Forecasting Customer Activities with Dilated Convolution Neural Networks: Use Case and Best Practices

This will be presented in English.

Tao Lu (Microsoft), Chenhui Hu (Microsoft)
11:1511:55 Thursday, June 20, 2019
模型与方法 (Models and Methods), 英文讲话 (Presented in English)
Location: 多功能厅6A+B (Function Room 6A+B)

必要预备知识 (Prerequisite Knowledge)

1. Basic understanding of time series forecasting 2. Basic understanding of neural network concept and deep learning modeling languages such as Keras or PyTorch

您将学到什么 (What you'll learn)

1. Deep understanding of the customer activities forecasting user case 2. Best practice for developing a dilated convolution neural networks for time series forecasting 3. How to achieve fair comparison with different algorithms

描述 (Description)

Forecasting user activities is one of the most common problems that business groups face. Various industries, such as cloud solution providers and online retailer, often need to forecast product consumptions and user engagement. The data scientist team under Azure Identity faces same challenges. With billions of user activities on the Azure cloud, being able to accurately forecast user behavior becomes essential in business planning. In this talk, we focus on forecasting the customer usage in daily basis. In this user case, statistical time series models and machine learning approaches like tree-based models failed to generate forecasts with satisfactory accuracies. Recently developed deep learning models could boost the performance yet they are hard to implement.
We will introduce how we improve 25% of forecasting accuracy with dilated convolutional neural networks (CNNs) and reduce 80% of the time in development with the best practices of time series forecasting. Dilated CNNs are recently proposed for modeling sequence data. They can achieve state-of-the-art performance in time series forecasting and are easier to train compared with recurrent neural networks. To speed up the model development process, we leverage the best practices of time series forecasting that another Microsoft team has created. This is a framework that provides standard workflows of various methods, e.g., statistical methods, traditional machine learning methods, and recently developed deep learning approaches. It also includes reusable utility functions for data standardization and feature engineering. With such utility functions, we convert our dataset into a standard format. Then, we implement the Dilated CNN model and other baseline models by following the templates in the framework. Moreover, we tune the hyperparameters of each model with HyperDrive in Azure Machine Learning to achieve a fair comparison of model performance in terms of accuracy, running time, and cost.

Photo of Tao Lu

Tao Lu


Tao Lu is a Data Scientist in the Cloud and AI organization at Microsoft. He has strong background in applying machine learning and deep learning techniques to forecasting problems. He has deep domain knowledge in cloud identity and financial services industry. He graduated from University of Washington with a master degree in Computational Finance.

Photo of Chenhui Hu

Chenhui Hu


Chenhui Hu is a Data Scientist in the Cloud AI organization at Microsoft. His current interests include retail forecast, inventory optimization, IoT data, and deep learning. He received his PhD degree from Harvard University with his PhD thesis focusing on biomedical imaging data mining. He also has research experience in wireless networks and network data analysis. He is a recipient of the third IEEE ComSoc Asia-Pacific Outstanding Paper Award. 

Leave a Comment or Question

Help us make this conference the best it can be for you. Have questions you'd like this speaker to address? Suggestions for issues that deserve extra attention? Feedback that you'd like to share with the speaker and other attendees?

Join the conversation here (requires login)


Picture of Tao Lu
2019-03-05 15:58 CST

Yes we’ll present in English.

011071b6 3e8a9859 | DATA SCIENTIST
2019-03-03 07:50 CST

Would none chinese audience be able to understand the talk?