Google AI Proposes Temporal Fusion Transformer for Multi-Horizon Time Series Forecasting Google AI Proposes Temporal Fusion Transformer for Multi-Horizon Time Series Forecasting
Time series forecasting is a useful data science tool for helping people predict what will happen in the future based on... Google AI Proposes Temporal Fusion Transformer for Multi-Horizon Time Series Forecasting

Time series forecasting is a useful data science tool for helping people predict what will happen in the future based on historical, time-stamped information. Google researchers recently explained how they developed and used the company’s Temporal Fusion Transformer (TFT) to achieve more progress with these types of predictions. Here’s a breakdown of what the company and its research teams have accomplished with this temporal fusion transformer.

What Does Google’s Temporal Fusion Transformer Offer?

Google’s Temporal Fusion Transformer is an attention-based deep neural network. That means it mimics how the human brain can focus on certain bits of information while filtering out others. Such events happen when a person concentrates on the words spoken by a friend sitting next to them while eating in a crowded cafeteria.

Another example could read people’s lip movements to dramatically increase speech-recognition capabilities in noisy environments. The researchers clarified that innovation was 75% more accurate than the best other options that use sound and lip movements to analyze speech. It’ll be interesting to see further developments in the coming years in this area. 

The TFT also splits processing into two parts when dealing with heterogeneous time-series data. It provides local processing that deals with specific event characteristics. Global processing then handles the collective attributes of all the associated time-series events. Moreover, it allows multi-horizon forecasting by predicting events in steps rather than giving only one overall conclusion. 

Interpretability is another advantage of the TFT. This aspect lets people gauge the importance of certain features found in the information. According to the abstract of a research paper written by Google’s team, “The TFT also utilizes specialized components for judicious selection of the relevant features and series of gating layers to suppress unnecessary components — enabling high performance in a wide range of regimes.”

How Does Google’s TFT Perform? 

Benchmark tests also showed that Google’s TFT performed better than traditional statistical models and even surpassed other deep learning neural models. One potential limitation is that the TFT might work differently in real-life situations than in laboratory settings. 

That’s always a risk with all kinds of technological progress. However, as the researchers expand its usage, they should learn valuable takeaways that allow them to adapt their methods and accommodate for any uncovered shortcomings. 


How Might People Use It?

There are plenty of possible scenarios that could benefit from time-series forecasting. Take the health care industry, for example. American state laws mandate one nursing staff member for every 20 residents at Pennsylvania long-term care facilities, while in Maine, the nurse-to-resident ratio differs depending on the time of day. Facility administrators could use time-series forecasting to determine their future staffing needs. 

Google’s researchers also provided real-life examples of using its TFT for multi-horizon time-series forecasting in a blog post. In one case, the team looked at how several factors affect store sales. There were static shop and item variables, and promotional periods and national holidays were the largest-weighted future variables. 

The team also looked at periods surrounding the 2008 financial crisis, choosing data spanning from 2002 to 2014 taken from the S&P 500 Index. The model paid increased attention to periods of high volatility and showed more normalized attention levels when the situation indicated more stability. Such outcomes show how the TFT could help data scientists cut through the noise and recognize truly significant events versus blips with little or no meaning. 

What Does the Future Look Like for Google’s Temporal Fusion Transformer?

The researchers closed their blog post with a brief mention of how retail and logistics companies have already used Google’s TFT approach to improve demand forecasting. If more entities follow suit, we should soon see more examples of how the TFT could help.

Google’s team also suggested TFT could play an important role in mitigating future climate change challenges. For example, they said it might assist with real-time supply and demand balancing of the electricity grid to reduce emissions. Alternatively, it might lead to improved accuracy in rainfall forecasting, improving preparedness for people living in areas of concern. 

 Google only published its findings in December 2021, so it’s too early to say how much of a prolonged impact the research team’s work will have. However, the examples given in the paper are undoubtedly promising. People in all areas of business and society are always looking for better ways to make accurate predictions. Google’s TFT algorithm could become an indispensable resource for some of them.

April Miller

April Miller

April Miller is a staff writer at ReHack Magazine who specializes in AI, machine learning while writing on topics across the technology sphere. You can find her work on ReHack.com and by following ReHack's Twitter page.