Supervised Learning without Explicit Rules

The conventional approach to detect implicit patterns is the development of rule sets in which every relevant pattern is represented by an explicitly coded rule. Rule-based methods suffer from the limitations imposed by: the strictness of the rules themselves, the need to estimate and tune weight values (a process that tends to devolve into guesswork), and the need for completeness in the rule set, which is often unachievable in practice given the variability and complexity of real data. Moreover, while the rules detect some genuine occurrences of the pattern that each rule represents, each rule is also a source of false positives. These are records that are reported as matches, but are not true matches. In general, the less rigid the rule, the greater the number of false positives it generates. Since rules must be coded directly, the addition of any new feature can involve time-consuming adjustments to the entire rule set.

The approach of TIBCO Patterns Machine Learning Platform is completely different. Rather than a programmer coding a large abstract rule set, a domain expert (who need not be a programmer, or even a technical user) is asked to make simple judgments of “Yes” or “No” when presented with a set of concrete examples. The saved "Yes" or "No" judgment is called a label of the feature vector.

The TIBCO Patterns machine learning algorithm uses these human judgments, together with the corresponding feature vectors, to generate a machine learning model that represents the “wisdom” of the human domain expert, without coding a single rule.

Since the scores output by TIBCO Patterns queries can be used as feature values, the “fuzziness” of features such as textual similarity is naturally taken into account.

The Learn model contains the result of the training. This takes into account all the dimensions of complexity identified in detecting implicit patterns, including the context-dependent relevance of features, and the way that human criteria change when features are missing.