Our next post in the series on how customers are gaining actionable insights on their data through the power of Microsoft advanced analytics – at scale and in the cloud.
Scientists and researchers log on to Mendeley’s web site to access reference content and manage their personal libraries at one of the world’s largest repositories of academic articles. The site also helps them discover and collaborate with fellow researchers, create groups for common research goals and author new scientific work. You can learn more about Mendeley's unique service through their “Getting Started” video below:
Engaging Researchers Through Predictive Analytics
The more users there are on Mendeley’s network, the more useful their network becomes. To meet their goal of serving as many academics and thinkers as possible, Mendeley wanted to know more about their users with a view to tailor their experiences to their unique usage patterns and needs. Specifically, they wanted to perform predictive modeling on historical user behavior data to forecast user intention and likely future activity.
Mendeley’s marketing team found that certain user behaviors – if they occurred within the first two weeks of adoption of the product – correlated with increased long-term usage of their site. For instance, if a scientist opened a PDF document on more than one device, or if a researcher looked for and used certain features on the interface, they were much more likely to find value in the network and get hooked on to the service. Armed with information of this nature, Mendeley was eager to build predictive models that could make accurate recommendations for boosting user activity, turning window shoppers into loyal customers who were successfully able to extract more value out of the service.
Being a startup with an infrastructure built primarily on open-source, Mendeley ran into some not-so-uncommon problems around predictive modeling – for instance, they specifically encountered challenges around:
Creating models at scale.
Effectively supporting collaboration between data scientists working at different locations.
Automated deployment of their models.
In practical terms, these challenges translated into greater “time to market” – creating and deploying an effective analytical model took them much longer than they desired. Thus, while making data-driven improvements to the site was deemed essential, doing so was bogging them down relative to other urgent product development goals.
Azure ML Rises to the Challenge
Given this situation, as Fernando Fanton, Senior Vice President of Product Development at Mendeley puts it, it was a no-brainer to try Microsoft Azure ML once it became available: “We welcome any tool that can help us be more productive and reduce lead time, and Azure ML has unique capabilities,” he said.
Working closely with Microsoft’s data science team, Mendeley was able to quickly implement the new technology and start using it almost immediately. With no hardware to procure, no software to install and no complex development environments to configure, Mendeley was really able to accelerated the process of getting their predictive solution deployed – this despite the fact that the tool was brand new to them. Although Mendeley uses various open-source technologies, Azure ML was operational in moments. “The beauty of Azure Machine Learning is that it integrates in a decoupled fashion,” says Fanton. “Whatever your infrastructure, Azure Machine Learning can deliver value.”
Mendeley’s data scientists found Azure ML easy to use, given that it supported the complete end-to-end workflow right from accessing and cleaning data to modeling and deployment. As Lili Tcheang, a Mendeley data scientist, said, "With the drag-and-drop functionality of Azure ML, we are able to set up and test multiple models in parallel." The cloud environment further simplified their iterative process, allowing experiments to be published as web services that were readily accessible to other team members.
Fanton says, “There is a perception that Azure ML doesn’t work without a Microsoft stack. But Azure ML is very refreshing because it is completely in the cloud, and it plays nicely with all the other tools. Plus, it has REST endpoints, which are common for the web, so there is nothing proprietary you need in order to use it. If we can do it, pretty much anybody can.”
Finding the Right Answers Right Away
Within weeks of deploying their solution, Mendeley was able to improve their predictive model to achieve 30 percent better recall. “That, for us, is where the value is,” says Fanton. “This improvement makes the model much more meaningful in how we target users or change features in the product. We want to build better capability and quickly put it in front of our customers.”
The new model is helping Mendeley tailor functionality to more users. “In a company like ours that is growing so rapidly, that is a material difference,” says Fanton. “We are able to identify and target significantly more users, where before they could have gone unnoticed.”
Data scientists at Mendeley are now able to build, share and deploy models more collaboratively and efficiently than ever before. Product development and marketing are able to make realize an accelerated return on such experiments and investments.
And what of the future? As Fanton says, “We have an almost never-ending list of use cases for machine learning. We’ll be able to create specific metrics and datasets, and we’ll be able to do it much faster and more efficiently with Azure ML.”
By optimizing their platform to best serve the needs of their customers, Microsoft Advanced Analytics is helping Mendeley fulfill its mission of advancing research all around the world.
ML Blog Team