A machine learning “pipeline” – similar to the pipeline for traditional business intelligence – is emerging. It should, in theory, enable developers to pick components off of a “Chinese menu” of building blocks, writes George Gilbert, big data & analytics analyst at the research firm Wikibon, a sister company of SiliconANGLE.
This pipeline (left) has four basic links: ingest, explore, process, transform and analyze-and-predict. In the abstract, they are similar to the steps in traditional business intelligence. However, they don’t map completely.
In BI, for example, the ingest stage involves a complex, extract, transform and load, or ETL, process to support answering pre-determined questions. In the machine learning process, however, data is loaded directly into a data lake in its native format.
This moves much of the complexity to the exploration phase, where the data must be rationalized, enriched and its sources and destinations identified. Processing refines the data. Finally, predictive analytics is the least mature part of the pipeline and the part that provides most of the business value.