Sentiment and Topic Modeling of Livin Reviews for Consumer Trust Insights

Abstract

The rapid growth of mobile banking in Indonesia has driven the adoption of financial super-apps

like Livin' by Mandiri, where consumer trust serves as a critical prerequisite for engagement.

While traditional survey-based methods dominate existing literature, they often suffer from

limited sample sizes and inherent biases. To address this gap, this study integrates Natural

Language Processing (NLP) techniques, specifically Support Vector Machine (SVM) and Latent

Dirichlet Allocation (LDA), to analyze user reviews and characterize consumer trust perceptions.

A dataset of 3,000 user reviews was scraped from the Google Play Store and underwent a

comprehensive text pre-processing pipeline. Sentiment classification using an SVM model with

a TF-IDF approach revealed a highly polarized distribution, consisting of 59.3% negative, 40.1%

positive, and 0.6% neutral reviews. While the SVM model achieved robust classification for the

positive and negative classes, it struggled with the severe class imbalance of the neutral group.

Subsequent LDA topic modeling isolated each sentiment category to uncover the underlying

thematic structures. The results indicate that ease of use (mudah) and functional helpfulness

(membantu) act as primary drivers for positive trust formation. Conversely, negative sentiment is

heavily clustered around system reliability and login failures (sulit, login). These findings imply

that resolving authentication stability should be prioritized to mitigate trust erosion, providing

actionable intelligence for digital financial service developers