Accuracy Report

Table name:
{{table_name}}
Report generated at:
{{time}}
Training parameters:
    {% for training_parameter_key, training_parameter_value in training_config.items() -%}
  • {{training_parameter_key}} = {{training_parameter_value}}
    • {% endfor %}
Inference parameters:
    {% for inference_parameter_key, inference_parameter_value in inference_config.items() -%}
  • {{inference_parameter_key}} = {{inference_parameter_value}}
    • {% endfor %}

Synthetic Data Quality

Median accuracy: {{accuracy_value}}

Bad
< 0.7
Low
0.7 - 0.8
Medium
0.8 - 0.9
High
> 0.9

The accuracy heatmap is calculated in the following way: for each pair of original columns a similarity between them (the Jensen-Shannon distance) is measured. The same operations is performed on the same pair of columns in the synthetic dataset. The resulting score is calculated as 1 - abs(original_score - synthetic_score).

Content

Correlations
Clustering
Utility
Univariate distributions
Bivariate distributions

Correlations

Median of differences of correlations {% if correlation_median < 0.0001 %} is less than 0.0001 {% else %} : {{round(correlation_median, 4)}} {% endif %}

GOOD
BAD

The correlation metric is calculated in the following manner: for each pair of numeric columns in original dataset a pairwise correlations is measured thus resulting in the correlations matrix. The same operation is performed for the synthetic dataset. The resulting matrix is the difference between original correlation matrix and the synthetic one.

Clustering metric

For clustering metric, both original and synthetic datasets are concatenated. After this, a clustering using K-means is performed on the concatenated dataset. The optimal number of clusters is chosen using the elbow rule. For the good synthetic data the proportion of original to synthetic records in each cluster should be close to 1:1. The mean clusters homogeneity is calculated as a mean of ratios of original to synthetic records in each cluster.

Median clusters homogeneity: {{clustering_value}}

Utility metric

{% if is_data_available %}

For utility calculation an automatic modeling process is engaged. For every binary, categorical and continuous column it builds a model using given column as a target and other columns as predictors. Then the best target in terms of accuracy is chosen. For each of the best targets for different modeling tasks a model is trained on the original dataset, then tested on synthetic one. For the good synthetic data the score of model on the generated data should be close to the original one.

{{table_name}}

{{utility_table}}
{% else %}

No data available

{% endif %}

Univariate distributions

The univariate distributions are either the density plots for original vs synthetic values overlapped for continous columns or barplots for original vs synthetic counts (in percent) for categorical columns.

{% for col, uni_img in uni_imgs.items() -%}

Column Name: {{col}}

{% endfor %}

Bivariate distributions

Bivariate distributions are actually the two-dimensional histograms. For each pair of columns, each combination of their corresponding values are plotted as a matrix, resulting in a grid. Each cell within the grid contains a ratio, computed as the number of records belonging to both classes simultaneously divided by the total number of records in corresponding dataset. For example, consider a column with values 'Monday', 'Tuesday', 'Sunday', and another column with values 'Working', 'Museum visiting', 'Walking'. In this case, the ratio representing the occurrence of each activity on each day is calculated and displayed in the appropriate grid cell. For good synthetic data the pattern of bivariate distribution should match the original distribution.

{% for title, bi_img in bi_imgs.items() -%}

Original: {{title}}

Synthetics: {{title}}

{% endfor %}