ipoca: Using Vertex AI to build a highly accurate AI model for predicting visitor traffic in just one year
About ipoca
Established in 2007, ipoca Inc. (ipoca) has been providing services that connect people and regions to create new value in the realm of "real shops" and "mobile." The company develops an information marketing platform designed to connect people with the community. It utilizes smartphone-based behavior data that includes location information, online search results, and purchase history, to attract new customers and improve repeat buyers, enabling clients to optimize costs.
Tell us your challenge. We're here to help.
Contact usipoca Inc. builds a development and operation environment for its new service, Asushiru, for retail stores in just a year on Google Cloud, with a highly accurate AI model by leveraging Vertex AI and BigQuery.
Google Cloud results
- Launches a new service with an advanced AI model in less than a year, starting from scratch
- Constructs an efficient ML Ops environment to achieve more accurate predictions on visitor traffic
- Enabling the training of developers seamlessly in remote environments with extensive teaching materials and documentation
Efficiently builds advanced AI models
ipoca Inc. (ipoca) is on a mission to "unify" regions, services, information, and people, to revitalize shopping districts. The company develops marketing support services for physical stores, based on the 3Cs – Customer, Competitor, and Company. It provides analytics for stores, and recently introduced a new service called Asushiru.
"Retail stores, especially supermarkets and convenience stores that handle FMCG goods such as food with short expiration dates are troubled with sales losses caused by discounts, disposals, and general lack of goods. ipoca's new service Asushiru aims to solve that problem," says Masanori Yamamoto, Executive Vice President at ipoca.
Yamamoto goes on to add that Asushiru is a name for the whole digital transformation (DX) service provided to retail companies that leverage machine learning (ML). In addition to being able to predict demand to solve food waste loss and shortage based on a simple AI-based forecast of visitor traffic to the store, it also provides sales forecasts when opening new stores.
Google Cloud provides all the necessary functions in one package
The president of ipoca, Taku Ichinose, started development of Asushiru when he saw a store staff sadly disposing expired food at a supermarket. Daiki Danno, Business Development Leader and Asushiru Project Manager, explains why Google Cloud was adopted as the development and operation platform.
"Asushiru is our first AI service. We didn't have in-house AI engineers or data engineers, so we had a wide range of things to catch up on, such as understanding business domains, statistics and time-series models," shares Danno. "Google Cloud was an ideal choice because it was easy to use, enabling us to quickly build an environment from end-to-end in one platform like a package."
"Asushiru is our first AI service. We didn't have AI engineers or data engineers in-house, so we had a wide range of things to catch up on, such as understanding business domains, statistics and time-series models. Google Cloud was an ideal choice because it was easy to use, and we could quickly build and package it."
—Daiki Danno, Business Development Leader and Asushiru Project Manager, ipoca"In addition, since we used hundreds of millions of records of purchase data, the high-speed features of BigQuery, and pay-as-you-go rates became a critical factor that led us to the decision to use Google Cloud. It is said that data preprocessing and exploratory data analysis account for 80% of the effort in building an AI model, so processing performance is key. The ease of linking with Google Colaboratory was also attractive," says Danno.
He adds that they focused on the ease of learning, since there were student interns assigned to the project, as well as the company became fully remote since April 2020 due to the COVID-19 pandemic. Google Cloud made it possible with learning tools such as Qwiklabs, Coursera, and YouTube. For example, Hideto Furugen, a project intern, started with zero experience, but is now able to use various Google Cloud products, mainly BigQuery.
"When I joined ipoca as an intern, I was still in my first year of university and had absolutely no knowledge of AI. I was grateful that BigQuery was simple to use, as I could easily figure out what to do next. In addition to Qwiklabs and other teaching materials, the documentation was substantial, and it was easy to find, so I was able to proceed with my studies smoothly," says Furugen.
Developing an AI model that exceeds the performance of an experienced store manager with Vertex AI
In June 2021, development of Asushiru began by analyzing hundreds of millions of records of actual purchase data provided by partner companies. "First, we started by analyzing the characteristics of the data on BigQuery, and from there, we proceeded with the study of ML models by connecting to Google Colaboratory. After that, we moved the production environment to Vertex AI and started full-scale development of AI models," says Danno.
"The advantage of Vertex AI is that Vertex AI Workbench, which has a managed notebook environment with pre-installed PyTorch and TensorFlow, allows you to start experimenting right from your browser. I also like that Vertex AI Workbench is highly available and powerful, and that it can easily connect to BigQuery and Cloud Storage for speedy exploratory data analysis and feature generation."
—Daiki Danno, Business Development Leader and Asushiru Project Manager, ipoca"The advantage of Vertex AI is that Vertex AI Workbench, which has a managed notebook environment with pre-installed PyTorch and TensorFlow, allows you to start experimenting right from your browser. I also like that Vertex AI Workbench is highly available and powerful, and that it can easily connect to BigQuery and Cloud Storage for speedy exploratory data analysis and feature generation. Also, at this time, Vertex AI was added with AutoML Tables' time series forecasting function (Vertex AI Forecast) and BigQuery ML components, so we were able to easily create, experiment, and implement many models. The time-series model, which is the core of Asushiru and is important for demand forecasting, becomes outdated every day. Therefore, it is essential to build ML Ops, such as monitoring and re-learning to maintain accuracy. With the announcement of the lineup for these production environments, we were able to have an image of actually building and operating."
In November 2021, the AI model for predicting the number of visitors for Asushiru was completed. The company was able to achieve a remarkable result in making more accurate predictions than manual methods at partner companies' stores that received purchase data.
"Predicting the number of visitors to a store is an extremely important feature for placing orders correctly in a retail store. By enabling Asushiru to predict the number of visitors at the same level as an experienced store manager, even inexperienced store managers can make appropriate predictions and place orders. Additionally, we have realized a sales forecast function using a regression model for opening a new store, which requires a large amount of investment. In the future, we will further improve forecast accuracy with ML Ops. We also plan to take on the challenge of new initiatives such as natural language analysis, including customer word-of-mouth, image analysis of in-store products and the flow of people," says Danno.
Although there are still issues to be solved in order to achieve stable operation with ML Ops, Danno believes that the fact that he was able to obtain an "experimental environment" with the introduction of Google Cloud was a major achievement.
"By gaining an environment where we can quickly analyze the big data we already have, such as visitor traffic data from GPS, and purchase data on receipts, and create AI models, we will accelerate the development of new services to support DX. I feel that we have done well."
"By gaining an environment where we can quickly analyze the big data we already have, such as visitor traffic data from GPS, and purchase data on receipts, and create AI models, we will accelerate the development of new services to support DX. I feel that we have done well."
—Daiki Danno, Business Development Leader and Asushiru Project Manager, ipocaYamamoto says that the success of this initiative will lead to further expansion of initiatives that utilize AI and Google Cloud. "While we are working to solve marketing issues with Asushiru and miseshiru, we are starting to see the organizational and cultural issues we will face as we promote DX. We have also started developing a communication service that will improve the engagement between the stores and shift workers. We would like to explore effective approaches through the Google Cloud data analysis platform to find out what factors lead to job satisfaction and improved performance through HR surveys," says Yamamoto.
ipoca Inc.
・Masanori Yamamoto, Executive Vice President
・Hideto Furugen, student intern
・Daiki Danno, Business Development Leader
Tell us your challenge. We're here to help.
Contact us
About ipoca
Established in 2007, ipoca Inc. (ipoca) has been providing services that connect people and regions to create new value in the realm of "real shops" and "mobile." The company develops an information marketing platform designed to connect people with the community. It utilizes smartphone-based behavior data that includes location information, online search results, and purchase history, to attract new customers and improve repeat buyers, enabling clients to optimize costs.