Machine learning isn’t magic – it needs strategy and a human touch

Yasmin Jia, Associate Director of Data Science at Blockthrough.

Data-driven thinkingIt is written by members of the media community and contains new ideas about the digital revolution in the media.

Today’s column was written by Yasmin Jia, Associate Director of Data Science at blockthrough.

term Machine learning’ seems to have a magical effect as a buzzword in sales. Couple that with the term Data science,” and a lot of companies believe they have a winning formula for attracting new customers.

Is it smoke and mirrors? Often times, the answer is yes.”

But what is all too real is the need for best practices in data science and for companies to invest in and fully support talent that can apply these principles effectively.

Laying the foundation for machine learning

Machine learning success starts with hiring talent that can benefit from machine learning – a team of skilled data scientists – which is very costly. Adding to the cost is time. It takes a lot to build a data science team and integrate them with other teams across operations.

A successful machine learning pipeline requires data cleaning, data exploration, feature extraction, model building, model validation, and more. You also need to maintain and develop this pipeline. And not only is the cost high, companies rarely have the patience and time to manage this process and still achieve their ROI goals.

Define best practices

With the right talent and the right pipeline in place, the next step is to create best practices. This is vital. Machine learning depends on how you implement it, what problem you use to solve it, and how deeply you integrate it with your company.

To paint a picture of how things could go wrong, just think of the times when unbalanced data sets led to what the media call racist robots” And the automatic racism. Or, on a lighter note, how about those Memes Demonstrate a machine learning confusing blueberry pancakes and a Chihuahua. Or mixing up pictures of baking with pictures of curled up puppies?

Best practices can prevent some of these common pitfalls, but it is essential to identify them for the entire data analysis process: before, during, and after the decision.

Leavess take this step by step.

Before: It is very common for companies to update an offer by adding a feature. But they often do so before meaningfully complete data collection and analysis. No one took the time and resources to answer, Why do we add this feature? “

Before answering this very important question, other questions must be addressed. Do you see users doing this behavior normally already? What would the potential lift be? Is it worth spending and time to utilize your engineering resources? What is the expected effect? What does this new feature ultimately mean for the future success of this product?

You areWe will need a lot of data to answer these queries. but letLet’s say you’ve culled everything and decided it pays to move on.

during: You areWe launched this feature. There should be a steady stream of data showing whether the new feature is having an impact at the network level, at the publisher level, and at the user level.

Do you see the same effect across the board? Sometimes the benefits to someone else can hurt. You should pay attention. Factor analysis is key. What are the influencing factors that influence the analysis? Once identified, you need to determine if it is physically significant or not.

after, after: At this point, there are more questions to be addressed. What exactly is the effect? If you use A/B testing, can those short-term experiments provide reliable long-term predictions? What lessons can you learn? Whether it was a failure or a success, how could it continue to evolve? What are the new opportunities? What behavioral changes are new to you?Re-vision.

Machine learning in the long run

There is a lot of data and oversight required to make machine learning software truly viable. He. SheNo wonder many do notYou have what it takes to implement them correctly and reap the benefits.

Here’s the kicker: the data team doesn’t make the decisions. The machine learning algorithm does not make decisions. People make decisions. You can hire an impressive group of data scientists, and they can build and optimize a machine learning model based on 100% accurate data sets. But in order to make any kind of difference to your business, you need to develop a solid workflow around it.

The best way to do this? Ensure data science teams are deeply integrated with different teams in your organization.

Create a well-established data science practice, and you’ll see that machine learning can make the magic happen.

Follow Blockthrough (Tweet embed) and AdExchanger (Tweet embed) on Twitter.

Leave a Reply

Your email address will not be published.