Conclusion

This study analyszed customer reviews from Flip.kz using VoC and RCA framework. Analysis included EDA, AI generated tag extraction, RCA Taxonomy developement and automated classification of product problems.

Several aproaches were evaluated for assigning issue classes to extracted tags. Unsupervised methods showed limited performance, indicating that semantic similarity alone is unsuffucuent for accurately identifying root cause categories. Semi-supervised methods significantly improved results by leveraging predefined issue classes and anchor examples. Among all methods, LinearSVC combined with BGE-M3 embeddings achived the best performance, reaching ARI of 0.51, NMI of 0.67 and overall accuracy 72%. Few-shot classification with Llama 3.1 8B showed competetive results and outperformed several embedding based methods.

The results also demonstrated that embedding quality has a substantial impact on RCA performance. Across almost all evaluated methods, BGE-M3 embeddings consistently outperformed MiniLM embeddings.

To prioritize product problems from business perspective, RCA impact scoring framework was introduced. The framework combines issue prevalence, customer satisfaction signals, review volume and business defined severity weights. Analysis showed that the most frequent issues are not always the most critical. While appearance and durability issues account for a large share of reported problems, categories such as health reactions, usability issues and performance related problems may require higher prioritization due to their potential impact on customer experience and business risk.

Overall, the proposed pipeline demonstrates that customer reviews can be automatically transformed into a structured RCA taxonomy, enabling scalable identification, classification, and prioritization of product issues. Such an approach can support product quality monitoring, category management, and data-driven decision making across large e-commerce catalogs.

A key limitation of this study is the lack of additional business context, such as refund counts, sales volume, revenue and other auxiliary features that could improve the accuracy of the analysis. Even without these features, the study provides useful insights into market behavior and supports decision making under limited data conditions.

Future work is more likely to include AI summary models and their evaluation. However, modern LLMs perform multitask processing and are not limited to summarization only. We can identify the main root causes while generating summaries, which may help reduce overall processing time.