Home » Risorse » Case History » Identificare le tendenze di acquisto dei consumatori: il caso studio PepsiCo
Case History Risorse

Identificare le tendenze di acquisto dei consumatori: il caso studio PepsiCo

tendenze consumatori pepsi

La statunitense PepsiCo, una delle aziende leader nel settore alimentare e delle bevande, aveva l’esigenza di identificare le tendenze di acquisto dei consumatori in migliaia di negozi degli Stati Uniti.

Per raggiungere questo obiettivo ha implementato Microsoft Azure Machine Learning ottenendo così informazioni utili dai suoi vasti repository di dati. In questo caso studio raccontiamo i benefici ottenuti da PepsiCo grazie all’implementazione di Azure Machine Learning.

PepsiCo uses Azure Machine Learning to identify consumer shopping trends and produce store-level actionable insights

As one of the world’s leading food and beverage companies, PepsiCo needs to balance consumer demand for its products with inventory on hand—in thousands of US stores. To move to a more machine learning, predictive analytics approach, the company deployed Microsoft Azure Machine Learning to gain actionable insights from its vast repositories of data, and PepsiCo has started using it for select markets in the United States. As the company learns from this pilot test and the models, its goal is to expand more broadly to other sectors and markets. In the pilot, PepsiCo field associates received daily lists of priorities for each store they visit, so they can better predict the snacks that individual stores need to stock to better meet the shopping needs of consumers. The data scientists generating the lists use the machine learning operations capabilities in Azure Machine Learning to streamline the process of developing and working with the models that inform those priorities. As a result, early signs estimate a shift of 4,300 days of work a year from routine tasks to value-added activities.

While refreshing Pepsi beverages, tasty Frito-Lay snacks, and wholesome Quaker foods drive the consumer-facing side of the business at PepsiCo, within the company, it’s information that keeps things moving.

“Data is the lifeblood of the company,” says Michael Cleavinger, Senior Director of Shopper Insights Data Science and Advanced Analytics at PepsiCo. “We have 23 billion-dollar brands across multiple product segments. We rely on insights from machine learning to bring together our knowledge of the industry, the market, and our in-depth understanding of the shopping habits and preferences of consumers. It enables us to make informed decisions that ensure consumers get the products they want, helping us consistently meet consumer demand and drive growth for PepsiCo.”

Part of that business growth depends on positive relationships with PepsiCo’s technology partners, and data plays a key role there, too. “Our retail partners operate with a daily mindset,” explains Evan Shaver, Vice President, Shopper Analytics and Insights at PepsiCo North America. “The days of getting them week-old or month-old data are largely gone. We believe that the vendors who can respond nimbly will be the ones that retailers trust to help them win in the marketplace.”

Adds Jeff Swearingen, Senior Vice President and Head of Global Demand Accelerator at PepsiCo, “The key is having insightful, meaningful data. In some ways, the proliferation of data makes decision making harder than ever. The ability to analyze data, quickly glean actionable insights, and convert those to compelling, customized programming is at a premium. Our retail partners, more than ever, value this capability.”

Finding new ways to collect and analyze data

PepsiCo generates huge amounts of data every day about sales volumes, inventory, and purchase patterns. The company wants to do more than just spend time organizing that data in traditional time-consuming ways—like maintaining it in thousands of separate spreadsheets, PDF files, Word documents, PowerPoint decks, databases, and emails, as the company did in the past. PepsiCo would rather devote time to focusing on big business questions and determining smart strategies.

“We want to level up what we’re doing with data,” says Cleavinger. “We’ve had a pretty traditional approach, doing things the same way for the past 20 years. With the advancement of technology, we started to look for alternative methods to harness computing power and gain valuable insights into critical business processes. That led us to AI and machine learning.”

PepsiCo moved its voluminous data to the Microsoft Azure cloud platform to bring the company’s many data sources together and look at the contents in more constructive ways, starting with implementing Azure Machine Learning. “We wanted to take a very granular, intelligent, forward-thinking approach, and Azure Machine Learning opened up a lot of possibilities for us,” says Cleavinger. “You can use whatever language you like, and whatever algorithms you like, and the platform supports you from beginning to end.”

Accelerating machine learning development with new tools

A typical machine learning life cycle involves a number of steps that repeat in a process of ongoing refinement. Data scientists develop models that help computers learn from past data to identify future outcomes. The scientists then train these models with existing datasets, package them, validate the behavior of the model for responsiveness and any required regulatory compliance, deploy the model to the cloud where it processes incoming data, monitor the model for behavior and business value, and then retrain or replace the model as necessary to improve performance. The entire process can be structured as part of a continuous integration and continuous delivery (CI/CD) pipeline. 

“In PepsiCo, we use Azure as a modern, next-generation advanced analytics platform. We started with foundational capabilities of Azure and worked closely with our business partners to add more advanced services and features,” says Atul Jain, Director, Advanced Analytics at PepsiCo. “No doubt, the machine learning operations capabilities in Azure Machine Learning have facilitated better coordination between the PepsiCo business team and PepsiCo IT teams. We have been able to move faster and stronger with our business partners by opening capabilities to the business team, creating a new way of working.”

Because machine learning has such a specific, well-defined process, it doesn’t always fit smoothly into traditional DevOps tools. For this reason, PepsiCo was pleased to find that using Azure Machine Learning provides access to the service’s machine learning operations (MLOps) capabilities, which are specifically designed for machine learning development.

“Microsoft has built Azure Machine Learning to support a machine learning–specific approach to DevOps, so our people can focus on agility,” says Cleavinger. “Our data scientists and developers easily monitor everything from a central portal, and the ability to manage, validate, and deploy models from one environment drastically decreases time spent, makes difficult processes easier, and reduces workflow complexity.”

The data science team at PepsiCo has also found that taking this machine learning operations approach to CI/CD using Azure Machine Learning has made the team much more effective, as members are able to iterate through machine learning development more easily than before.

“We’ve used the MLOps capabilities in Azure Machine Learning to simplify the whole machine learning process,” says Cleavinger. “That allows us to focus more on data science and let Azure Machine Learning take care of end-to-end operationalization. This makes a real difference, because we have people on our team who aren’t experienced developers. With Azure Machine Learning and its MLOps capabilities, our data scientists and data analysts get an entry point that traditional DevOps tools don’t provide.”

Using the machine learning operations capabilities in Azure Machine Learning for task automation not only helps PepsiCo focus more on data science, but it has also sped up development and cut time spent on routine tasks. It also helps developers and data scientists maintain accuracy and repeatability of models by ensuring that they aren’t making inadvertent mistakes. “Automation helps us lift and shift a process from one project to another, so that the general structure is already complete, and we know what to expect and what needs to be changed,” says Cleavinger. “This can remove weeks of work from a machine learning project.”

Gaining great benefits with sophisticated, easy-to-use features

PepsiCo employees appreciated that adopting the machine learning operations capabilities in Azure Machine Learning didn’t require a big shift in the way they’re used to working. “We essentially removed a lot of things we didn’t use often and now spend more of our time where we really want to—building, deploying, and monitoring models,” says Cleavinger. “Because the MLOps capabilities in Azure Machine Learning keep everything together in one spot, we don’t need to move between programs and platforms. And features like the rich model registry make our pipeline more straightforward and speed iterations.”

The company found other Azure resources helpful too. “With Azure Machine Learning, we can integrate our machine learning workflows into automated pipelines, scale those up, and have them launch new managed ‘computes’ for the entire process,” says Cleavinger. “It really helps automate the process and makes it easier for us to share code and models.”

For PepsiCo data scientists, the adoption of Azure Machine Learning and its machine learning operations capabilities provides a framework for working with big data and AI in a way that fosters collaboration, speeds time to market, and reduces the hassle associated with file sharing, file tracking, and versioning. “With the machine learning operations capabilities in Azure Machine Learning, several teams can work together in a coordinated way,” says Ying Feng, Senior Manager, Data Science, Marketing/Shopper Insight at PepsiCo. “Each team can work on part of the solution and then merge them together very smoothly. It saves us a lot of time. In the past, it could take a year to get a model to a production level, and now we can finish the process in as little as four months.”

Digging into stores’ DNA to better meet customer expectations

PepsiCo’s field associates visit more than 200,000 US retail locations each week to stock and organize products and displays. Traditionally, the field associates arrived at a store with a long list of things that needed to be done based on intuition and past years’ historical data. But they didn’t always know what to prioritize for that given week. And because each store caters to a variety of consumers across differing regions, the same guidelines that would apply to one store might not apply to another. Priorities are also affected by things like major sporting events, school schedules, and even weather.

“For example, when our field associate arrived at a store in North Texas recently, it was 85 degrees,” explains Cleavinger. “He wasn’t aware that the forecast for the next week was 105 degrees. If he had been, he would have known to stock up on water. We want to use machine learning to integrate and analyze data from both internal and external sources—like weather forecasts—to provide that sort of predictive advice, quantify the value of the recommendations, and customize priorities for each store.”

The priorities take into consideration PepsiCo’s internal information about its shoppers and their buying patterns. Different stores may sell different product amounts, and the company wants its field associates to know about these differences and set their expectations for what is likely to sell from week to week. The data challenge is huge, involving billions of rows and thousands of columns of data, and it incorporates demographics, attitudes, psychographics, purchase behavior, product affinity, local events, weather, and more.

To bring all this data together and provide appropriate recommendations, PepsiCo developed an application called Store DNA. The company used Azure Machine Learning and its machine learning operations capabilities to create machine learning models that analyze store and customer data and give field associates a prioritized list of the top three product actions to take when they reach each location, like stocking more of particular products. The company is also scaling up this granular data to identify broader areas of opportunity where there is potential unmet consumer demand, helping PepsiCo to shape strategy and focus areas and work toward the ultimate goal of meeting consumer needs.

“Now we can use data to tell our field associates which products will have the biggest impact that week, what the associates need to watch out for, and what they can do to make sure they really meet customer expectations—things they might not have been aware of before,” says Cleavinger. “They can better capitalize on sales opportunities thanks to insights drawn via machine learning from our vast data warehouses. Due to the scale and scope of the data, we wouldn’t have been able to even contemplate doing this sort of work until recently, with tools like Azure Machine Learning and its MLOps capabilities.”

Expanding the solution and delivering granular insights

PepsiCo did a Store DNA proof of concept in the North Texas market covering 700 large-format retail stores and received positive feedback from its field associates. They acted on more than 85 percent of the application’s recommendations, and the company determined that it was improving predictions by more than 40 percent. Based on the success of the North Texas rollout, the company is expanding Store DNA to 14 markets in four different regions in the central United States. Each market has its own machine learning models, trained to the market’s idiosyncrasies. 

“Using Azure Machine Learning and its MLOps capabilities made it easy for us to scale our solution to the additional markets,” says Feng. “Our process uses two models per market—one for training and one for scoring—so we are currently tracking 28 models across the deployment.”

The Store DNA rollout heralds efficiency and productivity gains for PepsiCo. Based on the proof of concept, early signs indicate that the company may be able to shift 4,300 days of work a year from mundane tasks to activities more centered around adding value to the business—that’s nearly 12 years’ worth of tedious work transformed into meaningful time spent. Field workers no longer spend time trying to find out or derive what needs to happen at a store, and they can focus on the work that creates the most impact. Data scientists are potentially saving weeks of time using the machine learning operations capabilities in Azure Machine Learning and the task automation it supports. This automation helped change the dynamic between the business and IT groups within PepsiCo. Says David Patron, Sr. Director Enterprise Data, “Azure Machine Learning automated a lot of IT processes, allowing us to focus on strategy with the business team. But beyond this, it also reduces the time to produce and deliver machine learning/AI solutions across new projects.”

The granular insights provided by Store DNA benefit not just field workers, but also PepsiCo executives. “We give our executives great dashboards modeled in Microsoft Power BI that provide insight into where the company should be spending time and effort,” says Cleavinger. “They can drill down into items in a specific store, and also in the surrounding stores, or scale up to see where opportunities lie and make appropriate high-level strategic decisions.”

With insights derived from Azure Machine Learning, PepsiCo gains three benefits that every company values greatly. “We place a real emphasis on delivering both speed and quality, while reducing costs,” says Swearingen. “Historically, this has been a difficult trifecta to achieve. However, we believe our new capabilities will shift the curve to enable a new, better baseline.”

As PepsiCo dives deeper into machine learning, the company is making sure that it adheres to responsible AI policies. “We use a consumer-centered approach to AI to better understand and meet customer needs,” explains Cleavinger. “With the power of AI, we can generate valuable insights without having to use personally identifiable information. We also have data governance councils to ensure that we use and secure data appropriately and that we understand the data provenance.”

At the moment, Store DNA focuses on PepsiCo’s beverage products, but as the company rolls out the solution to more geographic areas, it plans to add its snacks as well. “We’ll continue our efforts to save time for our people in the field and give them better, deeper insights,” says Cleavinger. “The support we’ve gotten from Microsoft as we’ve embarked on our machine learning journey has been great, and we look forward to continuing our work together to increase what we like to call our ‘share of stomach’ and our business success.”

Questo contenuto è stato gentilmente fornito da Microsoft



Clicca qui per inserire un commento


per ricevere aggiornamenti sui trend e le opportunità IOT per il tuo business

Podcast & Smartspeaker

IOTtoday su Spotify IOTtoday su Google Podcast IOTtoday su Apple Podcast IOTtoday su Amazon Alexa IOTtoday su Google Home