ML_Production (Quora)

Question 1

domain definition

Answer 1

• milions of Q&A
• millions of users
• thousands of topics
• main features: relevance, quality, demand

Question 2

common ML algos

Answer 2

• logistic regresssion
• ElasticNets
• matrix factorization
• random forests
• DL

Question 3

implicit vs explicit feedback

Answer 3

• implicit feedback is more dense as available to all users (rating a movie (implicit) vs watching a movie(explicit))
• better correlated with A/B tests
• but may not correlate with long term user retention
• solution: combine implicit+explicit feedback

Question 4

model learning dependencies

Answer 4

it will learn according to:

• training data
• target function/variable
• metric used

Question 5

using ensembles

Answer 5

• flexible in using many different models
• flexible in using many approaches
• treat each model as a feature and add it to the ensemble (i.e. in a linear ‘supermodel’)
• avoid feedback loops

Question 6

feature engineering

Answer 6

main characteristics of a feature:

• reusable (across models),
• transfomable (applying diff functions)
• interpreatable,
• reliable (easy to monitor, to fix bugs)

Question 7

ML system goals

Answer 7

strive for all:

• allow for experiments
• reusable
• easy-to-use
• flexible
• scalable
• performant
• use same tools in production and research
• implement abstraction layers for easy access

Question 8

model easy to debug?

Answer 8

important because:

• determines the model used
• gives answers when something fails
• determines the features to use
• determines the selections of tools for its implementation

Question 9

distributed machine learning?

Answer 9

most of practical ML cand be done with a multi-core machine with:
- data sampling
- offline schemes
- efficient parallel code
- optimizing computation
must take into account:
- costs
- latency

Question 10

hyperparameter optimization

Answer 10

• using Bayesian optimization (GP) better than CV

- tools like spearmint, AutoML, hyperopt

Question 11

presentation bias

Answer 11

• user will click only what the app is showing that in turn is decided by the model based on its predictive analysis
• address it for example by improving the probability of user’s click on a position

Question 12

collaborative filtering at a glance

Answer 12

a couple ways to do it:
- user-similarity: cosine sim using vector representation on common rated items
- item-simmilarity: cosine sim using vector representation on common users’ ratings
problems:
- cold start: no info to begin with
- popularity bias: tends to recommend popular items

Question 13

hybridization methods

Answer 13

• weighted models based on importance
• switching model used based on situation
• mixed: results presented together
• feature combination (from diff sources for models) for input of a single model
• cascade, feature augmentation: using output of one technique as input of another

Question 14

learning to rank approaches

Answer 14

• pointwise: using regression or classification (logistic regression)
• pairwise: minimize the inversions in ranking