---
license: mit
task_categories:
  - tabular-classification
language:
  - en
tags:
  - music
  - spotify
  - audio-features
  - eda
  - popularity
  - hit-prediction
size_categories:
  - 10K<n<100K
---

# 🎵 What Makes a Hit Song? — Spotify EDA

> **Assignment #1 — EDA & Dataset | March 2026**

---

## 📽️ Video Presentation

<video src="https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/assignment_1_DS.mp4" controls="controls" style="max-width: 720px;"></video>

---

## 📌 Overview

This project explores the **30,000 Spotify Songs** dataset to answer one core question:

> **How do audio features — danceability, energy, loudness, speechiness, tempo, acousticness, instrumentalness, liveness, valence, and duration — affect the popularity of a song?**

The goal is to identify the optimal range of audio features that music producers should aim for to increase the chances of their song becoming a hit.

---

## 📂 Dataset

| Property | Details |
|---|---|
| **Source** | [Kaggle — 30000 Spotify Songs](https://www.kaggle.com/datasets/joebeachcapital/30000-spotify-songs) |
| **Raw Size** | 32,833 rows × 23 features |
| **Final Size** | 32,790 rows × 25 features (after cleaning, feature engineering & column pruning) |
| **Target Variable** | `track_popularity` (0–100) |
| **Hit Definition** | `track_popularity > 75` (top ~8.1% of all tracks) |

### Audio Features Explained

| Feature | What It Means | Example |
|---|---|---|
| `danceability` | How easy it is to dance to (0.0–1.0) | Low = slow ballad, High = club banger |
| `energy` | Perceptual intensity and activity (0.0–1.0) | Death metal = high, Bach prelude = low |
| `loudness` | Average loudness in decibels (dB) | Closer to 0 = louder, −60 = almost silent |
| `speechiness` | Presence of spoken words (0.0–1.0) | >0.66 = mostly speech, <0.33 = music |
| `acousticness` | Confidence that a track is acoustic (0.0–1.0) | Guitar unplugged = high, EDM synths = low |
| `instrumentalness` | Likelihood the track has no vocals (0.0–1.0) | Jazz instrumental = high, pop singer = low |
| `liveness` | Presence of a live audience (0.0–1.0) | >0.8 = likely live recording |
| `valence` | Musical positiveness (0.0–1.0) | Happy/cheerful = high, sad/angry = low |
| `tempo` | Estimated tempo in BPM | 60 BPM = slow, 180 BPM = very fast |
| `duration_min` | Song length in minutes | 2 min = short, 6 min = long |

---

## 🛠️ Setup

```python
from google.colab import drive
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
```

Developed in **Google Colab** with the dataset loaded from Google Drive.

---

## 🧹 Data Cleaning

**Missing Values:** Found 5 rows with missing `track_name`, `track_artist`, and `track_album_name` — all removed.

**Duplicates:** None found — no action needed.

**Impossible Values:**
- `tempo = 0 BPM` → dropped (1 track). A song with 0 BPM cannot exist.
- `loudness > 0 dB` → dropped (6 tracks). Industry standard caps loudness at 0 dB.

**Final shape after cleaning: (32,790, 23)**

---

## ⚙️ Feature Engineering

Three new features were created:

- **`release_year`** — extracted from `track_album_release_date` (covers 1957–2020)
- **`is_hit`** — `True` if `track_popularity > 75` (top ~8.1% of tracks by popularity score)
- **`loudness_scaled`** — loudness rescaled from dB (−60 to 0) to a 0–100 range for readability
- **`duration_min`** — changed duration ms to duration in min.

## 🗑️ Column Pruning

Two columns were intentionally dropped after feature engineering:

- **`track_album_release_date`** — used only to derive `release_year`. Once extracted, the original date string is redundant.
- **`playlist_subgenre`** — all genre analysis operates at the broad `playlist_genre` level. This column was never computed, visualized, or referenced.

**Final dataset shape: (32,790, 25)**

---

## 📊 Visualizations & Key Findings

### 1. Outlier Detection — All 10 Audio Features

![Outlier Detection](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/OUTLIERS.png)

Box plots across all 10 features reveal the spread and outliers in the dataset. Key observations:
- **Tempo** clusters tightly between 100–130 BPM
- **Instrumentalness** is nearly flat near 0 — ~90% of tracks have vocals
- **Speechiness & Liveness** are heavily right-skewed — most tracks are studio recordings with minimal spoken word
- **Valence** is evenly spread, reflecting a healthy mix of happy and sad music

All outliers were **kept** — extreme values (e.g. 10-min songs, instrumental tracks) are valid artistic choices, not data errors.

---

### 2. Hit Rate Threshold Sensitivity

![Hit Rate Threshold](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/HIT_RATE_THRESH.png)

To define "hit" objectively, every threshold from 60 to 85 was tested. Results:

| Threshold | Hit Rate | Decision |
|---|---|---|
| 60 | 27.2% | Too generous — over a quarter of songs can't all be hits |
| 65 | 20.1% | Still too broad |
| 70 | 13.5% | Getting closer |
| **75** | **8.1%** | ✅ Selected — meaningful minority, large enough to analyze |
| 80 | 4.1% | Too strict — loses too much signal |
| 85 | 1.9% | Only mega-hits qualify |

**Threshold = 75** was selected as the sweet spot.

---

### 3. Popularity Distribution & Hit/Non-Hit Split

![Popularity Distribution](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/HIT_THRESH_VALID.png)

The popularity distribution shows a natural drop-off above 70, confirming that 75 sits right at the edge of the top tier. Result: **2,657 hits (8.1%)** vs **30,133 non-hits (91.9%)**. For every 1 hit, there are ~11 non-hit songs.

---

### 4. Hits vs Non-Hits — Median Feature Comparison

![Median Comparison](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/MEDIAN_COMP.png)

All 10 features normalized to 0–1 and compared between hits and non-hits:

- **↑ Danceability** — hits score consistently higher. Groove matters.
- **↑ Loudness** — hits are louder. Tight mastering is industry standard.
- **↑ Acousticness** — hits have a slightly more organic sound.
- **↑ Valence** — hits trend happier. Feel-good music has a streaming edge.
- **↓ Speechiness** — hits are more melodic, less spoken-word.
- **↓ Instrumentalness** — hits almost always have vocals.
- **↓ Liveness** — hits are studio recordings, not live performances.
- **↓ Tempo** — hits are slightly slower (~118 BPM vs ~122 BPM average).
- **↓ Duration** — hits are shorter. Streaming rewards conciseness.
- **~ Energy** — similar between groups; hits are slightly lower in energy.

---

### 5. Correlation Heatmap — All Audio Features

![Correlation Heatmap](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/HEATMAP.png)

Key relationships revealed by the heatmap:

- **Energy ↔ Loudness (`r = +0.68`)** — the strongest link. Louder songs almost always feel more energetic.
- **Energy ↔ Acousticness (`r = −0.54`)** — acoustic tracks feel calmer and less intense.
- **Loudness ↔ Acousticness (`r = −0.36`)** — louder songs are less acoustic.
- **Danceability ↔ Valence (`r = +0.33`)** — happier songs tend to be more danceable.
- **Instrumentalness ↔ Popularity (`r = −0.15`)** — the clearest predictor: vocals = popularity.
- **Duration ↔ Popularity (`r = −0.14`)** — shorter songs perform better.

All other features vs popularity show correlations below 0.10 — no single feature is a magic predictor on its own.

---

### 6. Audio Features vs Track Popularity — Scatter Plots

![Scatter Plots](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/SCATER.png)

Each feature plotted against popularity with a linear trend line (`r` values shown):

- **Instrumentalness (`r = −0.15`)** and **Duration (`r = −0.14`)** show the clearest negative trends.
- **Acousticness (`r = +0.09`)** is the strongest positive predictor — a touch of organic sound helps.
- **Tempo (`r ≈ 0.00`)** is essentially flat — tempo alone does not drive popularity.
- **Energy, Speechiness, Liveness, Valence** all show weak but directionally consistent trends.

---

### 7. Music Trends Over Time

![Music Trends Over Time](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/TIME.png)

Hit rates peaked in the mid-1960s (~17%) and surged again post-2015, reaching ~14% by 2019.

**Why recent songs score higher:** Spotify's popularity score is recency-weighted — it favors tracks with recent, high-volume streaming activity over all-time totals. Combined with the viral amplification of social media (TikTok, Instagram Reels), newer songs accumulate streams far faster than songs from previous decades could. This is a measurement artifact, not evidence that modern music is objectively better.

---

### 8. Songs count per year.


![Music Trends Over Time](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/track_count_per_year.png)

Clearly shows a difference in amount of data per each year.
### 9. Hit Ratio by Genre

![Hit Ratio by Genre](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/GENERE.png)

| Genre | Hit Ratio | Notes |
|---|---|---|
| **Pop** | **0.157** | Nearly double the overall average — designed for mass appeal |
| **Latin** | **0.138** | Strong global wave (Bad Bunny, J Balvin) shows in the data |
| **R&B** | 0.098 | Above average — loyal streaming audience |
| **Rap** | 0.060 | Below average |
| **EDM** | 0.044 | Lowest — EDM success lives in festivals, not Spotify streams |

---

### 10. Sweet Spot by Genre — KDE Analysis

![Sweet Spot by Genre](https://huggingface.co/datasets/Noam12345/my_first_asignment_ds/resolve/main/PHOTOS_NEW/SWEET_SPOT.png)

KDE plots showing where hit songs cluster for **Pop, Latin, and R&B** across all 10 features:

**Universal across all genres:**
- High energy (0.6–0.8), high loudness, near-zero instrumentalness and speechiness, low liveness, short duration (3–4 min)

**Genre-specific differences:**
- **Tempo:** Pop/R&B peak at 95–125 BPM; Latin shows a second peak at ~175 BPM (reggaeton)
- **Danceability:** Latin requires the highest (~0.78), R&B allows more variation
- **Valence:** Pop and Latin lean happier; R&B spans a wider emotional range
- **Acousticness:** R&B allows more organic elements than Pop



## ✅ Summary & Conclusions









#### General range for popular music

| Feature | Direction | Hit Sweet Spot |
|---|---|---|
| Danceability | ↑ positive | 0.70–0.85 |
| Energy | ~ neutral | 0.60–0.80 |
| Loudness | ↑ positive | 85–90 (scaled) |
| Speechiness | ↓ negative | < 0.10 |
| Tempo | ~ irrelevant | Genre-dependent |
| Acousticness | ↑ slight positive | Slightly above average |
| Instrumentalness | ↓ strong negative | ~0.0 (vocals essential) |
| Liveness | ↓ negative | < 0.15 (studio only) |
| Valence | ↑ slight positive | Slightly above average |
| Duration | ↓ negative | 3–4 minutes |


### Genre Sweet Spot Tables — Producer Cheat Sheet
Based on the KDE analysis of hit songs in each of the top 3 genres, here are the optimal feature ranges:


#### 🎤 Pop

| Feature | Sweet Spot | Notes |
|---|---|---|
| Danceability | 0.65–0.80 | High but not extreme — steady groove |
| Energy | 0.60–0.85 | Energetic, radio-ready feel |
| Loudness (scaled) | 85–90 | Loud, compressed mastering |
| Speechiness | < 0.10 | Melody-first — minimal spoken word |
| Tempo | 95–130 BPM | Standard pop range |
| Acousticness | 0.05–0.25 | Mostly produced, slight organic touch |
| Instrumentalness | ~0.00 | Vocals are essential |
| Liveness | < 0.15 | Studio production only |
| Valence | 0.40–0.75 | Leans happy and upbeat |
| Duration | 3.0–3.8 min | Short, replayable tracks |

#### 💃 Latin

| Feature | Sweet Spot | Notes |
|---|---|---|
| Danceability | 0.75–0.85 | Highest of all genres — dance is mandatory |
| Energy | 0.60–0.85 | High energy, rhythm-driven |
| Loudness (scaled) | 85–90 | Same loud standard as pop |
| Speechiness | < 0.10 | Singing over speaking |
| Tempo | 95–105 *or* 170–200 BPM | Bimodal — slow reggaeton *and* fast salsa/cumbia |
| Acousticness | 0.10–0.35 | Slightly more organic than pop |
| Instrumentalness | ~0.00 | Always vocal |
| Liveness | < 0.15 | Studio recordings dominate |
| Valence | 0.50–0.80 | Happy, feel-good energy |
| Duration | 3.0–4.0 min | Streaming-friendly length |

#### 🎶 R&B

| Feature | Sweet Spot | Notes |
|---|---|---|
| Danceability | 0.60–0.80 | Wider range — groove matters but isn't everything |
| Energy | 0.50–0.75 | Slightly lower — allows smooth, mellow tracks |
| Loudness (scaled) | 80–90 | A touch softer than pop/Latin |
| Speechiness | < 0.15 | Slightly higher tolerance (rap-influenced R&B) |
| Tempo | 95–125 BPM | No second peak — steady mid-tempo |
| Acousticness | 0.10–0.45 | Most acoustic-friendly genre — organic sound welcome |
| Instrumentalness | ~0.00 | Vocals remain essential |
| Liveness | < 0.15 | Studio quality standard |
| Valence | 0.25–0.70 | Widest emotional range — moody and happy both work |
| Duration | 3.0–4.2 min | Slightly longer than pop is acceptable |

> **Bottom line:** High energy, loud mastering, vocals, and 3–4 minutes are universal. The real genre fingerprint lives in **tempo** (Latin's bimodal peaks), **danceability** (Latin demands the most), **acousticness** (R&B allows the most), and **valence** (R&B allows the darkest mood).

---





---

## 👤 Author

**Noam Fuchs** — Data Science Assignment #1, March 2026