Data Science and Machine Learning for eCommerce and Retail

Dr. Andrei Lopatenko Director of Engineering,

Recruit Institute of Technology Recruit Holdings

former Walmart Labs, Google (twice), Apple (twice) andrei@recruit.ai

ML for eCommerce

• Search, Browse, for commerce sites and application

• Help users to find and discover items they will purchase

• Maximize revenue/profit per user session

Search

Search - ranking

ranking

Search - LHN

Left Hand

Navigation

Search spell correction

Search type ahead

Browse

Search data size

• Catalogue items • 8 M items now compare ~ 400 M

Amazon / eBay • X 10 in near future • 2 K text description per item + images • Several hundreds of structured attributes

per catalog

Search – user searches

• Tens of millions per day • Tens billions session per year • Online sales 13.2 B per year (http://

fortune.com/2015/11/17/walmart-ecommerce/)

• 500B per year sales offline stories (8% USA economy) in ~ 11K stores

• The number of transactions ~ 10B (public data)

ML addressable problems

• Learning to rank • Given a query, what’s the list of items

with the highest probability of conversion (purchase), ATC (add to card), page view

ML addressable problems

• Typeahead • Given a sequence of characters types by

user, what’s most probably competitions, what are most probable items users wants to buy

ML addressable problems

• Spell correction • Given a user query, what’s the query user

actually wanted to type

ML addressable problems

• Cold start • Given a new items with it’s set of

attributes and no history of sales or exposure on site, predict items sales and item sales per query

ML addressable problems

• Prediction of LHN • Given a user query, what’s the best set of

facet and facet values, which gives higher probability of users interacting with them and finally buying an item

ML addressable problems

• Query understanding • Given a query, build a semantic parse of

query, tag tokens with attributes: blue tshirts for teenagers -> blue:color tshirts:type for:opt teenagers:agerestriction10-20

• Classification: blue tshirts for teenagers: -> type:apparel, price preference: 10-30, releaseyearpreference: 2014-2016

ML addressable problems

• Related searches • Given a query, what are queries which are

either semantically close to this one, or represent coincidental users interests

• Nike shoes -> adidas shoes, sport shoes, • Coffee mugs -> travel mugs, photo coffee

mugs, cappuccino cups

ML addressable problems

• product discovery • help users to explore product assortment, • drive users to diverse products • reduce risk of selecting irrelevant items • help to find price,quality,brand etc

alternatives • reduce pigeonhole risk • provide relevant data to make a decision

ML addressable problems

• Image similarity • Given images of the items, give other

items such that images of those are visually appealing to the users which like the original item (appealing by shape? Color? Texture?) -> causing high conversion in recommendation

ML addressable problems

• Voice search • Given voice input, reply with a list of the

best items • “what are the cheapest samsung tvs in the

store” • “what is best deal on queen bed today?”

ML addressable problems

• extraction of item attributes • Given an item: what are item attributes:

brand, color, size (wheel, screen, height, S/M/XL, Queen/Twin/King/Full), Gender, Pattern, Shape, Features

ML addressable problems

• Representations of users : actions on websites/apps -> searches, clicks, browsing behaviour, product -> purchase preferences, reviews, ratings, return rates

ML addressable problems

• title generation: how to generate the title which will cause maximum conversion rate

• which product attributes select for the title?

What makes a good title?

What makes a good title?

Limits

• Most models should be served in production

• 50ms on prediction • Part of big system, memory limits ~ 10G

Retail

Retail

• Key directions which require machine learning:

• discounting tools • coupons and rewards • loyalty • inventory management

Inventory management

• Customer want to buy products • Customers have diverse needs • Products should be in stock, ideally in

warehouses close to customers • but it’s expensive to store products • Problem: How many products of each type

should be stored, when product supply should be refilled?

Customer intelligence

• Retail • analyze sales data, find anomalies, explain

them • low sales of umbrellas during last month in

North California’s stores • No rains? (integration with external data about

weather conditions) • Seasonal / the same as last year / time series • Competitors

Fraud detection

• identify fraudulent transactions online • Hundreds fraud schemas detected daily • Global retail shrinkage is $119 billion in

2011, an average of 1.45% of retail sales. • from stolen credit card to price tag

replaced, price discounts by high level managers to achieve personal goals

Propensity Modeling for Marketing Campaigns

• build effective email/facebook/google ads campaign addressing proper customer at proper time at proper costs

• behavior based customer segmentation and clusterization with demographics, lifestyle, attitudinal information

Online Grocery

• which items can be replaced by other items and by which items they can be replaced

• data are individual purchases in chain grocery, drug stores, online grocery shopping

• the problem - find which items can be replaced by other item if they are not in store to fulfill the order

Dynamic pricing

• define the best price • scrap continuously prices of competitors,

predict demand by price, know the expenses

• online commerce sites change prices every 10 minutes

Challenges

• Data volumes: transactions: Walmart: 10 Million per day

• Computations: complicated modeling techniques

Hardware platform

• Needs: • Data storage • Data processing • Serving online

Data storage

• Volumes of data: • 10 M transactions per day, 5 years - 18

billion transactions -> 1T • Catalog: 500 M items * 2K per each -> 1T

Data Storage

• but if go to video: petabytes of data, RetailNext 75P per year from 30000+ sensors

• Walmart 500P • eBay 40 P in 2013 (transactions + online

behaviours)

Data processing

• Rebuild model over fresh data: • typically daily: add daily data (millions of

transactions, hundreds of millions of behavior units) to year data store (billions of transactions, hundred billion/trillion behavior units)

• build a model to serve in production the next day

Data processing

• some models such as fraud detection,dynamic pricing should be almost online (10-15 minutes)

• build over data such as daily transactions or web crawl over competitors' sites

Serving online

• online commerce WML - thousands / tens thousands queries per second in peak times

• complicated algorithm of ranking, recommendation,

• 50ms limit

serving online

• price, in store availability - millions requests per second in peak times

• item informations - millions requests per second

• serving online - Solr/Lucene/Elastic Search, Cassandra, MongoDB, Oracle, CouchDB,Node.JS/Java solutions etc

Data processing

• Hadoop / Spark clusters • a lot of I/O • HDFS does the redundancy , RAID is not

necessary, RAID is slow to write, Hadoop writes a lot

• SAN, NAS are not good either • so bare metal with DAS Directly Attached

Storage

Data Processing

• more servers, cheaper servers • more smaller disks is better than large

disks • allocate cluster 100% to Hadoop

Data processing

• Hadoop Masters vs Workers • large clusters: Masters > 64G RAM, dual

Ethernet NIC, dual quad core CPU • Workers: memory 64G+, SAS 6Gb/s disk

controllers, 2 Ethernet cards, 2*6core processors, 15M cache, Intel’s Hyper-Threading and QPI good to have

Data Processing

• big models, deep learning • Nvidia DGX-1 and alike • Pascal GPUs , NVLink interconnect • Tesla k40, K80 work pretty well too • may require a lot of tuning http://

timdettmers.com/2015/03/09/deep-learning-hardware-guide/

• hard to buy: big data solutions are considered profit generators, HPC servers are not

Serving online

• Typically large memory, but not necessary (for example, Elastic Search/Solr degrades over 64G)

• CPUs: more cores rather than faster • Disks: SSD, RAID 0, no NAS, a lot of

conditions frequently optimize wrt how easy to change drivers rather than SSD endurance

ecommerce example

• Database servers • Unified hardware platform : from HP • HP DL line: • 4 cpu sockets • 256 GM RAM • network interfaces • not much HDD, data is in NAS

ecommerce example

• cloud servers: • purchased by racks: 40 in a rack • 2 CPU socket • 198G • 18 core CPU • SSD

network requirements

• 1 network card per server - a big mistake, 1 switch per rack

• 3 cards per servers: • typical three data flows: • production • “administrative” (dockers etc) • analytics

example

• application servers vs big data servers • application servers (java, node.js apps): • 1TB SSD, RAID 5 • Big data servers: • 5T SAS

Questions?

Dr. Andrei Lopatenko Director of Engineering,

Recruit Institute of Technology Recruit Holdings

• andrei@recruit.ai