Share
Embed
The Adaptive Commerce Experience With Machine Learning APIs
By Jason Lobel @ API Strategy & Practice Conference
San Francisco, October 23-24-25, 2013
API Strategy & Practice Conference
October 24, 2013
Transcript
29 slides
-
About Us Swift Access Data backend-as-a-service Import, store,
access and analyze data on-demand Power digital experiences and intelligence Swift Predictions Machine-learning-as-a-service Harness a catalog of adaptive algorithms Embed predictive APIs to make your apps smarter -
Why Do Marketers Generally Fail to Satisfy Us? Disparate Data
Silos Conflict in Team Goals Lack of Machine Readable Data (APIs) -
<2000 2000’s 2011 2012 2013 & Beyond Shopping Touchpoint In-store
Online Mobile Social Visual discovery Tablets & Mobile POS 50 Billion Internet- Connected Devices Commerce Is Complex and Fast Changing -
Why ML? Optimize Relevance Continuously Analyze Massive Data Volume, Velocity,
Variety Models are Not Static – User Purchase Habits Change Over Time Testing! Variables (e.g., Weather) Affect Businesses Differently Infrastructure for Large-Scale Distributed Computing is Available Open Source and Paid ML Algorithms Are Available Software Can Solve These Challenges Machine Learning Benefits Why Now -
Why APIs for Machine Learning Algorithms? Hard Eas{ier} Human
Finding a data scientist Technical Database selection Algorithm(s) selection Model training & iteration Embedding predictions into applications Security Query speed / caching Scaling On-Demand Access Human Finding an engineer that can use an API Training (if needed) Technical -
Why APIs? Interactive Big Data Visualization D3.js (d3js.org) JavaScript
library for manipulating documents using HTML, SVG and CSS -
Useful Machine Learning Concepts Recommendation: Analyzes users' preferences and
finds items users might like Frequent Pattern Mining: Discovers unique frequently co-occurring items in a transaction list Classification: Learns from existing categorized data and assigns a category to uncategorized data Clustering: Organizes items from a large volume of data into groups of similar items and features -
Common ML Applications for Commerce Item Recommendation: observes what
the user likes and finds similar items (“I like the Chicago Bulls, I may like the Chicago Bears”) User Recommendation: recommend items finding similar users and sees what they like (e.g., Kin and I are friends. He likes IPAs. I may like IPAs) Product Affinity: if X user wants X, what else is Y user likely to want based on the relationship between X and Y (men who buy diapers, also tend to buy beer) Predict Inventory: based on history, predict future sales (next 7, 30 days, etc.) Discover Customer Segments: examine purchasing habits to identify clusters of shopper segments Prevent Fraud: identify anomalies in cashier activity, such as voids (is this likely fraud? yes/no) -
Where Do I Collect Data? Web: JavaScript Tag
Product data: export catalog, affiliate feed, etc CRM: export Segmentation: third-party sources (e.g., Acxiom) Web Tag CRM Web Tag / Catalog Segmentation -
Item Recommendation Generate Model {Inputs} 1. Gather data Order
history (user ID, past purchases) Web behavior (page views, adds to cart) Segmentation (age, income) Search path (keywords, items bought) 2. Run algorithm 3. Query User ID Response Returns a score for items that are deemed relevant based on item/score pair -
Recommendations In Use (Amazon) Amazon’s product pages includes many
behavioral-based suggestions -
No Single Algorithm is a Magic Bullet ? 1 2
3 4 5 6 7 8 -
Frequent Pattern Mining Model Inputs Order history (POS data)
1 = Yes 0 = No Understand affinity between milk, bread, butter & beer Understand affinity between products (e.g., grocery localization) -
Frequent Pattern Mining API Response Pattern Count [milk, bread],1
[butter],1 [beer],1 [milk, bread, butter],1 [milk, bread, beer],1 [milk, bread],1 Frequency List :[(bread,4), (milk,4), (beer,2), (butter,2)] Number of unique items 4 Patterns found with… • Butter: 2 Patterns, butter:[butter]:2 [bread,butter,milk]:1 Bread: 3 Patterns, bread:[bread,milk]:4 [bread,butter,milk]:1 [beer,bread,milk]:1 -
Classification Predict inventory (continuous variables) Match search term
to a SKU (not continuous) Model Training {Inputs} Inventory: date, purchase history, other (offers) Evaluation: # instances, mean squared error Query a SKU Search: keywords, adds to cart, orders Evaluation: # instances, accuracy score (0.80) Query a keyword or feature (e.g., area rugs) Response 6.686672 -
Clustering: Discovering Shopper Segments Inputs Order history(average basket, purchases)
Set # of clusters? Set # of iterations? API Response Links to a file with evaluation tools (next page) -
How Would a Machine Classify These Guys? Differences Similarities
Visual appearance: Casual Job: Entrepreneur IP Address: San Francisco Buys: T-Shirts & Jeans Name Male Frequent traveler Love APIs Love IPAs Browses: tech sites Visual appearance: Formal Job: Public employee IP Address: Washington DC Buys: Suits & Tie Browses: Gov websites -
Compete With Adaptive Intelligence Commerce must be adaptive Commerce must
be relevant always in the now Adaptive Intelligence v1.0 = machine readable data (APIs) / collaboration v2.0 = automate decisions with machine learning