Property Recommendations for all Australians September 2016 Glenn Bunker Data Science Manager Ben Kuai Senior Developer - Data
Change the way the world experiences property Change the way the world experiences property Change the way the world experiences property Change the way the world experiences property > 5.9 million unique audience* > Hundreds of thousands of property listings** > 10,000 real estate agencies** > 160 million consumer-property interactions** *Source: Nielsen Digital Ratings (Monthly), July 2016 **Source: REA Internal Data (Buy and Rent), July 2016
Why build a recommendation engine? Passive or Passive or Passive or Passive or Properties to Properties to Properties to Properties to intangible intangible Serendipity Serendipity Serendipity Serendipity intangible intangible people people people people characteristics characteristics characteristics characteristics
Implicit interest ratings + > > > > No explicit No explicit No explicit No explicit Many more Many more Many more Many more More accurate More accurate More accurate More accurate functionality functionality functionality functionality implicit ratings implicit ratings implicit ratings implicit ratings ratings ratings ratings ratings required required required required
6
Collect implicit interest ratings + Consumer Consumer Consumer Consumer information information information information Property Property Property Property information information information information Consumer- Consumer Consumer Consumer - - -property property property property interest rating interest rating interest rating interest rating Consumer Consumer Consumer Consumer events events events events 7
Calculate implicit interest ratings 8
Collaborative Collaborative Collaborative Collaborative filtering filtering filtering filtering
Advantages of collaborative filtering Rewards market Rewards market Rewards market Rewards market Serendipity Serendipity Serendipity Serendipity Data simplicity Data simplicity Data simplicity Data simplicity leading audience leading audience leading audience leading audience
Item-based collaborative filtering process
Item-based collaborative filtering process
Item-based collaborative filtering implementation 13
Item-based collaborative filtering Spark implementation + Column similarity Distributed Distributed Distributed Distributed Consumer Consumer- Consumer Consumer - -property - property property property row matrix row matrix row matrix row matrix Property Property- -property property Property Property - - property property r rating ating RDD RDD r r ating ating RDD RDD similarity RDD similarity RDD similarity RDD similarity RDD Weighted sum 1. 1. 1. 1. Top N Top N Top N Top N 2. 2. 2. 2. prediction prediction prediction prediction 3. 3. 3. 3. 14
Item-based collaborative filtering dataflow Consumer events Property information Prediction & property similarity Consumer information 15
Content Content Content Content- - - -based based based based filtering filtering filtering filtering
Advantages of content-based filtering Understanding Understanding Understanding Understanding Cold Cold Cold- Cold - -start - start start start & trust & trust & trust & trust
Content profiles Property type Property type Property type Property type Location Location Location Location Price Price Price Price Bedrooms Bedrooms Bedrooms Bedrooms Bathrooms Bathrooms Bathrooms Bathrooms
Content-based recommendations • Search by property (more like this) • Search by Search by Search by Search by consumer consumer consumer consumer • Consumer profile with property features • Search property matching given consumer profile 19
Content-based consumer profiles Consumer Consumer Consumer Consumer information information information information Property Property Property Property Consumer Consumer Consumer Consumer information information information information profile profile profile profile Consumer Consumer Consumer Consumer events events events events 20
Content-based dataflow Consumer Consumer profile events Elastic search Property information Indexed property information 21
Blended recommendations Serendipity Cold-start new properties Rewards market leading audience Natural understanding builds trust Data simplicity Differentiated to search experience Cold-start new properties
Blended recommendations Blend Blend API Blend Blend API API API Collaborative filtering API Collaborative filtering API Collaborative filtering API Collaborative filtering API Content Content Content- Content - - -based API based API based API based API CF Predictions Property-property Indexed property Consumer similarity information profile 23
Suggested properties, and so much more… Suggested properties, and so much more… Suggested properties, and so much more… Suggested properties, and so much more…
What worked well Keep it simple Keep it simple Sampling is fine Sampling is fine Keep it simple Keep it simple Sampling is fine Sampling is fine Iterate, test & learn Iterate, test & learn Iterate, test & learn Iterate, test & learn Know the technique Know the technique Know the technique Know the technique Subject matter Subject matter Subject matter Subject matter Bigger picture Bigger picture Bigger picture Bigger picture expertise expertise expertise expertise
Related papers • Item-based Collaborative Filtering Recommendation Algorithms by Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl GroupLens • All-pairs similarity via DIMSUM twitter • Accurate Methods for the Statistics of Surprise and Coincidence by Ted Dunning 26
Thank you Thank you Thank you Thank you Glenn Bunker Glenn Bunker Glenn Bunker Glenn Bunker Ben Kuai Ben Kuai Ben Kuai Ben Kuai https://au.linkedin.com/in/glenn- https://au.linkedin.com/in/glenn -bunker bunker- -13003112 13003112 https://www.linkedin.com/in/ben https://www.linkedin.com/in/ben- -kuai kuai- -7b1aa73 7b1aa73 https://au.linkedin.com/in/glenn https://au.linkedin.com/in/glenn - - bunker bunker - - 13003112 13003112 https://www.linkedin.com/in/ben https://www.linkedin.com/in/ben - - kuai kuai - - 7b1aa73 7b1aa73
Recommend
More recommend