2-minute read

Search engines serve customers across a wide range of demographics. Businesses interested in improving their search engine performance can work with technology companies behind these search engines to analyze search data, find potent keywords, and refine their ad content strategy. For the search engine providers, this improves client relationships and provides insights into user behavior.

We recently completed a project in this space, providing a Fortune 100 technology company’s team with a new tool to increase analytics ROI.

The Problem

Though they were getting a high volume of client requests, the company’s advertising group was spending far too much time manually classifying search data. They were frequently using spreadsheet tools to import, sort, and label data by hand, and using “contains” queries and “if” statements to classify it. Misspelled words could be misclassified, and there was no way to reuse successful steps on later datasets. Significant effort was being expended with limited, sometimes flawed results.

The team knew that with machine learning a more efficient and reliable process was possible, and they turned to Logic20/20 to create a solution.

The Vision

Our client’s goal was clear: create a custom classification management tool to classify data on-demand, employ machine learning technology to expedite the process, and provide data analysis. This would save the Sales team time and help them more easily select effective ad categories for their clients, increasing revenue and freeing up resources. Thanks to our previous work with the client and on similar machine learning projects, the team trusted us to evaluate the situation and propose a tool that would meet their needs.

Making it Happen

Our team had a lot of factors to consider. The tool needed to be efficient, but also flexible and user-friendly. Not only that, it would be used by stakeholders across the United States and potentially other countries, so we needed to accommodate a wide range of users.

We began by interviewing local team members, then constructed short demos and solicited feedback. This process went back and forth as we refined our tool, considering multiple types of machine learning before finally settling on supervised learning. We considered unsupervised learning, but ultimately decided to implement supervised learning because users prefer to create their own taxonomies. We are considering using unsupervised learning in the future to help the client isolate data from a larger volume of raw queries. This would help them create training datasets.

We refined both the machine learning algorithm and infrastructure surrounding it. Our team used the concepts of microservices and continuous integration to create a responsive, comprehensive, user-friendly solution. We used Microsoft Azure, Kafka, Python, Kubernetes, Jenkins, and Git.

The Tool We Created

We delivered a custom machine learning tool and user-friendly web interface, enabling the team to easily create, reuse, and iteratively improve supervised machine learning models to classify datasets. We also created a centralized, accessible repository to store the models and classification results, providing single-source access for team members across the country.

Our tool allows technical and non-technical users to label data, place it into a repository, and run machine learning algorithms to create a new model that mimics human labeling behavior. The user can then select this model and classify new data with it. Classification models like these can be easily and continually improved, making the models a constantly evolving tool that can be rapidly adapted to new trends in the market.

As for data cleanliness, misspellings are no longer a problem. The machine learning creates frequencies based on spelling and can detect similar but misspelled items and label them correctly.

The Results

Since adopting our tool, the client’s data classification time has decreased, while data quality and visibility have increased. Processing time has decreased from weeks to hours, and the new tool delivers better results for less effort. The sales team can now spend more time on high-value tasks, and data visibility across the organization has improved.

Thanks to the overwhelming success of the project, we continue to manage the team’s backend tool and front-end user interface, as well as expand functionality by adding new features. We look forward to continuing our relationship with this client.

Anna Emmett is a manager in our Architecture practice with 12 years of experience in systems and business analysis.