Wearer landscape — new vs existing themes

Generated: 2026-06-26 13:55
Script: analyze_wearer_landscape.py

Report block convention

Each analytical section uses up to three executable layers:

Prose How we found this (math) gives the formulas. Cached artifacts live in cluster_output/.

What else we can do (total landscape roadmap)

This corpus supports a full VoC × product × people × time stack. Completed or in progress:

Next extensions (not yet built):

1. Causal mix decomposition — is category sentiment fall driven by OLENS/color mix shift or within-SKU deterioration? (Shapley or fixed-effects on segment×month.)
2. Competitive switching attribution — link detractor Cluster 3 to prior brand/SKU (from transition pairs).
3. Store × wearer cohort — are new wearers concentrated in outlier stores (H7249002)?
4. Prescription / toric / multifocal onboarding — first-review topic paths for medical vs cosmetic entry.
5. LLM re-prompt for wearer intent — add wearer_type: new|existing|unknown to classification schema (cleaner than regex).
6. Survival / repeat hazard — P(second review within 90d | first-review themes).
7. Price-regret vs product-failure split — separate mixed-sentiment price timing from fit/eye_health negatives.

Cohort definitions

[result:wearer_cohort_counts] | Cohort | Rule | N | Interpretation | | --- | --- | --- | --- | | declared_new_wearer | Scored novice / first-lens trial | 37 | Self-identified novice | | declared_existing_wearer | Scored repurchase / veteran | 3,885 | Self-identified repeat | | platform_first_review | User's 1st review (no declared signal) | 6,678 | New to reviewing on HKTV | | platform_repeat_reviewer | User's 2+ review (proxy) | 6,224 | Engaged repeat reviewer | [/result]

[ml:wearer_cohort-rules] library: wearer_cohort.py (regex scoring, not sklearn) script: classify_review(comment, m_topics, m_praise, review_ordinal) scores: S_n novice patterns, S_e veteran patterns, veto penalties gates: wearing context, ordinal caps for weak trial language audit: cluster_output/wearer_audit_sample.md [/ml]

How we found this (math): If S_e ≥ 3 and S_e ≥ S_n → existing; elif S_n clears thresholds → new; else proxy by ordinal o_u within user_pk.

(a) Top themes — new wearers (declared_new_wearer)

Sentiment: 54.1% positive, 27.0% negative, 16.2% mixed · avg ★3.78 · avg comment 37 chars

LLM topics (mention rate per review)

[result:new_wearer_topics] | Topic | Rate | vs existing (pp) | | --- | --- | --- | | repurchase | 0.0% | -59.5 | | customer_service | 0.0% | -48.8 | | delivery | 2.7% | -44.7 | | price | 5.4% | -42.0 | | expiry | 2.7% | -40.3 | | other | 0.0% | -39.6 | | packaging | 16.2% | -37.2 | | prescription | 10.8% | -33.8 | | brand | 16.2% | -33.6 | | eye_health | 21.6% | -29.5 | [/result]

[babysql:polars] SELECT cohort, topic, rate FROM read_csv('/home/sidmishra/ontologer/subprojects/review_sequencing_sentiment/cluster_output/wearer_topic_rates.csv') WHERE cohort = 'declared_new_wearer' ORDER BY rate DESC LIMIT 12 [/babysql]

[ml:scipy-chi2] library: scipy.stats.chi2_contingency script: analyze_wearer_landscape.py :: topic enrichment test: topic mention independent of cohort [/ml]

Top complaint phrases (LLM-extracted)

• 唔好expect 上眼會有驚喜效果 (2)
• 唔舒服 (2)
• 盒裝濕 (2)
• con packing破損 (2)
• 遇到開左就崩左 (1)
• 啲碎留係眼入邊係咩後果 (1)
• 對Con變細咗第一次買到咁嘅Con 好難想像長期帶會點 (1)
• 眼乾眼澀 (1)
• un著un著感覺 (1)
• 第一次試戴呢個牌子 (1)

Top praise phrases

• 舒服 (3)
• 舒適 (2)
• 顏色自然嘅全深啡色 (2)
• 唔會直徑太大嘅大眼仔 (2)
• 第一次買，度數算準，配戴也比較舒服 (1)
• 度數準確 (1)
• 配戴舒服 (1)
• 顏色唔错 (1)

New-wearer narrative: Themes cluster on fit_sizing, comfort, and eye_health — onboarding anxiety, not repurchase. Negative rate is 27% vs 8% for declared existing — failure modes are adaptation (异物感, dry eyes) not loyalty erosion.

(b) Top themes — existing wearers (declared_existing_wearer)

Sentiment: 86.2% positive, 7.8% negative · avg ★4.54

LLM topics (mention rate per review)

Top complaint phrases

• 冇有 (111)
• 唔乾 (10)
• 唔舒服 (7)
• not available in all sizes (6)
• 眼已覺攰及乾 (4)
• 不能久帶 (4)
• 冇有其他問題 (3)
• 唔好用 (3)

Top praise phrases

• 回購 (398)
• 多次回購 (263)
• 用開 (248)
• 已多次回購 (160)
• good quality (119)
• comfortable fit (94)
• good (82)
• 一直用開 (66)
• good fit (64)
• 已回購多次 (61)

Existing-wearer narrative: repurchase and brand dominate — advocacy and habit. Comfort/quality are table stakes; negatives are rare and often operational (delivery, expiry).

[babysql:polars] SELECT cohort, topic, rate FROM read_csv('/home/sidmishra/ontologer/subprojects/review_sequencing_sentiment/cluster_output/wearer_topic_rates.csv') WHERE cohort = 'declared_existing_wearer' ORDER BY rate DESC LIMIT 12 [/babysql]

Platform proxy comparison (no declared signal in text)

[babysql:polars] SELECT wearer_cohort, COUNT() AS n, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'positive' THEN 1 ELSE 0 END) / COUNT(), 1) AS pos_pct, ROUND(100.0 * SUM(CASE WHEN m.m_sentiment = 'negative' THEN 1 ELSE 0 END) / COUNT(*), 1) AS neg_pct FROM '/home/sidmishra/ontologer/subprojects/review_sequencing_sentiment/cluster_output/wearer_tags.parquet' AS w JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl ON w.review_id = cl.review_id 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 wearer_cohort IN ('platform_first_review', 'platform_repeat_reviewer') GROUP BY wearer_cohort [/babysql]

First-time reviewers on the platform show higher negative rate than repeat reviewers — even without saying "first time wearing lenses." Likely mix of trialists, disappointed switchers, and motivated complainers.

Product mix — what new vs existing wearers buy

Declared new — top modalities / lens types

modality: daily(100%) lens: spherical(68%), color(24%), multifocal(8%)

Declared existing — top modalities / lens types

modality: daily(93%), monthly(4%), 2-week(3%) lens: spherical(77%), color(14%), multifocal(8%), toric(1%), color_toric(0%)

[babysql:polars] SELECT w.wearer_cohort, v.v2_modality, v.v2_lens_type, COUNT() AS n, ROUND(100.0 * COUNT() / SUM(COUNT(*)) OVER (PARTITION BY w.wearer_cohort), 1) AS pct FROM '/home/sidmishra/ontologer/subprojects/review_sequencing_sentiment/cluster_output/wearer_tags.parquet' AS w JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/cl_reviews.parquet' AS cl ON w.review_id = cl.review_id JOIN '/home/sidmishra/ontologer/data/processed/hk_market/hktvmall/product_id_to_attributes_v2.parquet' AS v ON cl.product_id = v.product_id WHERE w.wearer_cohort IN ('declared_new_wearer', 'declared_existing_wearer') GROUP BY w.wearer_cohort, v.v2_modality, v.v2_lens_type ORDER BY w.wearer_cohort, n DESC [/babysql]

Topic enrichment — new vs existing (χ²)

How we found this (math): Per topic (t): (\hat\pi_t = #{\text{reviews mentioning }t}/N). Test (H_0): topic mention independent of cohort using χ² on 2×2 table (mention × {new, existing}).

Output files