Combine machine learning with search using the LTR plugin for elasticsearch.

1. LEARNING TO RANK
SEARCH RESULTS
Jettro Coenradie &
Byron Voorbach
#JFALL
COMBINE MACHINE LEARNING WITH SEARCH

2. Neptune

3. UranusSaturnNeptuneJupiter Mars VenusEarth Mercury
UranusSaturn NeptuneJupiter MarsVenusEarth Mercury
https://solarsystem.nasa.gov/resources/all/

4. LEARNING TO RANK SEARCH RESULTS
COMBINE MACHINE LEARNING WITH SEARCH

5. Jettro Coenradie
@jettroCoenradie
https://www.linkedin.com/in/jettro/
https://github.com/jettro
• Fellow at Luminis Amsterdam
• specialised in (Elastic) search
• experimenting with Machine Learning

6. Byron Voorbach
@byronvoorbach
https://www.linkedin.com/in/byronvoorbach/
https://github.com/byronvoorbach
• Search & Data Engineer at Luminis
Amsterdam
• building and optimising search engines

7. Search and Ranking
in eCommerce

12. Order Matters

13. DEMO

14. How do we get from
‘Call of duty’ to a list of
games?

15. Meet Elasticsearch

16. Content Pipeline
elastic
search
Document
Store
Inverted
Index
Analyse
{
query
}
Match

17. Recap: Inverted Index
Terms doc_ids ttf
fifa 1 1
call 2 1
of 2 1
duty 2 1
god 3 1
war 3 1
pes 4 1
doodle 5 1
Doc Id Title
1 Fifa
2 Call of Duty
3 God of War
4 PES
5 Doodle God
2,3 2
3,5 2

18. {
"title": “Call of Duty®: Black Ops 4",
"image": "rs-137178-883f5fe955b2745cd539.jpg",
"description": "<p>Digital Standard Edition includes: - 1,100
Call of Duty® Points* - Digital Edition Bonus Items: --
Specialist Outfit for all Specialists -- Gesture -- Calling Card,
Emblem, Sticker and Tag inspired by the iconic Call of Duty®:
Black Ops 4 skull. Black Ops is back! Featuring gritty,
grounded, fluid Multiplayer combat, the biggest Zombies
offering ever with three full undead adventures at launch, and
Blackout, where the universe of Black Ops comes to life in
one massive battle royale experience.</p>",
"rating": 4.5,
"numberOfRatings": 279,
"vendor": "Activision",
"price": 59.99,
"releaseDate": “2018-12-10“,
"id": 350640
}

19. curl -XGET "http://localhost:9200/ecommproduct/_search"
-H 'Content-Type: application/json' -d'
{
"query": {
"multi_match": {
"query": "call of duty",
"fields": [
"title",
"description"
]
}
}
}'

20. {
query
}
1
3
20
211
410
Matching
410
211
1
20
3
Ranking

21. Ranking with BM25

22. Change Ranking
of results
{
"query": {
"bool": {
"should": [
{
"match": {
“title": “call of duty",
"boost": 2
}
},
{
"match": {
“description”: “call of duty"
}
}
]
}
}
}
Field centric boosting

23. Change Ranking
of results
Function score: field value
{
"query": {
"function_score": {
"query": {
"match": {
“title”: “call of duty"
}
},
"functions": [
{
"field_value_factor": {
"field": “numberOfRatings",
"modifier": "log1p"
}
}
]
}
}
}'

24. Change Ranking
of results
Function score: decay
{
"query": {
"function_score": {
"query": {
"match": {
“title: “call of duty"
}
},
"gauss": {
“releaseDate": {
"scale": "1y",
"offset": "6m",
"decay": 0.5
}
}
}
}
}

25. Learning To Rank
Learning to rank or machine-learned ranking is the application of machine
learning, typically supervised, semi-supervised or reinforcement learning, in the
construction of ranking models for information retrieval systems.
~ Wikipedia
http://bit.ly/ltr-wp

26. Machine LearningSupervised
4 rooms
100 m2 € 200k
6 rooms
150 m2 € 350k
3 rooms
500 m2 € 750k
Input Labelled
Output
Model
Train
Predicted
€ 210k
€ 370k
€ 800k
Error

27. Learning to rank
X
Query 1
Query 2
Query 3
Query 4
Model
f(X)
5 15 67 3 17
23 3 7 88 45
3 27 25 23 5
6 99 22 27 33
YY’
5 3 67 15 17
23 3 17 8 45
3 27 25 6 15
6 9 32 27 33
Predict Error

28. Model Evaluation Types (Errors)
• Binary relevance (MAP, Precision)
• Graded relevance, position based (DCG, NDCG)
• Only discounts based on relevance
• Graded relevance, cascade based (ERR)
• Discounts based on user interaction with the results
http://bit.ly/eval-metric

29. MAP using Binary relevance
23
9
88
33
45
YI Average Precision
1
0.67
0.6
fifa
X
relevant
not relevant
= (1 + 0.67 + 0.6) / 3
= 0.76
MAP

30. DCG using Graded relevance
23 rel=3
9 rel=2
88 rel=4
33 rel=1
45 rel=0
YI
Discounted
Cumulative
Gain
3
5 + 4/log2(3) = 7.57
8.07
fifa
X
Documents ranked on 0 - 4 relevance scale
3 + 2/log2(2) = 5
7.57 + 1/log2(4) = 8.07

31. NDCG using Graded relevance
23 rel=3
9 rel=2
88 rel=4
33 rel=1
45 rel=0
YI DCG
3
7.57
8.07
fifa
X
Documents ranked on 0 - 4 relevance scale
5
8.07
23 rel=3
9 rel=2
88 rel=4
33 rel=1
45 rel=0
Y MaxDCG
4
8.28
8.77
7
8.77
NDCG
0.75
0.91
0.92
0.71
0.92

32. Learning to rank - Model
X
Model
Y
Y’
Algorithm
Parameters
Cost function

33. LTR: Steps to take
1. Create Judgement List (Ground Truth)
2. Define features for the model
3. Log features during usage
4. Training and testing the model
5. Deploying and using the model
6. Feedback loop

34. 1. Judgement List
Obtain labelled data to train the model

35. Judgement List: Expert Panel
• Time consuming and Expensive
• Error prone due to different judgements
http://bit.ly/ebay-hum-judge

36. Experiment: Pet or Not

37. Pet or Not

38. Pet or Not

39. Pet or Not

40. Judgement List: Implicit Feedback
• Log user behaviour
• Compare actual clicks versus expected clicks
• A click is not a relevance judgement

41. Judgement List: Implicit Feedback
• Use as a signal to the ranking algorithm -> Feature
• Use as Label to train the model -> Ground truth

42. Using the LTR Plugin and the
python scripts
https://github.com/o19s/elasticsearch-learning-to-rank

43. Our Judgement List
# grade (0-4) queryid docId title
#
# Add your keyword strings below, the feature script will
# Use them to populate your query templates
#
# qid:1: fifa
# qid:2: football
# qid:3: call of duty
# qid:4: marvel
# qid:5: basketball
# qid:6: god
#
4 qid:1 # 1538781503000 FIFA 18
2 qid:1 # 1536187840000 EA SPORTS FIFA 16
3 qid:1 # 1538107776000 FIFA 19 Ultimate Edition
4 qid:1 # 1538694141000 FIFA 19
3 qid:1 # 1536293937000 EA SPORTS FIFA 17 Standard Edition
1 qid:1 # 1536022097000 FIFA 15
3 qid:2 # 1538509479000 PRO EVOLUTION SOCCER 2018
4 qid:2 # 1535257488000 Pro Evolution Soccer 2018 FC Barcelona Edition
1 qid:2 # 1536293937000 EA SPORTS FIFA 17 Standard Edition
1 qid:2 # 1538781636000 2MD: VR Football

44. 2. features for the model
• Raw Term Statistics
• Document Frequency
• Total Term Frequency
• Also max, min, sum (in case of multiple terms, fields)
• Elasticsearch queries

45. 2. features for the model
{
"query": {
"match": {
"title": "{{keywords}}"
}
}
}
{
"query": {
"match": {
"description”: "{{keywords}}"
}
}
}

46. {
"query": {
"function_score": {
"functions": [
{
"field_value_factor": {
"field": "rating",
"missing": 0
}
}
]
}
}
}

47. {
"query": {
"function_score": {
"gauss": {
"releaseDate": {
"scale": "720d",
"decay": 0.25
}
}
}
}
}

48. {
"query": {
"nested": {
"path": "basket",
"query": {
"function_score": {
"query": {
"term": {
"basket.term": {
"value": "{{keywords}}"
}
}
},
"functions": [
{
"field_value_factor": {
"field": "basket.clicks",
"modifier": "log1p"
}
}
]
}
}
}
}
}

49. {
"query": {
"nested": {
"path": "clicks",
"query": {
"function_score": {
"query": {
"term": {
"clicks.term": {
"value": "{{keywords}}"
}
}
},
"functions": [
{
"field_value_factor": {
"field": "clicks.clicks",
"modifier": "log1p"
}
}
]
}
}
}
}
}

50. 3. Log features
2 qid:1 1:7.7917995 2:10.646265 3:4.5 4:0.99658567 5:1.0154246 6:1.475652 # 1538694141000 fifa
0 qid:1 1:7.7917995 2:12.036756 3:4.0 4:0.6522283 5:0.0 6:0.0 # 1538781503000 fifa
2 qid:1 1:5.3206625 2:10.542777 3:4.5 4:0.20758471 5:0.0 6:0.0 # 1536293937000 fifa
2 qid:1 1:7.7917995 2:9.694633 3:4.5 4:0.00240296 5:0.0 6:0.0 # 1536022097000 fifa
2 qid:1 1:6.3233795 2:11.195704 3:4.5 4:0.03137532 5:0.0 6:0.0 # 1536187840000 fifa
4 qid:1 1:6.3233795 2:11.144586 3:0.0 4:0.99658567 5:1.0154246 6:1.475652 # 1538107776000 fifa
4 qid:2 1:0.0 2:6.58323 3:4.5 4:0.99142087 5:1.0787467 6:1.1217693 # 1538101584000 basketball
0 qid:2 1:0.0 2:5.237752 3:4.0 4:0.99915665 5:0.0 6:0.0 # 1538152581000 basketball
1 qid:2 1:0.0 2:5.9230056 3:4.0 4:0.9891866 5:0.8548865 6:0.8889811 # 1538360829000 basketball
1 qid:2 1:0.0 2:4.897768 3:5.0 4:0.19836442 5:0.0 6:0.0 # 1537838613000 basketball
0 qid:2 1:0.0 2:5.237752 3:4.0 4:0.6326628 5:0.0 6:0.0 # 1537838447000 basketball
1 qid:2 1:0.0 2:5.3618174 3:3.0 4:0.259035 5:0.0 6:0.0 # 1538781628000 basketball
1 qid:2 1:0.0 2:7.574521 3:5.0 4:0.7088474 5:0.0 6:0.0 # 1535173832000 basketball
0 qid:2 1:0.0 2:5.237752 3:4.5 4:0.9410281 5:0.0 6:0.0 # 1538518103000 basketball
1 qid:2 1:0.0 2:7.3292704 3:4.5 4:0.99142087 5:1.0787467 6:1.1217693 # 1538619230000 basketball
1 qid:2 1:0.0 2:9.499237 3:4.0 4:0.27305022 5:0.0 6:0.0 # 1536946885000 basketball
1 qid:3 1:5.0056663 2:7.690156 3:5.0 4:0.18467863 5:0.0 6:0.0 # 1538509528000 god
1 qid:3 1:4.608784 2:7.9265423 3:5.0 4:0.9016471 5:0.5393733 6:0.8889811 # 1536297114000 god
1 qid:3 1:4.2702136 2:8.228076 3:5.0 4:0.34165436 5:0.0 6:0.0 # 1535775896000 god
1 qid:3 1:0.0 2:7.043888 3:0.0 4:0.7977039 5:0.0 6:0.0 # 1537244713000 god

51. 4. Train and test model
• Making use of Ranklib
• Can specify separate train, validation and test set
• Can normalise feature sets

52. Models using Ranklib
MART
Multiple Additive Regression Trees, a gradient boosting machine. Can be used for
regression as well as classification.
RankNet
Compare two feature vectors using stochastic gradient descent with the help of a cost
function.
RankBoost
Based on AdaBoost, combining many weak rankings into a single highly accurate ranking.
Is pairwise comparison
AdaRank
Combines a number of weak learners in a linear way. Builds on AdaBoost, but directed
more at ranking.
Coordinate Ascent
Optimises one parameter at a time, keeping the other constant. Done iteratively until
some convergence criteria is met.
LambdaRank
Optimisation of RankNet that only looks at the gradients represented by arrows
indicating how much they need to move up or down
LambdaMART Combines using the gradients of LambdaRank and the use of MART
ListNet
Uses a list wise loss function, a neural network and gradient descent. Similar to RankNet,
only difference is List versus Pair loss functions.
Random Forests
Number of trees to vote for the most popular class for a vector of features. One tree
would not be better than a random choice, but a forest is
Linear Regression
Most of the times to simplistic for the learning to rank problem with lots of features, but
good to have available to at least try

53. Evaluation metrics
MAP Mean Average Precision:The average of all P@k
DCG@k Discounted Cumulative Gain:Add relevance of all documents discounted
by the position of the document making the 1st document more important
NDCG@k Normalised DCG:A DCG with a value between 0 and 1, normalised by the
highest score.
P@k Percentage of relevant documents of this top K
RR@k Reciprocal Rank: 1/K where K is the first relevant document
ERR@k Expected Reciprocal Rank: discounts documents that are below a very
relevant document.

54. Training results
MAP NDCG@10 DCG@10 P@10 RR@10 ERR@10
MART 0.999 0.979 38.7047 0.9333 1.0 0.9618
RankNet 0.8806 0.8831 33.298 0.8933 0.9 0.5988
RankBoost 1.0 1.0 39.8699 0.9333 1.0 0.9629
AdaRank (List) 0.9493 0.8271 31.4686 0.8733 0.8 0.7597
Coordinate Ascent 0.9446 0.8973 36.3888 0.9133 1.0 0.96
LambdaRank 0.9562 0.7079 35.7277 0.8933 0.9 0.7598
LambdaMART 0.9725 0.9805 38.788 0.9333 1.0 0.9818
ListNet (List) 0.9496 0.8973 33.1995 0.8933 1.0 0.8229
Random Forests 0.999 0.979 38.7047 0.9333 1.0 0.9618
Linear Regression 0.9463 0.8602 34.1433 0.8933 0.9 0.7906

55. 5. Deploy the model
• The model including learned parameters is stored in elasticsearch.
• Using the plugin we can now re-rank the top-n results

56. GET ecommproduct/_search
{
"query": {
"multi_match": {
"query": “call of duty",
"fields": ["title", "description"]
}
},
"rescore": {
"window_size": 100,
"query": {
"rescore_query": {
"sltr": {
"params": {"keywords": “call of duty"},
"model": "test_6"
}
}
}
}
}

57. DEMO

58. 6. Feedback Loop
• Register clicks by users (and other events)
• Use click data for predicting labels

59. Click models
• Random Click Model -> Every document has the same chance of being
clicked
• Click Through Rate Model -> Uses the fact that the first document is
clicked far more than the second document
• Cascade Model -> A click in the third item also tells us something about
the first two items. Only one click per session is assumed.
• Dynamic Bayesian Network Model -> Supports multiple clicks in a search
session and the difference in actual relevance of a document.
http://bit.ly/click-model

60. Dynamic Bayesian Network
Ei-1 Ei Ei+1
Ci
Ai Si
au su
http://bit.ly/dbn-clickmodel
Ei - Did the user examine the url
Ai - Was the user attracted by the url
Ci - Did the user click the url
Si - Was the user satisfied with the landing page
au - Probability of being attracted by the url
su - Probability of being satisfied by landing page

61. Logs
Queries
Clicks
Learning
data
varepsilon
Judgments
Webshop

63. Query id 82369cad-7ca3-4b69-ba8c-2eaf16446e55
Query Text call of duty
Document id’s
["1538784192000","1537838838000","1538799761000","1
538800073000","1536295676000","1536295271000","153
6296646000","1536292134000","1536292118000","15362
95845000"]
Clicks [1,0,0,0,1,0,0,0,1,0]

64. Python
…
fifa:1538694141000:2
fifa:1538781503000:0
fifa:1536293937000:2
fifa:1536022097000:2
fifa:1536187840000:2
fifa:1538107776000:4
call of duty:1536295845000:3
call of duty:1537838838000:2
call of duty:1538784192000:4
call of duty:1536295271000:2
call of duty:1538799761000:2
call of duty:1538800073000:2
call of duty:1536296646000:2
call of duty:1536295676000:3
call of duty:1536292134000:4
call of duty:1536292118000:4
…
# qid:1: fifa
# qid:2: basketball
# qid:3: god
# qid:4: football
# qid:5: call of duty
# qid:6: marvel
#
#
2 qid:1 # 1538694141000
0 qid:1 # 1538781503000
2 qid:1 # 1536293937000
2 qid:1 # 1536022097000
2 qid:1 # 1536187840000
4 qid:1 # 1538107776000
4 qid:2 # 1538101584000
0 qid:2 # 1538152581000
1 qid:2 # 1538360829000
1 qid:2 # 1537838613000
0 qid:2 # 1537838447000
1 qid:2 # 1538781628000
1 qid:2 # 1535173832000
0 qid:2 # 1538518103000

65. DEMO

66. Questions
https://www.elastic.co
https://github.com/o19s/elasticsearch-learning-to-rank
https://github.com/varepsilon/clickmodels
https://www.linkedin.com/
in/byronvoorbach/
https://www.linkedin.com/
in/jettro/

67. “Please rate our
talk in the official
J-FALL app”
#JFALL