Skip to main content

Data Sources (First-Party + 4 External Types)

Published 2026-05-14Updated 2026-06-306 min read

First-party real data (PII masked) + synthetic data + 4 external types (social · weather · economy · competitor) explicitly separated by cohort_tag.


1. Data Scale at a Glance

ItemScaleDescription
First-party real data (PII masked)N = 500~5,000 members12 months of LG Members owned-mall transactions + estimated channel sell-through
~50K owned-mall transactionsAverage 100 transactions/member/year
~10K SKUs3 BU own brands + category tree
First-party synthetic data49.5K membersfaker + first-party distribution learning
~5M transactions12 months synthetic
External (social)Daily keywords · postsNaver · Google Trends + X · Instagram · Olive Young
External (weather)365 × N regionsKMA daily temperature/precipitation/air quality
External (economy)Monthly indicatorsKOSIS · Bank of Korea
External (competitor)Event-levelPublic appearances of new products · campaigns

Total ≈ 5M+ transactions, 50K members, 10K SKUs, 365 days of external signals, ~500K Neptune edges


2. cohort_tag Separation Strategy

All instance nodes and OpenSearch documents carry a cohort_tag attribute:

ValueMeaningUI Badge
realPII-masked first-party real dataGreen real
synthFirst-party synthetic data (49.5K)Yellow synth
externalExternal data (social · weather · economy · competitor)Blue external

Queries always use WHERE cohort_tag IN (...). Campaign-send tools only target real members.


3. First-Party Real Data (Hypothetical Scenario)

3.1 PII Masking Rules

Original FieldAfter Masking
Namehash → "Member_a4f2c1"
Phone"010--"
AddressOnly down to city/district
Date of birthAge band (5-year units)
EmailDomain only
Card number"--****-1234"

3.2 Transaction Data

  • 12 months (T-12M ~ T)
  • Owned mall OrderTransaction (web/app)
  • ChannelSellThrough — estimated for Mart/H&B/Convenience/QSR
  • Average 100 transactions/year/member

3.3 SKU Catalog (3 BUs)

BURepresentative BrandsRepresentative Categories
BeautyWhoo / Su:m37 / OHUI / belif / CNP / VDLSkincare / Makeup / Cleansing
HDBElastine / Perioe / Bamboo Salt / Dr.Groot / Saffron / SafranHair care / Oral care / Laundry detergent
RefreshmentCoca-Cola / Fanta / Sprite / Minute Maid / Toreta / PoweradeCarbonated / Juice / Sports drinks

4. First-Party Synthetic Data (49.5K)

4.1 Seed Learning

from sklearn.mixture import GaussianMixture

# Extract member · transaction features from first-party real data
gmm_customer = GaussianMixture(n_components=8).fit(real_customer_features)
gmm_txn = GaussianMixture(n_components=10).fit(real_txn_features)

4.2 Seasonal Variation

PeriodWeighted Categories
Holidays (Lunar New Year/Chuseok ±2 weeks)Gift sets · household goods +30%
Summer (Jun-Aug)Sunscreen · iced beverages · hand wash +25%
Winter (Nov-Feb)Hand cream · hot packs · hot drinks +20%
Rainy daysUmbrellas · indoor diffusers +15%

5. 4 External Data Types

SourceDataAPI/Method
Naver DataLabCategory/keyword search trends (weekly)datalab.naver.com
Google TrendsGlobal · domestic keyword trendstrends.google.com
X (Twitter)Keyword · hashtag · posting frequencyAPI v2
InstagramHashtag trendsPublic Display API
Shorts · TistoryVideo · blog trendsRSS / crawl
Olive Young reviewsCategory · SKU review keywordsCrawl (PDPA compliant)

→ Neptune SocialSignal node + OpenSearch idx_social_trend/idx_review

5.2 Weather · Environment

SourceDataAPI
KMA (Korea Meteorological Administration)Daily temp/precipitation/humidity/winddata.kma.go.kr
Air qualityPM10/PM2.5/Ozonedata.go.kr

→ Neptune WeatherSignal

5.3 Economy · Consumption

SourceDataAPI
Statistics Korea KOSISCPI · employment indicatorskosis.kr
Bank of Korea ECOSFX · interest ratesecos.bok.or.kr

→ Neptune EconomicSignal

5.4 Competitor Signals

SourceDataMethod
Amorepacific / Aimo / Yuhan-Kimberly etc.Public appearance of new products · campaignsOfficial press releases + public SNS
Department store · H&B new product roundupsCategory new product trendsMedia RSS

→ Neptune CompetitorSignal + OpenSearch idx_competitor


6. Data Ingestion Pipeline

Ingestion Order

  1. First-party raw → S3 → PII masking → S3 (real)
  2. Synthetic generator → S3 (synth)
  3. External APIs → Glue → S3 (external)
  4. Semantic Pipeline (Lambda) → Neptune Bulk + OpenSearch Bulk
  5. Cohere embed-v4 batch embedding → OpenSearch knn_vector

7. Data Refresh Cadence

DataCadenceTrigger
First-party real data (N=500~5K)One-time load for PoC, then fixedManual
First-party synthetic data1~2 regenerations during PoCWhen seasons added
Social trendsDaily/weeklyEventBridge cron
WeatherDailyEventBridge cron
EconomyMonthlyEventBridge cron
CompetitorWeeklyEventBridge cron

8. Data Quality Rules

RuleThreshold
Missing Customer cohort_tag0
Product price ≤ 00
OrderTransaction total = sum of lines100%
Neptune edge orphans0
OpenSearch embedding dimension = 1024100%
Cohort distribution (real:synth)1:99

9. Demo-Time Data Exposure Rules

  • All result cards display a cohort_tag badge (green real / yellow synth / blue external)
  • When send/payment tools are invoked with synthetic data → "Synthetic — no real sending" notice
  • Mandatory source attribution for external data (social · weather · economy · competitor)
  • Statistical displays offer cohort separation options (real only / synth only / combined)