How Lenovo uses text analytics for product quality and design
- 02 September, 2015 11:44
Social media, user forums, review sites, call centres and many other sources can offer a rich source of data to extract insights on what customers want and need from your product and service.
The problem is it’s unstructured, big and often messy. But that doesn’t mean it’s not manageable.
During his visit to Sydney this week to speak at a SAS event, Lenovo’s chief corporate analytics officer, Anthony Volpe, discussed how he and his team mined text for early detection on product quality issues, and to help shape their products based on customer wants and needs.
“When we do text mining, it’s not about whether people like us or not. It’s mining for specific insights. If it’s not for anything and everything, so for specific things, text mining is a much more manageable problem,” he said.
Lenovo’s product quality teams have many items they work on daily, but which ones are a priority, and what are the most common customer complaints? This is where Volpe found a strong case to build up his team’s skills and capability to carry out text mining to listen to the millions of customers voicing their problems around the world.
The team built a Lenovo Early Detection (LED) system to help product quality staff decide which issues to focus on first before they manifest into major problems, resulting in higher customer dissatisfaction rates, and possibly some brand damage.
“Before LED, we would sell customers our software and if customer satisfaction was bad, people would get mad at us, we would incur lots of costs – both real in terms of fixing them but also in terms of our brand.
“These things finally trickled through to the warranty claims system where somebody looks at warranty claims and says ‘wow, we must have a problem with G580 keyboards’ months too late.”
Volpe and his team applied a statistical process control (SPC) model on the data coming mostly from user forums and call centres where people provide greater detail on their problem.
“We know there’s randomness and there’s systemic migration from a desired need. SPC separates randomness from a systemic variation - that's where we want to be. We said ‘that sounds like our problem’,” he said.
“Who has ever done SPC on social media data?” he asked the audience. “My guess is nobody. But we did it.”
Volpe and his team looked at specific key words or products in comments -‘G580 keyboard’ instead of ‘Lenovo keyboards’, for example - along with the volume of comments. They then subset those down to comments related to quality. Thresholds were used to filter out ones that indicate an issue.
“Sometimes those things go out of control limits, [there'll be] some threshold that suggests we have a quality problem.
"If we start to see something coming out of the control [limits], the quality person who makes these decisions, we might recommend... 10 items today to go investigate for pervasive quality issues.”
Listening to how a product is performing, or how customers are interacting with it out in the field, is another main application of text mining at Lenovo, Volpe said.
“People say ‘I wish this had more USB ports’, ‘I wish the screen was brighter’,” he noted.
He also includes competitor names when mining text data, finding out which features most customers focus on when comparing Lenovo to others. Some examples of what he found include customers saying ‘I like my friend's Dell screen so much better’, ‘my friend’s screen is easy to clean, mine is always dirty’, and so on.
This can also lead to new feature or product ideas, said Volpe. “There are comments that say ‘why can’t I build my PC like I can use Moto Maker to build my phone?’”
One discovery Volpe and his team made was that a large volume of comments made on a particular laptop was in relation to gaming. Though the laptop was not designed intentionally for gamers, with text mining he discovered people were finding the laptop specifications ideal for gaming.
“So we actually figured out that if we marketed that machine to gamers, we would win. That’s a huge outcome of just listening. And we don’t just listen, we think about what we are listening to.”
Volpe said when carrying out text mining for product performance, he places comments into certain categories.
“We have four different people responsible for performance; there are different aspects of performance. So we are asking, which of these 25 million [comments] point to a performance issue that we attribute to vice president 1’s initiatives? Which ones go into vice president 2’s initiatives?”
Volpe said tapping into unstructured data is an “uncomfortable” area for a lot people, mainly because there aren’t lot of case studies or examples of how companies pull it off successfully. Skills and building capability is another issue.
“On my team I have a lot of mathematicians, statisticians, I have a physicist or two, some industrial engineers. They have never done text mining."
Furthermore, he said, there is no book of instructions or guide you can turn to for all the answers.
"I can’t go to ‘operations research’ textbook, chapter one, section six and read ‘here’s the formulation to optimise our inventory position’, or go to the metrics textbook and understand ‘here’s the type of forecast model I should have for lumpy demand'.”
Follow Rebecca Merrett on Twitter: @Rebecca_Merrett