22 tips for better Data Science

Published November 12, 2014   |   
Vincent Granville

These tips are provided by Dr. Granville, who brings 20 years of varied data-intensive experience working with successful start-ups, small companies across various industries, and eBay, Visa, Microsoft, GE and Wells Fargo.

Tips for better Data Science

1. Leverage external data sources: tweets about your company or your competitors, or data from your vendors (for instance, customizable newsletter eBlast statistics available via vendor dashboards, or via submitting a ticket)
2. Nuclear physicists, mechanical engineers, and bioinformatics experts can make great data scientists.
3. State your problem correctly, and use sound metrics to measure yield (over baseline) provided by data science initiatives.
4. Use the right KPIs (key metrics) and the right data from the beginning, in any project. Changes due to bad foundations are very costly. This requires careful analysis of your data to create useful databases.
5. Fast delivery is better than extreme accuracy. All data sets are dirty anyway. Find the perfect compromise between perfection and fast return.
6. With big data, strong signals (extremes) will usually be noise. Here’s a solution.
7. Big data has less value than useful data.
8. Use big data from third-party vendors, for competitive intelligence.
9. You can build cheap, great, scalable, robust tools pretty fast, without using old-fashioned statistical science. Think about model-free techniques.
10. Big data is easier and less costly than you think. Get the right tools! Here’s how to get started.
11. Correlation is not causation. This article might help you with this issue. Read also this blog and this book.
12. You don’t have to store all your data permanently. Use smart compression techniques, and keep statistical summaries only, for old data.
13. Don’t forget to adjust your metrics when your data changes, to keep consistency for trending purposes.
14. A lot can be done without databases, especially for big data.
15. Always include EDA and DOE (exploratory analysis/design of experiment) early on in any data science projects. Always create a data dictionary. And follow the traditional life cycle of any data science project.
16. Data can be used for many purposes:
– quality assurance
– to find actionable patterns (stock trading, fraud detection)
– for resale to your business clients
– to optimize decisions and processes (operations research)
– for investigation and discovery (IRS, litigation, fraud detection, root cause analysis)
– machine-to-machine communication (automated bidding systems, automated driving)
– predictions (sales forecasts, growth, and financial predictions, weather)
17. Don’t dump Excel. Embrace light analytics. Data + models + gut feelings + intuition is the perfect mix. Don’t remove any of these ingredients in your decision process.
18. Leverage the power of compound metrics: KPIs derived from database fields, that have a far better predictive power than the original database metrics. For instance, your database might include a single keyword field but does not discriminate between the user query and search category (sometimes because data comes from various sources and is blended together). Detect the issue, and create a new metric called keyword type – or data source. Another example is IP address category, a fundamental metric that should be created and added to all digital analytics projects.
19. When do you need true real-time processing? When fraud detection is critical, or when processing sensitive transactional data (credit card fraud detection, 911 calls). Other than that, delayed analytics (with a latency of a few seconds to 24 hours) is good enough.
20. Make sure your sensitive data is well protected. Make sure your algorithms cannot be tampered by criminal hackers or business hackers (spying on your business and stealing everything they can, legally or illegally, and jeopardizing your algorithms – which translates in severe revenue loss). An example of business hacking can be found in section 3 in this article.
21. Blend multiple models together to detect many types of patterns. Average these models. Here’s a simple example of model blending.
22. Ask the right questions before purchasing software.
23. Run Monte-Carlo simulations before choosing between two scenarios.
24. Use multiple sources for the same data: your internal source, and data from one or two vendors. Understand the discrepancies between these various sources, to have a better idea about what the real numbers should be. Sometimes big discrepancies occur when a metric definition is changed by one of the vendors or changed internally, or data has changed (some fields no longer tracked). A classic example is web traffic data: use internal log files, Google Analytics and another vendor (say Accenture) to track this data.

This article originally appeared here. Republished with permission. For more tips, visit Data Science Central. Submit your copyright complaints here.