<![CDATA[TikTok is a short-video-sharing platform with an extremely large user base in Indonesia, and as a result it also receives numerous negative reviews concerning both technical performance and platform policies on the Google Play Store. This study aims to identify and cluster the main topics of TikTok user complaints using a Machine Learning-based approach built on the Latent Dirichlet Allocation (LDA) algorithm. Data were collected through a quota-based web scraping procedure using the google-play-scraper library, targeting Indonesian-language reviews with a rating below three (rating 1 and 2) up to a target of 1,000 reviews, yielding a raw sample of 1,000 reviews (814 one-star and 186 two-star). The preprocessing stage included cleaning, case folding, normalization using a colloquial lexicon enriched with a TikTok domain-specific correction dictionary, stopword removal, and stemming using Sastrawi, which reduced the data to 814 clean documents. Document representation was constructed using a Bag of Words (BoW) approach through the Gensim library, with vocabulary filtering (no_below = 3, no_above = 0.6) that produced a dictionary consisting of 397 unique tokens. The optimal number of topics was determined through a multi-metric evaluation combining Coherence Score (u_mass), Coherence Score (c_v), and Perplexity for candidate models ranging from two to five topics, complemented by a manual check of topic interpretability. Despite mixed signals from the c_v and Perplexity metrics, the two-topic model was selected based on its u_mass score (-8.7713) and its more interpretable, non-fragmented thematic structure. The document distribution results show that Topic 1 is more dominant, comprising 445 documents (54.67%), representing technical application complaints such as bugs, slow performance, failure to open the application, and excessive advertisements. Meanwhile, Topic 2, with 369 documents (45.33%), relates to complaints about content moderation and restrictions on the live streaming feature. Visualization using pyLDAvis confirmed that the two topics are clearly separated without overlap, reinforcing the validity of the modelling results. As the sample is restricted to Indonesian-language, one- and two-star reviews collected at a single point in time, the findings are intended to reflect this specific segment of TikTok users rather than the full population of TikTok reviews. This study provides objective insights for application developers in prioritizing system improvements and evaluating content moderation policies more effectively.]]>
Copyrights © 2026
