AI models in the context of COVID: How to use machine learning without historical data
One of the biggest challenges business faced in 2020 was the very sudden disruption of supply chains and demand forecasting. Most people in the US can remember the effects when grocery store shelves began to empty as many began to stockpile everything from pasta to toilet paper.
Many of IBM’s enterprise clients felt the reverberations as well: demand forecasts began to break down as these models are built with traditional time series data. This type of historical data used in building forecasts had nothing like the patterns the pandemic produced. This is where IBM’s Global Business Services (GBS) AI Workflows team came in to help.
Clients had trouble in the early part of 2020 taking corrective actions in a market with such high demand volatility. In this post, I will discuss some of the approaches the GBS team took to solve this problem and how we helped drive key outcomes like more accurate demand forecasts, lower inventories and optimized operations for our clients.
Among the first things we created was a curated data lake and a series of analytical and AI accelerators. These accelerators can be leveraged as independent insights or they can be used to augment demand forecasting models. Our data lake consists of a variety of sources:
- COVID-19 statistics
- Policies and regulations such as information on government actions on quarantines and shutdowns
- News data
- Socio-economic data
- Weather data
- Location data
All these data sets are curated, pre-processed, and ready to go into any forecasting model.
To construct our models, we also bring in client data. This data varies by industry, but can include things like point of sale data, shipments data, and product and customer master tables. From all these data sets, we created a series of derived signals using deep learning models. One such model that we developed was a resurgence index, which is a predictive measure of where and when the next wave of the pandemic will hit.
Another model that the GBS team produced was the disruption index, where we measure the risk of doing business at a county level, which is reflected in the granularity of the model. This index gives a forward-looking view of economic disruption. We started with the premise that consumer demand is very strongly shaped by the local response to COVID. We created this model using a variety of external data, such as data on economic volatility and COVID, as well as economic and public health data and was built using an LSTM network with Python and R. This index gives us a way of quantifying the financial actions taken by consumers based on the local COVID conditions they are directly experiencing.
The consumer safety index is another index we built. This index identifies consumer behavior by looking at social media data. We wanted to look at this because even if retail stores are open in an area, we want to know the probability that consumers are going to feel safe to go shopping. One of the benefits of using social media data is that we do not have to rely on historical data, giving us a new method of building forecasting models with current data sources. Also, because this data is real-time, this technique can be adapted to many different sources of disruption beyond a pandemic, such as civil unrest or climate disruptions. After harvesting and processing the social media data, we loaded it into our algorithms and identified posts of interests using a nearest neighbor algorithm. After this, we used NLP transfer learning to identify the sentiment of a post and the probability of this sentiment. This enables clients to see across different contexts like emotion, consumer behavior, and location.
A graph of these results shows how the consumer safety sentiment changed over time as the case numbers increased. In January 2020, as the pandemic began, consumers were showing fear, and one post notes that the poster’s local Costco was very unusually empty. If you follow the model through until May, you will notice a rise in the sentiment, reflecting growing complacency and people were no longer wearing masks in stores.
We deploy these indices and models using IBM Cloud Pak for Data with cloud visualization that allows us to use both internal and external data sets. We combine both long-and-short-term forecasting which uses a method of time segmentation, and ensembling a milestone model and a time series machine learning model. We can then combine all these trends, get information about the consumption uplift, and present everything in a visualization layer. When we deploy this for clients, we have seen up to 95% accuracy — an up to 30% improvement over their current algorithms.