The Nanotechnology in ecommerce Tao Zhu, @WalmartLabs BDTC2014
The Analogy Nanotechnology ( nanotech ) is the manipulation of matter on an atomic, molecular, and supramolecular scale. - Wikipedia Big Data in ecommerce is the management of business on an individual customer scale.
The Analogy Objective Nanotech in material science To be stronger, lighter, cheaper Big Data in ecommerce Smarter, better customer experience, less cost and more profit Scale Nanometers Individual customer, individual item Method Electronic microscope, Lithography Hadoop, hive Data mining, optimization
Agenda Chapter 1: Atom/molecular interacts with its neighbors, so does our product Bundle recommendation in ecommerce Chapter 2: Manipulate at nano scale Personalized recommendation at individual customer level Chapter 3: Environment/context matters O2O initiatives at Walmart: pickup in store, WMX, savings catcher and mobile
Agenda Chapter 1: Atom/molecular interacts with its neighbors, so does our product Bundle recommendation in ecommerce Chapter 2: Manipulate at nano scale Personalized recommendation at individual customer level Chapter 3: Environment/context matters O2O initiatives at Walmart: pickup in store, WMX, savings catcher and mobile
Ch1: Interactions with neighbors Bundle recommendation in ecommerce
Ch1: Interactions with neighbors Why consider bundles? How to construct bundles (technical) The Bundle Recommendation Problem (BRP) Solving BRP Offline and online experiments The baseline models Relative improvement over baseline models Sensitivity to candidate set size Sensitivity to parameter lambda Online test result
Ch1.1-Why consider bundles Customer buy bundles
Ch1.1-Why consider bundles Customer buy bundles Contextual influence
Ch1.1-Why consider bundles Customer buy bundles Contextual influencer Product compatibility or consistency
Ch1.1-Why consider bundles Customer buy bundles Contextual influencer Product compatibility or consistency Cost saving
Ch1.1-Why consider bundles A real life case study:
Ch1.2-The Bundle Recommendation Problem Let x i be the indicator decision variable where x i =1 denotes item I is chosen and x i =0 denotes notchosen Find x that maximizes the expected total reward rate
Ch1.2-The Bundle Recommendation Problem Approximate the objective with the first two terms, we arrive the Bundle Recommendation Problem (BRP) as follows However BRP is in general NP-hard n is usually on the order of 10 6 We want to solve it for every individual customer
Ch1.2-The Bundle Recommendation Problem
Ch1.3-Online experiment
Agenda Chapter 1: Atom/molecular interacts with its neighbors, so does our product Bundle recommendation in ecommerce Chapter 2: Manipulate at nano scale Personalized recommendation at individual customer level Chapter 3: Environment/context matters O2O initiatives at Walmart: pickup in store, WMX, savings catcher and mobile
Chapter 2: Personalized recommendations Walmart.com Transaction Data Click Stream Data (Omniture and Hubble) InStore Transaction Data Top K recommendations of products or categories Other Labs Feeds (social, pricing, costs, strongmail)
Ch2.1 Matrix Factorization Matrix Factorization or Mainly driven by Netflix Prize Competition (M ~ 500k X 17k) Solve: Algos: Grad descent, Stochastic grad descent, Alternating least squares, SVD (only for median scale problem)
Ch2.1 Matrix Factorization Reality and challenges at Walmart.com no ratings (but action-notaction behaviors, e.g., transactions) Way more users than items Existing algo s have a lot of randomness we have no control of; and the solution is local optimal How to incorporate signals from multiple domains? e.g., browsing behaviors, in-store behaviors, searched keywords, gender, age, geo, etc. Multi-source matrix factorization model Binary representing action or not-action Partial Least Squares Matrix Factorization (PLSMF) algorithm Multi-source model incorporating multiple signals
Ch2.2 PLSMF algorithm The PLSMF algorithm Phase I: Compute Q, where is computed by Phase II: Compute P, for all user i. (this step essentially solves an ordinary least squares) Q only need to be computed once in a while Computing P is embarrassingly parallel and can be done by mapreduce
Ch2.2 PLSMF algorithm Interpretation of PLSMF The Coocurrence Matrix: M T M # of users bought item 3 # of users bought item 2 and item 3
Latent factor 2 Q2[2, pidx] -0.2 0.0 0.2 0.4 Ch2.2 PLSMF algorithm where are from Kung Fu Panda Bridesmaids Flush Away Rio Yogi Bear Gnomeo&Juliet Fast Five Transformers Star Wars Secretariat Inception The King s Speech -0.4-0.2 0.0 0.2 0.4 Q2[1, pidx]
Latent factor 2 Q2[2, pidx] -0.2 0.0 0.2 0.4 Ch2.2 PLSMF algorithm where are from Kung Fu Panda Bridesmaids Flush Away Rio Yogi Bear Gnomeo&Juliet Fast Five Transformers Star Wars Secretariat Inception The King s Speech -0.4-0.2 0.0 0.2 0.4 Q2[1, pidx]
Ch2.3 Multi-source matrix factorization Interpretation of M T M Transaction # of users bought item 3 # of users bought item 2 and item 3
Ch2.3 Multi-source matrix factorization Interpretation of M T M Transaction + browsing Buy and buy View and buy View and buy View and view # of users viewed item 3 # of users viewed item 1 and viewed item 3 # of users bought item 3 and viewed item 3
Ch2.3 Multi-source Interpretation of M T M Transaction + browsing + geographic Buy given buy Buy given view Buy given geo View given buy View given view View given geo Geo given buy Geo given view Geo given geo # of users from region 1 # of users viewed item 2 given from region 1 # of users bought item 2 given from region 1
Chapter 2.3 Multi-source matrix factorization
Ch2.3 Online validation
Agenda Chapter 1: Atom/molecular interacts with its neighbors, so does our product Bundle recommendation in ecommerce Chapter 2: Manipulate at nano scale Personalized recommendation at individual customer level Chapter 3: Environment/context matters O2O initiatives at Walmart: pickup in store, WMX, savings catcher and mobile
Ch3: Seamless experience from online to offline, offline to online Pickup in store Mobile applications Savings Catcher Vendor Marketing
Ch3.1: Pickup in store The last mile battle in US ecommerce 90% of all Americans live within 15 miles of a Walmart
Ch3.2: Mobile Help shoppers in store check price, rating and reviews Find local coupons Express checkout Savings catcher
Ch3.3: Savings Catcher
Ch3.4: Vendor Marketing Online marketing is all about measurement One of every four dollars Americans spend on groceries is spent at Walmart Bigger data with more challenges
The End - Q&A?