DSA210 term project

Entertainment Behavior Under Academic Pressure

Hello, my name is Nihat Ömer Karaca. I am a CS & IE double-major student studying at Sabancı University.

For our DSA210 course project, we were supposed to select a dataset and implement what the course required. I started by thinking about my daily life and what I actually do during ordinary days, midterms, and finals.

After brainstorming, I realized that I was spending a lot of time on some platforms, and that this might be affected by academic requirements such as finals and midterms. I also thought that the platforms I use, such as Spotify, streaming platforms, YouTube, and the academic calendar, might affect each other. Maybe they pave the way for each other, or maybe they create a blockade.

Because of that, I downloaded and considered several datasets: ChatGPT, Instagram, Netflix, Spotify, Twitter, and YouTube. After careful consideration, I decided to exclude Instagram and Twitter because they would have increased the complexity of the project too much. I also postponed ChatGPT because it is a completely different domain and would need a heavier privacy and parsing process. So I stuck with YouTube, Spotify, Netflix, Prime Video, and the academic calendar.

I also prefer looking at project content on a webpage rather than only inside notebook files. That is why I created this page as a public-facing explanation of the project. Webpage explanation will be filled later (TODO).

Raw and public datasetsProcessing scriptsIndividual EDACombined EDAHypothesesCurrent test resultsNotebook plot appendix

Data and scripts Interactive EDA Hypothesis results

1,424daily rows in the shared analysis window

4public behavior datasets

50EDA appendix figures

9basic result rows

Raw data and public outputs

Different exports, one daily frame.

The raw exports were not equally clean: some arrived as structured CSV/JSON, while others needed parsing and privacy decisions. The public version keeps only the fields needed for analysis and uses the day as the shared unit.

Short-form and mixed video behavior

YouTube

34,317 rows

2022-04-17 to 2026-03-15

Google Takeout activity converted into a public activity table. Channel names are masked.

Music listening duration

Spotify

178,202 rows

2019-07-27 to 2026-03-14

Music-focused streaming history with exact timestamps and listening duration.

Long-form viewing

Netflix

2,493 rows

2020-02-25 to 2026-03-10

Viewing history kept at row level. Titles stay visible; local source paths are not exposed.

Long-form viewing

Prime Video

719 rows

2022-03-30 to 2026-03-19

Watch history parsed into movie and episode records where possible.

YouTube

Raw shape: Google Takeout activity export with watch/search-style events and timestamps.

Kept: Action type, target kind, masked channel references, timestamps, and daily activity fields.

Excluded: Real channel names and direct source identifiers were removed or masked.

Spotify

Raw shape: Structured streaming-history JSON with exact timestamps and listening duration.

Kept: Track, artist, album, timestamp, milliseconds played, minutes played, and hours played.

Excluded: Account identifiers, IP-like fields, and raw source-file identifiers were removed.

Netflix

Raw shape: Viewing-history CSV with title and date-level viewing records.

Kept: Original title string and viewing date at row level.

Excluded: No forced show/season/episode split because some Netflix titles are too inconsistent to parse safely.

Prime Video

Raw shape: Watch-history export parsed into movie and episode records.

Kept: Title, record type, series/movie fields where available, and date-level watch records.

Excluded: Raw parse issues and local source details stay out of the public browser app.

Processing scripts

From export to public analysis.

Each platform follows the same logic: normalize the raw export, create shared `fine_*` fields, reduce the public schema, then aggregate later for EDA and testing.

1. Platform scripts normalize raw exports into processed row-level tables.2. fine_* builders create shared date, platform, source, and record-id fields.3. Public builders reduce the schema and mask source/account identifiers.4. EDA notebooks aggregate public rows to daily and monthly variables.5. The web app uses compact aggregate JSON and copied plot images, not full raw CSVs.

Interactive EDA

Interactive exploratory analysis.

These charts are SVG-based and interactive. Hover bars, points, and cells to see values and highlight the element being inspected.

YouTube EDA

YouTube activity overview.

YouTube is mostly watch activity. The estimated watch-time line is a conservative proxy: it sums gaps of up to 60 minutes between consecutive watched-video timestamps.

Uses YouTube action counts to show whether the dataset is mostly watch activity, search activity, or other actions.

Uses monthly watched and search counts to show how YouTube activity changes over time.

Uses gaps between consecutive watched-video timestamps to estimate rough monthly watch time.

Uses YouTube timestamps converted to Istanbul time to show which hours have more YouTube activity.

Uses daily watched counts grouped by weekday to inspect whether YouTube use changes by day of week.

Spotify EDA

Spotify listening overview.

The main EDA checks monthly listening, time of day, and artist concentration without using it as a final hypothesis directly.

Uses Spotify monthly listening hours and stream counts to show long-term listening patterns.

Uses Spotify timestamps and hours played to show which hours have more listening activity.

Uses total Spotify hours per artist to show the most repeated artists.

Uses total listening minutes per track to show the most repeated individual songs.

Uses daily Spotify hours grouped by weekday to inspect weekday differences.

Netflix + Prime Video EDA

Long-form streaming overview.

Netflix and Prime stay separate public datasets, but the EDA groups them as long-form streaming. Because most days have zero long-form activity, active-day charts are clearer than standard boxplots for this part.

Manual data check: in the common 1,424-day window, 1,160 days have zero Netflix + Prime records and only 264 days are active. Since at least 75% of each weekday group is zero, the standard boxplot has Q1 = median = Q3 = 0. The chart below excludes zero days so it shows which weekdays had actual long-form use. Top-title charts are also shown as EDA summaries; Netflix uses conservative title groups only for display, while the public file still keeps the original title text.

Uses Netflix and Prime Video row counts to compare their total contribution.

Uses monthly Netflix, Prime Video, and combined counts to show long-form viewing over time.

Uses Prime Video record type to show whether the records are mostly episodes or movies.

Uses Netflix title quality checks to show usable titles versus missing or malformed title rows.

Uses grouped Netflix titles to show the most repeated Netflix series or title groups.

Uses Prime Video series and movie titles to show the most repeated Prime Video titles.

Uses active Netflix + Prime days only, excluding zero days, to show which weekdays had actual long-form use.

Combined EDA

Combined daily panel.

These views support the hypothesis stage without mixing incompatible raw units.

Uses daily Spotify hours and YouTube watched count to inspect same-day co-usage.

Uses daily Netflix + Prime count and YouTube watched count to inspect whether long-form activity relates to YouTube activity.

Uses daily platform variables to show which variables move together more strongly.

Uses YouTube watched count grouped by academic period to compare ordinary days, finals, and other periods.

Uses Spotify daily hours grouped by academic period to compare listening across periods.

Uses active long-form days grouped by academic period, excluding zero days for readability.

Uses the number of active platforms per day to show whether platform variety changes by period.

Uses YouTube activity after 21:30 to inspect late-evening behavior across academic periods.

Uses Spotify listening after 21:30 to inspect late-evening listening across academic periods.

Combined EDA overview

Cross-platform checks.

They summarize coverage, active days, monthly movement, distributions, academic-period movement, and hourly activity before the formal hypothesis-testing section.

Uses each platform's date range to show the common window where all datasets overlap.

Uses daily activity flags to show how many days each platform appears in the common window.

Uses monthly YouTube watched count, Spotify hours, and Netflix + Prime count as separate trend panels.

Uses daily variables to show skew, zero-heavy behavior, and outliers.

Uses period averages divided by overall averages to compare platforms without mixing raw units.

Uses YouTube and Spotify timestamps to compare time-of-day behavior in Istanbul time.

Hypothesis testing

Hypothesis tests and current results.

The formal tests use daily variables from the same public-data panel. Final-exam comparisons use one-sided Mann-Whitney U tests; same-day Spotify/YouTube co-usage uses one-sided Spearman correlation.

Hypothesis table

Hypothesis	Null hypothesis (H0)	Alternative hypothesis (H1)	Method
H1 Entertainment usage during academic pressure	Platform activity does not differ between final-exam days and ordinary-term days.	Final-exam days have lower platform activity.	One-sided Mann-Whitney U
H2 Platform diversity during academic pressure	Platform diversity does not differ between final-exam days and ordinary-term days.	Platform diversity is lower during finals.	One-sided Mann-Whitney U
H3 Netflix + Prime and YouTube activity	YouTube watched count does not differ between Netflix + Prime active and inactive days.	YouTube watched count is lower on Netflix + Prime active days.	One-sided Mann-Whitney U
H4 Spotify and YouTube co-usage	Spotify hours are not associated with YouTube watched count.	Spotify hours are positively associated with YouTube watched count.	One-sided Spearman correlation
H5 After-9:30 PM entertainment during finals	Late-evening entertainment share does not differ between final-exam days and ordinary-term days.	Late-evening share is lower during finals.	One-sided Mann-Whitney U

Entertainment usage during academic pressure

H0: Platform activity does not differ between final-exam days and ordinary-term days.

H1: Final-exam days have lower platform activity.

Method: One-sided Mann-Whitney U

Platform diversity during academic pressure

H0: Platform diversity does not differ between final-exam days and ordinary-term days.

H1: Platform diversity is lower during finals.

Method: One-sided Mann-Whitney U

Netflix + Prime and YouTube activity

H0: YouTube watched count does not differ between Netflix + Prime active and inactive days.

H1: YouTube watched count is lower on Netflix + Prime active days.

Method: One-sided Mann-Whitney U

Spotify and YouTube co-usage

H0: Spotify hours are not associated with YouTube watched count.

H1: Spotify hours are positively associated with YouTube watched count.

Method: One-sided Spearman correlation

After-9:30 PM entertainment during finals

H0: Late-evening entertainment share does not differ between final-exam days and ordinary-term days.

H1: Late-evening share is lower during finals.

Method: One-sided Mann-Whitney U

Entertainment usage during academic pressure

Null hypothesis (H0): Platform activity does not differ between final-exam days and ordinary-term days.

Alternative hypothesis (H1): Final-exam days have lower platform activity.

H1: X_final < X_ordinaryOne-sided Mann-Whitney U

Platform diversity during academic pressure

Null hypothesis (H0): Platform diversity does not differ between final-exam days and ordinary-term days.

Alternative hypothesis (H1): Platform diversity is lower during finals.

D = count(active platforms per day)One-sided Mann-Whitney U

Netflix + Prime and YouTube activity

Null hypothesis (H0): YouTube watched count does not differ between Netflix + Prime active and inactive days.

Alternative hypothesis (H1): YouTube watched count is lower on Netflix + Prime active days.

YouTube | long-form active < YouTube | inactiveOne-sided Mann-Whitney U

Spotify and YouTube co-usage

Null hypothesis (H0): Spotify hours are not associated with YouTube watched count.

Alternative hypothesis (H1): Spotify hours are positively associated with YouTube watched count.

rho_s = corr(rank(Spotify), rank(YouTube))One-sided Spearman correlation

After-9:30 PM entertainment during finals

Null hypothesis (H0): Late-evening entertainment share does not differ between final-exam days and ordinary-term days.

Alternative hypothesis (H1): Late-evening share is lower during finals.

share_after_2130 = after_2130 / daily_totalOne-sided Mann-Whitney U

Complete result table

Hypothesis	Outcome	Test	Statistic	p-value	Means	Decision
H1 Entertainment usage during academic pressure	YouTube watched count	Mann-Whitney U	34,501	0.0908	24.4 / 30.8	Do not reject H0
H1 Entertainment usage during academic pressure	Spotify listening hours	Mann-Whitney U	34,130	0.0697	1.6 / 1.9	Do not reject H0
H1 Entertainment usage during academic pressure	Netflix viewing count	Mann-Whitney U	38,928.5	0.8512	1 / 0.8994	Do not reject H0
H1 Entertainment usage during academic pressure	Prime Video viewing count	Mann-Whitney U	37,300	0.3890	0.2165 / 0.5587	Do not reject H0
H2 Platform diversity during academic pressure	Distinct active entertainment platforms	Mann-Whitney U	35,360	0.1305	1.8 / 1.9	Do not reject H0
H3 Long-form streaming and YouTube activity	YouTube watched count	Mann-Whitney U	159,759.5	0.8731	26.4 / 21.4	Do not reject H0
H4 Spotify and YouTube co-usage	Spotify hours and YouTube watched count	Spearman correlation	0.0934	< 0.001	1.9 / 22.3	Reject H0
H5 After-9:30 PM entertainment during finals	YouTube after-21:30 activity share	Mann-Whitney U	32,124	0.0059	0.1786 / 0.2732	Reject H0
H5 After-9:30 PM entertainment during finals	Spotify after-21:30 listening-hour share	Mann-Whitney U	33,480	0.0361	0.1764 / 0.2168	Reject H0