Decoding Thai Cosmetics: A Low-Cost Text Mining Approach to Customer Satisfaction

TL;DR

In the hyper-competitive "Red-ocean" of the Thai cosmetics market, understanding the nuance of customer feedback is a survival requirement. This research proposes a budget-friendly pipeline using Hierarchical Clustering and Word Clouds to distill over 10,000 social media reviews into actionable marketing insights, identifying the specific keywords that drive or destroy brand loyalty.

Background Positioning

While Silicon Valley focuses on trillion-parameter LLMs, this work addresses a practical "Value-First" problem: How do local startups in Southeast Asia leverage big data without a massive R&D budget? Positioned as an applied computing study, it bridges the gap between complex text mining and practical digital marketing strategy.

Problem & Motivation: The High Cost of Insight

Small-to-Medium Enterprises (SMEs) face a paradox—they need sophisticated data insights to compete with global giants, but platforms like SentiGeek or BirdEye are priced for enterprises. The authors argue that the "Red-ocean" nature of the cosmetics market requires a specific understanding of Customer Experience (CX). If a brand doesn't know that "inflammation" is the trending negative keyword for their new mask, they cannot pivot their R&D or marketing fast enough to survive.

Methodology: From Raw Text to Semantic Clusters

The core of this paper lies in its transition from unstructured Thai social media text (Facebook, Pantip) to a structured mathematical representation.

1. The Pre-processing Pipeline

The authors utilize a standard NLP pipeline:

• Tokenization: Splitting sentences into words.
• Standardization: Converting text to lower case and removing special characters.
• Data Cleansing: Removing stop words and indexing.

2. Hierarchical Clustering (Linkage Algorithm)

To move beyond simple word counts, the research employs Hierarchical Clustering. This groups related words to reveal deeper sentiment structures. The system calculates the distance between terms using Pearson’s correlation coefficient:

$Distance(X,Y)=1-(extcov(X,Y)/(extstd(X)imesextstd(Y)))$

This ensures that words frequently appearing together in reviews (e.g., "gel" and "texture") are clustered, providing a more context-aware visualization.

3. Visual Encoding with Linear Normalization

To ensure the Word Cloud isn't just a mess of characters, a linear normalization formula is applied to font sizes ($S$):

$S=S_{max}imes(C_{max}-C)/(C_{max}-C_{min})$

(Image: High-level overview of the analysis focus on Thai Cosmetics)

Experiments & Results: What Customers Truly Think

The study analyzed 10,564 samples over a 4-year period (2015-2018). The findings were granularly broken down by product category:

• Masks/Oils: Positive reviews focused on "acne removal" and "moisture." Negative feedback was dominated by "inflammation" and "ingredients."
• Powder/Foundation: Customers valued "thickness," "UV sunblock," and "smoothness." Negative sentiment centered on "allergies" and "clogged pores."
• Gels/Lotions: Fragrance ("smell") and "price" were the strongest satisfaction drivers.

(Image: Positive sentiment clusters for powder products, highlighting keywords like 'thickness' and 'sunblock')

(Image: Negative word clusters for powder products, showing risks like 'allergy' and 'cause of acne')

Critical Analysis & Conclusion

Takeaway

The research demonstrates that Word Clouds are not just "visual fluff." When backed by Hierarchical Clustering and Pearson correlation, they become a legitimate diagnostic tool for business policy. A startup can immediately see which ingredients trigger negative "inflammation" clusters and adjust their formula before a PR crisis occurs.

Limitations

• Temporal Lag: The data ends in 2018. Cosmetic trends and consumer language (especially slang on Thai social media) evolve rapidly.
• Sentiment Ambiguity: Word Clouds struggle with sarcasm or complex linguistic structures where a "positive" word is used in a "negative" context.

Future Outlook

The next logical step for this line of research is the integration of Aspect-Based Sentiment Analysis (ABSA) using lightweight models like DistilBERT, which could provide even higher precision while remaining accessible to SMEs with limited computing resources.