Read this article about Charter Communications, Finalist in the Best Achievement in Process Automation (RPA, Intelligent Automation, Digital Process Automation, Machine Learning, Cognitive Learning, Blockchain etc), selected by the independent judging panel, for the 2021 Digital Transformation & Operational Excellence Industry Awards program.
Identity Graphing for Network Genomes (Node Genome Model)
“Our Network is in a constant state of change. Nodes as well as various other network components are continually being added and reconfigured to optimize network performance. The Node Genome Model employs a unique process combining novel data engineering techniques, a multi-stage Ensemble ML Algorithm, and a Graph Data Structure that provides the capability to maintain state, learn from previous iterations of itself, and track changes over time.
1. The Node Genome Model employs data engineering techniques that leverage Natural Language Processing and Unsupervised Learning to parse out and process valuable components of the data as well as to determine any underlying structure. Specifically, -
o The data engineering employs natural language processing, geographic and string distance calculations, and fuzzy matching to deal with unstructured text, mismapped ids, and duplication.
o We develop a learned corpus of delimiters and invalid designations based on word and character frequency that are used for data cleaning, parsing, and purging.
o We employ a K-Means Clustering Algorithm to estimate Node Counts that can be used as a sanity check and enable a machine-learned grading mechanism to evaluate the quality of our data processing.
2. We also utilize a multi-stage training approach to probabilistically assign Node Counts and Node Names. Our training techniques take advantage of Ensemble Iterative Modeling to first identify Node Counts or the number of available slots while a second model is trained on the validity of potential Node Names that are available to be assigned to those slots.–
o Our Graph data structure will leverage Machine Learning Models to assign Node Counts, score our Node Count selections, and then probabilistically assign Node Names to the available Node Count slots.
o Training Stage 1: After leveraging natural language processing and data engineering to prepare our data we employ AutoML to select the most performant Multi-Label Classifier to help predict Node Counts for each CMTS + MacDomain combination in the Network. Our current top performing algorithm is a Light Gradient Boosted Trees Classifier.
o Training Stage 2: Our second phase trains a classifier to score the validity of candidate Node Names by bringing in other various sources to determine if a candidate name has been seen before. This will allow our model to learn from previous generations of itself as well as the flexibility to incorporate the addition of new data sources/assets.
3. Finally, the Node Genome Model combines a Graph Data Structure with ML probability scores to construct and maintain a digital Network Structure over time. This innovative data structure enables us to detect changes to Network as well as prove the capability to summarize and run calculations across the daily views of the network.
If we can train the Node Genome Model to learn from previous iterations of itself and enable it to automatically detect and alert due to anomalies in its structure then we are taking steps towards AI”
Digital Transformation & Operational Excellence Industry Awards 2021