Commerce

Framework

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

  1. Title/Desc

uninformative; misleading; spelling/grammar errors

  1. Images

low quality

  1. Location

Postal

user provided -> low coverage; incorrect

Lat-Long

gps inferred -> high coverage; incorrect; upload location might be different than product availability

  1. Price

incorrect; misleading/scam

  1. Category

mostly missing; possibly incorrect

  1. Tags

category dependent; mostly missing; possibly incorrect

  1. Attributes

colour; size

  1. Condition

new; refurbished

  1. Style

minimalistic; vintage; casual

  1. Use-case

gift-ideas; travel friendly

  1. Occasion

wedding; office; gym

  1. Catchphrases

huge discount

open-ended; clickbaity

Product Understanding

Taxonomy classification

Attribute extraction

Entity linking

Product Quality & Integrity

Duplicate detection

Moderation

Product Recommendation

Similar listings recommendation

Homepage recommendation