Machine Learning/AI
AdobeStock_205110625
As consumer expectations grow for more personalized, relevant, and assistive experiences, Machine Learning (ML) is becoming an invaluable tool to meet those demands. It’s helping marketers across dozens of ML use cases in advertising with brands using the technology to intelligently identify and segment audiences, build ad creatives, test variations, improve performance, and optimize spend automatically, in real-time and at scale.
The golden rule of machine learning is that an algorithm is only as good as the data it’s fed. 95% of advertisers, per the Interactive Advertising Bureau’s Data Center of Excellence, have terabytes of personal data, location information and user interest they can use to target prospects. But to effectively use ML, advertisers and marketers must have the right data for the problems they are trying to solve when building predictive models. Machine learning also require that data be properly formatted, cleaned, and organized to enable data scientist to develop predictive models. Cleaning, maintaining, and governing data is just one part of the process of developing ML models.
To get more insights on this topic, I reached out to Jamal Robinson whose experience includes being a Director of Artificial Intelligence (AI), data science professor at University of California Berkeley, ML author and who currently leads AI/ML Business Development at Amazon Web Services.
Gary Drenik: What is the role that data plays when enterprises are developing machine learning models?
Jamal Robinson: It’s no secret that data is fueling digital transformation across many use cases. This is why 90% of enterprises told Forrester improving their use of data to gain better business insights is a top priority. With Machine Learning (ML) the quality of data used to build predictive models heavily influences those model’s accuracy making the role of data in ML extremely important.
Drenik: So, is a significant portion of the ML development cycle is spent focusing on data?
Robinson: Correct. Data science teams can spend up to 70% of their time acquiring, loading and transforming the data before the machine learning process starts. Even advanced data science teams spend a significant portion of their time on data operations, which I assume is the value customers like Twitter, Facebook and Google see in consuming Prosper’s curated customer insights data?
MORE FOR YOU
Drenik: Yes, our customers have communicated the data sets helped them get to predictive insights faster.
Robinson: Makes sense as a substantial portion of ML operations and governance is spent focusing on management of data and starting with curated data set can help to simplify things.
Drenik: I keep hearing about ML operations or MLOps, what is it, what is the connection to governance and why do these topics matter?
Robinson: There are many definitions out there but the simplest way to define governance and MLOps is best practices and policies for businesses to successfully use machine learning in an explainable, repeatable and production ready manner.
Drenik: After I hear about MLOps, the next thing I hear is enterprises are struggling to implement an MLOps strategy, is that accurate and if so why?
Robinson: What I’ve seen in current and previous roles when speaking with executives, data scientists and IT managers across startups, Fortune 500 and Global 2000 companies is similar to Gartner’s findings that 50% of ML projects never see the light of day. ML is still an emerging technology and while the technology is maturing, understanding of governance structures required to make ML explainable, repeatable and production ready is still lacking. Even with MLOps, which should cover end-to-end operations and governance of ML, you often only see technology diagrams showing how different components connect. Businesses still have to figure out ML project management, put cost controls in place that don’t stifle experimentation, adopt a robust platform that satisfies all personas (e.g. data scientist, business analyst), makes model development process explainable and create model and data governance frameworks.
Drenik: What are some of the considerations when building a data and model governance framework for enterprise ML?
Robinson: In addition to data quality, data governance requires an understanding and development of best practices for metadata management, storage, integration, interoperability and securing access to data, amongst other things. Reviewing the Data Management Body of Knowledge (DMBoK) and similar materials are a good place to start to see what a comprehensive data governance framework looks like. For models, companies need to govern how they development, validate, promote to production, remove from production, inventory and retire the models. Marketing-specific industry-wide standards on model governance is not well developed at the moment but there are materials advertisers can use like Ernst & Young’s Model Risk Management for AI & ML, Model AI Governance Framework from Singapore’s PDPC or Amazon Web Services’ Machine Learning Best Practices in Financial Services as starting points.
Drenik: Those are great starting points, are there other considerations leaders should be thinking about when considering use of ML to drive business value?
Robinson: Developing a clear way to categorize ML projects as it relates to ability to execute and expected ROI, before starting projects, is extremely helpful.
Drenik: Can you speak to some approaches?
Robinson: To start, leaders can put ML projects into one of three buckets based on their desired business outcome; efficiency improvements, automating decision making or innovation. For example, an advertiser can use ML to make the ad-viewing experience more efficient, resulting in average uplifts of 10.5% for brand awareness, 19.2% for ad recall, 9.7% for brand perception and 10.3% for recommendation intent according to Cedato. Advertisers can also use ML to automate creation of ad creative, copyright and variations based on past campaign performance, JP Morgan Chase did this and saw as high as a 450% lift in click-through rates on ads when compared to human-created ads. Innovative ML projects transform old business models to help companies enter new markets resulting in the most significant ROI. For an advertisement firm, this could be changing from promoting other business’s products online to selling a subscription to 3D printers consumers embed in their homes which uses ML on consumer behavioral data to predict and print customized items as users need them. New market, more revenue, significant ROI.
Drenik: That sounds futuristic and amazing. Our Prosper Media Behaviors & Influence survey has shown that digital media has a very high influence on purchasing across a wide range of categories and, influence to purchase leads to ROI. With the larger ROI, shouldn’t everyone just focus on innovative projects?
Prosper – Media Influence to Purchase
Prosper Insights & Analytics
Robinson: Innovating your business model is most transformational but also the most challenging to undertake. Companies that are less mature with ML should start with efficiency improvements or automating decisions, which is where most organizations currently invest and are able to see success.
Drenik: Agreed. Well this has been extremely helpful, any parting words for those on the journey to successfully adopt ML and create governance structures for their models or data?
Robinson: Adopting, governing, and scaling ML is still an emerging area for many organizations so don’t stress if you don’t get everything right initially. The process is iterative but definitely well worth the pursuit as validated by McKinsey who showed companies that fully absorb AI can double their cash flow and companies who don’t will see a 20% decline by 2030. Apply a little patience, proper governance and in no time you’ll have a system in place to effectively use ML to drive desired business outcomes successfully.
Drenik: Well said. Jamal, thanks for your insights on how AI/ML tools can help create personalized relevant and assistive experiences for both consumers and marketers. Prosper Insights & Analytics suite of models on AWS (make the word a link) is an example of using the right type of data with AI/ML tools for optimizing and improving targeting at scale.
Source: https://www.forbes.com/sites/garydrenik/2021/06/17/importance-of-data-governance-and-mlops-when-using-machine-learning-to-drive-successful-business-outcomes/
