KDD Cup 2022 Workshop - Winning Solutions

[manisnesan]

ESCI Challenge for Improving Product Search - https://amazonkddcup.github.io/

[manisnesan]

A simple but effective solution for Task 1 of KDD Cup 2022 Challenge on improving product search

Task Given a user specified query and a list of matched products, the goal of this task is to rank the products so that the relevant products are ranked above the nonrelevant ones[1]. Relevance is broken down into four classes, named Exact(E), Substitute(S), Complement(C) and Irrelevant(I)

Problem Definition Given a list of query-result paired with annotated E/S/C/I labels, predict the gain (regression) for the input <query, product>. This can then be used for ranking.

Relationship b/w product type & gain	Product Type	Gain
Exact	1.0
Substitute	0.1
Complement	0.01
Irrelevant	0.0

Input: query + + product title + ‘ ’ + product description + ‘ ’ + product bullet point + ‘ ’ + product brand + ‘ ’ + product color name

Output: A value, which is regarded as the predicted gain for the input <query, product>

Techniques Used

• Supplementary Data from task 2 + k-fold ensemble (avg of predictions from each fold)
• BERT Encoder (Debertav3-large | mDebertav3-base) + Regression layer

Paper | Poster | Slides

[manisnesan]

Solution of Team GraphMIRAcles in the KDD Cup 2022 Query-Product Ranking Task

Techniques Used

• xlm-roberta-large
• data preprocessing (re cleanup | keyword extraction using YAKE(??) & Textrank)
• Masked Language Model Pretraining on the Domain Corpus for Domain Adaptation
• Fine tuning using Ranking Loss(??) and Fast Gradient Method (??) adversarial training.

Poster | Paper

[manisnesan]

BM25 + Cross Encoder as Reranker is a strong baseline.

as per BEIR benchmarks