Data Quality & Preprocessing


Exploratory Data Analysis & Unsupervised Learning

     

Lecturer: Dr. HAS Sothea

📋 Outline

  • Data Sources

  • Data Quality

  • Data Preprocessing

  • Real Examples

Data Sources

Data Sources

Primary

  • Data collected directly from the source for a specific purpose.
  • Example:
    • Surveys or Questionnaires 🗳️
    • Interviews 🎙️
    • Observations 🧐
    • Experiments 🔬

Secondary

  • Data that has already been collected, processed, and made available by others.
  • Example:
    • Government publications or reports 📄
    • Books and articles 📚
    • Online databases and repositories 🌐
    • Industry/NGO reports 🏭

Data Sources

Format

Structured

  • Highly organized and easily searchable in databases using predefined schemas.
  • Structure: typically stored in tables with rows and columns.
  • Example:
    • Spreadsheets: Excel
    • CSV files

Unstructured

  • Lacks a predefined format or schema and is typically stored in its raw form.
  • Structure: Free-form and can be text, images, videos…
  • Example:
    • Emails/Documents (e.g., Word files, PDFs)
    • Social media posts, images, audio, videos
    • Web pages…

Data Quality

Data quality

  • Someone in 60s said Garbage In, Garbage Out (GIGO)!.
  • Data quality is the most important thing in Data Analysis.

Data quality

Data quality

  • Timeliness: how up-to-date the data is for its intended use.
  • Ex: Temperature of 60s wouldn’t help forecasting tomorrow.

Data quality

  • Uniqueness: data should not be duplicated.
  • Ex: Recording the same female heart disease patient many times may lead to a conclusion that females have a higher likelihood of developing heart disease.

Data quality

  • Validity: data should take values within its valid range.
  • Ex: Height & weight should not be 0 nor negative!

Data quality

  • Consistency: data should be uniform and compatible (format, type…) across different datasets and over time.
  • Ex: 15/03/2004 & 03/15/2004, Male & M…

Data quality

  • Accuracy: data should be accurate and reflects what it is meant to measure.
  • Ex: You cried and said ‘I like Data Analysis Course SO MUCH 😭!’ when I ask nicely (😠).

Data quality

  • Completeness: data should not contain missing values.
  • Ex: In Lab1, Cabin column of Titanic dataset contains mostly NaN.

Data quality

  • Data quality includes these 6 factors.

Data quality

  • Data quality includes these 6 factors.
  • If there is a problem with any of these, you may ☝️
  • For secondary sources, Incompleteness is the common one.

Data Preprocessing

Data preprocessing

Consider an example

Sex Pclass Fare Cabin
0 male 3 7.8292 NaN
1 female 3 7.0000 NaN
2 male 2 9.6875 NaN
3 male 3 8.6625 NaN
4 female 3 12.2875 NaN
5 male 3 9.2250 NaN
6 female 3 7.6292 NaN
7 male 2 29.0000 NaN
8 female 3 7.2292 NaN
9 male 3 24.1500 NaN
10 male 3 7.8958 NaN
11 male 1 26.0000 NaN

Data preprocessing

Consider an example

Sex Pclass Fare Cabin
0 male 3 7.8292 NaN
1 female 3 7.0000 NaN
2 male 2 9.6875 NaN
3 male 3 8.6625 NaN
4 female 3 12.2875 NaN
5 male 3 9.2250 NaN
6 female 3 7.6292 NaN
7 male 2 29.0000 NaN
8 female 3 7.2292 NaN
9 male 3 24.1500 NaN
10 male 3 7.8958 NaN
11 male 1 26.0000 NaN 
data.dropna(inplace=True)

Data preprocessing

Consider an example

Sex Pclass Fare Cabin
0 male 3 7.8292 NaN
1 female 3 7.0000 NaN
2 male 2 9.6875 NaN
3 male 3 8.6625 NaN
4 female 3 12.2875 NaN
5 male 3 9.2250 NaN
6 female 3 7.6292 NaN
7 male 2 29.0000 NaN
8 female 3 7.2292 NaN
9 male 3 24.1500 NaN
10 male 3 7.8958 NaN
11 male 1 26.0000 NaN 
data.drop(columns = ['Cabin'])

Data preprocessing

Missing values

  • Data of \(4\)-\(7\) years old kids.
Gender Age Height Weight
F 68 0 20
F 68 0 18
F 65 105 0
F 63 0 15
F 68 112 0
F 66 106 0
  • What’s wrong with this data?

  • These are probably missing values (NA, nan, NaN…) in disguise.

  • Question: how do we handle it: Drop or Impute?

  • Answer: we should at least know what kind of missing values are they: MCAR, MAR or MNAR?

Data Preprocessing

Missing values

Missing Completely At Random (MCAR)

  • They are randomly missing.
  • Easy to handle with imputation or dropping methods if not so many.
  • They don’t introduce bias.
  • Removing them does not affect other columns.
  • Ex: Missing values do not affect Age. They are MCAR in this case.
Code 
import plotly.graph_objects as go
import plotly.express as px
from plotly.subplots import make_subplots
data_dropped_NA = data_kids.loc[(data_kids.Height > 0) & (data_kids.Weight > 0)]
fig_kid1 = go.Figure(go.Histogram(
    x=data_kids.Age, 
    name="Before dropping NA", 
    showlegend=True))
fig_kid1.add_trace(
    go.Histogram(
        x=data_dropped_NA.Age, 
        name="After dropping NA", 
        showlegend=True, 
        visible="legendonly"))
fig_kid1.update_layout(barmode='overlay', 
                       title="Distribution of Age", 
                       xaxis=dict(title="Age"),
                       yaxis=dict(title="Count"),
                       width=500,
                       height=400)
fig_kid1.update_traces(opacity=0.5)
fig_kid1.show()

Data Preprocessing

Missing values

Missing At Random (MAR)

  • The missingness is related to other columns.
  • One can try using those related columns to impute.
  • If not too many, model-based imputation often work well: KNN…
  • Ex: Most of the missing values are from female children. They are MAR in this case.
Code 
count = data_kids.Gender.value_counts()
fig_kid2 = go.Figure(
    go.Bar(
        x=count.index, 
        y=count, 
        name="Before dropping NA"))
count_NA = data_dropped_NA.Gender.value_counts()
fig_kid2.add_trace(
    go.Bar(x=count_NA.index, 
    y=count_NA, 
    name="After dropping NA", 
    visible="legendonly"))
fig_kid2.update_layout(barmode='overlay', 
                       title="Distribution of Gender", 
                       xaxis=dict(title="Gender"),
                       yaxis=dict(title="Count"),
                       width=500,
                       height=400)
fig_kid2.update_traces(opacity=0.5)
fig_kid2.show()

Data Preprocessing

Missing values

Missing Not At Random (MNAR)

  • These are the trickiest, as the missingness is related to the missing values themselves.
  • It may require domain-specific knowledge or advanced techniques (more data, external info…).
  • It’s hard to judge if missing values are actually MNAR.
  • Ex: Very high or very low salaries are often missing from a survey if it’s optional.
  • If not so many, dropping is a common solution.

Data Preprocessing

Rules of Thumb

Proportion of NA Rules of thumb
\(< 5\%\) Drop/remove rows.
\(5-10\%\) Can be dropped but must be cautious about the type of missing.
\(10-20\%\) Better to be imputed according to their types.
\(20-30\%\) Remove the entire column, if it’s not so important.
\(>30\%\) Remove the entire column.

Data Preprocessing

Outliers

  • Data points that deviate significantly from the majority of observations in a dataset.
  • It can influence our analyses: insightful or problematic!
  • We can hunt them down using:
    • Graphs: Scatterplots, Boxplots or histograms…
    • They often fall outside \([\text{Q}_1-1.5\text{IQR},\text{Q}_3+1.5\text{IQR}]\).

Data Preprocessing

Handling outliers

  • Not all outliers would affect the analysis.
  • We can apply capping (limiting outliers to some values) or Trimming (completely remove them).
  • Some transformations may reduce the effect of outliers:
    • Z-score: \(x\to \frac{x-\overline{x}}{\sigma_{x}}\) (centered by mean, scaled by std).
    • Min-Max scaling: \(x\to\frac{x-\min}{\max-\min}\in [0,1]\).
    • If the data are positive: \(x\to \log(x)\) or \(x\to \sqrt{x}\)
  • No absolute solution! It depends on the analysis.

Data Preprocessing

One-hot encoding

  • This is taken from gapminder dataset (2007).
Code 
from gapminder import gapminder
import numpy as np
from sklearn.preprocessing import OneHotEncoder as onehot
encoder = onehot()
encoded_data = encoder.fit_transform(gapminder.loc[gapminder.year == 2007, ['continent']]).toarray()

# encoded dataset
X_encoded = pd.DataFrame(encoded_data, columns=[x.replace('continent_', '') for x in encoder.get_feature_names_out(['continent'])])
df_encoded = X_encoded.copy()
df_encoded['lifeExp'] = gapminder.lifeExp.loc[gapminder.year==2007].values
sorted_order = gapminder[gapminder.year == 2007].groupby('continent')['lifeExp'].median().sort_values().index
fig_cont = px.box(data_frame=gapminder.loc[gapminder.year==2007,:],
                  x="continent", y="lifeExp", color="continent",
                  category_orders={'continent': sorted_order})
fig_cont.update_layout(title="Life Expectancy vs Continent", height=250, width=500)
fig_cont.show()
  • Sometimes, categorical data are very informative and useful especially for building predictive models.
  • One-hot encoding is a way to covert them into numbers.
  • For example:
    [‘Asia’, ‘Europe’, ‘Africa’, ‘Africa’] 👇
Africa Americas Asia Europe Oceania
0 0.0 0.0 1.0 0.0 0.0
1 0.0 0.0 0.0 1.0 0.0
2 1.0 0.0 0.0 0.0 0.0
3 1.0 0.0 0.0 0.0 0.0

Real Example

Real Example

Titanic Dataset (891 rows, 12 columns)

Survived Pclass Sex Age SibSp Parch Fare Cabin Embarked
0 0 3 male 34.5 0 0 7.8292 NaN Q
1 1 3 female 47.0 1 0 7.0000 NaN S

Data types:

Survived Pclass Sex Age SibSp Parch Fare Cabin Embarked
0 int64 int64 object float64 int64 int64 float64 object object

Missing values:

Survived Pclass Sex Age SibSp Parch Fare Cabin Embarked
0 0 0 0 86 0 0 1 327 0
  • Question: What should we do in the preprocessing step?
  • Answer: We should:
    • Convert Survived and Pclass to be object.
    • Handle missing values: drop 1 NA of Fare, remove column Cabin and study Age.

Real Example

Titanic Dataset (891 rows, 12 columns)

  • Convert data types:
col_to_be_converted = ['Survived', 'Pclass']
for col in col_to_be_converted:
    data[col] = data[col].astype(object)
data[col_to_be_converted].dtypes.to_frame().T
Survived Pclass
0 object object
  • Drop column Cabin:
data.drop(columns = ["Cabin"], inplace = True)
  • Drop 1 row with NA in Fare:
data.dropna(subset = ['Fare'], inplace = True)
  • Study missing values in Age:
    • Impact on qual. columns:
Code 
import seaborn as sns
import matplotlib.pyplot as plt
sns.set(style="whitegrid")
fig, axs = plt.subplots(2, 4, figsize=(6, 3.5))
col_qual = ['Survived', 'Pclass', 'Sex', 'Embarked']
for i, va in enumerate(col_qual):
    sns.countplot(data, x=va, ax=axs[0,i], stat = "proportion")
    axs[0,i].bar_label(axs[0,i].containers[0], fmt="%0.2f")

    sns.countplot(data.dropna(), x=va, ax=axs[1,i] , stat = "proportion")
    axs[1,i].bar_label(axs[1,i].containers[0], fmt="%0.2f")
    if i == 0:
        axs[0,i].set_ylabel("Before remove NA")
        axs[1,i].set_ylabel("After remove NA")
    else:
        axs[0,i].set_ylabel("")
        axs[1,i].set_ylabel("")
plt.tight_layout()
plt.show()

Real Example

Titanic Dataset (891 rows, 12 columns)

  • Impact on quan columns:
Code 
sns.set(style="whitegrid")
fig, axs = plt.subplots(2, 3, figsize=(6, 3.5))
col_quan = ['SibSp', 'Parch', 'Fare']
for i, va in enumerate(col_quan):
    sns.histplot(data, x=va, ax=axs[0,i], kde=True)
    sns.histplot(data.dropna(), x=va, ax=axs[1,i], kde=True)
    if i == 0:
        axs[0,i].set_ylabel("Before remove NA")
        axs[1,i].set_ylabel("After remove NA")
    else:
        axs[0,i].set_ylabel("")
        axs[1,i].set_ylabel("")
plt.tight_layout()
plt.show()

  • Do you think that removing NA greatly affects other columns?
  • Study missing values in Age:
    • Impact on qual. columns:
Code 
import seaborn as sns
import matplotlib.pyplot as plt
sns.set(style="whitegrid")
fig, axs = plt.subplots(2, 4, figsize=(6, 3.5))
col_qual = ['Survived', 'Pclass', 'Sex', 'Embarked']
for i, va in enumerate(col_qual):
    sns.countplot(data, x=va, ax=axs[0,i], stat = "proportion")
    axs[0,i].bar_label(axs[0,i].containers[0], fmt="%0.2f")

    sns.countplot(data.dropna(), x=va, ax=axs[1,i] , stat = "proportion")
    axs[1,i].bar_label(axs[1,i].containers[0], fmt="%0.2f")
    if i == 0:
        axs[0,i].set_ylabel("Before remove NA")
        axs[1,i].set_ylabel("After remove NA")
    else:
        axs[0,i].set_ylabel("")
        axs[1,i].set_ylabel("")
plt.tight_layout()
plt.show()

Real Example

Titanic Dataset (891 rows, 12 columns)

  • As removing NA barely impacts other columns, we can
    • Drop rows with NA or
    • Impute with mean or median.
data.fillna(value = data[['Age']].mean(), inplace = True)
data.iloc[:,[1,2,4,5,6,7,9,10]].isna().sum().to_frame().T
Survived Pclass Sex Age SibSp Parch Fare Embarked
0 0 0 0 0 0 0 0 0
  • Age after imputation:

🥳 Yeahhhh….









Let’s Party… 🥂