Skip to main content

Knowledge Graph 25 Classes

Neptune (openCypher) 기반 25 클래스 KG. 추정 ~500K edges. 3 BU(Beauty + HDB + Refreshment) 통합 + 외부 4종 시그널 반영.

1. 25 클래스 개요​

2. 그룹별 명세​

2.1 고객/회원 (5)​

클래스핵심 속성주요 관계
Customercustomer_id · gender · age_band · region · signup_date · cohort_tag(real/synth)-[BELONGS_TO]→ Household · -[HAS_MEMBERSHIP]→ Membership
Membership (LG 멤버스)grade(일반/실버/골드/VIP) · points · joined_at · opted_in_marketing-[ON_BEHALF_OF]→ Customer
Householdhousehold_id · household_size · household_type(1인/2인/4인+/기타)-[CONTAINS]→ Customer
Personapersona_id · name(키즈맘/골드미스/1인가구/시니어/트렌드세터) · traits-[CLASSIFIES]→ Customer
Segmentsegment_id · definition · dynamic_query · snapshot_date-[INCLUDES]→ Customer

2.2 상품/카탈로그 (5)​

클래스핵심 속성주요 관계
Product (SKU)sku · name · price · attributes(JSON)-[IN_CATEGORY]→ Category · -[OF_BRAND]→ Brand
Categorycategory_id · name · parent_id · depth-[CHILD_OF]→ Category
Brandbrand_id · name(후·숨37·오휘·엘라스틴·페리오·코카콜라·환타·...) · launched_at-[OWNS]→ Product · -[BELONGS_TO_BU]→ BU
BUbu_id · name(Beauty/HDB/Refreshment)-[GROUPS]→ Brand
Bundlebundle_id · theme · season(설/추석/여름/겨울 등)-[INCLUDES_SKU]→ Product

2.3 거래/행동 (5)​

클래스핵심 속성주요 관계
OrderTransaction (자사몰)order_id · total_amount · channel(web/app) · order_at-[BY]→ Customer · -[CONTAINS]→ Product
ChannelSellThrough (외부 채널 추정)channel_id · sku · units · period · estimated_at-[OF]→ Product · -[AT]→ Channel
CartEventcart_event_id · action(add/remove/abandon) · at-[BY]→ Customer · -[ON]→ Product
SearchEventsearch_event_id · query_text · click_sku · at-[BY]→ Customer
ReviewRatingreview_id · rating(1-5) · text · source(자사몰/네이버/올리브영/X) · at-[ABOUT]→ Product · -[BY]→ Customer

2.4 채널/캠페인 (5)​

클래스핵심 속성주요 관계
Channelchannel_id · name(자사몰·이마트·롯데마트·올리브영·CU·GS25·QSR) · type-[HOSTS]→ Transaction
Campaigncampaign_id · name · start_at · end_at · budget · target_bu-[USES]→ Touchpoint · -[OFFERS]→ Promotion
Promotionpromotion_id · type(할인/1+1/적립) · conditions-[APPLIES_TO]→ Product
Touchpointtouchpoint_id · channel(SMS/카톡/푸시/SNS광고) · template-[REACHES]→ Customer
Couponcoupon_id · value · expires_at · redeemed_at-[ISSUED_TO]→ Customer · -[FROM]→ Campaign

2.5 운영/외부 (5)​

클래스핵심 속성주요 관계
SocialSignaldate · keyword · source(네이버/구글/X/인스타/올영) · volume · sentiment(외부)
WeatherSignaldate · region · temp_c · precipitation_mm · pm10(외부, KMA·대기질)
EconomicSignaldate · indicator(환율·물가·고용) · value(외부, 통계청·한국은행)
CompetitorSignaldate · competitor(아모레/아이모/유한킴벌리) · event_type · text(외부)
Compliancerule_id · topic(약관/마케팅동의/PII/연령/미성년화장품) · enforced_by-[GOVERNS]→ Customer

3. 핵심 관계 (예상 ~500K edges)​

엣지 추정:

  • Customer × OrderTransaction (~50K)
  • Customer × CartEvent / SearchEvent / ReviewRating (~150K)
  • ChannelSellThrough × Product × Channel (~80K)
  • OrderTransaction × Product (~150K, 평균 4 SKU/주문)
  • Product × Category × Brand × BU (~10K)
  • 외부 시그널 (~50K daily snapshots)
  • 기타 (~10K)

→ 약 500K edges

4. openCypher 예시 쿼리​

4.1 S1 시맨틱 검색 — Beauty BU 신상 + 회원 매칭​

MATCH (b:Brand {name: '숨37'})-[:OWNS]->(p:Product)
WHERE p.attributes.is_new = true
MATCH (c:Customer)-[:PLACED]->(o:OrderTransaction)-[:CONTAINS]->(p2:Product)
      -[:OF_BRAND]->(b2:Brand)-[:BELONGS_TO_BU]->(bu:BU {name: 'Beauty'})
WHERE o.order_at > datetime() - duration('P90D')
RETURN p.sku, p.name, count(DISTINCT c) AS potential_buyers
ORDER BY potential_buyers DESC LIMIT 20

4.2 S5 캠페인 ROAS — 옴니채널 어트리뷰션​

MATCH (camp:Campaign {target_bu: 'HDB'})
      -[:USES]->(:Touchpoint)-[:REACHES]→(c:Customer)
      -[:PLACED]->(o:OrderTransaction)-[:CONTAINS]→(p:Product)
      -[:OF_BRAND]->(:Brand {name: '엘라스틴'})
WHERE o.order_at > camp.start_at AND o.order_at < camp.end_at
RETURN c.customer_id, sum(o.total_amount) AS attributed_revenue

4.3 S7 옴니채널 여정 — 단일 회원 BU 가로지르는 행동​

MATCH (c:Customer {customer_id: $cid})
OPTIONAL MATCH (c)-[:TRIGGERED]->(s:SearchEvent)
OPTIONAL MATCH (c)-[:TRIGGERED]->(ce:CartEvent)
OPTIONAL MATCH (c)-[:PLACED]->(o:OrderTransaction)-[:CONTAINS]->(:Product)
                  -[:OF_BRAND]->(:Brand)-[:BELONGS_TO_BU]->(bu:BU)
OPTIONAL MATCH (c)-[:RECEIVED]->(tp:Touchpoint)
WITH c, [s, ce, o, tp] AS events
UNWIND events AS e
RETURN labels(e)[0] AS event_type, e.at AS timestamp,
       coalesce(e.channel, e.source, 'unknown') AS source
ORDER BY timestamp

4.4 S6 외부 시그널 융합 — 기온 vs 자외선차단제 GMV​

MATCH (w:WeatherSignal)
WHERE w.date > date() - duration('P90D') AND w.region = '서울'
WITH w
MATCH (o:OrderTransaction)-[:CONTAINS]->(:Product)-[:IN_CATEGORY]->(:Category {name: '자외선차단'})
WHERE date(o.order_at) = w.date
RETURN w.date AS date, w.temp_c AS temp, sum(o.total_amount) AS sun_care_gmv
ORDER BY date

5. cohort_tag 분리 전략​

값의미UI 배지
realPII 마스킹 실데이터 (N=500~5K)🟢 실데이터
synth합성 데이터 (49.5K)🟡 합성
external외부 데이터 (소셜·기상·경제·경쟁사)🔵 외부

쿼리에서 항상 WHERE cohort_tag IN (...). 캠페인 발송 도구는 real 회원에만.

6. OpenSearch 인덱스 매핑​

인덱스도큐먼트분석기임베딩
idx_productSKU 메타·설명·자사 리뷰 요약NoriCohere embed-v4 (1024d)
idx_customer회원 프로필·태그·세그먼트NoriCohere embed-v4
idx_review리뷰 본문 (자사몰/네이버/올영/X)NoriCohere embed-v4
idx_campaign캠페인 메타·문구Nori(없음)
idx_social_trend소셜 키워드·게시글 (네이버·구글·X·인스타)NoriCohere embed-v4
idx_competitor경쟁사 신제품·이벤트 텍스트NoriCohere embed-v4

→ BM25 + KNN → RRF → Cohere rerank-v3