With the increase in publicly available user-generated content due to the proliferation of internet-assisted communication, researchers have developed several automated approaches to identify, summarize, and classify the available information [26,36]. The development of new tools allows researchers to obtain more information about users’ opinions and sentiments in their writing. There is a trend to shift the focus of opinion mining from studying long texts to shorter user posts on various social media platforms and websites [22]. The rapid development and increase in the efficiency and capabilities of text mining and natural language processing (NLP) algorithms is evident, and they are becoming an important part of social science research in the study of user-generated content [37,38]. Speed, reproducibility, and reliability are considered some of the most important advantages of text mining when it comes to classifying and categorizing text [39].

In this study, we applied two text mining methods to our dataset: topical n-grams identification and topic modeling. Both methods work by identifying and grouping words that occur simultaneously in the text (user reviews in our case). Data analysis required preprocessing of the raw data, which was performed through several procedures commonly used in data preparation and preprocessing for text mining analysis [40].

Data preprocessing is a data mining technique that transforms raw data into an understandable format. Real-world data are often incomplete, inconsistent, contain a substantial amount of redundant information, and are likely to include many errors [41]. This is a critical and time-consuming process as the output depends on the quality of the data.

Since both the topical n-grams identification and topic modeling approaches have the same preprocessing steps, the same preprocessed dataset was used in both methods. For these tasks, we used the Python programming language in combination with its data science–specific tools (ie, libraries) that made this process possible given the large amount of data. The Python libraries numpy and pandas, which are well known in the data science community, were used extensively throughout the process, along with several other libraries, each specialized for a particular task. The following data preprocessing steps were applied.

First, we removed all non-English reviews. We were only interested in English reviews at this point since our dataset contains reviews in different languages such as Portuguese, Spanish, and German. These reviews make up about 12% of our dataset (9576 reviews), and it was safer to remove them than to translate them into English. A total of 72,084 user reviews in English were identified in this step using the Python library langdetect. These reviews were used for all analyses performed in this study.

We then converted the text to lowercase, performed an extensive spell check of every review, and made necessary corrections using the Speller Python library. Words such as “I,” “are,” “and,” and “the” were considered “stop words” and removed, as such common words tend to dominate the results. We further removed any special characters and numbers from the reviews.

Second, lemmatization was performed, which is a process of grouping the inflected forms of words so that they can be analyzed as a single item. These forms are identified by the lemma of the word (ie, both “tracking” and “tracks” share the same lemma and become “track”). The lemmatization algorithm considers the morphological analysis of the words during data preparation [42]. Finally, we applied part-of-speech tagging, which determines the category of the word (eg, noun, adjective, adverb, verb) based on both its definition and context [43].

After data preprocessing, a new dataset was obtained with cleaned data that could be used for both topic modeling and n-grams identification. This new dataset comprised a collection of arrays of words. For example, the item in the previous dataset ‘’I love this app and it’s easy to use’’ is converted into the word array “love app easy use,” which was used as such for further analysis.

The use of topical n-grams is common in text and topic mining, as is the NLP approach when tracking word or phrase frequencies [44,45]. For the purpose of this study, we implemented the analysis of the most frequent trigrams (ie, groups of three words used together) occurring in user reviews (textual user evaluation of the app). For example, applying this to our previously mentioned word array from the cleaned dataset “love app easy use,” we could extract exactly two trigrams: (love, app, easy) and (app, easy, use).

These two trigrams would then be added to the trigrams extracted from other reviews, resulting in a total of 744,808 trigrams from 72,084 reviews.

Evaluation of the word combination was then used in conjunction with the users’ numerical app assessment (ie, rating), which is often used as a proxy for sentiment (eg, [46,47]). User ratings for apps in Google Play Store are scored on a scale from 1 to 5, where 1 is the worst and 5 is the best rating an app can receive. In our study, scores below 3 were associated with a negative app assessment, whereas scores of 4 and 5 were considered positive. This approach is common in studies in this field (eg, [46]). The ratings were averaged by grouping ratings from reviews containing given trigrams.

Topic modeling is another text-mining and NLP method that is commonly used to discover latent topics in a corpus of text. Topic modeling has been shown to be useful for clustering documents or text, and is considered a probabilistic statistical technique for semantic structures [48]. In this study, we used the Python Gensim library, which is commonly used in NLP, for topic modeling analysis. This library helped us to build a mathematical model that could classify each review by topic. The list of possible topics was determined during model training, and we predetermined the number of possible topics. To find the most appropriate number of topics, we used the coherence score. Topic coherence measures the degree of semantic similarity between the highly scored words in the topic, which can help to distinguish between topics that are semantically interpretable and topics that are artifacts of statistical inference [49]. This value is given after each model training process and helped us determine the performance of our trained model.

Finding the best number of topics that would give optimal results required several trials, starting with a randomly selected number of topics until we narrowed down to the model with the best score. We would simply select this model and apply it to our dataset. For example, if our model found 11 topics in the dataset, for each review in our dataset, the model would provide us with the probabilities of how likely the review is to belong to each of the 11 topics.

Since topic modeling is an unsupervised method, it was not constrained by certain predefined standards (ie, number of topics). Instead, for the first run, we programmed the script to start with only 2 topics, repeat and increase by 4 (since this is a computationally intensive and demanding process, we had to minimize the number of runs) until reaching 30 (ie, finding optimal number of topics anywhere between 2 and 30 topics). This high number was randomly chosen to find the optimal range for our topic number. This analysis revealed that the number of topics with the best coherence score was between 6 and 13 (Figure 1). For the second iteration, we again reprogrammed the script to repeatedly run in this range, this time increasing by 1 to more accurately determine the best score. The best coherence (0.646) was achieved with 11 topics (Figure 2, Table 2).

Most of the identified topics included the use of positive words when describing apps in the reviews. In their feedback, users often use words such as “love,” “nice,” “easy,” “good,” and “amaze” to describe the apps. Positively rated topics were more common than negatively rated topics. Users who leave feedback for diet-tracking apps positively rate the possibility to track their food intake; use food scanners and create/access food databases in the apps; and consider the apps to be user-friendly, convenient, and easy to use overall. Weight loss was another important topic, appearing in 10% of user reviews (Table 3).

By pairing the topics with the average ratings of user reviews from the topic, we found that difficulties with cancellations, payment plans, and charges seem to bother users the most (topic average rating 2.32). One of the users described her experience as follows:

In addition, technical difficulties appear to create issues in using the app (average topic rating 2.84):

Similar to the topic modeling results, our overall trigram analysis suggested that, on average, users rate the diet-tracking apps positively in their reviews (Table 4). Trigrams indicating the apps’ ease of use and helpfulness (21/50 top trigrams by frequency of mention), the expression of liking/loving the app (13/50 top trigrams by frequency of mention), and the apps’ tracking option (19/50 top trigrams by frequency of mention) dominated the top 50 trigrams identified from the pool of user reviews. These topics were also frequently associated with high average app ratings; 49 of the top 50 trigrams by frequency attained ratings over 4.

The number of mentions of positive and negative trigrams in user reviews also showed a trend of positive evaluation dominance among users leaving reviews. The top 50 most frequent positive trigrams appeared 12,723 times, while the top 50 most frequent negative trigrams were mentioned 1270 times in our dataset of 72,084 user reviews.

With respect to positively valenced user ratings, we identified a significant presence of reviews praising the apps in general. Most of the top 50 positively rated user reviews refer to the apps’ ease of use and helpfulness (21/50 top positive trigrams by frequency of mention), intake and calorie tracking (20/50 top positive trigrams by frequency of mention), and weight loss (6/50 top positive trigrams by frequency of mention) (Table 5). Users generally characterized the apps as “very easy to use and intuitive” and “easy to use and pretty straightforward.” The ability to track their intake is described as “keeps me accountable” and “very informative about the diets you choose.” Some users also viewed the apps as contributors to their better and healthier food choices:

Similar comments were found for weight loss:

Based on our results, an aggressive approach to advertising premium app options, and unclear policies of subscription charges and cancellations seem to be particularly problematic (15/50 top negative trigrams by frequency of mentions). The reviews also indicate that constant display of ads and reminders about premium app options often lead users to delete the app, as illustrated by the following excerpts from the reviews:

Given the size of the apps in question and their numbers of users, it is easy to see how the ability to upload content could be difficult to manage from a technical perspective. Nevertheless, 14 of the 50 most frequently mentioned negative trigrams refer to technical issues experienced by users (Table 6), which include app crashing, inability to use or open the app, and similar:

Some content-related complaints could also be found in reviews, such as “need to be able to add new food with more than the 100 grams” or “needs an option to let users easily add new foods and correct scanned foods that have incorrect nutritional data.”

Adding new foods creates additional issues for users (5/50 top frequent negative trigrams). Namely, users’ complaints in this area usually refer to the inability to add a product to the database due to technical challenges: