series and subsequent maintenance of a new level.

These anomalies present as clear discontinuities in the time series graph, where the IMTP, after experiencing a sharp change, stabilizes around a new value, indicating a reestablishment of the topic’s prevalence.

Two primary types of level anomalies can be identified:

• (1)

  Upward level anomaly: This occurs when the IMTP experiences a sudden increase from one stable baseline to a higher one, without returning to the original level.

• (2)

  Downward level anomaly: Conversely, a downward level anomaly is observed when there’s a sharp decrease in the IMTP, with the new lower level persisting over time.

For instance, as illustrated in Figure 3, an upward level anomaly is observed when the $IMTP$ shifts from a stable trend with values around 0.12 to a new stable trend with values around 0.3.

Trend Anomalies are characterized by a gradual yet enduring shift in the $IMTP$ from one predictable pattern to another, which may be stable, increasing, or decreasing. These are not abrupt spikes but rather a series of data points that collectively signify a change in the underlying trend.

The difference from spike anomalies is the persistence and direction of this change. While spike anomalies are transient and may return to the previous level, trend anomalies suggest a more fundamental shift in the discourse, marking a new phase in the discussion around a particular topic. These anomalies imply that the change in topic prevalence is not a fleeting occurrence but rather a new trend that is likely to continue over time.

Figure 4 details the various characteristics that can be observed in trend anomalies, outlining the nature of the shifts and the manner in which they diverge from prior trends.

Seasonal Anomalies in Digital VoC analysis pertain to recurrent, periodic fluctuations in the $IMTP$ that diverge from established seasonal patterns. These anomalies are characterized by variations in the $IMTP$ that do not align with the expected periodicity based on known seasonal trends.

Typically, seasonal patterns exhibit predictable and regular movements that correspond to specific time frames, such as months or seasons. Seasonal anomalies occur when these patterns are disturbed. The most common forms of seasonal anomalies are as follows:

• (1)

  Intensity variations: Variations in the intensity of seasonal discussion peaks provide insight into the changing levels of consumer interest or market dynamics (see an example in Figure 5a).

• (2)

  Timing discrepancies: These occur when customer discussions about seasonally popular topics arise at unexpected times (see an example in Figure 5b).

• (3)

  Emerging or fading seasonal trends: The appearance of new seasonal trends or the decline of established ones can alert businesses to changes in consumer preferences or societal shifts (see an example in Figure 5c).

To address the second research question concerning the identification of effective procedures for quality improvement based on Digital VoC anomalies, the following methodological steps were considered:

• (1)

  Quality tracking methodologies and conceptual alignment: the first step involved a review of literature on quality tracking methodologies (Barlow and Møller, 1996; Nasr et al., 2018; Tang, 2017; Yang et al., 2019) and on Quality 4.0 general guidelines (Carvalho and Lima, 2022; Goecks et al., 2020; Ranjith Kumar et al., 2022).

• (2)

  Case study analysis: Following the theoretical framing, real-world case studies were analyzed to identify patterns in Digital VoC anomalies and their corresponding impact on product quality.

• (3)

  Practical procedures: The case study findings and the theoretical insights were synthesized to develop a set of structured procedures for addressing each type of anomaly.

The following subsections detail the interpretation of digital VoC anomalies. Additionally, they present four case studies, each of which illustrates the practical aspects of anomaly identification, Case studies data were extracted from review aggregator platforms and e-commerce sites. The Structured Topic Model (STM) algorithm was employed to detect the underlying topics discussed by customers (Roberts et al., 2014, 2019). This methodological approach aligns with the framework detailed by Mastrogiacomo et al. (2021). For the sake of brevity, each case study concentrates on a single anomalously behaving topic.

The interpretation of spike anomalies in Digital VoC demands a discerning analysis, particularly when differentiating between sharp increases and decreases in topic discussion.

When there is an upward spike anomaly, i.e. an abrupt rise in the discussion of a certain topic, it typically reflects an increase in customer interest or the emergence of concerns. The root cause of these spikes might be diverse, ranging from the surfacing of potential issues to the launch of a new significant updates. To differentiate between positive and negative drivers behind these spikes, an analysis of the MRP profile corresponding to the topic during the period in question can be conducted:

• (1)

  Upward spike anomaly with positive MRP profile: could indicate robust customer engagement or approval.

• (2)

  Upward spike anomaly with negative MRP profile, this could signal problems or growing customer dissatisfaction, requiring swift attention and possibly urgent remedial actions.

Interpreting a downward spike anomaly could be more difficult. Since IMTP values of various topics are interrelated—where the sum of all IMTP values is constrained to equal one—an increase in one topic’s IMTP naturally results in a compensatory decrease in others. If the discussion on other topics does not display anomalous behavior, a sharp decline in the discussion of a specific topic could suggest waning customer interest or the resolution of a previously prevalent issue. For instance, a rapid decrease in conversations about a product problem could indicate that the issue has been effectively resolved, leading to a corresponding reduction in customer complaints.

Table 5 outlines procedures to implement for quality improvement based on the interpretation of spike anomalies in Digital VoC.

A case study is presented to illustrate the process of identification and exploitation of a spike anomaly. This case study focuses on the analysis of Digital VoC data related to a smartwatch, spanning a 100-weeks period. Among the topics identified through the application of a topic modeling algorithm, the topic of “battery life” was notable, concerning battery performance and longevity (keywords: battery, charge, duration, last, life, speed, capacity, power, day).

The trend of the $IMTP$ over the 100-weeks period is depicted in Figure 6a. The line represents the weekly $IMTP$ values, with a significant spike anomaly highlighted around week 60. This anomaly, marked by a sudden and significant increase in $IMTP$ values within a short timeframe, indicates an abrupt shift in customer interest or concern. This shift necessitates further investigation to understand the root causes and to formulate an appropriate response.

A detailed review of the $MRP$ profiles during this period unveiled a dominance of negative ratings associated with discussions about the smartwatch’s battery life, pointing to widespread customer dissatisfaction with battery performance (refer to Figure 6b).

Following the identification of the upward spike anomaly in discussions on battery life, marked by a negative MRP profile indicating customer dissatisfaction, the company implemented a structured approach for quality improvement. Initially, the team prioritized a rapid investigation into the feedback, pinpointing the firmware update as the primary cause of the decreased battery performance. Table 6 lists some reviews with a high Topic Prevalence for “battery life”.

To explore deeper and ensure a comprehensive understanding of customer concerns, a direct engagement strategy was employed. This involved reaching out to affected customers through surveys and social media platforms, providing them with a channel to voice their specific experiences and issues encountered post-update.

With a clear understanding of the problem, the technical team worked on developing a corrective firmware update aimed at resolving the battery life reduction. This solution was tested to ensure it addressed the identified issues without introducing new ones.

By effectively implementing the outlined procedure, trust was restored among the smartwatch users, and a positive shift in customer satisfaction was observed in subsequent VoC analyses.

Interpreting Level Anomalies within Digital VoC, offers significant insights into lasting shifts in consumer discussions. Distinct from spike anomalies that highlight temporary changes, level anomalies reveal a deep and enduring transformation in customer engagement. This lasting change is critical for businesses to comprehend and act upon for strategic adjustments.

An Upward Level Anomaly, characterized by a positive MRP profile, typically signifies increased customer satisfaction, evolving needs, or shifting preferences. This might result from successful product improvements, efficient marketing strategies, or new features meeting emerging customer demands. Here, the opportunity lies in enhancing these positive perceptions, further refining products or services to leverage the identified customer preferences.

In contrast, an Upward Level Anomaly with a negative MRP profile may signal the emergence of a consistent problem or an unaddressed need, culminating in customer dissatisfaction. This condition necessitates a detailed investigation to uncover the dissatisfaction’s root causes, potentially leading to product or service redesign or the introduction of new solutions to bridge the identified gaps.

A Downward Level Anomaly might imply different outcomes based on the context. If the decrease is accompanied by an increase in discussions on other topics, it may suggest a shift in customer focus rather than a problem with the topic experiencing the decrease. However, if the decrease stands alone without an uptick in other areas, it could indicate that a previous issue has been resolved or a need has diminished among the customer base. In such instances, it’s imperative for companies to reevaluate their offerings and communication strategies, ensuring they align with the latest customer needs and expectations.

Table 7 outlines procedures for quality improvement for each variant of a level anomaly, aiming to operationalize actions for leveraging these insights.

To illustrate the process of detecting and leveraging a level anomaly, a case study is presented. The analysis of digital VoC data from a major airline over a period of 100 days is the focus of this case study. The topic of “in-flight Wi-Fi service” was one of the topics identified through the application of a topic modeling algorithm (keywords: Wi-Fi, connection, speed, stable, login, flight, e-mail, streaming, password). Historically, customer discussions about this topic were consistently moderate, but a sustained increase in the discussion level suggested a shift in customer experiences and expectations.

Figure 7a shows the $IMTP$ trend for in-flight Wi-Fi service discussions over 100 days. The line represents the daily $IMTP$ values, which exhibit a noticeable and sustained step-like increase in the level of discussion around day 38. This change in level remains consistent throughout the remainder of the observed period.

A comparison of the $MRP$ profiles for the in-flight Wi-Fi service topic across periods before and after the anomaly indicated a shift in sentiment. Initially, the $MRP$ profile was neutral during the first 38 days of analysis. However, it transitioned to a positive profile in the subsequent period. This shift suggests increasing customer satisfaction with the Wi-Fi service’s reliability and speed (refer to Figure 7b).

Upon identifying an upward level anomaly in discussions related to the in-flight Wi-Fi service, accompanied by a transition to a positive MRP profile, the airline recognized this as an opportunity to further enhance and capitalize on the positive customer sentiment. This shift indicated not just a temporary spike in interest or satisfaction, but a lasting change in customer perceptions and expectations regarding the quality of Wi-Fi services offered during flights.

To leverage this positive shift, the airline embarked on a multi-faceted strategy aimed at reinforcing the improvements that had led to increased customer satisfaction. Initially, the airline conducted an in-depth analysis to pinpoint the specific enhancements or changes that had contributed most significantly to the positive customer feedback. This involved a detailed examination of customer reviews and feedback.

Table 8 lists a sample of selected reviews with a high Topic Prevalence for in-flight Wi-Fi service. Reviews extracted from the latter period reflect positive sentiments and appreciation for the onboard Wi-Fi service, highlighting improvements and increased satisfaction among passengers.

Armed with these insights, the airline took steps to further improve the Wi-Fi service, focusing on areas such as connection stability, speed, and ease of access, which were frequently mentioned in positive reviews. Investments were made in upgrading technology and infrastructure to ensure these improvements were not only maintained but also enhanced over time.

In parallel, the marketing team developed campaigns to highlight the upgraded in-flight Wi-Fi service, showcasing real customer testimonials and data demonstrating the service’s reliability and speed.

Interpreting Trend Anomalies can offer insights into customer interests, preferences, and behaviors. These anomalies, characterized by a gradual and sustained transition in the IMTP from one established pattern to another, can serve as an indicator for changing customer dynamics.

Exploring the six types of trend anomalies within the context of Digital VoC analysis provides a granular understanding of how customer discussions can evolve over time. These anomalies can be interpreted as follows:

• (1)

  Plateau (From Ascending Trend to Stable Trend): this anomaly suggests that the initial surge in interest or discussion intensity has leveled off, possibly due to the topic reaching market saturation or customers having their immediate needs met. For companies, recognizing a plateau provides an opportunity to innovate or introduce new features to stimulate interest and discussion.

• (2)

  Inversion–Mountain (From Ascending Trend to Descending Trend): this turnaround could be due to emerging issues with a product or service, a shift in market interest, or the natural lifecycle of a topic’s relevance. Identifying an inversion is critical for businesses to quickly address any underlying problems or to manage on customer interest.

• (3)

  Escalation (From Stable Trend to Ascending Trend): this shift could be driven by new product features, effective marketing strategies, or external factors boosting interest. For companies, escalations signal areas of growing customer engagement that could be leveraged for further growth.

• (4)

  Decline (From Stable Trend to Descending Trend): it may highlight customer dissatisfaction, the emergence of competitive offerings, or a decrease in the topic’s relevance. Companies need to investigate declines to address potential issues and adapt their strategies to customer expectations.

• (5)

  Inversion–Valley (From Descending Trend to Ascending Trend): this reversal could indicate successful resolution of previous issues, effective marketing recovery efforts, or a resurgence in interest due to external influences.

• (6)

  Stabilization (From Descending Trend to Stable Trend): this change of pattern may result from corrective actions taken by the company, changes in customer perceptions, or the topic reaching a natural point of lesser but consistent discussion.

Table 9 outlines procedures to implement for quality improvement based on the interpretation of trend anomalies in Digital VoC.

A case study is presented to provide a practical example of a trend anomaly and subsequent quality improvement. This case study examines digital VoC data concerning a mobile phone across a 24-week span. Among the topics identified through topic modeling algorithms, “Camera Performance” emerged as a significant theme, described by the keywords: camera, photo, picture, light, image, social, save, video, and pro.

Analysis of the indicator $IMTP$ revealed a distinct trend anomaly (see Figure 8a). Initially, there was a negative trend, which resulted in a decline in discussions about the camera performance. However, a notable inversion anomaly was observed around week 14. This shift resulted in a marked increase in discussions regarding the camera in subsequent weeks.

Comparing the $MRP$ profiles before and after the anomaly indicated a clear transition from a neutral to a positive profile, suggesting increased customer appreciation for the smartphone’s camera features (see Figure 8b).

A deeper examination of the review content (see an extract in Table 10) highlighted enhanced customer satisfaction with the camera’s performance. Further investigation into the smartphone’s operating system updates pinpointed a camera software update as the catalyst for the improved performance. This update enhanced the camera’s functionalities, influencing customer perceptions and discussions around the topic.

In response to the identification of the inversion (valley) anomaly related to “Camera Performance,” the company moved to capitalize on this newfound positive sentiment. Recognizing the impact of the camera software update on this shift, the first step involved a thorough analysis to understand how the update improved user experiences and perceptions regarding the camera functionalities.

With the positive influence of the update confirmed, the company then focused on widespread communication of these improvements. This involved crafting messages highlighting the enhancements made to the camera, how these changes contribute to a superior photography experience, and instructions for users to update their devices if they hadn’t already done so.

Additionally, the marketing and product teams worked to leverage this positive sentiment in promoting the smartphone.

Interpreting and leveraging seasonal anomalies in Digital VoC analytics involve understanding how these deviations from expected seasonal patterns can offer actionable insights for quality and product development. Here’s how businesses can interpret and exploit the different types of seasonal anomalies:

• (1)

  Increased intensity: An upsurge in seasonal discussion intensity can highlight a range of factors influencing consumer engagement. This could be attributed to more effective marketing strategies capturing public interest, enhancements in product features meeting seasonal demands more aptly, or external influences heightening product relevance. Increased intensity might also signal problems that become more pronounced in specific conditions, such as seasonal defects in products. The MRP analysis sheds light on the underlying sentiment associated with the anomaly:

  • •

    Positive MRP profile: This indicates enhanced consumer appreciation or heightened interest. Such a profile suggests that the increase in discussion intensity is tied to positive aspects of the product or service, such as successful feature implementations or satisfying consumer needs effectively during the season.

  • •

    Negative MRP profile: This signals potential issues that become more pronounced or noticeable during the relevant season. A negative profile denotes consumer dissatisfaction or critical problems with the product or service, necessitating attention to mitigate negative impacts and address consumer concerns.

• (2)

  Decreased intensity: a decline in discussion intensity might suggest a drop in consumer interest, potentially due to a saturated market, heightened competition, or a disconnect between the product offerings and consumer expectations. The MRP analysis provides critical insights into the change in consumer sentiment:

  • •

    Positive MRP profile: might reveal a diminished appreciation for previously favorable aspects of the product or service. It indicates that while discussions have lessened in intensity, the remaining conversations still lean towards a positive sentiment, perhaps reflecting a shift in consumer priorities or emerging competition.

  • •

    Negative MRP profile: the reduction in discussion intensity alongside a negative MRP profile suggests a decrease in previously significant issues. It could indicate successful resolution of past problems or a natural decline in their relevance, leading to reduced consumer complaints or negative discussions.

• (3)

  Timing discrepancies–Earlier than expected: seasonal discussions that start sooner than usual may reveal shifts in consumer planning behaviors or the influence of external factors that prompt an earlier start to the season. This discrepancy might suggest an anticipatory interest in seasonal products or a market response to environmental changes.

• (4)

  Timing discrepancies–Later than expected: when discussions emerge later in the season, it may indicate delayed consumer interest or demand. Reasons could include prolonged effects from the previous season, a crowded market landscape, or other events capturing consumer attention, thereby delaying the typical seasonal engagement.

• (5)

  Emerging trends: the emergence of new seasonal discussion for a topic point to evolving consumer interests or untapped market opportunities. These discussions can uncover shifts in consumer preferences, societal trends, or innovations driving new consumer behaviors. Identifying these early can provide a clear indication of where consumer interests are heading and highlight new areas for potential engagement.

• (6)

  Fading trends: a decline in established seasonal discussions reflects changing consumer preferences or a landscape increasingly competitive, where certain products or services lose their previous appeal. This trend might indicate areas where consumer interest is shifting away, possibly due to advancements in technology, changes in societal values, or better alternatives becoming available.

Table 11 outlines strategic procedures to implement for quality improvement based on the interpretation of different variants of seasonal anomalies in Digital VoC.

To provide a real-world illustration of a trend anomaly and the corresponding step response, a case study is presented. This case study investigates the digital VoC concerning an electric scooter, focusing on a four-year span of collected data. The primary subject of this case study is the topic of “overheating” which revolves around discussions of electric motor overheating issues in scooters. Keywords defining this topic include engine, overheating, heat, warm, temperature, malfunction, speed, cool, battery and led.

Historically, conversations about “overheating” have shown distinct seasonal trends, peaking during summer months due to higher scooter usage and elevated temperatures. However, a shift was observed in the most recent year’s data, with discussions on overheating markedly intensifying compared to previous years. These discussions are predominantly associated with complaints and issues, with $MRP$ profiles consistently negative over the four years analyzed (see Figure 9b). The textual content of the reviews did not indicate a change in the topics discussed (see Table 12); rather, it was the analysis of the $IMTP$ that showed a pronounced increase in discussions during the summer months (see Figure 9a).

This significant uptick in overheating discussions suggests a potential decline in the cooling system’s performance, leading to a higher occurrence of overheating-related issues.

Following the identification of an increased intensity seasonal anomaly with a negative MRP profile, relating to discussions about overheating issues in electric scooters, the manufacturing company initiated a targeted approach for immediate quality improvement. This anomaly, marked by a pronounced rise in customer complaints during the summer months, highlighted a need for intervention to address the cooling system’s decline in performance.

The first step involved a detailed analysis to isolate and verify the root causes of the overheating issues. The company’s engineering team conducted a review of the scooter’s design, focusing on the cooling system’s effectiveness under varying temperature conditions.

Upon identifying the technical factors contributing to the increased overheating incidents, the company moved to develop and test solutions aimed at enhancing the cooling system’s efficiency. This involved exploring design modifications and material changes to improve heat dissipation and motor cooling during extended use or high temperatures.

Once the necessary improvements were finalized and validated for their effectiveness, the company implemented these changes in the production of new scooters. To communicate these corrective actions and reassure customers of its commitment to product quality and reliability, the company launched a communications campaign to inform them of the steps taken to address the overheating issues, including details of how existing scooter owners could benefit from the improvements made.

This study presents an innovative framework for categorizing and responding to anomalies in Digital VoC data, providing a systematic approach to improve quality management processes in alignment with the principles of Quality 4.0. By introducing a structured taxonomy for identifying various types of anomalies (spike, level, trend, and seasonal anomalies) the research addresses a critical gap in both theory and practice. This approach enables businesses to not only understand the evolving dynamics of customer feedback but also apply targeted interventions that align with shifting consumer needs.

Various studies highlight the importance of uncovering underlying themes in customer feedback. However, most of these analyses have been limited to a static approach, focusing on a snapshot of customer feedback at a particular moment in time. In cases where studies did identify evolving trends in topic discussions over time, they typically noted these changes without examining their nature or categorizing the types of trends observed. In essence, the dynamic aspects of customer feedback were acknowledged but not explored or interpreted.

A vast body of literature on the concept of anomalies in time series is available, encompassing methods for detecting, analyzing, and interpreting deviations from expected patterns. However, despite the maturity of this field, the concept of anomalies has not been thoroughly applied to the analysis of Digital VoC. While time series analysis is a powerful tool for identifying trends and patterns over time, its application to customer feedback in digital environments remained underexplored.

From a practical standpoint, by categorizing anomalies in customer feedback and offering detailed procedures for addressing them, organizations can enhance their quality management systems in several ways, including: (1) early detection of quality issues: the identification of spike anomalies enables companies to react to emerging customer complaints, preventing issues from escalating into major quality problems; (2) strategic product improvements: level and trend anomalies offer insights into long-term changes in customer expectations, enabling companies to adjust their products and services to align with evolving market needs; (3) product feature refinement and development: seasonal anomalies provide insights into shifting customer priorities or pain points during specific periods, allowing businesses to adjust product features or services to meet evolving customer needs.

This study contributes to the expanding literature on Digital VoC by introducing a novel application of anomaly detection techniques to customer feedback analysis. The taxonomy of anomalies and the corresponding quality improvement procedures offer a structured approach for scholars studying customer behavior and sentiment in the digital age. Furthermore, the integration of advanced text mining techniques with quality management practices contributes to the field of Quality 4.0, particularly the role of real-time customer feedback in continuous improvement systems (Antonino et al., 2022; Broday, 2022; Sader et al., 2021).

Despite its contributions, this study presents some limitations. First, the study relies on historical data to analyze customer feedback anomalies, assuming that past patterns are indicative of future behavior. This assumption may not hold in fast-evolving markets or during disruptive events such as economic crises, where customer behavior can change rapidly and unpredictably. Second, the study’s reliance on text mining techniques may not fully capture the breadth of customer sentiment, particularly in cases where feedback is implicit or not clearly articulated in textual form. This limitation can result in a partial understanding of customer perceptions.

Building on the limitations identified, future research should focus on developing structured methodologies and tools for incorporating real-time data analysis techniques to overcome the reliance on historical data, enabling businesses to adapt to rapidly changing markets and disruptive events. Dynamic anomaly detection models (Forti et al., 2021) would allow for quicker identification of emerging trends or shifts in customer sentiment. To address the limitations of text mining, future studies should explore multimodal approaches that capture implicit feedback. Finally, expanding the range of text-mining techniques beyond topic modeling, such as integrating clustering algorithms, sentiment analysis, and advanced neural networks like BERT, could improve the depth and accuracy of anomaly detection, offering more actionable insights into customer feedback.