Tuning options for term assignment

Depending on the data in your project and the use cases for metadata enrichment, you might want to improve term assignment results by tuning some term-assignment options.

Various factors affect the accuracy of term assignments, for example, the data that is analyzed, the term-assignment algorithms that are used, or how well the machine learning model is trained. So sometimes, algorithms can suggest too many terms because the analyzed data is not uniform. Dissimilar values can also result in too many matches for data classes and thus to an disproportionately long list of term suggestions. Sometimes, data classes can be too generic so that they are assigned to too many columns, which also results in many term suggestions.

On the other hand, algorithms might come up with the same term suggestions, which makes a suggestion more trustworthy.

If you see too many business terms suggested for individual assets or columns or if you want to give extra weight to terms that are suggested by more than one algorithm, you can work with these tuning options. However, finding the right settings is an iterative process. You will probably need to test different settings to find out which ones work best for your use case.

After you toggle the Tuning options setting to On, you can select from different tuning presets or create a custom configuration based on your needs. If you select the Disabled option, no tuning configuration is applied. This selection is equivalent to switching tuning options off.

Tuning options

Tolerance for suggestions

Define how many suggested terms above the average number of terms suggested by all selected term-assignment algorithms are allowed for a single algorithm before the confidence scores are reduced. Specify the percentage for this reduction with the Lowering factor for the confidence score of surplus suggestions setting.

For example, you might want to limit the number of suggestions to not exceed the average by more than 4. The average number of term suggestions is 5. If an algorithm now returns 10 or more suggestions, you can reduce the confidence of term number 10 and any further terms that are suggested by that algorithm.

Lowering factor for the confidence score of surplus suggestions (%)

This value is used for calculating a lower confidence score for any term suggestions that exceed the tolerated number of suggestions. Set this number with the Tolerance for Suggestions value.

If one of the selected algorithms suggests more terms than are suggested on average by all selected algorithms, you can limit the number of tolerated suggestions and then reduce the confidence of further terms that might be suggested by that algorithm. This reduction might lead to the score falling below the specified suggestion or assignment threshold.

For example, you might want to reduce the confidence score of term suggestions that exceed the tolerated number by 5%.

Maximum number of suggested data classes without impact on term confidence

Define how many data classes can be suggested for a column before term confidence is recalculated. If the number of suggested data classes exceeds the specified value, the confidence score for any term that might be assigned based on the assigned data class is reduced by the value calculated based on the value for Lowering factor for the confidence score of class-based term assignments.

If a lot of data classes are matched to a single column, the number of suggested terms can become high. You can limit the number of data classes that are considered for term assignment without changes to the data quality score.

For example, you might want to accept terms from 4 data classes at maximum. If more data classes match the column, the terms that are tied to these data classes get a lower confidence.

Lowering factor for the confidence score of class-based term assignments (%)

The specified value is used for recalculating the confidence score for terms that might be assigned based on data class assignment if the number of suggested data classes for the same column exceeds the threshold. This threshold is defined in Maximum number of suggested data classes without impact on term confidence.

To filter out terms that might result from too many data-class matches, you can reduce the confidence score of terms that are defined on any surplus data classes. Thus, the confidence of such terms might fall below the defined suggestion or assignment threshold.

For example, you might want to reduce the confidence score of terms that are suggested based on data-class matches beyond the specified limit by 10%.

Raising factor for the confidence score of suggestions (%)

This value is used for calculating a higher confidence score for terms that were suggested by more than one of the selected algorithms.

For example, you might want to raise the confidence score for a term that is suggested by 2 or more algorithms by 10%.

Presets

Balanced

This preset is aimed at balancing precision and recall of term assignments. You might want to start with this preset to get a baseline for further tuning.

Precision focused

This preset is aimed at reducing false positives.

If one of the selected algorithms suggests 4 terms more than are suggested on average by all selected algorithms, the confidence of further terms that might be suggested by that algorithm is reduced by 10%. This reduction can help filter out irrelevant terms.

To filter out terms that might result from too generic data classes, a maximum of 4 data class suggestions is considered. The confidence score of terms that are defined on any surplus data classes is reduced by 10%.

For any term that is suggested by more than one of the selected algorithms, the confidence score is increased by 5%.

Recall focused

This preset is aimed at selecting all relevant terms.

Learn more