Protect your Analytics Intellectual Property

Clearly analytics algorithms are big money as proven by a report today - acquisition of BuyingIQ by CarDekho in India -

We've acquired BuyingIQ for the team and their technology. In the age of big data, innovation for developing customer-driven solutions comes from analytics... BuyingIQ has developed an algorithm which scores products for prospective customers by analyzing millions of curated product reviews from all over the internet. It then puts a fair score, based on which it tells a user whether to buy a product or skip it

Good predictive analytics model can be a short term competitive advantage. But can you protect that model or algorithm from being leaked to a rival?

Should you protect the model at all from theft and what are the protection mechanisms used?

The Problem

There are four main aspects of the prescriptive model - the subject matter expertise which gets codified (algorithm), the code itself (implementation, including technology platform), underlying data (and quality) and customer test based refinements (continuous improvements). Out of these, data or data quality cannot be leaked. Software platform is somewhat similar across and code itself is automatically protected by Copyright. However, remember that Copyright protects the implementation and not really porting the same model into some other language. The customer testing bit needs to be done either way. Which brings us to the crux - codified subject matter expertise (SMEs). What prevents any contractual "SME" to build something similar for your rival by just adding another "insignificant" variable to make the equation unique. The trick is obviously to choose that "insignificance" so as not to affect technical or behavioral performance. Even the programmer who wrote the code could go and implement it elsewhere. So to take a general perspective - "How to protect an algorithm?" An example from past was Sergey and court guidance in the Bilski Case

Goldman Sachs computer programmer, Sergey Aleynikov, was arrested by the FBI for stealing a trading model he had created for the bank. The proprietary computer code in question facilitates “sophisticated, high-speed and high-volume trades on various stock and commodities markets.

Problem gets complicated with outsourcing model development to remote location Data Scientist e.g. does law punish theft of Trade Secret like in USA or does it punish breach of trust and confidence like in home country of Data Scientist. So depending on local law within outsourced country- legal recourse can be taken against that person or outsourcing entity.

The Protection

Many protection options exist but none are foolproof.

• Ideally your analytics software should allow for capturing meta data properties, comprehensive change logs and password protection at a model level to both protect and prove theft.
• Trade secret is most commonly used option but they can have multiple jurisdiction issue e.g. Outsourcing case above or like in USA where individual state jurisdictions apply. During development of an analytics model, a proper contract should be signed by all parties as soon as possible covering all jurisdiction issues. Algorithms or models can also be theoretically reverse engineered.
• Patents may be used but you will then have to register the invention with the patent office and disclose the secret.
• Copyrights have already been discussed as valid for protecting the implementation but not core model

Analytics models have an inbuilt theft protection mechanism as well 

The Model Health Deterioration

Like people, the health of a predictive model also deteriorates with time: if proper care is not taken by the team of data scientists. So if someone blindly copies the model without the process of keeping it healthy, the IP loss is only temporary (which does not mean it cannot be devastating).

If a company does not look after the health of its models, the results would lead to definite losses.It is extremely important to build processes and capability around improving health of analytics model for long term competitive advantage.

As always, I have just touched the tip of the iceberg and would love to hear from you!

NB: Predictive and Prescriptive analytics may have been used interchangeably in the article. They are however distinct.