Turning Reviews into Revenue: Sentiment Analysis as a Sustainable Competitive Edge

TL;DR

In the hyper-competitive global fashion market, "listening" to customers is no longer optional. This study leverages a dataset of 23,486 e-commerce reviews to demonstrate that by combining text-mined sentiment scores with machine learning (Logistic Regression and MLP Neural Networks), businesses can predict customer recommendations with up to 94.5% accuracy. The research highlights that while numerical ratings are vital, the sensitivity of written text is a critical secondary indicator of brand loyalty.

Contextualizing Social Media in Business Strategy

Social media has evolved from a mere communication tool into a "research medium." The authors position this work within the Web 2.0 framework, where user-generated content (UGC) acts as the primary driver for brand awareness. The core intuition is simple: if a company can quantitatively decode the "feeling" behind a review, it can identify specific product flaws and strengths that numerical stars alone might miss.

Methodology: From Words to Weights

The researchers employed a two-stage analytical pipeline:

1. Lexicon-Based Sentiment Mining: Using the Harvard-IV Dictionary, they calculated three primary metrics: PosGI (Positivity), NegGI (Negativity), and SentGI (Overall Sensitivity). They also calculated normalized versions (per 100 words) to eliminate bias from longer reviews.
2. Supervised Classification: These scores were used as independent variables to predict the "Recommended" (0 or 1) status.

Figure 5: The architecture of the applied MLP neural network showing the flow from input variables to the hidden layers.

Deep Dive: Logistic Regression vs. Neural Networks

The study compared two distinct approaches:

• Logistic Regression: Proved that Rating, SentGI, and Age are statistically significant. One unit increase in Rating increases the odds of recommendation by a staggering 27.57倍.
• MLP Neural Network: Utilized hyperbolic tangent and softmax activation functions. While Logistic Regression removed certain variables like PosGI as insignificant, the Neural Network successfully utilized them by identifying non-linear relationships, resulting in a higher accuracy (94.5%).

Experimental Insights & Results

The "Word Cloud" and frequency analysis identified that positive terms like "love," "perfect," and "comfortable" dominate recommendations, while "returned" and "disappointed" are the primary churn indicators.

Fig 7: Relative importance of factors. Note how Rating dominates, but SentGI provides significant additional predictive value.

The ROC curve (AUC = 0.976) confirms that the model is exceptionally robust at distinguishing between those who would recommend a product and those who wouldn't.

Critical Analysis & Takeaways

Key Takeaway: Sentiment Analysis isn't just a "nice-to-have" metric—it’s a predictor of future business success. The SentGI (Sensitivity Index) acts as a bridge between qualitative feedback and quantitative strategy.

Limitations: The authors acknowledge that sentiment is context-dependent. A "tight" fit might be positive for leggings but negative for a summer dress. Future research must look into domain-specific lexicons or LLM-based embeddings to better capture these linguistic nuances.

Conclusion: For modern retailers, the path to a sustainable competitive advantage lies in the transition from "broadcasting" to "listening." By automating the analysis of sentiment, companies can adapt their marketing and product development in near real-time.