Adtrac’s customers require precise predictions of customer flows to efficiently sell campaigns to advertisers. This is particularly difficult because the number of people often varies greatly on a seasonal basis. That is why we have developed a new approach to make the predictions more accurate.

Our approach focuses on encoding calendar_dt into new features using the density function of a Gaussian mixture of three normal distributions. Based on this curve, we can create features that represent how spring-like, how summer-like, how autumn-like, and how winter-like a given date is. As a result, days that are close together will have similar features, while days that are far apart will look quite different. This effect then helps the neural network to recognize and predict such seasonal patterns.

Another advantage of this approach over conventional methods is that less data is required

 

For more details, see our blog post on Medium: https://medium.com/@raphael.schoenenberger_95380/encoding-temporal-features-part-2-440336afae13