Data Quality

Search…
Data Quality
TL;DR: The report provides a detailed dataset overview.
Calculates base statistics for numerical, categorical and datetime features
Displays interactive plots with data distribution and behavior in time
Plots interactions and correlations between features and target
The report works for a single dataset or compares the two.
Summary
The Data Quality report provides detailed feature statistics and a feature behavior overview.
It can also compare any two datasets. You can use it to compare train and test data, reference and current data, or two subgroups of one dataset (e.g., customers in different regions).
Requirements
If you want to run this report for a single dataset, you need to prepare a pandas.DataFrame or csv file with features you want to explore. Pass it as reference data.
If you have a datetime column and want to learn how features change with time, specify the datetime column in the column_mapping parameter.
If you have a target column and want to see features distribution by target, specify the target column in the column_mapping parameter.
To compare two datastes, you need two DataFrames or csv files. The schema of both datasets should be identical.
Feature types (numerical, categorical, datetime) will be parsed based on pandas column type. If you work with csv files in CLI, or want to specify a different feature mapping strategy, you can explicitly set the feature type using column_mapping.
The report contains the section that plots interactions between the features and the target. It will look slightly different for classification and regression tasks. By default, if the target has a numeric type and has >5 unique values, Evidently will treat it as a regression problem. Everything else is treated as a classification problem. If you want to explicitly define your task as regression or classification, you should set the task parameter in the column_mapping object.
You can read more to understand column mapping and data requirements for Evidently reports in the corresponding sections of documentation.
How it looks
The default report includes 3 widgets. All plots are interactive.
1. Summary widget
The table gives an overview of the dataset, including missing or empty features and other general information. It also shows the share of almost empty and almost constant features. This applies to cases when 95% or more features are missing or constant.
2. Features widget
For each feature, this widget generates a set of visualizations. They vary depending on the feature type. There are 3 components:
2.1. Feature overview table
The table shows relevant statistical summaries for each feature based on its type and a visualization of feature distribution.
Example for a categorical feature:
Example for a numerical feature:
Example for a datetime feature:
2.2. Feature in time
If you click on "details", each feature would include additional visualization to show feature behavior in time.
Example for a categorical feature:
Example for a numerical feature:
Example for a datetime feature:
2.3. Feature by target
Categorical and numerical features include an additional visualization that plots the interaction between a given feature and the target.
Example for a categorical feature:
Example for a numerical feature:
3. Correlation widget
This widget shows the correlations between different features.
3.1. Insights
This table shows a summary of pairwise feature correlations.
For a single dataset, it lists the top-5 highly correlated variables from Cramer's v correlation matrix (categorical features) and from Spearman correlation matrix (numerical features).
For two datasets, it lists the top-5 pairs of variables where correlation changes the most between the reference and current datasets. Similarly, it uses categorical features from Cramer's v correlation matrix and numerical features from Spearman correlation matrix.
3.2. Correlation heatmaps
This section includes four heatmaps.
For categorical features, Evidently calculates the Cramer's v correlation matrix. For numerical features, Evidently calculates the Pearson, Spearman and Kendall matrices.
If your dataset includes the target, the target will be also shown in the matrix according to its type.
JSON Profile
If you choose to generate a JSON profile, it will contain the following information:
1
{
2
  'data_quality': {
3
    'data': {
4
      'cat_feature_names': [],
5
      'datetime_feature_names': [],
6
      'metrics': {
7
        'current': {
8
          'feature_name': {
9
            'count': count,
10
            'feature_type': 'num',
11
            'infinite_count': 0,
12
            'infinite_percentage': 0.0,
13
            'max': max,
14
            'mean': mean,
15
            'min': min,
16
            'missing_count': 0,
17
            'missing_percentage': 0.0,
18
            'most_common_value': most_common_value,
19
            'most_common_value_percentage': most_common_value_percentage,
20
            'percentile_25': percentile_25,
21
            'percentile_50': percentile_50,
22
            'percentile_75': percentile_75,
23
            'std': std,
24
            'unique_count': unique_count,
25
            'unique_percentage': unique_percentage
26
            },
27
          },
28
        'reference': {
29
          'feature_name': {
30
            'count': count,
31
            'feature_type': 'num',
32
            'infinite_count': 0,
33
            'infinite_percentage': 0.0,
34
            'max': max,
35
            'mean': mean,
36
            'min': min,
37
            'missing_count': 0,
38
            'missing_percentage': 0.0,
39
            'most_common_value': most_common_value,
40
            'most_common_value_percentage': most_common_value_percentage,
41
            'percentile_25': percentile_25,
42
            'percentile_50': percentile_50,
43
            'percentile_75': percentile_75,
44
            'std': std,
45
            'unique_count': unique_count,
46
            'unique_percentage': unique_percentage
47
            }
48
          }
49
        },
50
        'num_feature_names': [],
51
        'target_names': None,
52
        'utility_columns': {
53
          'date': 'dteday',
54
          'id': None,
55
          'prediction': 'prediction',
56
          'target': 'target'
57
          }
58
        },
59
      'correlations': {
60
        'reference': {
61
          'pearson': {
62
            'feature_name_1': {
63
              'feature_name_2': value,
64
              'feature_name_3': value
65
              },
66
            'feature_name_2': {
67
              'feature_name_1': value,
68
              'feature_name_3': value
69
              },
70
            'feature_name_3': {
71
              'feature_name_1': value,
72
              'feature_name_2': value
73
              }
74
            },
75
          'spearman': {
76
            'feature_name_1': {
77
              'feature_name_2': value,
78
              'feature_name_3': value
79
              },
80
            'feature_name_2': {
81
              'feature_name_1': value,
82
              'feature_name_3': value
83
              },
84
            'feature_name_3': {
85
              'feature_name_1': value,
86
              'feature_name_2': value
87
              }
88
            },
89
          'kendall': {
90
            'feature_name_1': {
91
              'feature_name_2': value,
92
              'feature_name_3': value
93
              },
94
            'feature_name_2': {
95
              'feature_name_1': value,
96
              'feature_name_3': value
97
              },
98
            'feature_name_3': {
99
              'feature_name_1': value,
100
              'feature_name_2': value
101
              }
102
            },
103
          'cramer_v': {
104
            'feature_name_4': {
105
              'feature_name_5': value
106
              },
107
            'feature_name_5': {
108
              'feature_name_4': value
109
              }
110
            }  
111
          }
112
        }  
113
    'datetime': '2022-02-22 16:35:15.529404',
114
    'name': 'data_quality'
115
    },
116
  'timestamp': 'timestamp'
117
}
Copied!
When to use this report
Here are a few ideas on how to use the report:
1.Exploratory data analysis. You can use the visual report to explore your initial training dataset and understand which features are stable and useful enough to use in modeling.
2.Dataset comparison. You can use the report to compare two datasets to confirm similarity or understand the differences. For example, you might compare training and test dataset, subgroups in the same dataset (e.g. customers from Europe and from Asia), or current production data against training.
3.Data profiling in production. You can use the report to log and store JSON snapshots of your production data stats for future analysis. You can combine this with testing data distributions for drift using Data Drift report.
4.Production model debugging. If your model is underperforming, you might use this report to explore and interpret the details of changes in the input data.
Data Quality Report Examples
Browse our example notebooks to see sample Reports.