Commerce¶
Table of Contents
Features¶
Item Embedding¶
Ranker¶
1. User features¶
2. Item features¶
3. Context features¶
4. Statistical features¶
5. Fusion features¶
Index |
Feature Type |
Description |
|---|---|---|
1 |
dense_score |
Cosine or dot similarity with query embedding |
2 |
bm25_score |
Raw or normalized BM25 score |
3 |
query_term_match_ratio |
# query tokens matched / total query tokens |
4 |
category_match |
1 if same category; 0 otherwise |
5 |
brand_match |
1 if same brand; 0 otherwise |
6 |
image_sim_score |
Optional: if query is image |
7 |
retrieval_source |
One-hot: [dense only; sparse only; both] |
8 |
retrieved_rank_dense |
Position in dense top-k list |
9 |
retrieved_rank_sparse |
Position in sparse top-k list |
10 |
dense_to_sparse_rank_gap |
Sparse rank - Dense rank |
Domain Understanding¶
Listings¶
Attribute |
Sub-attribute |
Examples |
Characteristics |
|---|---|---|---|
|
uninformative; misleading; spelling/grammar errors |
||
|
low quality |
||
|
|||
Postal |
user provided -> low coverage; incorrect |
||
Lat-Long |
gps inferred -> high coverage; incorrect; upload location might be different than product availability |
||
|
incorrect; misleading/scam |
||
|
mostly missing; possibly incorrect |
||
|
category dependent; mostly missing; possibly incorrect |
||
|
colour; size |
||
|
new; refurbished |
||
|
minimalistic; vintage; casual |
||
|
gift-ideas; travel friendly |
||
|
wedding; office; gym |
||
|
huge discount |
open-ended; clickbaity |
Product Search¶
Problem Understanding¶
Use-case - System:
text queries
system returns a list of listings
sorted to maximise engagement
filtered by geolocation
[*] personalisation
[*] contextualisation
available across different surfaces
Actions (users): - click -> product details page
save to wishlist
contact seller
scroll past
Actors: - users, sellers, platform
Interests: - users: find most relevant results - sellers: increase coverage of their listings - platform:
[out of scope] quality: results should not contain listings that violate policies
user engagement
Business KPIs - CTR, CVR, coverage, QBR, DwellTime
Scale - 1M sellers, 50M listings, 1M/day new listings - 1B users, 95% on mobile device - low latency req (50ms for retrieval, 200ms for rerank)
Signals - Search logs
events: click, dwell-time, contacted-seller, added-to-wishlist - clicks: 10-20%, noisy (weak signal - curiosity, clickbaits) - dwell-time - added-to-wishlist: 1-3%, (stronger - delayed feedback, sparse, niche/personalised) - contacted-seller: 0.1-0.5% (delayed feedback)
depends on: - platform: surface, display-pos - seller: listing-quality, seller reputation, previous engagement with seller - user: user’s click propensity overall/query-specific/category-specific/attribute-specific
baseline - kw search
Misc - subsystems
listings side - kw extraction - taxonomy classification - attribute extraction
query side - query segmentation - query intent - browse, buy, brand - query rewrite/expansion