SLIDE 1
Collaborative Search
Haystack 28.10.2019
Sadat Anwar and Matthieu Pons
We are...
- Matthieu Pons
○ Backend engineer at reBuy
- Sadat Anwar
○ Backend Search Engineer (ex- reBuy) ○ Search engineer Delivery Hero
2
Collaborative Search Haystack 28.10.2019 Sadat Anwar and Matthieu - - PowerPoint PPT Presentation
Collaborative Search Haystack 28.10.2019 Sadat Anwar and Matthieu - - PowerPoint PPT Presentation
Collaborative Search Haystack 28.10.2019 Sadat Anwar and Matthieu Pons We are... Matthieu Pons Backend engineer at reBuy Sadat Anwar Backend Search Engineer (ex- reBuy) Search engineer Delivery Hero 2 1. Context
SLIDE 2
SLIDE 3
- 1. Context
- 2. Sequential Rules Method
- 3. Applications
SLIDE 4
reBuy
- re-commerce shop
- Media and consumer
electronics
- Volatile availability
4 4
SLIDE 5
Context
- Replace our 3rd party
provider
- Very little product reviews
- Anonymous user sessions
- Difgerent type of
recommendations
5 5
SLIDE 6
Context
- First tried Product/User embeddings
○ Satisfying results ○ Problems with implicit and sparse signals
- Moved on to Prod2Vec with gensim
○ Worked much better ○ No more problem with implicit signals
Product User ... 6 6
SLIDE 7
Problem
- Prod2Vec is “too good” for
- ur add-to-cart sessions
- Comes at a cost (time, GPUs,
- ps)
- Sequence based is the right
track
7 7
SLIDE 8
Sequential Rules Method - Intuition
A -> R -> F -> C -> G -> E -> H A -> C -> H -> G -> E -> D -> I
8 8
1st order Markov chain: p(G|C) = .5 p(H|C) = .5
SLIDE 9
Sequential Rules Method - Intuition
9 9
A -> R -> F -> C -> G -> E -> H A -> C -> H -> G -> E -> D -> I C: [A, H, G, E, R, D, I] Simple Co-occurrences:
SLIDE 10
Sequential Rules Method - Intuition
10
SLIDE 11
Sequential Rules Method - Intuition
5 4 3 5 4 3 2 1
A -> R -> F -> C -> G -> E -> H A -> C -> H -> G -> E -> D -> I
11 11
C: [G: 9, H: 8, E: 7, D: 2, I: 1] Sequential Rules:
SLIDE 12
Sequential Rules - results at rebuy
Hit Rate on add to cart sessions:
HR @5 HR @10 Provider .230 .359 Prod2Vec .167 .241 Sequential Rules .245 .372
12 12
SLIDE 13
Sequential Rules - improvements
Using a SR on itself: A: [R, F,..., C, Q] B: [G, H,..., C, E] D: [M,..., C, K, N] D: [M, C, R,..., G]
Original recommendations for D D re-arranged, based on session Recommendations for one user session
13 13
SLIDE 14
Sequential Rules - improvements
Results of using a SR on itself:
HR @5 HR @10 Provider .230 .359 Prod2Vec .167 .210 SR - simple .245 .372 SR - many2many .257 .416
14 14
SLIDE 15
Sequential Rules - pros and cons
Pros:
- Explainable
- Only one pass through data 0(n)
○ Fast training (~10 min. for 10M sessions)
- Can be written as a map-reduce
- Versatile
15 15
SLIDE 16
Sequential Rules - pros and cons
Cons:
- Overfit on small datasets
- No seasonality
- Not good at replaying sequences
A -> R -> F -> C -> Q A -> G -> H -> C -> E A -> L -> M -> C -> S A: [C: 9, ...]
16 16
SLIDE 17
Sequential Rules - original research
- M. Ludewige, D. Jannach: Evaluation of Session-based Recommendation
Algorithms, Oct.2018, https://arxiv.org/pdf/1803.09587.pdf I.Kamehkhosh, M. Ludewige, D. Jannach: A Comparison of Frequent Pattern Techniques and a Deep Learning Method for Session-Based Recommendation, 2017, http://ceur-ws.org/Vol-1922/paper10.pdf
17 17
SLIDE 18
SR method - original research
18
- M. Ludewige, D. Jannach: Evaluation of Session-based Recommendation Algorithms, 5.1.1 Table 3
SLIDE 19
Search Applications
SLIDE 20
Search Applications
1. (Collaborative) Spell Check 2. Similar Search 3. Hybrid Search and Recommendation
20 20
SLIDE 21
Status Quo
- Bad query and results
- Ambiguous situation
- Unsatisfying customer
experience
21
SLIDE 22
Status Quo
- Bad query and results
- Incorrect spellcheck
- Unsatisfying customer
experience
22
SLIDE 23
Collaborative Spell Check
- Solr/Elasticsearch only do text spell check
- min-prefix is too constrictive
○
mintendo -> nintendo
- Word2Vec clustered spelling errors
23 23
SLIDE 24
24
Collaborative Spell Check
SLIDE 25
How to use SR for search
huwai p20 pro -> angebot handy -> handy angebot -> huawei -> huawei p20 pro huwai p20 pro -> huwai -> huawei 5 4 3 2 5 4 huwai p20 pro -> [huawei: 7, ...] Would give
25
SLIDE 26
- Applied on enough
sessions
- Works on long queries
- Correct error at first
position
26
Collaborative Spell Check
SLIDE 27
Status Quo
- Correct query and results
- Disappointing from a
customer point of view
- Dead-end
27
SLIDE 28
Similar searches
28 28
SLIDE 29
29
SLIDE 30
30
SLIDE 31
Similar Search
- Very little data preprocessing
needed
- Increase user engagement
- How to difgerentiate between
spellcheck and similar search?
31
SLIDE 32
Status Quo
- Correct query and results
- Wasted real estate
- Looks broken
- Unsatisfying customer
experience
32
SLIDE 33
Hybrid Search and Recommendation
Nash bridges | louis de funes | p#1735350 | p#3112824 | p#1492633 | p#1492632 | p#7238 | p#9177463 | p#1166173 | Rabbi jakob | rabbi jacob | p#7483 | james bond
33
- Consider not only search terms, but also products
- Boost products that customers interact with
(Learn-to-Rank?)
SLIDE 34
Hybrid Search and Recommendation
34
Recommend related products when page is not full
SLIDE 35
Hybrid Search and Recommendation
35
Boost products on ambiguous queries
SLIDE 36
Hybrid Search and Recommendation
36
Recommend products on bad queries
SLIDE 37
Q & A
37
SLIDE 38
Thank You!
SLIDE 39
39