Photo By: David Doubilet
CIKM AnalytiCup
Lazada Product Title Quality Challenge
1
$6,000
2$2,000
3$1,000
$2,000
Team Members
Tam T. Nguyen
nthanhtam@gmail.com
Postdoctoral Research Fellow
Ryerson University
Kaggle Grandmaster
Hossein Fani
hosseinfani@gmail.com
PhD Student
University of New Brunswick
Gilberto Titericz
giba1978@gmail.com
Machine Learning Expert
AirBnb Inc.
Kaggle Grandmaster
Ebrahim Bagheri
ebrahim.bagheri@gmail.com
Associate Professor
Ryerson University
Photo By: Justin Hofman
“hot sexy red clutch rug sack travel backpack unisex cheap with free gift”
𝑦1
clarity
𝑦2
conciseness
“Hot Sexy Tom Clovers Womens Mens Classy Look Cool Simple Style Casual
Canvas Crossbody Messenger Bag Handbag Fashion Bag Tote Handbag Gray”
Problem Setting
Photo By: David Doubilet
Clarity if within five seconds one can understand the title, what the product is, and quickly figure out the key
attributes (color, size, model, ...).
Conciseness if it is short enough to contain all the necessary information. Otherwise, i.e., the title is
too long with many unnecessary words, Or it is too short such that it is unsure what the product is.
Data Set
ML-DM
1. Cleansing
• Noise
• Missing Values
• Outliers
2. Flirting
• Attributes
• Labels (if any)
• Augmentation
3. Feature Eng.
• Extraction
• Reduction
• Selection
4. Model Eng.
• Selection
• Tuning
• Evaluation
1. Cleansing
• Noise
• Html tags in ‘short_description’ (%94)
• Missing Values
• ‘product_type’ (less than %1)
• ‘category_lvl_3’ (about %6) → assign ‘category_lvl_2’
• ‘description’ (less than %1)
• Outliers
• ‘price’ {-1, 999999, 9999999},
• ‘price’ Normalization based on country
2. Flirting
• Attributes
• Color
• Brand
• Non-English
• <img> Image
• <li> enumeration
• 𝒚: Labels
• Disagreement in labels!(label noise)
• Augmentation
• Cloning  color, brand
Label Noise
multi-class
𝑓: 𝑋1 × 𝑋2 × … × 𝑋 𝑑 → 𝑦: 𝑐1, 𝑐2, … , 𝑐 𝑘
binary(boolean) classifier: 𝑦: 0,1
multi-output(label)
𝑓: 𝑋1 × 𝑋2 × … × 𝑋 𝑑 → 𝑦1: 𝑐1, 𝑐2, … , 𝑐 𝑘1
× 𝑦2: 𝑐1, 𝑐2, … , 𝑐 𝑘2
× ⋯ × 𝑦𝑟: 𝑐1, 𝑐2, … , 𝑐 𝑘r
multi-output binary(boolean) classifier: 𝑦1: 0,1 × 𝑦2: 0,1
Targets correlation: (single, fast model for all targets)
Only 3 combinations for (Clear,Concise):
(1,0), (1,1), (0,0)  |~Clear & Concise|= 0
if ~Clear then ~Concise
if Concise then Clear
3. Feature Eng.
• Extraction
• Reduction
• LSA,T-SNE,PCA,SVD
• Selection
• STD
• Correlation X~y
• Linear(t-test, chi2)
• Non-linear(mi)
• Model-driven
• LinearSVM
Feature Engineering
Feature Importance
Linear SVM
10-Fold Set 1 10-Fold Set 2 10-Fold Set 3 10-Fold Set 4
Base Model
Ensemble Model
Final Prediction
Fold Bagging
Fold Bagging
Set Fold Bagging
BLENDBLEND BLEND BLENDSTACK STACK STACK STACK
BLENDBLEND BLEND BLEND
BLEND
Bagging Models
Performance Evaluation
SGD: stochastic gradient descent
LOR: logistic regression
RDG: ridge regression
NBC: naive bayes classifier
XGB: extreme gradient boosting
LGB: light gradient boosting
W2V: word2vec
Model Importance
clarity conciseness
CIKM AnalytiCup 2017: Bagging Model for Product Title Quality with Noise

CIKM AnalytiCup 2017: Bagging Model for Product Title Quality with Noise

  • 1.
  • 2.
    CIKM AnalytiCup Lazada ProductTitle Quality Challenge 1 $6,000 2$2,000 3$1,000 $2,000
  • 3.
    Team Members Tam T.Nguyen nthanhtam@gmail.com Postdoctoral Research Fellow Ryerson University Kaggle Grandmaster Hossein Fani hosseinfani@gmail.com PhD Student University of New Brunswick Gilberto Titericz giba1978@gmail.com Machine Learning Expert AirBnb Inc. Kaggle Grandmaster Ebrahim Bagheri ebrahim.bagheri@gmail.com Associate Professor Ryerson University
  • 4.
  • 5.
    “hot sexy redclutch rug sack travel backpack unisex cheap with free gift” 𝑦1 clarity 𝑦2 conciseness “Hot Sexy Tom Clovers Womens Mens Classy Look Cool Simple Style Casual Canvas Crossbody Messenger Bag Handbag Fashion Bag Tote Handbag Gray” Problem Setting
  • 6.
  • 7.
    Clarity if withinfive seconds one can understand the title, what the product is, and quickly figure out the key attributes (color, size, model, ...). Conciseness if it is short enough to contain all the necessary information. Otherwise, i.e., the title is too long with many unnecessary words, Or it is too short such that it is unsure what the product is. Data Set
  • 9.
    ML-DM 1. Cleansing • Noise •Missing Values • Outliers 2. Flirting • Attributes • Labels (if any) • Augmentation 3. Feature Eng. • Extraction • Reduction • Selection 4. Model Eng. • Selection • Tuning • Evaluation
  • 10.
    1. Cleansing • Noise •Html tags in ‘short_description’ (%94) • Missing Values • ‘product_type’ (less than %1) • ‘category_lvl_3’ (about %6) → assign ‘category_lvl_2’ • ‘description’ (less than %1) • Outliers • ‘price’ {-1, 999999, 9999999}, • ‘price’ Normalization based on country
  • 11.
    2. Flirting • Attributes •Color • Brand • Non-English • <img> Image • <li> enumeration • 𝒚: Labels • Disagreement in labels!(label noise) • Augmentation • Cloning  color, brand
  • 12.
  • 14.
    multi-class 𝑓: 𝑋1 ×𝑋2 × … × 𝑋 𝑑 → 𝑦: 𝑐1, 𝑐2, … , 𝑐 𝑘 binary(boolean) classifier: 𝑦: 0,1 multi-output(label) 𝑓: 𝑋1 × 𝑋2 × … × 𝑋 𝑑 → 𝑦1: 𝑐1, 𝑐2, … , 𝑐 𝑘1 × 𝑦2: 𝑐1, 𝑐2, … , 𝑐 𝑘2 × ⋯ × 𝑦𝑟: 𝑐1, 𝑐2, … , 𝑐 𝑘r multi-output binary(boolean) classifier: 𝑦1: 0,1 × 𝑦2: 0,1 Targets correlation: (single, fast model for all targets) Only 3 combinations for (Clear,Concise): (1,0), (1,1), (0,0)  |~Clear & Concise|= 0 if ~Clear then ~Concise if Concise then Clear
  • 16.
    3. Feature Eng. •Extraction • Reduction • LSA,T-SNE,PCA,SVD • Selection • STD • Correlation X~y • Linear(t-test, chi2) • Non-linear(mi) • Model-driven • LinearSVM Feature Engineering
  • 17.
  • 19.
    10-Fold Set 110-Fold Set 2 10-Fold Set 3 10-Fold Set 4 Base Model Ensemble Model Final Prediction Fold Bagging Fold Bagging Set Fold Bagging BLENDBLEND BLEND BLENDSTACK STACK STACK STACK BLENDBLEND BLEND BLEND BLEND Bagging Models
  • 20.
    Performance Evaluation SGD: stochasticgradient descent LOR: logistic regression RDG: ridge regression NBC: naive bayes classifier XGB: extreme gradient boosting LGB: light gradient boosting W2V: word2vec
  • 21.

Editor's Notes

  • #4 On Lazada, we have millions of products across thousands of categories. To stand out from the crowd, sellers employ creative, sometimes disruptive efforts to improve their search relevancy or attract the attention of customers. Product titles like this degenerate user experience by cluttering the site with irrelevant, misleading titles. In this challenge, we provide you with a set of product titles, description, and attributes, together with the associated title quality scores (clarity and conciseness) as labeled by our internal QC team. Your task is to build a product title quality model that can automatically grade the clarity and the conciseness of a product title. ‘judging a book by its cover’
  • #6 On Lazada, we have millions of products across thousands of categories. To stand out from the crowd, sellers employ creative, sometimes disruptive efforts to improve their search relevancy or attract the attention of customers. Product titles like this degenerate user experience by cluttering the site with irrelevant, misleading titles. In this challenge, we provide you with a set of product titles, description, and attributes, together with the associated title quality scores (clarity and conciseness) as labeled by our internal QC team. Your task is to build a product title quality model that can automatically grade the clarity and the conciseness of a product title. ‘judging a book by its cover’
  • #15 Contraposition Use one target as a feature for the other one. But has problem in practice since we don’t have the validation or test sets’ label.
  • #17 Plus the attributes, we extract more features from the textual attributes, title and short_description stability selection recursive feature elimination and cross-validation