Initial Assessment Analytics — HKTVMall Contact Lens Reviews

Author: Composer (Cursor AI Agent)
Date: 2026-06-26
Status: Initial exploratory assessment — LLM-enriched VoC analysis

Sources

Methodology notes:

• Analysis performed with Polars (not pandas).
• Base data deduplicated on review_id (keep last) before join → 16,835 rows (77 duplicate IDs in raw file).
• LLM labels: m_sentiment, m_score, m_topics, m_complaints, m_praise.
• Topic sentiment heatmap cells = % of mentions of that topic falling into each sentiment bucket (rows sum to ~100% per topic).

Descriptive formulas used in this document:

• Positive rate (year t): (\hat\pi_{pos,t} = N_{pos,t} / N_t)
• Brand share (period P): (s_b = N_{b,P} / \sum_{b'} N_{b',P})
• Period share shift: (\Delta s_b = s_{b,\text{late}} - s_{b,\text{early}})
• Star–sentiment decoupling: count of rows with rating ≥ 4 AND m_sentiment = negative

For multivariate segment tests, reviewer clustering, store anomalies, and controlled logistic models, see product_people_analytics.md (sections a–e, each with explicit math).

1. Executive summary

This corpus is a high-volume, superficially healthy category (mean ★4.42, 76.5% LLM-positive) that exhibits clear structural deterioration since 2022 and a systematic decoupling of star ratings from textual sentiment.

Three findings dominate:

1. Category erosion — positive sentiment fell from ~80% (2020) to ~70% (2026); negative sentiment doubled from ~6% to ~21%.
2. Star inflation — 342 reviews with ★≥4 carry LLM-negative sentiment; 240 five-star reviews tag price as negative/mixed (price-timing regret, not product failure).
3. Brand bifurcation — incumbents (ACUVUE, 博士倫, Bausch + Lomb) dominate volume but OLENS and 博士倫 show acute 2025–2026 sentiment collapse; Bausch + Lomb's review share collapsed even as its per-brand sentiment stayed relatively stable.

The dimensions extracted by the LLM (comfort, fit, eye_health, price, repurchase, etc.) are actionable: negative reviews fingerprint to fit_sizing and eye_health, not comfort; mixed reviews cluster on price; repurchase is the strongest advocacy signal.

2. Dataset overview

[babysql:polars] SELECT COUNT(*) AS reviews, COUNT(DISTINCT brand) AS brands, COUNT(DISTINCT product_id) AS products, AVG(rating) AS avg_rating, AVG(LENGTH(comment)) AS avg_comment_len, MIN(LEFT(date, 4)) AS earliest_year, MAX(LEFT(date, 4)) AS latest_year FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' [/babysql]

Sentiment landscape

[babysql:polars] SELECT m.m_sentiment, COUNT() AS n, ROUND(100.0 * COUNT() / SUM(COUNT(*)) OVER (), 2) AS pct, AVG(cl.rating) AS avg_rating, AVG(LENGTH(cl.comment)) AS avg_len FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id GROUP BY m.m_sentiment ORDER BY n DESC [/babysql]

Interpretation: Negative reviewers write ~4.7× longer comments. Volume-weighted dashboards understate pain. Complaints are near-perfectly diagnostic: 100% of negative rows carry an extracted complaint; 98% of mixed rows do; only 5.4% of positive rows mention a complaint.

3. Temporal dynamics — category health

[babysql:polars] SELECT CAST(LEFT(cl.date, 4) AS INT) AS year, COUNT() AS reviews, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'positive' THEN 1 ELSE 0 END) / COUNT(), 1) AS positive_rate, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) / COUNT(*), 1) AS negative_rate, AVG(cl.rating) AS avg_rating FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id GROUP BY year ORDER BY year [/babysql]

Long-run shift (2018 → 2026): positive −5.1 pp, negative +10.7 pp, avg ★ −0.51.

Interpretation: 2020 marks a peak health year. Decline accelerates from 2022 onward — coinciding with OLENS share growth and ACUVUE share concentration. This is consistent with (a) category mix shift toward cosmetic/color lenses, (b) post-COVID price sensitivity, and/or (c) platform review cohort change — not yet decomposed causally here.

4. Brand time-series composition

4.1 Volume mix shift — early vs late period

Share of total review volume among the top 10 brands:

[babysql:polars] WITH tagged AS ( SELECT brand, CASE WHEN CAST(LEFT(date, 4) AS INT) BETWEEN 2018 AND 2020 THEN 'early' WHEN CAST(LEFT(date, 4) AS INT) BETWEEN 2024 AND 2026 THEN 'late' END AS period FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' WHERE CAST(LEFT(date, 4) AS INT) BETWEEN 2018 AND 2020 OR CAST(LEFT(date, 4) AS INT) BETWEEN 2024 AND 2026 ), totals AS ( SELECT period, COUNT() AS n_period FROM tagged GROUP BY period ), brand_period AS ( SELECT period, brand, COUNT() AS n FROM tagged GROUP BY period, brand ) SELECT bp.brand, ROUND(100.0 * SUM(CASE WHEN bp.period = 'early' THEN bp.n ELSE 0 END) / MAX(CASE WHEN t.period = 'early' THEN t.n_period END), 1) AS share_early, ROUND(100.0 * SUM(CASE WHEN bp.period = 'late' THEN bp.n ELSE 0 END) / MAX(CASE WHEN t.period = 'late' THEN t.n_period END), 1) AS share_late FROM brand_period bp CROSS JOIN totals t GROUP BY bp.brand HAVING share_early IS NOT NULL AND share_late IS NOT NULL ORDER BY share_late DESC LIMIT 10 [/babysql]

Key narrative:

• ACUVUE is consolidating share of voice — from ~1-in-4 reviews to nearly 2-in-5. Category health metrics increasingly are ACUVUE metrics.
• Bausch + Lomb's review volume collapsed (−19.4 pp) without disappearing from the market — likely reflects SKU/listing shift on HKTVMall rather than brand exit. Per-brand sentiment for remaining B+L reviews stays relatively strong.
• OLENS gained share (+4.5 pp) into a period of worsening category sentiment — cosmetic lens brands disproportionately contribute to fit/eye_health negative topics (see §6).

4.2 ACUVUE — dominant but not immune

ACUVUE's negative rate tripled from 4.0% (2020) to 17.3% (2026 YTD) while its share of all reviews hit 50.8% in 2026. The category decline is not happening despite ACUVUE dominance — it is increasingly driven by ACUVUE's review base.

4.3 OLENS — share growth, sentiment collapse

OLENS shows the steepest sentiment cliff in the dataset. 2026 YTD (small n=65, treat cautiously) suggests nearly coin-flip negative sentiment — consistent with cosmetic lens expectation gaps (appearance, fit, eye irritation).

4.4 博士倫 — volume leader under stress

[babysql:polars] SELECT CAST(LEFT(cl.date, 4) AS INT) AS year, cl.brand, COUNT() AS reviews, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'positive' THEN 1 ELSE 0 END) / COUNT(), 1) AS pos_rate, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) / COUNT(*), 1) AS neg_rate, AVG(cl.rating) AS avg_rating FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id WHERE cl.brand IN ('ACUVUE', 'OLENS', '博士倫') GROUP BY year, cl.brand ORDER BY cl.brand, year [/babysql]

博士倫's negative rate quadrupled from 7.8% (2020) to 31.1% (2026 YTD). This aligns with high-volume polarized SKUs (e.g. Lacelle Enchanting Gold, 909 reviews, 34% dissatisfaction pressure in product-level analysis).

4.5 Bausch + Lomb — the silent exit from review share

Bausch + Lomb went from 24.5% of 2020 reviews to 3.6% of 2024–2026 reviews, while brands like ACUVUE absorbed share. Remaining B+L reviews retain relatively high sentiment (e.g. 2024: 83.8% pos, 5.0% neg on n=80). Interpretation: this is a distribution/channel story on HKTVMall, not necessarily a product quality crisis.

5. Brand equity snapshot (full period, n≥100)

[babysql:polars] SELECT cl.brand, COUNT() AS reviews, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'positive' THEN 1 ELSE 0 END) / COUNT(), 1) AS pos_rate, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) / COUNT(*), 1) AS neg_rate, AVG(cl.rating) AS avg_rating FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id GROUP BY cl.brand ORDER BY reviews DESC LIMIT 10 [/babysql]

* Net sentiment = positive rate − negative rate
* Loyalty index = share of all brand reviews that are repurchase topic + positive sentiment*

OLENS paradox: 73.6% positive but lowest loyalty index (30.5%) among major brands — high trial, lower repeat. Repurchase language is a stronger retention signal than raw positive %.

6. Topic–sentiment heatmap

Each cell = percentage of that topic's mentions tagged with the given sentiment.
Only topics with ≥200 mentions shown. Sorted by volume.

How to read the heatmap

Hygiene factors (high volume, high positive skew):
comfort, appearance, quality, repurchase, brand — mentioned constantly, overwhelmingly positive. These are table stakes, not differentiators.

Failure modes (elevated negative share):
fit_sizing and eye_health tie at 9.9% negative — the highest negative rates among high-volume topics. When contact lens customers are unhappy, they talk about fit and eye damage, not generic "discomfort."

Ambivalence channel:
price has the lowest negative share (3.2%) but elevated mixed (6.6%) — customers express price-timing regret ("買貴左", "價格已減") rather than pure hatred. This is structurally invisible to star ratings.

Operational risk:
expiry (6.4% neg) and packaging (7.9% neg) point to fulfillment/QC issues — batch expiry and damaged packaging appear in negative fingerprint analysis.

Negative review topic fingerprint (absolute counts)

When m_sentiment = negative, which topics appear?

7. Rating–sentiment decoupling

Calibration matrix (selected cells)

[babysql:polars] SELECT CASE WHEN cl.rating >= 4 THEN 'stars_ge_4' WHEN cl.rating <= 2 THEN 'stars_le_2' ELSE 'mid' END AS star_bucket, m.m_sentiment, COUNT(*) AS n FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id WHERE cl.rating >= 4 OR cl.rating <= 2 GROUP BY star_bucket, m.m_sentiment ORDER BY star_bucket, n DESC [/babysql]

9.0% of all reviews sit in rating–sentiment tension.

Hidden detractors — 5★ + price topic + negative/mixed sentiment

[babysql:polars] SELECT cl.brand, COUNT(*) AS n FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id WHERE cl.rating >= 4 AND m.m_sentiment = 'negative' GROUP BY cl.brand ORDER BY n DESC LIMIT 10 [/babysql]

These are customers who award five stars but encode price regret in text. Star averages overstate true advocacy.

8. Product-level hotspots (n≥30 reviews)

[babysql:polars] SELECT cl.product_title, cl.brand, COUNT() AS n, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) / COUNT(), 1) AS neg_rate, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'mixed' THEN 1 ELSE 0 END) / COUNT(), 1) AS mixed_rate, ROUND(100.0 * ( SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) + SUM(CASE WHEN m.m_sentiment = 'mixed' THEN 1 ELSE 0 END) ) / COUNT(), 1) AS dissatisfaction_pressure FROM '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet' AS m ON cl.review_id = m.review_id GROUP BY cl.product_title, cl.brand HAVING COUNT(*) >= 30 ORDER BY dissatisfaction_pressure DESC LIMIT 10 [/babysql]

* Dissatisfaction pressure = neg rate + mixed rate

Lacelle Enchanting Gold (博士倫, n=909) is the highest-volume polarized SKU — a priority for qualitative review (color expectation vs reality).

9. Voice of customer — extracted language

Top complaint themes

Dominant complaint class: price timing / promotion regret — a pricing architecture problem, not purely product quality.

Top praise themes

10. Strategic implications

11. Recommended next analyses

1. Causal decomposition of 2022+ decline — brand mix vs within-brand deterioration (Shapley or Oaxaca-style)
2. Join to products.parquet — lens type (daily/monthly/color), BC, diameter as covariates
3. Complaint topic modeling — embed m_complaints for finer-grained price vs fit vs health clusters
4. True Sentiment Score — composite of LLM sentiment, topic weights, and star-rating penalty for hidden detractors
5. Dedup fix upstream — 77 duplicate review_ids in cl_reviews.parquet should be resolved at ETL

Appendix A — Reproducing this analysis

cd /home/sidmishra/ontologer/subprojects/review_sequencing_sentiment
source venv/bin/activate
pip install polars   # if not installed
python analyze_reviews.py

Data inputs hardcoded in analyze_reviews.py:

• Base: ~/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet
• Enrichment: ~/ontologer/data/processed/hk_market/hktvmall/6091e_cl_reviews_model_enriched.parquet

Appendix B — LLM classification provenance

See README.md for full pipeline reproduction.

Appendix C — Extended analytics (product × people)

Follow-on analysis with per-section mathematical methodology:

Covers: reviewer clusters in product-mix space, Mar–May 2026 trend deltas, store z-scores, k-NN coherence, χ² topic heterogeneity, logistic regression (ORs with modality/lens/pack/brand controls).

End of initial assessment.