Crystal Structure analysis and detailed information pptx
Deep Learning for Personalized Search and Recommender Systems
1. Deep Learning for Personalized
Search and Recommender
Systems
Ganesh Venkataraman
Airbnb
Nadia Fawaz, Saurabh Kataria, Benjamin Le, Liang Zhang
LinkedIn
1
2. Tutorial Outline
• Part I (45min) Deep Learning Key concepts
• Part II (45min) Deep learning for Search and Recommendations at Scale
• Coffee break (30 min)
• Deep Learning Case Studies
• Part III (45min) Jobs You May Be Interested In (JYMBII) at LinkedIn
• Part IV (45min) Job Search at LinkedIn
Q&A at the end of each part
2
3. Motivation – Why Recommender Systems?
• Recommendation systems are everywhere. Some examples of impact:
• “Netflix values recommendations at half a billion dollars to the company”
[netflix recsys]
• “LinkedIn job matching algorithms to improves performance by 50%” [San Jose
Mercury News]
• “Instagram switches to using algorithmic feed” [Instagram blog]
3
4. Motivation – Why Search?
4
PERSONALIZED SEARCH
4
Query = “things to do in halifax”
Search view – this is a classic IR problem
Recommendations view – For this query,
what are the recommended results?
5. Why Deep Learning? Why now?
• Many of the fundamental algorithmic techniques have existed since
the 80s or before
2.5 Exobytes of data produced per
day
Or 530,000,000 songs
150,000,000 iPhones 5
6. Why Deep Learning?
Image classification
eCommerce fraud
Search
Recommendations
NLP
Deep learning is eating the world
6
7. Why Deep Learning and Recommender
Systems?
• Features
• Semantic understanding of words/sentences possible with embeddings
• Better classification of images (identifying cats in YouTube videos)
• Modeling
• Can we cast matching problems into a deep (and possibly) wide net and learn
family of functions?
7
8. Part I – Representation Learning and Deep
Learning: Key Concepts
8
9. Deep Learning and AI
http://www.deeplearningbook.org/contents/intro.html 9
10. Part I Outline
• Shallow Models for Embedding Learning
• Word2Vec
• Deep Architectures
• FF, CNN, RNN
• Training Deep Neural Networks
• SGD, Backpropagation, Learning Rate Schedule, Regularization, Pre-Training
10
12. Representation learning for automated feature generation
• Natural Language Processing
• Word embedding: word2vec, GloVe
• Sequence modeling using RNN’s and LSTM’s
• Graph Inputs
• Deep Walk
• Multiple Hierarchy of features for varying granularities for semantic meaning
with deep networks
12
13. Example Application of Representation
Learning - Understanding Text
• One of the keys to any content based recommender system
is understanding text
• What does “understanding” mean?
• How similar/dissimilar are any two words?
• What does the word represent? (Named Entity
Recognition)
• “Abraham Lincoln, the 16th President ...”
• “My cousin drives a Lincoln”
13
14. How to represent a word?
• Vocabulary – run, jog, math
• Simple representation:
• [1, 0, 0], [0, 1, 0], [0, 0, 1]
• No representation of meaning
• Cooccurrence in a word/document matrix
14
15. How to represent a word?
• Trouble with cooccurrence matrix
• Large dimension, lots of memory
• Dimensionality reduction using SVD
• High computational time nxm matrix => O(mn^2)
• Adding new word => redo everything
15
16. Word embeddings taking context
• Key Conjecture
• Context matters.
• Words that convey a certain context occur together
• “Abraham Lincoln was the 16th President of the United States”
• Bigram model
• P (“Lincoln”|”Abraham”)
• Skip Gram Model
• Consider all words within context and ignore position
• P(Context|Word)
16
18. Word2Vec: Skip Gram Model
• Basic notations:
• w represents a word, C(w) represents all the context around a word
• 𝜃 represents the parameter space
• D represent all the (w, c) pairs
• 𝑝 𝑐 𝑤; 𝜃 represents the probability of context c given word w
parametrized by 𝜃
• The probability of all the context appearing given a word is given by:
• 𝑐∈𝐶(𝑤) 𝑝(𝑐|𝑤; 𝜃)
• The loss function then becomes:
• 𝑎𝑟𝑔𝑚𝑎𝑥 𝜃 𝑤,𝑐 ∈𝐷 𝑝(𝑐|𝑤; 𝜃)
18
19. Word2vec details
• Let 𝑣 𝑤 and 𝑣𝑐 represent the current word and context. Note that
𝑣𝑐 and 𝑣 𝑤 are parameters we want to learn
• p c w; 𝜃 =
𝑒 𝑣 𝑐∗𝑣 𝑤
𝑑∈𝐶 𝑒 𝑣 𝑑∗𝑣 𝑤
• C represents set of all available contexts
19
20. Negative Sampling – basic intuition
p c w; 𝜃 =
𝑒 𝑣 𝑐∗𝑣 𝑤
𝑑∈𝐶 𝑒 𝑣 𝑑∗𝑣 𝑤
• Sample from unigram distribution instead of taking all contexts into
account
• Word2vec itself is a shallow model and can be used to initialize a
deep model
20
22. Neuron: Computational Unit
• Input vector: x = [x1, x2 ,… ,xn]
• Neuron
• Weight vector: W
• Bias: b
• Activation function: f
• Output
a = f(WT x + b)
x1
x2
x3
x4
W
b
f
a = f(WTx + b)
Input x Neuron Output a 22
24. Layer
• Layer l: nl neurons
• weight matrix: W = [W1,…, Wnl]
• bias vector: b = [b1,…, bnl]
• activation function: f
• output vector
• a = f(WT x + b)
x1
x2
x3
x4
W1
b1
f
a1 = f(W1
T x + b1)
W2
b2
f
a2= f(W2
T x + b2)
Input x Layer Output a
W3
b3
f
a3= f(W3
T x + b3)
24
25. Layer: Matrix Notation
• Layer l: nl neurons
• weight matrix: W
• bias vector: b
• activation function: f
• output vector
• a = f(WT x + b)
• more compact notation
• fast-linear algebra routines for
quick computations in network
x1
x2
x3
x4
Input x Layer Output a
a = f(WT a + b)
W , b , f
25
27. Why CNN: Convolutional Neural Networks?
• Large size grid structured data
• 1D: time series
• 2D: image
• Convolution to extract features from image (e.g. edges, texture)
• Local connectivity
• Parameter sharing
• Equivariance to translation: small translations in input do not affect output
33. CNN example for image recognition: ImageNet [Krizhevsky et al., 2012]
Pictures courtesy of [Krizhevsky et al., 2012], http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf
33
1st GPU
2nd GPU
filters learned by first CNN layer
34. Why RNN: Recurrent Neural Network?
• Sequential data processing
• ex: predict next word in sentence: “I was born in France. I can speak…”
• RNN
• Persist information through feedback loop
• loop passes information from one step to the next
• Parameter sharing across time indexes
• output unit depends on previous output units through same
update rule.
xt
ht
ht-1
35. Unfolded RNN
• Copies of NN passing feedback to one another
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
35
36. LSTM: Long Short Term Memory [Hochreiter et al., 1997]
• Avoid vanishing or exploding gradient
• Cell state updates regulated by gates
• Forget: how much info from cell state to let
through
• Input: which cell state components to update
• Tanh: values to add to cell state
• Output: select component values to output
picture courtesy of http://colah.github.io/posts/2015-08-Understanding-LSTMs/
Cell state
• Long term dependencies
• large gap between relevant information and
where it is needed
• Cell state: long-term memory
• Can remember relevant information over long
period of time
36
37. Examples of RNN application
• Speech recognition [Graves et al., 2013]
• Language modeling [Mikolov, 2012]
• Machine translation [Kalchbrenner et al., 2013][Sustkever et al., 2014]
• Image captioning [Vinyals et al., 2014]
37
39. Cost Function
• m training samples (feature vector, label)
(𝑥 1 , 𝑦 1 ), … , (𝑥 𝑚 , 𝑦 𝑚 )
• Per sample cost: error between label and output from prediction layer
𝐽 𝑊, 𝑏; 𝑥 𝑖 , 𝑦 𝑖 = 𝑎(𝐿) 𝑥 𝑖 − 𝑦(𝑖) 2
• Minimize cost function over parameters: weights W and biases b
𝐽 𝑊, 𝑏 =
1
𝑚
𝑖=1
𝑚
𝐽(𝑊, 𝑏; 𝑥 𝑖
, 𝑦(𝑖)
) +
𝜆
2
𝑙=1
𝐿
𝑊(𝑙)
𝐹
2
Average error Regularization 39
40. Gradient Descent
• Random parameter initialization: symmetry breaking
• Gradient descent step: update for every parameter Wij
(l) and bi
(l)
𝜃 = 𝜃 − 𝛼𝛻θ 𝔼[𝐽(𝜃)]
• Gradient computed by Backpropagation
• High cost of backpropagation over full training set
40
41. Stochastic Gradient Descent (SGD)
• SGD: follow negative gradient after
• single sample
𝜃 = 𝜃 − 𝛼𝛻𝜃J(θ; 𝑥 𝑖
, 𝑦(𝑖)
)
• a few samples: mini-batch (256)
• Epoch: full pass through training set
• Randomly shuffle data prior to each training epoch
41
42. Backpropagation [Rumelhart et al., 1986]
Goal: Compute gradient numerically
Recursively apply chain rule for derivative of composition of functions
Let 𝑦 = 𝑔 𝑥 and 𝑧 = 𝑓 𝑦 = 𝑓(𝑔(𝑥))
then
𝜕𝑧
𝜕𝑥
=
𝜕𝑧
𝜕𝑦
𝜕𝑦
𝜕𝑥
= 𝑓′
𝑔 𝑥 𝑔′(𝑥)
Backpropagation steps
1. Feedforward pass: compute all activations
2. Output error: measures node contribution to output error
3. Backpropagate error through all layers
4. Compute partial derivatives
42
43. Training optimization
• Learning Rate Schedule
• Changing learning rate as learning progresses
• Pre-training
• Goal: training simple model on simple task before training desired model to perform desired task
• Greedy supervised pre-training: pre-train for task on subset of layers as initialization for final network
• Regularization to curb overfitting
• Goal: reduce generalization error
• Penalize parameter norm: L2, L1
• Augment dataset: train on more data
• Early stopping: return parameter set at point in time with lowest validation error
• Drop out [Srivatstava, 2013] : train ensemble of all subnetworks formed by removing non-output units
• Gradient clipping to avoid exploding gradient
• norm clipping
• element wise clipping
43
44. Part II – Deep Learning for Personalized
Recommender Systems at Scale
44
48. item j from a set of candidates
User i
with
<user features, query
(optional)>
(e.g., industry,
behavioral features,
Demographic features,……)
(i, j) : response yijvisits
Algorithm selects
(action or not, e.g. click, like, share, apply…)
Which item(s) should we recommend to the user?
• The item(s) with the best expected utility
• Utility examples:
• CTR, Revenue, Job Apply rates, Ads conversion rates, …
• Can be a combination of the above for trade-offs
Personalized Recommender Systems
48
50. User
Interaction
Logs
Offline Modeling
Workflow + User /
Item derived
features
User
User Feature
Store
Item Store +
Features
Recommendation
Ranking
Ranking
Model Store
Additional Re-
ranking Steps
1
2
4
5
Offline System Online System
3
An example of Recommender System
Architecture
Item
derived features
50
51. User
Interaction
Logs
Offline Modeling
Workflow + User /
Item derived
features
User
Search-based
Candidate
Selection &
Retrieval
Query
Construction
User Feature
Store
Search Index
of Items
Recommendation
Ranking
Ranking
Model Store
Additional Re-
ranking Steps
1
2
3
4 5
6
7
Offline System Online System
Item
derived features
An example of Personalized Search
System Architecture
51
52. Key Components – Offline Modeling
• Train the model offline (e.g. Hadoop)
• Push model to online ranking model store
• Pre-generate user / item derived features for online systems
to consume
• E.g. user / item embeddings from word2vec / DNNs based
on the raw features
52
53. Key Components – Candidate Selection
• Personalized Search (With user query):
• Form a query to the index based on user query annotation [Arya et al., 2016]
• Example: Panda Express Sunnyvale +restaurant:panda express
+location:sunnyvale
• Recommender system (Optional):
• Can help dramatically reduce the number of items to score in ranking steps
[Cheng, et al., 2016, Borisyuk et al. 2016]
• Form a query based on the user features
• Goal: Fetch only the items with at least some match with user feature
• Example: a user with title software engineer -> +title:software engineer for
jobs recommendation
53
54. Key Components - Ranking
• Recommendation Ranking
• The main ML model that ranks items retrieved by candidate selection based
on the expected utility
• Additional Re-ranking Steps
• Often for user experience optimization related to business rules, e.g.
• Diversification of the ranking results
• Recency boost
• Impression discounting
• …
54
55. Integration of Deep Learning Models
into Personalized Recommender
Systems at Scale
55
56. Literature: Deep Learning for Recommendation Systems
• RBM for Collaborative Filtering [Salakhutdinov et al., 2007]
• Deep Belief Networks [Hinton et al., 2006]
• Neural Autoregressive Distribution Estimator (NADE) [Zheng, 2016]
• Neural Collaborative Filtering [He, et al., 2017]
• Siamese networks for user item matching [Huang et al., 2013]
• Deep Belief Networks with Pre-training [Hinton et al., 2006]
• Collaborative Deep Learning [Wang et al., 2015]
56
57. User
Interaction
Logs
Offline Modeling
Workflow + User /
Item derived
features
User
Search-based
Candidate
Selection &
Retrieval
Query
Construction
User Feature
Store
Search Index
of Items
Recommendation
Ranking
Ranking
Model Store
Additional Re-
ranking Steps
1
2
3
4 5
6
7
Offline System Online System
Item
derived features
57
58. Offline Modeling + User / Item Embeddings
User Features Item Features
User Embedding
Vector
Item Embedding
Vector
Sim(U,I)
User Feature
Store
Item Store / Index
with Features
58
59. Query Formulation & Candidate Selection
• Issues of using raw text: Noisy or incorrect query tagging due to
• Failure to capture semantic meaning
• Ex. Query: Apple watch -> +food:apple +product:watch or +product:apple watch?
• Multilingual text
• Query: 熊猫快餐 -> +restaurant:panda express
• Cross-domain understanding
• People search vs job search
59
60. Query Formulation & Candidate Selection
• Represent Query as an
embedding
• Expand query to similar
queries in a semantic
space
• KNN search in dense
feature space with
Inverted Index [Cheng,
et al., 2016]
Q = “Apple Watch”
D = “iphone”
D = “Orange Swatch”
D = “ipad”
60
61. Recommendation Ranking Models
• Wide and Deep Models to capture all possible signals [Cheng, et
al., 2016]
https://arxiv.org/pdf/1606.07792.pdf
61
62. Challenges & Open Problems for Deep
Learning at Recommender Systems
• Distributed training on very large data
• Tensorflow on Spark (https://github.com/yahoo/TensorFlowOnSpark)
• CNTK (https://github.com/Microsoft/CNTK)
• MXNet (http://mxnet.io/)
• Caffe (http://caffe.berkeleyvision.org/)
• …
• Latency Issues from Online Scoring
• Pre-generation of user / item embeddings
• Multi-layer scoring (simple models => complex)
• Batch vs online training
62
63. Part III – Case Study: Jobs You May Be
Interested In (JYMBII)
63
64. Outline
• Introduction
• Generating Embeddings via Word2vec
• Generating Embeddings via Deep Networks
• Tree Feature Transforms in Deep + Wide Framework
64
66. Introduction: Problem Formulation
• Rank jobs by 𝑃 User 𝑢 applies to Job 𝑗 𝑢, 𝑗)
• Model response given:
66
Careers History, Skills, Education, Connections Job Title, Description, Location, Company
66
67. Introduction: JYMBII Modeling- Generalization
Recommend
• Model should learn general rules to predict which
jobs to recommend to a member.
• Learn generalizations based on similarity in title, skill,
location, etc between profile and job posting
67
68. Introduction: JYMBII Modeling - Memorization
Applies to
68
• Model should memorize exceptions to the rules
• Learn exceptions based on frequent co-
occurrence of features
69. Introduction: Baseline Features
• Dense BoW Similarity Features for Generalization
• i.e: Similarity in title text good predictor of response
• Sparse Two-Depth Cross Features for Memorization
• i.e: Memorize that computer science students will transition to entry engineering roles
Vector BoW Similarity Feature
Sim(User Title BoW,
Job Title BoW)
Sparse Cross Feature
AND(user = Comp Sci. Student,
job = Software Engineer)
Sparse Cross Feature
AND(user = In Silicon Valley,
job = In Austin, TX)
Sparse Cross Feature
AND(user = ML Engineer,
job = UX Designer)
69
70. Introduction: Issues
• BoW Features don’t capture semantic similarity between user/job
• Cosine Similarity between Application Developer and Software Engineer is 0
• Generating three-depth, four-depth cross features won’t scale
• i.e. Memorizing that Factory Workers from Detroit are applying to Fracking
jobs in Pennsylvania
• Hand-engineered features time consuming and will have low coverage
• Permutations of three-depth, four-depth cross features grows exponentially
70
71. Introduction: Deep + Wide for JYMBII
• BoW Features don’t capture semantic similarity between user/job
• Generate embeddings to capture Generalization through semantic similarity
• Deep + Wide model for JYMBII [Cheng et al., 2016]
Semantic Similarity Feature
Sim(User Embedding,
Job Embedding)
Global Model Cross Feature
AND(user = Comp Sci. Student,
job = Software Engineer)
User Model Cross Feature
AND(user = User 2,
job = Job Latent Feature 1 )
Job Model Cross Feature
AND(user = User Latent Feature,
job = Job 1)
71
Sparse Cross Feature
AND(user = Comp Sci. Student,
job = Software Engineer)
Sparse Cross Feature
AND(user = In Silicon Valley,
job = In Austin, TX)
Sparse Cross Feature
AND(user = ML Engineer,
job = UX Designer)
Vector BoW Similarity Feature
Sim(User Title BoW,
Job Title BoW)
72. Generating Embeddings via Word2vec:
Training Word Vectors
• Key Ideas
• Same users (context) apply to similar jobs (target)
• Similar users (target) will apply to the same jobs (context)
Application Developer => Software Engineer
• Train word vectors via word2vec skip-gram architecture
• Concatenate user’s current title and the applied job’s title as input
User Title Applied Job Title
72
73. Generating Embeddings via Word2vec:
Model Structure
Application, Developer Software, EngineerTokenized Titles
Word Embedding Lookup
Pre-trained Word
Vectors
Entity Embeddings
Via Average Pooling
Word Vectors
Response Prediction (Logistic Regression)
Cosine Similarity
User Job 73
74. Generating Embeddings via Word2vec:
Results and Next Steps
• Receiver Operating Characteristic – Area Under Curve for evaluation
• Response prediction is binary classification: Apply or don’t Apply
• Highly skewed data: Low CTR for Apply Action
• Good metric for ranking quality: Focus on discriminatory ability of model
• Marginal 0.87% ROC AUC Gain
• How to improve quality of embeddings?
• Optimize embeddings for prediction task with supervised training
• Leverage richer context about user and job
74
75. Generating Embeddings via Deep Networks:
Model Structure
User Job
Response Prediction (Logistic Regression)
Sparse Features (Title, Skill,
Company)
Embedding Layer
Hidden Layer
Entity Embedding
Hadamard Product (Elementwise Product)
75
76. Generating Embeddings via Deep Networks:
Hyper Parameters, Lots of Knobs!
• Optimizer Used
• SGD w/ Momentum and exponential decay vs. Adam [Kingma et al., 2015] (Adam)
• Learning Rate
• 10−5
to 10−3
(𝟏𝟎−𝟒
)
• Embedding Layer Size
• 50 to 200 (100)
• Dropout
• 0% to 50% dropout (0% dropout)
• Sharing Parameter Space for both user/job embeddings
• Assumes communitive property of recommendations (a + b = b + a) (No shared parameter space)
• Hidden Layer Sizes
• 0 to 2 Hidden Layers (200 -> 200 Hidden Layer Size)
• Activation Function
• ReLU vs. Tanh (ReLU)
76
77. Generating Embeddings via Deep Networks:
Training Challenges
• Millions of rows of training data impossible to store all in memory
• Stream data incrementally directly from files into a fixed size example pool
• Add shuffling by randomly sampling from example pool for training batches
• Extreme dimensionality of company sparse feature
• Reduce dimensionality of company feature from millions -> tens of thousands
• Perform feature selection by frequency in training set
• Hyper parameter tuning
• Distribute grid search through parallel modeling in single driver Spark jobs
77
78. Generating Embeddings via Deep Networks:
Results
Model ROC AUC
Baseline Model 0.753
Deep + Wide Model 0.790 (+4.91%***)
*** For reference, a previous major JYMBII
modeling improvement with a 20% lift in ROC
AUC resulted in a 30% lift in Job Applications
78
79. Response Prediction (Logistic Regression)
The Current Deep + Wide Model
Deep Embedding Features (Feed Forward NN)
• Generating three-depth, four-depth cross features won’t scale
• Smart feature selection required
Wide Sparse Cross Features (Two-Depth)
79
80. Tree Feature Transforms: Feature Selection via
Gradient Boosted Decision Trees
Each tree outputs a path from root to leaf encoding
a combination of feature crosses [He et al., 2014]
GDBT’s select the most useful combinations of
feature crosses for memorization
Member Seniority: Vice
President
Yes
No
Member Industry:
Banking
Yes
No
Member Location:
Silicon Valley
Member Skill:
Statistics
Yes No
80
Yes No
Job Seniority:
CXO
NoYes
Job Title: ML
Engineer
Yes No
81. Response Prediction (Logistic Regression)
Tree Feature Transforms: The Full Picture
How to train both the NN model and GBDT model
jointly with each other?
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GBDT)
81
82. Tree Feature Transforms: Joint Training via
Block-wise Cyclic Coordinate Descent
• Treat NN model and GBDT model as separate block-wise coordinates
• Implemented by
1. Training the NN until convergence
2. Training GBDT w/ fixed NN embeddings
3. Training the regression layer weights w/ generated cross features from GBDT
4. Training the NN until convergence w/ fixed cross features
5. Cycle step 2-4 until global convergence criteria
82
83. Response Prediction (Logistic Regression)
Tree Feature Transforms: Train NN Until
Convergence
Initially no trees are in our forest
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GDBT)
83
84. Response Prediction (Logistic Regression)
Tree Feature Transforms: Train GDBT w/ NN
Section as Initial Margin
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GDBT)
84
85. Response Prediction (Logistic Regression)
Tree Feature Transforms: Train GDBT w/ NN
Section as Initial Margin
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GDBT)
85
86. Response Prediction (Logistic Regression)
Tree Feature Transforms: Train Regression
Layer Weights
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GDBT)
86
87. Response Prediction (Logistic Regression)
Tree Feature Transforms: Train NN w/ GDBT
Section as Initial Margin
Deep Embedding Features (Feed Forward NN) Wide Sparse Cross Features (GDBT)
87
88. Tree Feature Transforms: Block-wise
Coordinate Descent Results
Model ROC AUC
Baseline Model 0.753
Deep + Wide Model 0.790 (+4.91%)
Deep + Wide Model w/ GBDT Iteration 1 0.792 (+5.18%)
Deep + Wide Model w/ GBDT Iteration 2 0.794 (+5.44%)
Deep + Wide Model w/ GBDT Iteration 3 0.795 (+5.57%)
Deep + Wide Model w/ GBDT Iteration 4 0.796 (+5.71%)
88
89. JYMBII Deep + Wide: Future Direction
• Generating Embeddings w/ LSTM Networks
• Leverage sequential career history data
• Promising results in NEMO: Next Career Move Prediction with Contextual
Embedding [Li et al., 2017]
• Semi-Supervised Training
• Leverage pre-trained title, skill, and company embeddings on profile data
• Replace Hadamard Product for entity embedding similarity function
• Deep Crossing [Shan et al., 2016]
• Add even richer context
• i.e. Location, Education, and Network features
89
90. Part IV – Case Study: Deep Learning Networks
for Job Search
90
94. Introduction: Query Understanding -
Segmentation and Tagging
• First divide the search query into
segments
• Tag query segments based on
recognized entity tags
Oracle
Java
Application Developer
Oracle
Java Application Developer
Query Segmentations
COMPANY = Oracle
SKILL = Java
TITLE = Application Developer
COMPANY = Oracle
TITLE = Java Application
Developer
Query Tagging
94
95. Introduction: Query Understanding –
Expansion
• Task of adding additional
synonyms/related entities to the
query to improve recall
• Current Approach: Curated dictionary
for common synonyms and related
entities
COMPANY = Oracle OR NetSuite OR
Taleo OR Sun Microsystems OR …
SKILL = Java OR Java EE OR J2EE
OR JVM OR JRE OR JDK …
TITLE = Application Developer OR
Software Engineer OR
Software Developer OR
Programmer …
Green – Synonyms
Blue – Related Entities
95
96. Introduction: Query Understanding - Retrieval
and Ranking
COMPANY = Oracle OR NetSuite OR Taleo OR
Sun Microsystems OR …
SKILL = Java OR Java EE OR J2EE OR JVM
OR JRE OR JDK …
TITLE = Application Developer OR
Software Engineer OR
Software Developer OR
Programmer …
Title
Title
Skills
Company
96
97. Introduction: Issues – Retrieval and Ranking
• Term retrieval has limitations
• Cross language retrieval
• Softwareentwickler Software developer
• Word Inflections
• Engineering Management Engineering Manager
• Query expansion via curated dictionary of synonyms is not scalable
• Expensive to refresh and store synonyms for all possible entities
• Heavy reliance on query tagging is not robust enough
• Novel title, skill, and company entities will not be tagged correctly
• Errors upstream propagates to poor retrieval and ranking
97
98. Introduction: Solution – Deep Learning for
Query and Document Representations
• Query and document representations
• Map queries and document text to vectors in semantic space
• Robust to Handle Out of Vocabulary words
• Term retrieval has limitations
• Query expansion via curated dictionary of synonyms is not scalable
• Map synonyms, translations and inflections to similar vectors in semantic space
• Term retrieval on cluster id or KNN based retrieval
• Heavy reliance on query tagging is not robust enough
• Compliment structured query representations with semantic representations
98
99. Representations via Word2vec:
Leverage JYMBII Work
• Key Ideas
• Similar users (context) apply to the same job (target)
• The same user (target) will apply to similar jobs (context)
Application Developer => Software Engineer
• Train word vectors via word2vec skip-gram architecture
• Concatenate user’s current title and the applied job’s title as input
User Title Applied Job Title
99
100. Representations via Word2vec:
Word2vec in Ranking
Application, Developer Software, EngineerTokenized Text
Word Embedding Lookup
Pre-trained Word
Vectors
Entity Embeddings
Via Average Pooling
Word Vectors
Learning to Rank Model (NDCG Loss)
Cosine Similarity
JobQuery 100
101. Representations via Word2vec:
Ranking Model Results
Model Normalized Cumulative
Discounted Gain@5 (NDCG@5)
CTR@5(%)
Baseline Model 0.582 +0.0%
Baseline Model + Word2Vec Feature 0.595 (+2.2%) +1.6%
101
102. Representations via Word2vec:
Optimize Embeddings for Job Search Use Case
• Leverage apply and click feedback to guide learning of embeddings
• Fine tune embeddings for task using supervised feedback
• Handle out of vocabulary words and scale to query vocabulary size
• Compared to JYMBII, query vocabulary is much larger and less well-formed
• Misspellings
• Word Inflections
• Free text search
• Need to make representations more robust for these free text queries
102
103. Robust Representations via DSSM:
Deep Structured Semantic Model [Huang et al., 2013]
Query Applied Job (Positive)
Application Developer Software EngineerRaw Text
#Ap, App, ppl… #So, Sof, oft…Tri-letter Hashing #Ha, Hai, air…
Hairdresser
Randomly Sampled
Applied Job (Negative)
Hidden Layer 3
Hidden Layer 2
Hidden Layer 1
Cosine Similarity
Softmax w/ Cross Entropy Loss
103
104. Robust Representations via DSSM:
Tri-letter Hashing
• Tri-letter Hashing Example
• Engineer -> #en, eng, ngi, gin, ine, nee, eer, er#
• Benefits of Tri-letter Hashing
• More compact Bag of Tri-letters vs. Bag of Words representation
• 700K Word Vocabulary -> 75K Tri-letters
• Can generalize for out of vocabulary words
• Tri-letter hashing robust to minor misspellings and inflections of words
• Engneer -> #en, eng, ngn, gne, nee, eer, er#
104
105. Robust Representations via DSSM:
Training Details
105
• Parameter Sharing Helps
• Better and faster convergence
• Model size is reduced
• Regularization
• L2 performs better than dropout
• Toolkit Comparisons (CNTK vs TensorFlow)
• CNTK: Faster convergence and better model quality
• TensorFlow: Easy to implement and better community support.
Comparative model quality
Training performance with/o parameter sharing
106. Robust Representations via DSSM:
Lessons in Production Environment
106
+ 100%
+ 70%
+ 40%
• Bottlenecks in Production
Environment
• Latency due to extra computation
• Latency due to GC activity
• Fat Jars in JVM environment
• Practical Lessons
• Avoid JVM Heap while serving the
model
• Caching most accessed entities’
embedding
107. Robust Representations via DSSM:
DSSM Qualitative Results
Software Engineer Data Mining LinkedIn Softwareentwickler
Engineer Software Data Miner Google Software
Software Engineers Machine Learning
Engineer
Software Engineers Software Engineer
Software Engineering Microsoft Research Software Engineer Engineer Software
For qualitative results, only top head queries are taken to analyze similarity to each other
107
108. Robust Representations via DSSM:
DSSM Metric Results
Model Normalized Cumulative
Discounted Gain@5 (NDCG@5)
CTR@5 Lift (%)
Baseline Model 0.582 +0.0%
Baseline Model + Word2Vec Feature 0.595 (+2.2%) +1.6%
Baseline Model + DSSM Feature 0.602 (+3.4%) +3.2%
108
109. Robust Representations via DSSM:
DSSM Future Direction
• Leverage Current Query Understanding Into DSSM Model
• Query tag entity information for richer context embeddings
• Query segmentation structure can be considered into the network design
• Deep Crossing for Similarity Layer [Shan et al., 2016]
• Convolutional DSSM [Shen et al., 2014]
109
110. Conclusion
• Recommender Systems and personalized search are very similar
problems
• Deep Learning is here to stay and can have significant impact on both
• Understanding and constructing queries
• Ranking
• Deep learning and more traditional techniques are *not* mutually
exclusive (hint: Deep + Wide)
110
113. Difference between parameter sharing in 1-D
convolution and RNN?
• CNN Kernel: output unit depends on small number of neighboring input units
through same kernel
• RNN update rule: output unit depends on previous output units through same
update rule. Deeper computational graph.
115. Introduction: JYMBII Modeling -
Personalization
Applies to
Recommend jobs that are similar to jobs user has
previously applied to
115
116. Introduction: JYMBII Modeling - Collaboration
Applies to
Recommend jobs that similar users have previous
applied to
116
117. Introduction: Generalized Linear Mixed
Models
• Mixture of linear models into an additive model [Zhang et al., 2016]
• Fixed Effect – Population Average Model
• Random Effects – Subject Specific Models
Response Prediction (Logistic Regression)
User 1
Random Effect Model
User 2
Random Effect Model
Personalization
Job 2
Random Effect Model
Job 1
Random Effect Model
Collaboration
Global Fixed Effect Model
Content-Based Similarity
117
118. Introduction: Features
• Dense Vector BoW Similarity Features in global model for Generalization
• i.e: Similarity in title text good predictor of response
• Sparse Cross Features in global,user, and job model for Memorization
• i.e: Memorize that computer science students will transition to entry engineering roles
Vector BoW Similarity Feature
Sim(User Title BoW,
Job Title BoW)
Global Model Cross Feature
AND(user = Comp Sci. Student,
job = Software Engineer)
User Model Cross Feature
AND(user = User 2,
job = Software Engineer)
Job Model Cross Feature
AND(user = Comp Sci. Student,
job = Job 1)
118
119. Introduction: GLMM Formulation
• Generalized Linear
Mixed Models
• Per-User Random Effect
• Per-Job Random Effect
𝑃 𝐴𝑝𝑝𝑙𝑦 𝑚, 𝑗) = 𝜎(𝐵𝑓𝑖𝑥𝑒𝑑 𝑋cos 𝑚,𝑗 , 𝑋 𝑚, 𝑋𝑗, 𝑋 𝑚,𝑗 + 𝐵 𝑚 𝑋cos 𝑚,𝑗 , 𝑋𝑗 + 𝐵𝑗 𝑋cos 𝑚,𝑗 , 𝑋 𝑚 )
Notation Meaning
Xcos(m, j) Dense Cosine Similarity Features for sample pair m,j.
(i.e. - Cosine similarity between title BoW)
Xm Sparse Features of user m. (i.e. – User is a software
engineer)
Xj Sparse Features of job j. (i.e. – Job is from company
LinkedIn)
Xm,j Sparse Cross product feature transformation of Xm,
Xj. (i.e. – User is software engineer and job is from
company LinkedIn)
Bfixed Fixed effect model coefficients.
Bm User model coefficients for user m.
Bj Job model coefficients for job j.
Fixed Effect – User-Job Affinity Per-User Model Per-Job Model 119
120. Introduction: Issues
• BoW Features don’t capture semantic similarity between user/job
• Cosine Similarity between Application Developer and Software Engineer is 0
• Sparse features in user and job specific models make it difficult to fit
• Current linear model is unable to share learning across similar titles
• Large number of sparse features does not scale to infrastructure
• Billions of User and Job Model Cross features
120
121. Introduction: Proposed Solution
• Learn dense semantic embeddings of user and job entities
• Integrate embeddings into GLMM model as a set of latent features
Semantic Similarity Feature
Sim(User Embedding,
Job Embedding)
Global Model Cross Feature
AND(user = Comp Sci. Student,
job = Software Engineer)
User Model Cross Feature
AND(user = User 2,
job = Job Latent Feature 1 )
Job Model Cross Feature
AND(user = User Latent Feature,
job = Job 1)
121
122. Generating Embeddings Via Deep Networks:
Data and Features
• 4 weeks (Training) and 1 week (Test) of Click Log Data
• 15M Applies
• 2.5M Dismisses
• 30M Skips
• 15M Random Negatives
• 5M Users
• 1M Jobs
• Input Features
• Sparse Standardized Title - |Title Taxonomy| = 25K
• Sparse Standardized Skills - |Skill Taxonomy| = 35K
• Sparse Standardized Company - |Companies| = 1M+
122
123. Generating Embeddings via Deep Networks:
GLMM Integration Results
Improved model performance w/ 10x
reduction in number of model parameters
Model ROC AUC
Baseline GLMM Model w/ Sparse Features 0.800
GLMM Model w/ Dense Embedding Features 0.811 (+1.38%)
123
125. Robust Representations via DSSM:
DSSM Training Details
• Training Data: 6 months of apply data from job search
• Training Tuples: (query, applied job)
• Network Size: (3 Layers:- 300, 300, 300)
• Parameter Sharing: Shared Parameters among query and job
• Activation Function: Tanh
125
126. References
• [Rumelhart et al., 1986] Learning representations by back-propagating errors, Nature 1986
• [Hochreiter et al., 1997] Long short-term memory, Neural computation 1997
• [LeCun et al., 1998] Gradient-based learning applied to document recognition, Proceedings of
the IEEE 1998
• [Krizhevsky et al., 2012] Imagenet classification with deep convolutional neural networks, NIPS
2012
• [Graves et al., 2013] Speech recognition with deep recurrent neural networks, ICASSP 2013
• [Mikolov, 2012] Statistical language models based on neural networks, PhD Thesis, Brno
University of Technology, 2012
• [Kalchbrenner et al., 2013] Recurrent continuous translation models, EMNLP 2013
• [Srivatstava, 2013] Improving neural networks with dropout, PhD Thesis, University of Toronto,
2013
• [Sustkever et al., 2014] Sequence to sequence learningg with neural networks, NIPS 2014
• [Vinyals et al., 2014] Show and tell: a neural image caption generator, Arxiv 2014
• [Zaremba et al., 2015] Recurrent Neural Network Regularization, ICLR 2015
126
127. References (continued)
• [Arya et al., 2016] Personalized Federated Search at LinkedIn, CIKM 2016
• [Cheng et al., 2016] Wide & Deep Learning for Recommender Systems, DLRS 2016
• [He et al., 2014] Practical Lessons from Predicting Clicks on Ads at Facebook, ADKDD 2014
• [Kingma et al., 2015] Adam: A Method for Stochastic Optimization, ICLR 2015
• [Huang et al., 2013] Learning Deep Structured Semantic Models for Web Search using Clickthrough Data, CIKM 2013
• [Li et al., 2017] NEMO: Next Career Move Prediction with Contextual Embedding, WWW 2017
• [Shan et al., 2016] Deep Crossing: Web-scale modeling without manually crafted combinatorial features, KDD 2016
• [Zhang et al., 2016] GLMix: Generalized Linear Mixed Models For Large-Scale Response Prediction, KDD 2016
• [Salakhutdinov et al., 2007] Restricted Boltzmann Machines for Collaborative Filtering, ICML 2007
• [Zheng, 2016] http://tech.hulu.com/blog/2016/08/01/cfnade.html
• [Hinton et al., 2006] A fast learning algorithm for deep belief nets, Neural Computations 2006
• [Wang et al., 2015] Collaborative Deep Learning for Recommender Systems , KDD 2015
• [He et al., 2017] Neural Collaborative Filtering, WWW 2017
• [Borisyuk et al. 2016]. CaSMoS: A Framework for Learning Candidate Selection Models over Structured Queries and
Documents, KDD 2016
127
Welcome to our tutorial on Deep Learning for Personalized Search and Recommender Systems
I am Ganesh and I work for Airbnb. I will be presenting this with my former colleagues from LinkedIn – Nadia, Ben and Liang. Saurabh could not make it
Between the 5 of us we have worked on various aspects of search/recommendations/machine learning and deep learning.
We will be releasing these slides online. We want this to be an interactive tutorial, so please feel free to interrupt any time.
But, just to get a sense of the audience, show of hands to how many of you are from academia?
How many from industry?
How many have worked on some sort of production deep learning system before?
It is a fairly long tutorial and we want to make sure we have your attention
As we have noticed we have a fairly diverse audience and we want to ensure that we cater to all your interests which may be a challenge.
We start off with some foundational topic in deep learning. One may consider these like understanding the lego pieces
We then move on to explaining deep learning for search and recommendations at scale, followed by case study at LinkedIn
TODO – add logistics here
In Recys 2015, Netflix mentioned that recommender systems add about half a billion dollars to the company
It is almost a given these ,days that you open up a site or an app and it presents an experience that is tailor made for you, knowing your behavior and your preferences whether or not you made them explicit
There’s always been this friendly turf between search and recommendations.
Take a look at this query. A search person would treat this as classic IR problem, a recommendation person would treat this as a set of recommendations returned by the search engine for your query. Well, these two world’s aren’t that different and we believe they actually got married via personalized search.
Personalized search is where each result is tailored to your own preferences and not necessarily same for everyone. There are of course various levels of personalization
That brings us to the logical next question? Why deep learning
If you talk with a mathematician, they would point out that a lot of foundational techniques have been around since the 80’s or even before. And they would be right.
So, why now?
Two major things need to happen for deep learning to really work.
Computing cost
Availability of data – lot’s of it
TODO – add GPU
In several domains, deep learning has already made a dent. These include self driving cars and machine translation and a few others just to name some
deep learning is a particular case of representation learning.
It learns additional layers of more abstract features.
This slide should motivate why understanding text is critical
Understanding text is critical to any recommender system that works with underlying text data (example – news recommendations, jobs recommendations etc)
What are the components of “understanding”?
-- Similar or dissimilar words
-- concept of similarity can range from true synonyms to more ‘fuzzy’ types
-- Entity the word represents (Named Entity Recognition)
-- “Abraham Lincoln, the 16th President”, “My cousin drives a Lincoln”
Explain why the function isn’t scalable
Note on negative sampling
shallow models to learn embeddings, such as word2vec, are often used as initialization in deeper network architectures.
sigmoid used to convert real to (0,1), for example into a probability.
Used in output layer with log loss (cross entropy) in optimization, rather than squared loss, to avoid vanishing gradient issue when it saturates
discouraged use in hidden layer because of saturation
tanh resembles identity close to 0, can be trained easily as long as activation remains small. otherwise, saturation makes gradient-based learning difficult.
ReLU, also know as positive part
piecewise linear, easy to optimize
does not saturate on the positive side.
Initialize ReLU with small positive bias to make it likely for the ReLU to be active for most inputs in the training set
FF also known as Multi Layer Perceptron MLP
goal of the network: approximate a function y = F(x)
- non-linearities
output prediction layer:
linear unit for linear regression
sigmoid for logistic regression
softmax for multinoulli multi-class classification
CNN automates filters and feature extraction from images that used to be highly tailored and manual
Equivariance to translation: small translations in input do not affect output
Large image NxN pixels: N^2 input units
Hidden layer: K features
Number of parameters: ~KN^2
2 GPs to train large deep neural net (layers split between the 2 GPUs)
5 CNN layers with pooling + 3 fully connected layers
ReLU
DropOut against overfitting
Large dataset with 1000 image categories
image, label, classification. top 4: correctly classified, bottom 4: misclassified
63% accuracy
correct label was in top 5 predicted 85% of time
RNN update rule: output unit depends on previous output units through same update rule. Deeper computational graph.
1-D CNN Kernel: output unit depends on small number of neighboring input units through same kernel
chain structure with repeating module (parameter sharing across time)
cell state: conveyor belt that runs through the chain and stores long term memory.
cell state has a linear self-loop that allows product of derivatives close to one, weights of the self-loop are control by gates.
Problem with gradient descent for standard RNNs: error gradients vanish exponentially quickly with the size of the time lag between important events.
gradients propagated over many stages tend to either vanish (often) or explode
sigmoid gate lives in [0,1]: controls amount of information that it lets through from each component of cell state
tanh: create a vector of candidate values to update the cell state or from the cell state to update the output
speech recognition: acoustic modeling for mapping from acoustic signal to sequences of word (phonetic state)
machine translation: LSTM trained on concatenation of source sentences and their translation [Sustkever et al., 2014]
maximum likelihood: negative log likelihood = cross entropy between data distribution p(x,y) and model distribution p(y|x)
if model distribution is gaussian p(y|x) = N(y; a^L(x), I), then cross entropy boils down to squared loss.
alpha: learning rate
Dataset augmentation for image: translate, rotate, scale images to generate new samples.
Dropout:
- Dropout operator corrupts the information carried by the units, forcing them to perform their immediate computations more robustly
for LSTM, apply dropout only the non-recurrent connections, to maintain memorization ability see [Zaremba et al., 2015]
Gradient norm clipping maintains the direction of the gradient.
During time interval t
User visits to the webpage
Some item inventory at our disposal
Decision problem: for each visit, select an item to display
each display generates a response
Goal: choose items to maximize expected clicks
- Introduce My Self
- Passive Job Seekers: Allow them to discover what is avaliable in the marketplace.
- Not Alot of Data for Passives, show them jobs that make the most sense to them given their current career history and experience
- Active Job Seekers: Reinforce their job seeking experience. Show them similar jobs that they applied to that they may have missed. Make sure they don't miss opportunities
- Powers alot of modules including jobs home, feed, email, ads
- Goal: Get people hired - Confirmed Hires
- Time Lag on signal
- Proxy in Total Job Applies
- Metric: Total Job Applies
- Optimize probability of Apply. Not View. Showing users popular/attractive jobs not as important as showing them actual good matches
- User, Job, Activity
- Leverage our rich context on a user and job
- Generalize rules that similarity in title is a good thing
- Memorize exceptions
- I.E. for people in ML who have replaced our jobs w/ ML itself, we need to move to design
- Example of features in each of the model. Generalization, on most of the time to share learning across examples in a linear model
- Sparse Features for memorization
- Need to choose good features to memorize
- Random effect sparse features model personal affinity
- No semantic similarity do well for the most part
- Difficult fitting because no shared learning
- Don't scale to infrastructure. Too many parameter updates via SGD costly to run w/o too much ROI. Alot of sparse features need to be computed online which makes for expensive serving
- Dimensionality reduction
- Learn dense embeddings for job and user entities. Semantic embeddings
- Key things are to defined context and target. You shall know a word by the company it keeps
- Leverage word2vec existing libraries which take sentences as input
- Random Effects are not shown here
- Cross Features are not shown here
- Word vectors trained seperately
- ROC is good
- Accuracy won't be as good metric
- Marginal Gain from ROC AUC over BoW
- Leverage our large amount of labeled data
- Two Towers
- 3 Layers
- One tower for embedding member
- One for Job
- One similarity layer
- Non Linear Interactions. Important Skills for a Job
- Adam is adaptive learning rate. Momentum
- Momentum good for smoothing out noisy gradient
- Decay is good so that we don't overshoot the optium
- Dropout wasn't good for us. Model wasn't complex enough given our amount of data
- Sharing parameter space doesn't make sense for us
- Similar to Tensorflow batch queue implementation
- Millions of parameters * 100 in embedding size
- No distributed training so just train multiple models in parallel. Tensorflow supports distributed training but not on YARN
- Do well just knowing the member's profile
- Deep model does better due to all the non linear interactions introduced
- Gets close to the Baseline GLMM model, despite there being no random effects in the deep model to handle personalization and collaboration
- Can see how this translates
- Visual representation
- Model Driven way like for deep model?
- Random effects not shown
- Output of each tree is a path used as a categorical feature
- GDBT learns to select the most useful combinations of feature crosses for memorization
- Example of features
- Vice Presidents in Banking are not a normal VP. Seniority level is lower than other VP's in comparision
- Data Engineers who don't know programming but know statistics are still very good matches for Data engineering jobs
- Can't run SGD through graph
- SGD to train network
- Greedy split on learning decision rules
- Coordinant Descent is single coordinate optimzation (Iterative 1 dimensional optimization fixing the other coordinants)
- Train deep model first because we want to learn generalizations before exceptions. Can't have exceptions if we don't have rules first
- Deep Model Initialization. Train Deep Model first since need to learn the generalizations first before learning the exceptions
- Logit gets uses as an intial prediction for GDBT to start boosting from
- Tree learns to generate features to help classify cases where the NN is wrong
-
- Train Regression Layer weights on the features generated by tree layer transformation
- GDBT generates better features than sparse. Iterative training improves slightly
- AUC might not be best way to visualize how tree features help given that they won't move the needle as much since they handle exceptions
- Profile Data. Word2vec type of networks w/ career history
- Hadamard product is easy to deploy in production. Only need to do a vector operations rather than a full matrix multiply
- Location good to learn interactions between industry and location.
- Education good for students
- Search Jobs active job searchers
- Personalized Component to Job Search
- We will focus on the query to job matching though
- Extracting structured information from a user's raw query
- Query Understanding most important part
- Talk about indexer first, then talk about user query
- Reference LinkedIn Economic graph and taxonomy
- Add structure to query
-- Exact terms matching for retrieval will miss very relevant documents containing similar terms
- Expansion allows us to retrieve those very relevant documents
- Curate dictionary through domain knowledge or knowledge base
- Can also mine synonyms through query logs
- Give enough example for it
Google -> LinkedIn -> Apple
- Leverage query understanding for matching phrases rather than terms. Phrase match with distance
- Leverage query tags to match w/ appropriate fields for higher precision (Job Application -> Application Developer)
- BoW features (Title, query title sim)
- Might not be handled in synonyms
- Curating is not scalable given all the possible inflections and languages around. Large dictionary will be expensive to store all the mappings
- If query tagging fails, then our retrieval and ranking will not be as good
- Pinterest as a Skill, Hedge Fund as a company
-
- Good representations that map to dense semantic space
- Robust to OOV words
- Robust to issues w/ current segmentation/query tagging
- Map on job and query side for embeddings
- Compliment them. Query tagging gives us good information about the phrases and handles ambiguity for us but potentially, taxonomy not might be refreshed often enough to keep up w/ the professional world
- Embeddings learned from query log data should reflect new entities automatically as we see more new data
Same slide as before
- Show differences
NDCG Normalized Discounted Cumulative Gain
- Describe NDCG
- Relative Ranking
- Penalized lower rankings for relevant results
- Normalized across query
- CTR@5 goes up
Emphasize OOV words. Job search vocabulary alot more different from JYMBII vocab
- Skills, companies, titles, can search anything
- Talk about tri letter hashing later. BoTrigrams vs. BoW
- Extra Negatives. 10 Negatives
- 3 layers
- Softmax layer, learn positives and negatives at the same time
- Triletter hashing achieves goal of robustness at OOV
- Parameter Sharing for Full Siamese
- Reuse more
- Some Regularization used. Didn't really do too much though. Parameter sharing already adds some form of regularization
- Tensorflow for the flexibility
- Latency, model size, model delivery was issue
- Store model off heap to reduce GC activity
- Cache most access entities for speed
- Emphasize head queries
- Even better lift
- Also try to use for candidate selection
Recommendations and Search are more in common that what people may realize.
Online phase of any recommender or search consists of two very board stages and deep learning has impact on both