big wins with small data: predictionio in ecommerce

Post on 29-Jan-2018

2.409 Views

Category:

Software

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Big Wins with Small Data:PredictionIO in Ecommerceby David Jones (@d_jones) - Technical Director - Resolve Digital

Comic by @d_jones. Inspired by xkcd.

Machine learning & Ecommerce

better. together.

Predictive APIsmake it feel like you have your own

team of ML experts

Personalised recommendations

Abandoned cart emails

Lead scoring

Email marketing

Benefits of ML in EcommerceMore visitors with longer session duration

Improved retentionHigher conversion rates

Larger order sizeBottom line, more revenue.

machine learning in ecommercedeeply enhances customer experiences

Customers expect ML featuresEven if they don’t know what ML is

Real world example:United Cellars

Goal: Increase revenueSolution: Product recommendationsImplementation: Predictive APIs

16k product views60k orders

3k product ratings

79000 rowsSmall data

Any data that linkscustomers to products

Optimise data collectionInitially had no data for logged out users

PredictionIO

Open source machine learning server

Ecommerce templatesEvaluation metrics

Production ready

Why PredictionIO?

Active Open Source project

Productivity focusedBuilt-in support for Apache Spark MLlib

Opinionated DASE architecture:

Data Source and Preparator, Algorithm, Serving, Evaluator

Storefront <-> Processing Queue <-> PredictionIO

PAPI Production TipsMonitor performance

Cache where realtime is not neededDerived data is low risk

Bricks & mortar retailinspires an ecommerce experience with ML

With winetaste preferences are diversepeople usually prefer red or white

A/B TESTwith Google Analytics

Filter PredictionsAvoid anything that lessens relevanceDon’t show out of stock products

A/B Test Results45% longer average session22% increase in conversion rate37% increase in average order size

71%More Revenue

Where to next?Predictive features in more channelsIntroduce new predictive features

Conclusion

1Small data can be enough

Quality is key

2Predictive APIs

are awesome

3Machine Learning

is highly effective withEcommerce

resolve.digital/papis2015David Jones, @d_jones

top related