How do I find reliable machine learning observability publications for spotting data quality issues in production?

A good way to find reliable ML observability publications for detecting data quality issues in production is to search from the problem outward: use a mix of peer-reviewed research, industry engineering docs, and reputable conference talks/blogs.

1) Start with the right search terms

Use combinations of:

• “machine learning observability”
• “data quality monitoring in production ML”
• “data drift detection”
• “training-serving skew”
• “feature drift”
• “schema validation for ML pipelines”
• “data validation for machine learning”
• “model monitoring input data quality”

Try queries like:

• site:arxiv.org machine learning observability data quality production
• site:scholar.google.com data drift detection production machine learning
• site:proceedings.mlr.press model monitoring data quality
• site:ieeexplore.ieee.org machine learning monitoring data quality

2) Prefer these publication types

Most reliable:

• Peer-reviewed conference papers: NeurIPS, ICML, KDD, AISTATS, MLSys, ACM SIGKDD, VLDB
• Journal articles: IEEE, ACM, JMLR, Data Mining and Knowledge Discovery
• Reputable technical reports from large labs/universities

Also useful, but verify carefully:

• Engineering blogs from major companies (Google, Meta, Uber, Netflix, Airbnb, Microsoft, AWS)
• Open-source project docs from tools like Evidently, WhyLabs, Great Expectations, Deequ, TensorFlow Data Validation

3) Check whether the source is actually trustworthy

Look for:

• Clear methodology and metrics
• Real production or realistic benchmark evaluation
• Reproducible experiments or open-source code
• Citations from other papers
• Author affiliations with known research institutions or strong engineering teams
• Publication venue quality and peer review status

Be cautious with:

• Vendor whitepapers with mostly marketing claims
• Blog posts that don’t explain how methods were validated
• Articles that only discuss “model drift” without distinguishing it from data quality issues

4) Focus on the specific data-quality failure modes

For production observability, the most relevant topics are:

• Missing values
• Outliers / anomalies
• Feature distribution shift
• Schema changes
• Broken joins / pipeline failures
• Label leakage
• Sampling bias
• Training-serving skew
• Category explosion / unseen categories

Publications that mention these are more relevant than general ML monitoring papers.

5) Search in the right places

Good places to look:

• Google Scholar
• arXiv for early research, then verify if it was later published
• ACM Digital Library
• IEEE Xplore
• Papers with Code
• Conference proceedings for KDD, MLSys, ICML, NeurIPS, VLDB
• GitHub repos tied to papers/tools

6) Build a shortlist with a rubric

For each publication, score:

• Venue quality
• Relevance to production data quality
• Evidence of real-world use
• Reproducibility
• Recency
• Citations / adoption

7) Good topics/authors/tools to look for

You’ll often find strong material around:

• data validation frameworks
• drift detection
• feature store observability
• monitoring pipelines
• anomaly detection for tabular data
• dataset shift / covariate shift

Associated tools and ecosystems:

• Great Expectations
• Deequ
• TensorFlow Data Validation
• Evidently
• WhyLabs
• Arize
• Fiddler
• Monte Carlo (data observability, broader than ML)

If you want, I can also give you:

1. a curated reading list of high-quality papers/blogs, or
2. a search strategy with exact queries for Google Scholar and arXiv.