Data Sampling and Error Rates in Selected Audience Manager Reports
A summary of the sampling methodology used for some reports, sampling error rates, and a list of reports that return information based on sampled data.
Data Sampling Ratio
Some Audience Manager reports display results based on a sampled set of the total amount of available data. The sampled data ratio is 1:54. For reports that use sampled data, this means your results are based on 1 record out of every set of 54 records.
These reports use statistical sampled data because they need a tremendous amount of computing power to generate results. Sampling helps strike a balance between reduced computational demands, maintaining system performance, and providing accurate results.
Error Rates
Errors can occur in reports that generate overlap data. An error is defined as the percentage of records that:
- Should not have been included in a report but were added anyway.
- Should have been included in a report but were left out.
It’s important to note that our tests and models show that the error rate decreases in an inverse proportion to the number of records in your data set. Data sets that have a lot of records generate fewer errors than sets with a small number of records. Let’s look at this assertion in a more quantitative manner. As shown in the following table, for a set number of records, 95% of your report results will be below a specific error rate.
Using the Minhash Sampling Methodology
Based on the Minhash sampling methodology, Audience Manager uses a novel method to compute trait and segment estimators on top of a One Permutation Hashing data sketch. This new method produces a lower variance than the standard estimator for Jaccard similarity. See the section below for the reports that use this methodology.
Reports That Use Sampled Data
The Audience Manager reports that use statistical sampled data and the Minhash sampling methodology include: