Anyone Can Build A Recommendation Engine With Solr: Presented by Doug Turnbull, OpenSource Connections

OCTOBER 11-‐14, 2016 • BOSTON, MA

Anyone can build a Recsys w/ Solr! Doug Turnbull

Relevance Consultant, OpenSource ConnecIons

I’m now available in book form!

https://www.manning.com/books/relevant-search Discount code: relsearch (38% off)

http://opensourceconnections.com/about-us/doug-turnbull/

Me The company...

field Body term laser doc 2

<metadata> doc 4

<metadata> term light doc 2 <metadata> term lightsaber doc 0

How do search engines work? The answer can be found in your textbook…

OpenSource Connections

Book Index: •  Topics -> page no •  Very efficient tool – compare to

scanning the whole book! Lucene uses an index: •  Tokens => document ids: laser => [2, 4]

light => [2, 5] lightsaber => [0, 1, 5, 7]

What's the point?


Solr:

-  A general purpose system for looking up content based on features that describe them

Tokens aren't really words! doc0: "I like the bananas"

Analysis

Analysis

term I: doc 0 term lik doc 0 term banan: doc 0

[lik] [banan] Search: "liked banana?"

[I] [lik] [banan]

TF*IDF -- measuring feature weight


term I: doc 0: freq: 5 doc 1: freq: 7 doc 3: freq: 4 term banan: doc 0: freq: 2

"Banana-ness" is pretty special

"I-ness" is not special

doc0: tf==5 df==3 (raw) TF*IDF = 5/3 = 1.6667


Search is really distributed feature matching and similarity (text-oriented)

Search often stands in for human interactions

I have a craving for a nice juicy cut of meat. What might you recommend?

I have JUST the thing!

Searching the market q=(juiciness:juicy meatiness:meaty)

Modeling arbitrary feature strength


term juicy: steak: juiciness: 5 grapefruit: juiciness: 7 orange: juiciness: 4 term meaty: burger: meatiness: 2

What you want:

{ item: "steak", juiciness: ["juicy", "juicy", "juicy"], meatiness: ["meaty"] }

Use term frequency as feature strength:

{ item: "grapefruit", juiciness: ["juicy", "juicy", "juicy", "juicy", "juicy"], meatiness: [""] }

(remember, Solr doesn't need to store this)

TF*IDF -- measuring feature weight


term juicy: doc 0: freq: 5 doc 1: freq: 7 doc 3: freq: 4 term meaty: doc 0: freq: 2

"meaty-ness" is pretty special

"juicy-ness" is pretty non-special



Search is really distributed feature matching and similarity

Requesting something from my grocer

More juicy Less juicy

More meaty Less meaty

q=meaty juicy

Results: 1.

2.

3.

Recsys also stands in for human interactions Hi Jane, Recommend me something?

Hmm… <Tom likes limes, what is similar to limes?>

"Content Based" recommendations

Use existing properties of thing to recommend similar things

juicy

citrus

More like this for unstructured data

What features/tokens are most representative of this thing?

http://solr.quepid.com/solr/tmdb/select?q={!mlt%20qf=overview}97&fl=title,id,overview (movies like Tron)

juicy

citrus (search)

Here's some ideas...

{ item: "lime", juiciness: ["juicy", "juicy", "juicy"], citrusness: ["citrus", "citrus", "citrus"], meatiness: [""], partyness: ["party"] }

"Content Based" more-like-these

Use existing properties of thing to recommend similar things

juicy

meaty citrus



Jane knows a few more things that Tom likes...

Personalization metadata

Index extra data alongside your products { item: "hamburger", preferred_by_genders: ["m", …], preferred_by_ages: ["30_40"] }

age:30_40

gender:m



Jane knows a few things about Tom (30 yr old male)

But, Jane's intuition transcends words!

age:30_40

gender:m

Currently we're stuck with predefined labels:

citrus juicy

meaty

We're curating using known vocabularies

(can we describe everything?)

What we like often transcends words There are emergent properties of our world that don't have names

Relative flarglewharbliness

More flarglewharbily Less flarglewharbily

Diet Coke

What's a flarglewharble?


fruit orange lemon banana mentos diet coke

tom X

sue X X X

charlie X X

clare X X

hal x x

Goes together

Diet Coke

Can search find the flargles?

q=(flargliwharbliness:very)

term flarglewharble: diet-‐coke: flargleness: 4 mentos: flargleness: 3 banana flargleness: 1

Can we somehow build?

Diet Coke

person\food orange lemon banana mentos diet coke

tom X X

sue X X X X

charlie X X

clare X X

hal x x X

Goes together

flarglewharble!

Babies often use made-up words based on emergent patterns in their universe They are less committed to our language

What's the point? Collaborative filtering

Latent vocabulary (the flarglewharbles)

Pure Search Content-based Recs

Predefined vocabulary

Can Solr discover the latent/emergent vocabularies?

Can Solr discover the latent/emergent vocabularies? Well first let's tell Solr about our users

{ user: "Sue" foods_bought: ["lemon", "banana", "mentos", "diet coke"] } { user: "Charlie" foods_bought: ["banana", "mentos", "diet coke"] }

Faceting? We need a way to look across users and look for patterns (analyze all the baskets that contain mentos) q=foods_bought:mentos&facet=true&facet.field=foods_bought

facets: mentos: 3 diet-coke: 3 banana: 2

Hmm: -  Bananas are globally popular -  Diet-coke is probably what matters

Counts don't work: importance of significance q=foods_bought:mentos&facet=true&facet.field=foods_bought

facets: mentos: 3 diet-coke: 3 banana: 2

Diet Coke: Global popularity: diet coke (3) Local popularity: 3 Score: 3/3 = 1

Banana: Global popularity: banana

(4) Local popularity: 2

Score: 2/4 = 0.5

by-significance: diet-coke: 1 banana: 0.5

Streaming Expressions

/select?q=*:*&facet=true&facet.field=liked_movies

But there's a new sheriff in town!

One option: we could go about and gather global doc freqs & compare those locally.

Terms component another option… plugins...

Streaming expressions -- distributed stream computation system on top of Solr Cloud You must ALWAYS cross the streams!

Streaming Expressions /stream?expr=scoreNodes(facet(...)...)

facet(movielens, q="*:*", buckets="liked_movies", bucketSorts="count(*) desc", bucketSizeLimit="100", count(*))

Faceting with Streaming Expressions:

Output:

{ "result-set": {"docs":[ { "count(*)":55807, "liked_movies":"318"}, { "count(*)":52352, "liked_movies":"296"}, { "count(*)":50114, "liked_movies":"593"}

Nodes to be transformed

Significance with streaming expr /stream?expr=scoreNodes(facet(...)...)

scoreNodes( select( facet(movielens, q="liked_movies:2571 OR liked_movies:4993", buckets="liked_movies", bucketSorts="count(*) desc", bucketSizeLimit="100", count(*)), liked_movies as node, "count(*)", replace(collection, null, withValue=movielens), replace(field, null, withValue=liked_movies)) )

1.  facet (just like above, just with streaming expr) 2.  select to format data for scoreNodes 3.  scoreNodes to score using TF*IDF

Banana occurs in 2 documents here, 4 globally -- 2/4 = 0.5 Diet coke occurs 2 documents here, 2 globally -- 2/2 = 1.0

Thinking back on my shoppers behaviors, here's some other items you might like:

(thanks Joel Bernstein!)

Diet Coke

Lots of power here /stream?expr=scoreNodes(facet(...)...)

scoreNodes( select( facet(movielens, q="juiciness_pref:juicy", buckets="liked_movies", bucketSorts="count(*) desc", bucketSizeLimit="100", count(*)), liked_movies as node, "count(*)", replace(collection, null, withValue=movielens), replace(field, null, withValue=liked_movies)) )

Find users that like juicy things, what do they like? Perhaps bucket over the aisle they like? Construct our query to focus on a date range? Many insights (thanks Joel Bernstein!)

Only glimpsing the underlying pattern...

We're not enumerating the flarglewharbles, and the schlumblefumbles


Diet Coke

More schlumblewumbly Less schumblewumbly

Diet Coke

Coming soon (Solr 6.3) http://yonik.com/solr-6-3/ https://issues.apache.org/jira/browse/SOLR-9258

-  Models for training classifiers -  Then in turn updating documents Progress is being made! -  Clustering?

Questions? The Flarglewharbles

Anyone Can Build A Recommendation Engine With Solr: Presented by Doug Turnbull, OpenSource Connections

Technology