Blog

Frequently Asked Questions about Machine Learning

By Carmel Dibner

Q: How is Applied Marketing Science’s approach to machine learning different from other approaches?

A: Two reasons: First, our approach to machine learning is cutting-edge and proven. Second, our approach combines the power of machine learning and human analysis.

Developed by researchers at MIT, Applied Marketing Science’s approach rests on rigorous academic research. You can read the academic article about the approach here.

Machine learning works using an algorithm to identify key insights in the data. First, we train the algorithm to distinguish between informative content that contains an important customer insight and uninformative content that does not. The algorithm uses convolutional neural networks (CNN) and clustering to return a sample of the dataset that it has identified as containing an accurate representation of all the insights in the database. Our algorithm identifies hidden opportunities for product development that are frequently missed by others. We’ve tested it on a range of categories, from vehicles to fast-moving consumer products.

In addition, our approach combines machine learning with thoughtful human analysis. This is important for ensuring the results are insightful and actionable. Machine learning uses algorithms, but algorithms are not a black box. Humans must train the algorithm to identify what’s important and unimportant to each study. This way, you benefit from what machines and humans do best: the machine is systematic and can sift through a large volume of data, and humans train the machine to focus on the right topics. But, human involvement doesn’t end there. Highly trained analysts then convert the output of the machine into a full database of insights.

Q: What kind of research questions are best answered with machine learning?

A: Machine learning at its core is a data reduction technique. It can analyze thousands, even millions of records. Large data sets that you want to explore are the best source for machine learning. Therefore, we recommend machine learning for a company that has broad questions to answer. For example, what needs exist in the marketplace? What insights in our category could we have missed with prior research? What insights can we gather from adjacent categories that might help inform our innovation strategy?

Q: I’d like to use machine learning for broad exploration of the category, but there is a range of topics I know I want to explore. How can I be sure those topics will be covered?

A: We often recommend that clients brainstorm a list of topics that they’d like to make sure are covered by machine learning. Then, we can identify data sources that cover each of the topics. We might mine for a separate forum thread that covers each of the different topics of interest in detail to ensure that we gather insights on all topics of interest.

Q: What kind of sources of data work well for machine learning?

A: Large sources of text-based data you want to investigate are the best sources. For example, public data sources include product reviews on e-commerce sites, product review sites, and online discussion forums. In contrast, proprietary data could include answers to open-ended survey questions or customer call center data.

Q: Do clients typically suggest the sources of data to mine or does AMS suggest them?

A: Often clients have a few sources of data in mind at the start of the project. AMS reviews those sources, assesses the research topics you’d like to explore, and suggests a list of other sources to include in the analysis. If you don’t have sources in mind, AMS will recommend a list.

Q: I have call center data, but it’s in the form of notes, not transcripts. Can I still use machine learning?

A: Yes, while transcripts are ideal, we’ve successfully used call center notes in the past. The key is that the notes are sufficiently detailed.

Q: I have call center data in the form of audio recordings. How can I use machine learning?

A: Audio recordings can be inexpensively transcribed and then the machine processes transcripts of the recordings.

Q: What is the advantage of having AMS do the analysis as an external third-party vs. purchasing a commercially available tool?

A: Our machine learning process is based on academic research and has been rigorously tested. Working with AMS as an external third party, we can handle the full end-to-end research process. We can identify the sites to mine based on our experience, mine the data, clean the data to prepare it for the machine, train the dataset to optimize the power of machine learning, and run the machine. Once the machine has been run, we take the output and our trained analysts convert it into detailed insights. We also present the findings in a rich qualitative report with quotations that bring the insights to life. Our output is similar in richness to a full qualitative primary research report but takes substantially less time and effort compared to a traditional research project.

Q: How often should I refresh a machine learning project?

A: This depends on several factors, including industry dynamics. If your industry is fast-changing, you might want to refresh the data every six months. If it’s slower changing, you may want to update the data every 1-3 years. Also, if there is a major change in the market such as the entry of a new player or a game-changing new product, you may want to update the data more frequently.

Q: How can I efficiently scale machine learning across my company?

A: Once you’ve conducted a machine learning project, each subsequent project requires less time and effort to complete as the machine has already been trained in the category. So, adding additional sources of data, refreshing the data or other similar requests are faster and more cost efficient.

Learn more about cutting-edge consumer-research methods in our one-day Consumer Insights for Innovation course.

Find out more

 

Machine Learning Big Data

Share this: