Hugging Face's logo

Datasets:
benjac8
/
OKCUPID-EDA

Modalities:
Image
Tabular
Formats:
csv
Size:
10K - 100K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
age
int64
status
int64
sex
int64
orientation
int64
body_type
int64
diet
int64
drinks
int64
drugs
int64
education
int64
ethnicity
int64
height
float64
income
int64
job
int64
location
int64
pets
int64
religion
int64
sign
int64
smokes
int64
days_since_online
int64
num_languages
float64
22
3
1
2
0
0
4
0
3
0
75
43,840
19
1
14
0
4
1
2
1
35
3
1
2
2
3
2
2
3
8
70
80,000
8
1
14
0
2
0
1
3
38
0
1
2
10
0
4
1
4
5
68
181,707
8
1
5
2
7
0
3
3
23
3
1
2
10
5
4
0
3
8
71
20,000
18
1
10
2
7
0
2
2
29
3
1
2
1
0
4
0
3
0
66
68,730
0
1
14
3
0
0
3
1
29
3
1
2
2
0
4
0
3
8
67
60,954
3
1
10
1
10
0
1
2
32
3
0
2
4
0
4
0
3
8
65
0
10
1
14
4
11
0
5
1
31
3
0
2
2
0
4
0
3
8
65
1,074
0
1
14
4
8
0
1
2
24
3
0
2
4
0
4
1
3
8
67
34,651
10
1
14
4
4
3
1
1
37
3
1
2
1
0
1
0
2
8
65
0
18
1
14
1
2
0
2
1
35
0
1
2
2
0
4
1
3
8
70
69,131
8
1
11
3
10
4
26
1
28
2
1
2
2
0
4
0
3
8
72
40,000
1
1
10
4
5
0
39
2
24
3
1
2
3
1
2
1
3
8
72
148,616
6
1
11
8
10
1
33
1
30
3
0
2
9
0
4
0
1
8
66
30,000
16
1
9
4
6
0
17
1
29
3
0
2
10
0
4
0
3
2
62
50,000
12
1
13
3
10
0
2
1
39
3
0
2
4
0
4
0
3
8
65
0
10
1
13
1
0
0
1
2
33
3
1
2
4
1
4
1
4
8
70
148,890
6
1
14
3
7
1
1
4
26
3
0
2
2
1
4
0
3
2
64
71,697
1
1
11
3
1
0
1
1
31
3
1
2
2
5
3
0
3
8
71
202,060
10
1
14
4
6
0
2
1
33
3
1
2
1
0
4
0
4
8
72
62,470
17
1
14
3
7
0
3
1
27
3
0
2
2
0
4
0
3
8
67
0
16
1
12
3
6
0
4
1
22
3
0
2
1
1
4
0
3
5
67
31,328
18
1
14
3
10
0
15
3
30
3
1
2
4
0
4
0
3
8
69
12,762
7
1
14
0
8
0
2
2
30
3
1
2
10
4
1
0
3
8
71
165,819
5
1
14
1
0
0
44
1
33
3
1
2
10
0
4
2
3
8
73
212,831
0
1
10
8
9
1
1
1
28
3
1
2
4
0
3
0
3
0
70
139,295
10
1
6
3
4
0
5
1
22
3
1
2
4
0
4
0
3
7
72
74,439
2
1
12
3
5
2
21
3
22
3
1
2
2
0
4
0
3
0
67
121,883
12
1
7
0
5
0
5
3
30
3
1
2
4
1
4
0
4
7
74
120,524
10
1
14
1
6
0
12
2
32
3
1
2
4
1
4
2
3
8
68
122,559
12
1
12
0
5
0
2
1
27
3
0
2
2
0
4
1
3
8
64
4,634
12
1
11
0
4
2
233
2
27
3
1
2
1
0
4
0
3
6
72
41,304
17
1
11
1
6
0
3
1
38
3
0
2
2
0
4
0
3
8
67
18,508
3
1
11
4
6
0
37
1
20
3
0
2
2
5
4
0
3
4
60
193,307
8
1
9
4
2
0
1
2
27
3
1
2
10
0
4
0
3
8
69
31,650
6
1
14
1
8
3
6
1
26
3
1
2
1
1
4
0
3
8
69
84,325
1
1
12
3
6
0
4
2
32
3
1
2
1
0
4
0
3
8
69
13,885
17
1
12
2
2
3
2
2
25
3
1
0
4
0
4
1
3
2
69
54,782
12
1
14
3
6
1
2
2
27
3
0
2
3
1
4
0
3
5
63
73,331
6
1
11
3
6
0
1
1
35
3
1
2
4
1
4
0
3
8
74
99,454
3
1
11
3
6
0
1
1
30
3
1
2
2
1
2
0
4
6
76
173,443
3
1
10
0
6
0
1
5
35
3
1
2
1
0
4
0
3
1
72
66,501
9
1
11
0
7
0
1
1
30
3
1
2
2
0
2
2
4
8
75
240,761
0
1
12
8
4
0
103
2
40
3
1
2
4
1
4
0
3
8
71
60,000
4
1
11
0
4
0
1
4
29
3
0
0
3
0
4
2
4
8
66
196,619
12
1
5
8
0
0
20
3
27
3
1
2
2
0
4
0
3
7
69
53,618
5
1
11
3
0
0
2
2
27
3
1
2
1
0
4
0
4
6
73
128,007
10
1
12
3
6
0
2
3
28
3
1
2
2
0
2
2
3
8
70
121,405
0
1
12
0
6
0
5
1
26
3
0
2
1
0
4
0
3
8
66
12,850
17
1
14
8
10
0
3
2
33
3
0
1
4
2
2
2
3
8
66
166,176
0
1
12
4
10
0
4
2
38
3
0
1
1
0
4
0
3
6
65
0
10
1
11
4
10
0
99
1
31
3
1
2
2
0
4
0
3
8
69
5,196
17
1
14
0
4
0
7
2
36
3
1
2
1
0
4
0
3
0
69
137,919
3
1
6
3
6
0
10
1
33
3
1
2
4
0
3
0
3
8
69
19,857
3
1
14
1
6
0
6
1
33
3
1
1
1
0
4
0
3
8
70
14,917
10
1
13
3
7
0
1
2
26
3
1
2
4
0
4
0
3
8
71
51,470
18
1
12
1
8
0
1
3
21
3
1
2
10
1
4
1
3
8
72
143,036
3
1
12
2
6
3
1
1
22
3
1
2
4
0
4
2
3
2
70
171,912
18
1
8
0
7
1
262
1
31
3
1
2
2
0
4
1
3
6
71
66,216
9
1
14
0
6
3
25
2
31
2
1
2
4
0
4
0
3
0
67
85,974
3
1
12
3
4
0
55
3
31
3
1
2
10
0
4
2
4
8
70
164,531
14
1
12
0
2
3
1
1
28
3
1
2
1
0
4
0
3
2
72
81,512
13
1
11
3
11
0
3
2
22
3
1
2
1
0
3
0
3
0
65
20,000
5
1
12
2
11
0
1
1
30
3
0
2
10
1
3
2
3
1
66
201,885
18
1
13
8
1
1
33
1
31
3
0
0
2
0
4
1
3
4
69
34,327
10
1
14
8
11
1
52
1
29
3
1
1
10
1
4
2
4
8
76
201,597
12
1
14
0
9
0
66
2
25
3
0
2
4
1
3
0
3
8
62
36,392
17
1
11
3
7
0
2
1
33
3
0
2
1
0
4
0
3
8
67
31,774
6
1
11
4
4
0
137
4
35
3
1
2
1
0
4
2
3
5
73
150,000
0
1
14
0
7
0
2
2
31
3
0
2
3
0
4
0
5
3
61
50,000
9
1
12
5
6
0
2
5
28
3
1
2
4
0
4
0
4
8
72
124,892
5
1
13
3
8
0
2
3
29
0
0
0
3
5
4
1
3
7
67
159,183
14
1
11
4
4
1
2
2
25
2
0
0
3
2
4
0
3
8
65
0
12
1
13
0
5
0
254
1
34
3
0
2
3
0
4
0
3
8
67
0
16
1
11
0
6
0
4
3
21
3
1
2
4
0
3
0
3
8
71
20,000
6
1
14
3
5
0
193
1
28
3
0
2
10
2
2
0
4
8
67
125,547
7
1
11
1
2
0
7
2
31
3
1
1
2
0
4
0
4
8
69
45,829
10
1
14
1
6
0
1
3
34
3
1
2
4
0
4
0
4
8
69
100,000
7
1
13
0
6
0
2
2
32
3
0
1
2
1
4
0
3
0
64
35,849
17
1
14
4
7
0
126
3
30
3
1
2
1
0
4
0
3
6
73
60,543
17
1
11
3
4
0
188
1
28
3
1
2
2
1
4
0
3
8
68
78,879
10
1
10
1
9
0
70
3
40
3
0
2
4
0
4
0
3
6
66
9,017
9
1
14
4
10
0
24
2
27
3
0
2
0
0
2
0
3
4
61
21,086
16
1
11
4
1
0
4
2
43
3
1
2
0
1
4
0
2
6
71
25,844
9
1
11
4
6
0
1
3
37
3
0
2
2
0
4
0
3
6
64
23,372
10
1
13
8
5
0
93
1
24
3
1
2
2
0
4
0
3
1
74
84,168
16
1
10
3
2
4
155
1
29
3
1
2
3
1
3
0
3
8
74
83,886
17
1
11
3
9
1
248
1
29
0
0
2
0
1
3
2
2
8
67
50,000
10
1
10
4
3
1
339
1
29
3
1
2
4
0
4
2
3
8
67
40,000
0
1
14
0
6
0
33
1
25
3
1
2
1
0
4
0
3
8
78
80,000
16
1
14
3
1
0
4
2
39
3
1
2
4
0
4
0
3
0
70
73,011
12
1
11
0
6
0
45
2
31
3
0
2
3
0
4
2
4
8
66
131,508
7
1
11
1
7
0
2
1
46
3
1
2
2
0
4
0
3
1
72
107,195
4
1
11
4
10
0
6
1
38
3
1
2
10
1
4
0
4
8
71
156,824
5
1
12
3
0
0
2
1
30
3
1
2
4
0
4
0
3
4
68
124,224
1
1
12
8
6
0
12
5
39
3
1
2
1
3
4
2
3
8
70
100,000
3
1
11
1
5
0
42
1
31
3
0
2
10
0
4
0
4
8
63
68,307
12
1
11
3
9
0
5
1
42
3
0
2
4
0
4
0
4
8
65
93,164
5
1
11
7
1
0
1
4
45
3
0
2
4
0
4
0
4
0
64
67,544
7
1
11
0
1
0
6
1
28
3
0
2
4
1
4
0
5
7
64
160,849
9
1
11
3
1
0
1
3
End of preview. Expand in Data Studio

OKCupid Dating Profiles — Exploratory Data Analysis

Building a Foundation for a Matchmaking Recommendation System

Watch the demo

The Goal

Can we group OKCupid users into distinct "dating personas" based on their lifestyle metrics, and use those groupings as the foundation for a matchmaking recommendation system?

This notebook takes 57,428 raw OKCupid user profiles and transforms them into a clean, fully numeric dataset ready for clustering. Every decision — from which columns to drop to how to handle missing values — is made with one question in mind: will this help us find meaningful groups of people?

Dataset

Property Detail
Source OKCupid Profiles — Kaggle
Raw profiles 57,428 users
Raw columns 31 (demographics + lifestyle + 10 free-text essay fields)
Cleaned columns 19 (all numeric, all meaningful)
Missing values after cleaning 0

Notebook Structure

The notebook is divided into four main acts:

Act 1 — Data Collection The dataset is loaded directly from Kaggle using kagglehub — no manual download or API key file needed.

Act 2 — Data Wrangling The raw data is transformed step by step into a clean, fully numeric dataset.

Act 3 — Exploratory Visualization Individual features are explored through histograms, bar charts, and box plots. Relationships between features are uncovered through scatter plots, heatmaps, and stacked bar charts.

Act 4 — Research Questions Two targeted questions are posed and answered visually to test whether distinct user groups actually exist in the data.

Data Wrangling — Step by Step

The raw data needed significant cleaning before any analysis was possible. Here is every transformation applied, in order:

1 — Dropping Irrelevant Columns

Essay columns (essay0–essay9): Ten free-text "about me" fields were removed. These require natural language processing to be useful and are out of scope for a numeric clustering EDA.

Offspring column: Removed because it has a high rate of missing values and does not meaningfully contribute to lifestyle-based persona grouping.

2 — Parsing Dates

The last_online column was a formatted string (YYYY-MM-DD-HH-MM). It was converted into a single number: days_since_online — the number of days since each user was last active. This turns an unanalysable string into a real engagement metric.

3 — Cleaning Free-Text Categoricals

Three columns had the actual category value glued to a user-written commentary:

Column Raw example Cleaned result
religion "agnosticism and very serious about it" "agnosticism"
sign "pisces but it doesn't matter" "pisces"
diet "mostly vegetarian" "vegetarian"

4 — Simplifying High-Cardinality Columns

Column Before After
ethnicity 193 unique values (multi-listed) First ethnicity only → 9 unique
speaks 5,765 unique language combos num_languages count (1–7)
location 166 city–state pairs State/region only → 32 unique

5 — Ordinal Encoding Education

Education has a natural order. It was mapped to a 0–6 integer scale:

Number Meaning
0 Dropped out of high school
1 High school graduate
2 Two-year college
3 College / University (Bachelor's)
4 Master's program
5 Law school / Med school
6 PhD program

6 — Label Encoding All Remaining Columns

Every other text column (sex, religion, job, status, orientation, etc.) was sorted alphabetically and assigned an integer starting at 0. A full reverse-mapping dictionary was saved so any number can be translated back to its original label.

7 — Handling the Income Sentinel

The income column used -1 to mean "not stated." This was replaced with NaN before imputation so the model treated it as genuinely missing rather than as a real value of negative one dollar.

8 — Predictive Imputation

Instead of deleting rows with missing data or filling with column averages, IterativeImputer was used. This trains a regression model on all other columns to predict each missing value — capturing real relationships like the link between religion and diet, or between job and education level. The dataset went from 14.7% missing cells to 0%.

9 — Outlier Removal

IQR analysis identified three categories of extreme values that would distort clustering:

Feature Filter Reason
Age Removed > 70 Outside the realistic dating pool
Height Removed < 50 in or > 85 in Biologically implausible — data entry errors
Income Removed > $250,000 Extreme right skew distorts distance calculations

Result: A fully numeric, zero-missing-value dataset ready for clustering.

Outlier Removal

Exploratory Visualizations

Univariate — Who Are the Users?

Age Distribution The user base peaks sharply between ages 25 and 32, with a long right tail. The platform is clearly dominated by young adults in their late 20s and early 30s.

Age Distribution

Sex Distribution Approximately 60% of users are male and 40% female — a notable imbalance that has implications for any recommendation system.

Distribution of Sex

Sexual Orientation The dataset includes users across multiple orientations. Straight users make up the large majority, with gay and bisexual users also represented.

Orientation Distribution

Relationship Status The vast majority of users list themselves as "single," confirming this is an active-use dataset rather than a collection of abandoned profiles.

Relationship Status

Religion Distribution Agnosticism and atheism are the two most common religions in this dataset — reflecting the platform's San Francisco Bay Area user base.

Religion Distribution

Job Distribution Tech, student, and "other" dominate. This is consistent with the geographic and demographic concentration of the dataset.

Job Distribution

Bivariate — How Do Features Relate?

Sex vs Income Men significantly outnumber women in every income bracket above $50,000. The income gap between sexes is visible and consistent across age groups.

Income by Gender

Income Distribution The income distribution is heavily right-skewed, with the majority of users earning under $100,000 and a long tail of high earners.

Income Distribution

Sex vs Education Both sexes show similar education distributions overall, but men are slightly overrepresented at the college/university level.

Education by Sex

Age vs Median Income Median income rises steadily from the 18–25 group up to the 41–50 bracket, then levels off — a classic career-progression pattern.

Age Group vs Median Income

Religion vs Drinking Habits A stacked bar chart reveals that religions associated with abstinence (Islam, specific Christian denominations) show markedly lower drinking rates, while atheists and agnostics cluster toward "socially" and "often."

Religion vs Drinking Habits

Multivariate — Where Do Patterns Converge?

Full Correlation Heatmap The complete correlation matrix across all 19 numeric features shows that drinks, smokes, and drugs form a cluster of moderate positive correlations — users who drink heavily are more likely to smoke and use drugs.

Correlation Heatmap

Research Questions

Question A — Which combination of 3 lifestyle features produces the largest user groups?

Method: Count the number of users in every unique drinks × smokes × drugs combination. Display as a treemap (rectangle size = group size) and a ranked bar chart.

Plot: Treemap + horizontal bar chart

Finding: A small number of lifestyle combinations account for the majority of users. The dominant group is socially-drinking, non-smoking, non-drug-using users — the "mainstream social" proto-persona. This concentration confirms that the data is not uniform: users genuinely cluster around a few lifestyle profiles.

Question A — Lifestyle Groups

Verdict: ✅ Supports the persona theory — dominant groups exist.

Question B — Which 3 features contribute so little that removing them won't affect grouping?

Method:

  1. Coefficient of Variation (CV) bar chart — features with low CV are nearly constant across all users and cannot separate them into groups.
  2. Correlation heatmap (lower triangle) — features that are highly correlated with another are redundant; removing one loses no information.

Plot: Side-by-side: CV bar chart + correlation heatmap

Finding: Features with the lowest variance are the weakest separators. Any feature that is both low-variance and strongly correlated with another feature is a prime candidate for removal before clustering. This gives a principled, data-driven answer to feature selection.

Question B — Feature Selection

Verdict: ✅ Identifies 3 removable features — reduces noise before clustering.

Conclusion

The research question is supported by the data.

Four pieces of evidence back the hypothesis that distinct dating personas exist in this dataset:

  1. Lifestyle habits are correlated — drinks, smokes, and drugs move together, forming a natural habit-cluster axis.
  2. A few combinations dominate — the treemap shows users concentrate around a small number of lifestyle profiles rather than spreading uniformly.
  3. Features have high separating power — age, income, and days_since_online show high variance, giving any clustering algorithm meaningful axes to work with.
  4. Demographic patterns are systematic — the gender income gap and generational smoking patterns confirm that the data has internal structure, not random noise.

The cleaned dataset produced by this notebook is the direct input for the next step: K-Means or hierarchical clustering to formally identify and label the personas.

How to Run This Notebook

  1. Open in Google Colab
  2. Run Cell 3 (!pip install kagglehub) — no API key needed
  3. Run all cells top to bottom
  4. The cleaned CSV is saved to /mnt/user-data/outputs/okcupid_cleaned.csv

Dependencies: pandas, numpy, matplotlib, seaborn, scikit-learn, squarify, kagglehub

File Structure 

EDA_OKCUPID_personas.ipynb    ← Main notebook
KAGGLE_README.md              ← This file

Benjamin — Data Science Assignment 1, 2026
Downloads last month
74
Number of rows:
57,428
Total file size:
17.2 MB