Exploring Television Data With Spark - Part 2

From Questions to Visualizations
python
spark
machine learning
datasets
visualization
Published

January 15, 2023

Moving From Spark to Visualizations.

In the last post, we reviewed some data cleaning and how to use Apache spark. We’re going to move on to visualizations; while I like the data tables and numbers, when showing the conclusions to others then visualizations are the way to go. Numbers without context are meaningless and the reason they mean something to the individual making the analysis is that you have the data as context. But, those you’re showing the conclusinos too do not - and they didn’t spend hours cleaning and thinking about the data like you did.

We’ll add all the cleaning from before just for that context I was talking about; I would recommending seeing the part one of this series if your curious about that process.

# Define the schema:
shows = StructType([
    StructField('Ignore', IntegerType(), True),
    StructField('Name', StringType(), True),
    StructField('Year', StringType(), True),
    StructField('Guide', StringType(), True),
    StructField('Duration', StringType(), True),
    StructField('Genres', StringType(), True),
    StructField('Ratings', FloatType(), True),
    StructField('Actors', StringType(), True),
    StructField('Votes', StringType(), True),
])

data = spark.read.csv(
    str(Path(
        '_data',
        'AllThrillerSeriesListClean.csv')),
        header=True,
    schema=shows)
# Cleanup
data = data.drop('Ignore')

data = data.withColumn('isOngoing',     # Target this column
    F.udf(                              # define a udf
        lambda x: x.strip()[-1] == '–') # use the logic we already wrote.
    ('Year')                            # pass the column we're applying this to.
)

data = data.withColumn('startYear', F.udf( lambda x: x.split('–')[0])('Year'))
data = data.withColumn('Votes', F.udf( lambda x: int(x.replace(',', '')))('Votes'))
data = data.withColumn('Votes', F.cast(IntegerType(), data['Votes']))

data = data.withColumn('Genre', F.split(data['Genres'], ','))
data = data.withColumn('Genre', F.explode(data['Genre']))

data = data.withColumn('Actor', F.split(data['Actors'], ','))
data = data.withColumn('Actor', F.explode(data['Actor']))

# Gotta clean up the inputs:
data = data.withColumn('Genre', F.udf(lambda x: x.strip())('Genre'))
data = data.withColumn('Actor', F.udf(lambda x: x.strip())('Actor'))
data = data.withColumn('Votes', F.udf( lambda x: int(x) )('Votes'))

cData = data.select(
    ['Name', 'Year', 'Guide', 
    'Duration', 'Ratings', 'Votes', 
    'isOngoing', 'startYear', 'Genre', 'Actor']
)

cData.show(10)
+-----+------+------+--------+-------+-----+---------+---------+---------+------------------+
| Name|  Year| Guide|Duration|Ratings|Votes|isOngoing|startYear|    Genre|             Actor|
+-----+------+------+--------+-------+-----+---------+---------+---------+------------------+
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|   Action|        Diego Luna|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|   Action|       Kyle Soller|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|   Action| Stellan Skarsgård|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|   Action|Genevieve O'Reilly|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|Adventure|        Diego Luna|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|Adventure|       Kyle Soller|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|Adventure| Stellan Skarsgård|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|Adventure|Genevieve O'Reilly|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|    Drama|        Diego Luna|
|Andor|2022– |TV-14 | 40 min |    8.4|82474|     true|     2022|    Drama|       Kyle Soller|
+-----+------+------+--------+-------+-----+---------+---------+---------+------------------+
only showing top 10 rows
                                                                                

Let’s Start Asking Some Questions.

Looking at the data we have, we’ll be asking: 1. Are the numeric columns Normally Distributed? 2. What is the Distribution of Ratings for the top Genres? 3. How does a show’s Ongoing status affect it’s ratings?

QQ plot of Numeric Values

A Quantile-Quantile Plot is a plot which takes the data and compares it against commonly another distribution. While it is less well known, it is most commonly used as a visual indicator of whether a dataset is Normally Distributed. Since matplotlib and seaborn don’t have this built in we’ll need to use a different library this time: statsmodels.

import numpy as np
import statsmodels.api as sm
import pylab as py

To plot the data we’ll need to pull it out of Spark and convert it to pandas; we’ve done this enough at this point. But, we’ve never used the sql api like this before. I quite like this interface since SQL queries are simply to read.

# To use this, we'll need what is called a View; this is an idea from SQL
cData.createOrReplaceTempView("TMP")
tmp = spark.sql("SELECT Ratings,Duration,Votes FROM TMP WHERE Duration IS NOT NULL")
tmp = tmp.withColumn('nDuration', F.udf( lambda x: x.replace(' min', '').replace(',', ''))('Duration'))

qqData = tmp.select(['Ratings', 'nDuration', 'Votes']).toPandas()
qqData
Ratings nDuration Votes
0 8.4 40 82474
1 8.4 40 82474
2 8.4 40 82474
3 8.4 40 82474
4 8.4 40 82474
... ... ... ...
26228 6.4 45 25
26229 6.4 45 25
26230 6.4 45 25
26231 6.4 45 25
26232 6.4 45 25

26233 rows × 3 columns

Are Ratings Normally Distributed?

Now we’ll check if the Ratings are Normally Distributed.

# Add the the y=x line for comparison.
sm.qqplot(qqData.Ratings, line ='45')
plt.show()
#py.show()

From the above, this is not a normally distributed dataset. When we have situations like this, it is not uncommon to try a log tranform if the data does not fit. The reason for this is that taking the logarithmic of the data maintains the underlying mathematical relationship. So, let’s try it!

plt.rcParams["figure.figsize"] = (6,6)
sm.qqplot(np.log(qqData.Ratings), line ='45')
plt.show()

Still No. We’ll want to avoid any model which has linear prerequisites.

Is Duration Normally Distributed?

The duration of a show is very likely to not be normally distributed. The length is not random and there are very few posssible choices. But, we’re going to check anyways.

plt.rcParams["figure.figsize"] = (6,6)
sm.qqplot(qqData.nDuration.astype(int), line ='45')
plt.show()

And, expected no.

Are Votes Normally Distributed?

We would expect votes to be normally distributed as this is the kind of process where as you would select samples it should approach normal.

plt.rcParams["figure.figsize"] = (6,6)
sm.qqplot(qqData.Votes.astype(int), line ='45')
plt.show()

Still even this is not normally distributed. This is starting to get worrying because at least this would be expected to be normally distributed.

Ridge Plot of Genre Ratings

Looking at the distribution of numeric variables can be useful. Normally, we could do a Distribution Plot or a Histogram to visualize this. But, I thought I would do something more interesting: a Ridge Plot. So, a Ridge Plot shows the distribution of values but conditioned on a category. And, we’re going to use it to check the ratings for the top categories and how they break down. First we’ll need to get our data.

from pyspark.sql.functions import mean as sMean, col

# Use results from previous post:
results = cData.select(['Genre', 'Ratings', 'Votes']).groupBy('Genre').agg(
    sMean('Ratings').alias("Mean"),
    sMean('Votes').alias('Votes')
    ).orderBy(F.desc("Mean")).toPandas()

results.head(10)
Genre Mean Votes
0 Biography 7.795238 33137.952381
1 Music 7.475000 413.625000
2 History 7.416667 3330.541667
3 Animation 7.270712 8447.388742
4 Fantasy 7.231312 18573.573756
5 Adventure 7.228473 8999.588235
6 Crime 7.225607 15978.954034
7 Comedy 7.223730 4206.774311
8 Family 7.207407 635.759259
9 Drama 7.203856 17351.695414

We’ll make a new SQL View with Apache Spark to filter out the top Genres.

cData.createOrReplaceTempView("Explore")
# Using our results, collect the categories to filter for.
values = ','.join([f"'{x}'" for x in results.head(8).Genre.tolist()])

ridgeSlice = spark.sql(f"SELECT Genre,Ratings FROM Explore WHERE Genre IN( { values } )").toPandas()
# This is just to stop some spam.
import warnings
warnings.filterwarnings("ignore")


# Theming:
sns.set_theme(style="white", rc={"axes.facecolor": (0, 0, 0, 0)})
pal = sns.dark_palette("seagreen", n_colors=10)

plt.rcParams["figure.figsize"] = (14,7)
g = sns.FacetGrid(ridgeSlice, row="Genre", hue="Genre", aspect=15, height=.5, palette=pal)

# Draw the densities in a few steps
g.map(sns.kdeplot, "Ratings",
      bw_adjust=.5, clip_on=False,
      fill=True, alpha=1, linewidth=1.5)
g.map(sns.kdeplot, "Ratings", clip_on=False, color="w", lw=2, bw_adjust=.5)

# passing color=None to refline() uses the hue mapping
g.refline(y=0, linewidth=2, linestyle="-", color=None, clip_on=False)

# Define and use a simple function to label the plot in axes coordinates
def label(x, color, label):
    ax = plt.gca()
    ax.text(0, .2, label, fontweight="bold", color=color,
            ha="left", va="center", transform=ax.transAxes)

g.map(label, "Genre")

# Remove axes details that don't play well with overlap
g.set_titles("")
g.set(yticks=[], ylabel="")
g.despine(bottom=True, left=True);

Music quickly stands out; people either like or dislike the show and there is little in between. Fantasy and Biography are clumped a little under 8 but mostly concentrated. The others are closer to what we’d expect from a normal distribution. Comedy, Animation and History are interesting since there is more spread among the ratings

JointPlot

A Joinplot is a nicer looking scatterplot. It includes some of the same information but also you can add distibution information. Personally, I like the Hexagons so we’re going to use those. And, we’re going to see how the year affects the Rating over time.

try:
    results = cData.select(['startYear', 'Ratings']).toPandas()
    results = results.assign(
        Year = results.startYear.astype(int)
    )
    results.head(15)
except Exception as e:
    print( e )
invalid literal for int() with base 10: 's2018'

Uh oh, looks like we’ve got some more cleaning to do. Let’s look at what is affected and see how widespread this is.

results.loc[ results.startYear == 's2018']
startYear Ratings
24638 s2018 6.1
24639 s2018 6.1
24640 s2018 6.1
24641 s2018 6.1
24642 s2018 6.1
24643 s2018 6.1
24644 s2018 6.1
24645 s2018 6.1
24646 s2018 6.1
24647 s2018 6.1
24648 s2018 6.1
24649 s2018 6.1

This might be the same movie/show which is being exploded out. We’re going to just fix it for our analysis at this point. Again, if we were getting this data from a Data Engineer then we’d want to discuss this with them to clean this upstream.

results['startYear'] = results.startYear.str.replace('s2018', '2018')
results = results.assign(
    Year = results.startYear.astype(int)
)
results.head(15)
startYear Ratings Year
0 2022 8.4 2022
1 2022 8.4 2022
2 2022 8.4 2022
3 2022 8.4 2022
4 2022 8.4 2022
5 2022 8.4 2022
6 2022 8.4 2022
7 2022 8.4 2022
8 2022 8.4 2022
9 2022 8.4 2022
10 2022 8.4 2022
11 2022 8.4 2022
12 2022 8.0 2022
13 2022 8.0 2022
14 2022 8.0 2022

So, let’s see the Data!

sns.jointplot(data=results, x="Year", y="Ratings", kind="hex", color='seagreen');

Honestly, I expected there to be a stronger relationship here. There is definitely a relationship here about the latest years beginning to fall though. If this was for some kind of report maybe it would be work breaking these into a Ridge Plots conditioned on the year.

Violin Plot

Violin Plots are similar to Box and Whisker plots but they also add a visual element to show the distribution. We’re going to check out whether a show being finished or not has an impact on the Ratings.

plt.rcParams["figure.figsize"] = (14,7)
sns.set_theme(style="white", rc={"axes.facecolor": (0, 0, 0, 0), 'legend.facecolor':'white'})
pal = sns.dark_palette("seagreen", n_colors=2)

sns.violinplot(data=violinData, x="Genre", y="Ratings", split=True, hue="isOngoing", palette=pal);

There is something really strange going on with the Biography category. In a real analysis we’d want to dig into Biography to pull this apart more.

Conclusions

We’ll end this here. Hopefully this was an interesting to read over as it was to explore.