Why use knowledge graphs?
Posted: Thu Feb 13, 2025 4:53 am
Mike Dillinger has a very direct take on the need for knowledge graphs to make AI work somewhat better:
“For computer and data scientists, one way to motivate the use of knowledge graphs is to position them as a way to overcome the many shortcomings of representing data and knowledge in relational databases and manipulating it with linear machine learning models.
One big, bad, and dramatic simplifying assumption of databases is that columns are treated as independent or orthogonal. Machine learning techniques like classifiers make the same assumption: there israel whatsapp number data are weights for each feature/variable but there are no terms to represent the covariance or interdependence between two or more features. The target classes for classifiers are also assumed to be disjoint or uncorrelated, which is why classifiers perform poorly in deciding between hierarchically related classes – they’re not disjoint, rather one subsumes the other. Making believe that variables are unrelated when they actually are related simply inflates error variance to intolerable levels.”
Also, from one of Dillinger’s slides: “ Because math is literally, intentionally, absolutely meaningless. And logic is, too.”
Making business impact is where it starts and where it ends.
AI must produce reliable propositions. Why not ask for certification?
More Informatics, Less Tech
The following is not a big issue, but unprecise terminology seems to have infected our “guild.”
“For computer and data scientists, one way to motivate the use of knowledge graphs is to position them as a way to overcome the many shortcomings of representing data and knowledge in relational databases and manipulating it with linear machine learning models.
One big, bad, and dramatic simplifying assumption of databases is that columns are treated as independent or orthogonal. Machine learning techniques like classifiers make the same assumption: there israel whatsapp number data are weights for each feature/variable but there are no terms to represent the covariance or interdependence between two or more features. The target classes for classifiers are also assumed to be disjoint or uncorrelated, which is why classifiers perform poorly in deciding between hierarchically related classes – they’re not disjoint, rather one subsumes the other. Making believe that variables are unrelated when they actually are related simply inflates error variance to intolerable levels.”
Also, from one of Dillinger’s slides: “ Because math is literally, intentionally, absolutely meaningless. And logic is, too.”
Making business impact is where it starts and where it ends.
AI must produce reliable propositions. Why not ask for certification?
More Informatics, Less Tech
The following is not a big issue, but unprecise terminology seems to have infected our “guild.”