Stage 3: Intelligence Layer

Raw data only tells part of the story. A product might have 1,000 five-star reviews, but if half are from paid reviewers, that's not genuine love.

This stage adds three layers of AI-powered intelligence to every review.

The Three Models

Model	What It Detects	Output
Quality Scoring	High-quality vs. low-quality reviews	Score 0-1
Fake Detection	Suspicious/incentivized reviews	Probability 0-1
Sentiment Analysis	Emotional tone of text	Score 0-1

Model A: Review Quality Scoring

The Problem

Not all reviews are equal. Compare these two 5-star reviews:

Review 1: "Great!"

Review 2: "I've been using this for 3 months now. My oily T-zone is finally under control without feeling tight. The texture is lightweight and absorbs quickly. Only downside is the price."

To a naive algorithm, both are "5 stars." But Review 2 is obviously more valuable.

How It Works

We train a machine learning model to predict which reviews are high-quality. The clever trick: we use incentivized reviews as a training signal.

Why? Incentivized reviews tend to be:

More detailed (people feel obligated to write more for free stuff)
More structured (they follow prompts)
But also biased (they received something)

By learning to detect these patterns, the model also learns what makes a review substantive.

Features Used (11 total)

Text length and word count
Rating and recommendation
Helpful votes received
User's total review count
User's average rating history
Whether they included photos

The Output

Every review gets a quality score from 0 to 1. We also aggregate to product level:

Average quality across all reviews
Percentage of high-quality reviews
Percentage of low-quality reviews

Model B: Fake Review Detection

The Problem

Some reviews are suspicious:

Brand employees reviewing their own products
Paid reviewers giving 5 stars for compensation
Competitors leaving fake negative reviews
Bot-generated reviews with generic text

A product with 500 reviews where 200 are fake is not as good as one with 300 genuine reviews.

How It Works

We use 68+ different signals to detect fake reviews:

61+ features across 8 categories

The Ensemble Approach

We combine two models for better accuracy:

Traditional Machine Learning — Fast, interpretable, good at numerical patterns (like "always gives 5 stars")
BERT Language Model — Understands text nuance, catches sophisticated fakes that write well but feel off

Final score: 50% traditional + 50% BERT

The Output

Every review gets a "fake probability" from 0 to 1:

0.0 = Almost certainly genuine
0.5 = Uncertain
1.0 = Almost certainly fake/incentivized

🔍Not Binary

We don't label reviews as "fake" or "not fake." Reality is nuanced. A review might be genuine but low-quality, or incentivized but still honest. The probability lets downstream systems make their own decisions.

Model C: Sentiment Analysis

The Problem

Star ratings are blunt instruments. Compare:

4 stars: "This is actually my holy grail product, I'd give it 5 but I'm strict with ratings"

4 stars: "It's okay I guess, nothing special, might repurchase if on sale"

Both are 4 stars, but the sentiment is completely different.

How It Works

We use a pre-trained language model (DistilBERT fine-tuned on sentiment) that reads the review text and outputs a sentiment score.

The Output

0.0 = Very negative sentiment
0.5 = Neutral
1.0 = Very positive sentiment

✅Combined Power

When we combine sentiment with fake detection, we can find products where genuine reviewers express genuine enthusiasm — not just high star counts.

How the Models Work Together

Review Text ────────────────────┐
                                │
Rating & Metadata ──────────────┼───► Quality Model ───► quality_score
                                │
User History ───────────────────┤
                                │
All Features (68+) ─────────────┼───► Fake Detection ──► fake_probability
                                │         │
                                │         ├─► Traditional ML
                                │         └─► BERT (ensemble)
                                │
Text Only ──────────────────────┴───► Sentiment ───────► sentiment_score

Each review now has three scores that travel with it to the ranking stage.

Training Data

For Quality & Fake Detection

We use the is_incentivized flag as a proxy label. Incentivized reviews are:

Not necessarily "bad"
But systematically different from organic reviews
A learnable signal

For Sentiment

We use a pre-trained model (no custom training needed). DistilBERT was already fine-tuned on millions of sentences.

What's Next?

Three scores per review. Now we combine them into one number that captures genuine product love.

Next: Product Ranking → — The Love Score formula.

2. Data Cleaning 4. Product Ranking