When to use Machine Learning?

When not to use Machine Learning?

• Clear specifications are available
• Simple heuristics are good enough
• Cost of building and maintaining the system outweighs the benefits (see technical debt paper)
• Correctness is of utmost importance
• ML is used only for the hype, to attract funding

Examples?

Consider Non-ML Baselines

• Consider simple heuristics -- how far can you get?
• Consider semi-manual approaches -- cost and benefit?
• Consider the system without that feature
• Discuss Examples
  • Ranking apps, recommending products
  • Filtering spam or malicious advertisement
  • Creating subtitles for conference videos
  • Summarizing soccer games
  • Controlling a washing machine

When to use Machine Learning

• Big problems: Many inputs, massive scale
• Open-ended problems: No single solution, incremental improvements, continue to grow
• Time-changing problems: Adapting to constant change, learn with users
• Intrinsically hard problems: Unclear rules, heuristics perform poorly

Examples?

Additional Considerations for ML

• Partial solution is acceptable: Mistakes are acceptable or mitigable
• Data for continuous improvement is available
• Predictions can have an influence on system objectives: Does it actually contribute to organizational objectives?
• Cost effective: Cheaper than other approaches, or benefits clearly outweigh costs

Examples?

Spotify: Use of ML?

Big problem? Open ended? Time changing? Hard? Partial solution acceptable? Data continuously available? Influence objectives? Cost effective?

Recidivism: Use of ML?

Big problem? Open ended? Time changing? Hard? Partial solution acceptable? Data continuously available? Influence objectives? Cost effective?

Organizational Objectives

Innate/overall goals of the organization

• Business
  • Current revenue, profit
  • Future revenue, profit
  • Reduce business risks
• Non-Profits
  • Lives saved, animal welfare increased
  • CO2 reduced, fires averted
  • Social justice improved, well-being elevated, fairness improved
• Often not directly measurable from system output; slow indicators

Accurate models themselves are not the ultimate goal!

AI may only very indirectly influence such organizational objectives; influence hard to quantify; lagging measures

Leading Indicators

Measures correlating with future success, from the business perspective

• Customers sentiment: Do they like the product? (e.g., surveys, ratings)
• Customer engagement: How often do they use the product?
  • Regular use, time spent on site, messages posted
  • Growing user numbers, recommendations

Indirect proxy measures, lagging, bias

Can be misleading (more daily active users => higher profits?)

User Outcomes

How well the system is serving its users, from the user's perspective

• Users receive meaningful recommendations, enjoying content
• Users making better decisions
• Users saving time due to system
• Users achieving their goals

Easier and more granular to measure, but only indirect relation to organization objectives

Model Properties

Quality of the model used in a system, from the model's perspective

• Model accuracy
• Rate and kinds of mistakes
• Successful user interactions
• Inference time
• Training cost

Not directly linked to business goals

What is Measurement?

• Measurement is the empirical, objective assignment of numbers, according to a rule derived from a model or theory, to attributes of objects or events with the intent of describing them. – Craner, Bond, “Software Engineering Metrics: What Do They Measure and How Do We Know?"

• A quantitatively expressed reduction of uncertainty based on one or more observations. – Hubbard, “How to Measure Anything …"

Everything is Measurable

• If X is something we care about, then X, by definition, must be detectable.
  • How could we care about things like “quality,” “risk,” “security,” or “public image” if these things were totally undetectable, directly or indirectly?
  • If we have reason to care about some unknown quantity, it is because we think it corresponds to desirable or undesirable results in some way.
• If X is detectable, then it must be detectable in some amount.
  • If you can observe a thing at all, you can observe more of it or less of it
• If we can observe it in some amount, then it must be measurable.

But: Not every measure is precise, not every measure is cost effective

On Terminology

• Quantification is turning observations into numbers
• Metric and measure refer a method or standard format for measuring something (e.g., number of mistakes per hour)
  • Metric and measure synonymous for our purposes (some distinguish metrics as derived from multiple measures, or metrics to be standardizes measures)
• Operationalization is identifying and implementing a method to measure some factor (e.g., identifying mistakes from telemetry log file)

Decomposition of Measures

Often higher-level measures are composed from lower level measures

Clear trace from specific low-level measurements to high-level metric

Specifying Metrics

• Always be precise about metrics
  • "measure accuracy" -> "evaluate accuracy with MAPE"
  • "evaluate test quality" -> "measure branch coverage with Jacoco"
  • "measure execution time" -> "average and 90%-quantile response time for REST-API x under normal load"
  • "assess developer skills" -> "measure average lines of code produced per day and number of bugs reported on code produced by that developer"
  • "measure customer happiness" -> "report response rate and average customer rating on survey shown to 2% of all customers (randomly selected)"
• Ideally: An independent party should be able to independently set up infrastructure to measure outcomes

Risks with Measurements

• Bad statistics: A basic misunderstanding of measurement theory and what is being measured.
• Bad decisions: The incorrect use of measurement data, leading to unintended side effects.
• Bad incentives: Disregard for the human factors, or how the cultural change of taking measurements will affect people.

Measurement Validity

• Construct: Are we measuring what we intended to measure?
  • Does the abstract concept match the specific scale/measurement used?
  • e.g., IQ: What is it actually measuring?
  • Other examples: Pain, language proficiency, personality...
• Predictive: The extent to which the measurement can be used to explain some other characteristic of the entity being measured
  • e.g., Higher SAT scores => higher academic excellence?
• External validity: Concerns the generalization of the findings to contexts and environments, other than the one studied
  • e.g., Drug effectiveness on test group: Does it hold over the general public?

Correlation vs Causation

https://www.tylervigen.com/spurious-correlations

Correlation vs Causation

• In general, ML learns correlation, not causation
  • (exception: Bayesian networks, certain symbolic AI methods)
  • For more details: See causal inference
• Be careful about interpretation & intervention based on correlations
  • e.g., positive correlation between exercise and skin cancer
  • Exercise less => reduce chance of skin cancer?
• To establish causality:
  • Develop a theory ("X causes Y") based on domain knowledge & independent data
  • Identify relevant variables
  • Design a controlled experiment & show correlation
  • Demonstrate ability to predict new cases

Confounding Variables

Confounding Variables

• To identify spurious correlations between X and Y:
  • Identify potential confounding variables
  • Control for those variables during measurement
    • Randomize, fix, or measure + account for during analysis
    • e.g., control for ''smoke'', check whether ''drink coffee" => "pancreatic cancer"
• Other examples
  • Degree from top-ranked schools => higher salary
  • Age => credit card default rate
  • Exercise => skin cancer
  • and many more...

Streetlight effect

• A type of observational bias
• People tend to look for something where it’s easiest to do so
  • Use cheap proxy metrics that only poorly correlate with goal
  • e.g., number of daily active users as a measure of projected revenue

Risks of Metrics as Incentives

• Metrics-driven incentives can:
  • Extinguish intrinsic motivation
  • Diminish performance
  • Encourage cheating, shortcuts, and unethical behavior
  • Become addictive
  • Foster short-term thinking
• Often, different stakeholders have different incentives

Make sure data scientists and software engineers share goals and success measures

Example: University Rankings

• Originally: Opinion-based polls, but complaints by schools on subjectivity
• Data-driven model: Rank colleges in terms of "educational excellence"
• Input: SAT scores, student-teacher ratios, acceptance rates, retention rates, alumni donations, etc.,

Example: University Rankings

• Who are different stakeholders? What are their incentives? Can they be misused or cause unintended side effects?

For more, see Weapons of Math Destruction by Cathy O'Neil