Title: Lite-AI: Learning from Sparse and Incomplete Data
Funding: This project is sponsored by Ericsson Research, Silicon Valley, USA.
Start Date: July 1, 2020.

Project Description

Ericsson Research and the AI4Networks Research Center at the University of Oklahoma are collaborating to address the data scarcity/sparsity challenge in mobile networks for enabling AI based zero touch automation.

AI/ML based solutions are being perceived as inevitable to enable the much-needed automation of the increasingly complex, difficult and expensive to manage emerging mobile networks. Some common examples of AI/ML driven telco use cases include but not necessarily limited to: user mobility prediction, traffic (load) forecasting, modeling relationship between system level key performance indicators and network configuration parameters, anomaly detection, root cause analysis (diagnosis), network parameters optimization, ML modeling conventional optimization approaches etc. The telco domain use cases fueled by machine learning algorithms require large amounts of labeled real training data as presence of more data usually results in better and accurate models. Hence availability of pertinent rich telco domain data sets is pre-requisite for achieving impactful gains from AI/ML based solutions. However, unlike image or natural language processing datasets that required only collecting and the labeling readily available data, the creation of relevant telco domain data sets is a far more challenging task. The factors that make this task challenging include:

1. Scarcity Challenge: Given that collection of real labeled data requires advanced domain (expert) knowledge and the operators cannot share this data openly for reasons that range from technical difficulty of identifying and extracting data trapped in silos to protecting user privacy and trade secrets, the telco labeled data is general scarce.
2. Sparsity Challenge: Although the big data in telco network is known for its Volume, however it has generally far less distinct sample points. The main reason being that the operators only try a limited range of configuration parameters in live network and non-uniform data probing (user) distribution, the real data is expected to be sparse or unevenly distributed.

Due to above reasons, the real data in Telco domain remains too scarce and when available too sparse to unlock the full potential of AI for mobile network automation.

Project Goals

To investigate/identify promising approaches to tackle data scarcity/sparsity challenge in context of different use cases in mobile networks and develop solutions that have:

1. Scalability to generate high dimensional data.
2. Robustness to highly dynamic real environments.
3. Capability to take conditional context of required network conditions into account.