MindMap Gallery Improve AI
The advancement of AI technology has greatly improved various aspects of our lives. From enhancing productivity to transforming industries, AI has made significant contributions.
Edited at 2021-06-21 10:21:14Improve AI
Generate understanding
Understand data drift on daily basis
What part of inventory customer see
AI vs bestseller
Can be online monitored
Predictability
Retailer
Newsletter
Marketing sales
Price and discount
External
Weather
Social media trends
Does past predict future
How many new/dropped/recurring customers
How good is 10-list expert of last 10 days today
Change of list
RBO
Coverage
Recall
Plot performance today with models of last 10 days
Customer behaviour
Understand anonimity
Number of customers vs number of past orders
First time customers vs returning customers ratio
Recency/Frequency/Monetary diagram
Historic knowledge vs impact point diagram
Session length
Dependence on category
Collecters vs hunters
Impression, viewing time
Popularity
Categories opened
Tab opening pattern
Tab shoppers vs back and forth shoppers
Multi expert list to understand potential for personalization
How would this impact?
People clones but not similar to bestseller still need AI
People only buying from bestseller - online monitoring
Customer flow
How do they visit product detail page
How many sessions before purchase
Only convert in sales
Understand cold vs hot customers
Returns
First purchase
Cart size
Purchases per year
Price range
Which features are used
Carousel
Search
Category pages
Checkout behaviour
Anonymous then guest
Anonymous then new account
Anonymous then login
Returning then purchasing
Returning then change account
Limitations of current AI
Categories or New/Sales/Regular
Category wise AI report
Sales vs Normal day
Analyze what is different in these days fundamentally
More anonymous customers
People only buy more discounted items
Hourly bestseller list
Customer segments
Before first purchase
After first purchase
Superbuyers
New collections
Age of product vs AI performance
AI behaviour
Change in ranking in session - RBO
Impact of dominating colorgroups and customers
Customers that act a lot and act atypical
Generate AI report while training to understand stability
How?
Online metric to compare if AI is good - compared to bestseller
How are conflicting examples handled
what does this mean?
New features
Improve Resorting
Improve model's raw performance
Better modelling
Customers behaviour
Analyse the Recency/Frequency/Monetary diagram of customersAnalyse how different types of behaviour translates to impactAnonimityPredictability of past behaviour in futureSession length
Impressions and viewing time
Tab shoppers vs back and forth shoppers
Session length
Anonimity
Items behaviour
Analyse how new/trending/popular items behaveSee how it impacts revenueHow many customers buy from theseOnline monitor how AI is different from these static lists
Newness/Popularity/Trending
Purchasability
Stock
Price/Discounts
Embedding improvements
High probability this will improve performance
Photo
Category
Tags
Text description
Item cold start
Analyze performance of AI based on age of product
One list per category
Analyze model's recall graphs category wise
Improve model stability
Everday
Training improvement
Train more frequently
SWA
More regularization
Dropout
Data Augmentation
Random skip nodes in input sequence - more robust to noise and improve performance in realistic setting
Early stopping
Improve normalization scheme
Analyse which type of events hurt model performance the most and come up with normalization strategyAnonymous events (or less purchase history customers)Change in prices/discountsBehaviour change of people on sales days/weekends
Generic bestseller for anonymous
Compare our bestseller to CBRWill need for SRGNN anyways
Confidence of model's ratings
If model is not confident switch to generic bestseller?
Ensemble
Dynamic adaption classifier
Train on different customer behaviour
Sales
Sales Affinity
Dynamic switch to generic bestseller
Online monitoring can help us automatically switch to bestseller if model is worse than it on sales daysConfidence can help when user is anonymous or is behaving in different manner - might also be better to switch to generic bestseller
Based on confidence
Based on online monitoring
Current price and discounts
Hourly bestseller
This could be strong baseline for sales days if AI has very low confidence or is worse than bestseller for most of the time during sales
Online monitoring of AI
High valueWill tell us about model's performance in different scenarios (anon vs returning, regular vs sales, cold vs hot items) and help us prioritize features
Average recommendation popularity
Compare with bestseller after each purchase
RBO between AI and bestseller on each resort call
Improve long term retention
Serendipity
Customer in a box
Influence behaviour of customer
RFM analysis
Newsletter
Feed news-letter to AI
Personalized newsletter
Dynamic discount/promos
Only change list price
Change current price
Personalized category - "For you"
Refactoring/Technical Improvements
Improving AI Report
Onboarding New Retailers