README.md

---
license: apache-2.0
task_categories:
- tabular-classification
language:
- en
pretty_name: Google Play Store Apps 2020 - Cleaned
size_categories:
- 1K<n<10K
---

# 📱 Google Play Store 2020 — What Made an App Go Viral?

> **A data-driven EDA of 9,436 apps released in 2020, exploring the patterns behind viral success on the Play Store.**

---

## 🎬 Video Walkthrough

<!-- Replace the src with your actual video link after uploading -->
<video src="https://www.loom.com/share/3b2dc51bcf384a339d3d3328766e4705" controls="controls" style="max-width: 720px;"></video>

*Can't see the video? [Click here to watch](https://www.loom.com/share/3b2dc51bcf384a339d3d3328766e4705)*

---

## 🔬 Research Question

> **"What made an app released in 2020 go viral?"**

Defined as reaching **1M+ installs by June 2021** — within 6–18 months of launch.

2020 was chosen deliberately: the COVID-19 pandemic drove unprecedented mobile app adoption, and since the data was collected in June 2021, apps from 2020 had a consistent 6–18 month measurement window.

---

## 📊 Key Visualizations

### 1. The Winner-Takes-All Market
86% of apps never exceed 10K installs. Viral apps (<1%) are a thin but real right tail.

![Installs Distribution](https://huggingface.co/datasets/yonilev/GooglePlayStoreApps/resolve/main/viz_01_installs_distribution.png)

---

### 2. Does Quality = Success? (Surprising Answer: No)
Rating and Installs show a weak **negative** linear correlation (r = −0.31).
Viral apps attract polarising reviews — millions of users means more critics.

![Rating vs Installs](https://huggingface.co/datasets/yonilev/GooglePlayStoreApps/resolve/main/viz_06_rating_vs_installs.png)

---

### 3. Monetization Strategy Matters
Ad-based and Hybrid (ads + IAP) apps reach significantly higher install counts.
The price barrier for Premium apps dramatically limits reach.

![Monetization vs Installs](https://huggingface.co/datasets/yonilev/GooglePlayStoreApps/resolve/main/viz_04_monetization_violin.png)

---

### 4. The Gap Between Tiers is Enormous
Each tier is roughly **1,000x** the previous — a textbook power-law gap.

| Tier | Median Installs | n apps |
|------|----------------|--------|
| Low | ~100 | 8,135 |
| Medium | ~10,000 | 1,227 |
| Viral | ~1,000,000 | 74 |

![Success Tiers](https://huggingface.co/datasets/yonilev/GooglePlayStoreApps/resolve/main/viz_08_success_tiers.png)
---

## 🔑 Key Findings

1. **The market is winner-takes-all.** 86% of apps sit in the Low tier. Viral apps represent less than 1% of 2020 launches.
2. **Monetization is the strongest predictor.** Ad-based and Hybrid models (ads + IAP) significantly outperform Pure Free and Premium.
3. **Higher ratings do NOT predict more installs** (r = −0.31). Quality alone is not the driver — distribution and visibility are.
4. **Having a rating at all is a strong proxy for traction.** Rated apps contain virtually all Medium and Viral tier apps.
5. **Fresher apps win.** Regular updates correlate negatively with days_since_update (r = −0.25).
6. **The Rating–Installs relationship is category-specific.** A global r = −0.31 masks very different dynamics per category.

---

## 🗂️ Dataset

| Property | Value |
|----------|-------|
| **Source** | [Kaggle — gauthamp10/google-playstore-apps](https://www.kaggle.com/datasets/gauthamp10/google-playstore-apps) |
| **Original size** | 2.31M rows × 24 columns |
| **Subset** | Apps released in 2020 only |
| **Sample size** | 9,436 rows (Cochran's formula + stratified by Category) |
| **Scraped** | June 2021 |

### Features in this dataset

| Feature | Type | Description |
|---------|------|-------------|
| `App Name` | string | Name of the app |
| `Category` | categorical | App category (48 unique) |
| `Rating` | float | Average user rating (NaN if unrated) |
| `Rating Count` | int | Number of user ratings |
| `Installs_clean` | int | Parsed install count |
| `log_installs` | float | log(1 + Installs) — for modeling |
| `Size_MB` | float | App size in MB (winsorized at 65.8MB) |
| `Min_Android_ver` | float | Minimum Android version required |
| `Released` | datetime | Launch date (all in 2020) |
| `Last Updated` | datetime | Last update date |
| `Free` | bool | Whether app is free |
| `Ad Supported` | bool | Whether app shows ads |
| `In App Purchases` | bool | Whether app has IAP |
| `Editors Choice` | bool | Whether app is Editor's Choice |
| `has_rating` | bool | Whether app has crossed rating threshold |
| `app_age_days` | int | Days from release to scrape date |
| `days_since_update` | int | Days from last update to scrape date |
| `engagement_ratio` | float | Rating Count / Installs |
| `log_engagement` | float | log(1 + engagement_ratio) |
| `developer_app_count` | int | Total apps by same developer in 2020 |
| `monetization_model` | categorical | Pure Free / Ad-based / Freemium / Hybrid / Premium |
| `success_tier` | categorical | **Target variable** — Low / Medium / Viral |

---

## 📓 Notebook

The full analysis notebook is available here:
👉 [Open The Notebook in Google Colab](https://colab.research.google.com/drive/1YxqN2Urjli1ToxtYO9LtUztXxb5SeEcb?usp=sharing)
---

## ❓ Questions & Answers

**Q1: Is the app market democratic — can any app go viral?**
No. 86% of apps released in 2020 never exceeded 10K installs. The market is winner-takes-all: less than 1% of apps reached the Viral tier (1M+ installs), and their median install count is roughly 1,000x that of Medium-tier apps.

**Q2: Does a higher rating mean more installs?**
Surprisingly, no. Rating and Installs show a weak negative linear correlation (r = −0.31). Viral apps attract polarising reviews — millions of users means more critics. Quality alone is not the driver; distribution and visibility are.

**Q3: Does monetization strategy matter?**
Yes — it's the strongest predictor in the dataset. Ad-based and Hybrid (ads + IAP) apps reach significantly higher install counts than Pure Free or Premium apps. The price barrier of Premium apps dramatically limits reach.

**Q4: Does having a rating at all matter?**
Yes, dramatically. Apps with a visible rating contain virtually all Medium and Viral tier apps. This reflects a chicken-and-egg dynamic: installs drive ratings, ratings drive visibility, visibility drives more installs.

**Q5: Does the Rating–Installs relationship hold across all categories?**
No. The global r = −0.31 masks very different dynamics per category. In Tools and Business, there is almost no relationship. In Entertainment and Games, high-install apps cluster at specific rating bands.

---

## 🔧 Key Decisions

| Decision | What I did | Why |
|----------|-----------|-----|
| **Subset to 2020** | Filtered 544,882 apps released in 2020 | Consistent 6–18 month measurement window for all apps |
| **Stratified sampling** | Cochran's formula → 9,436 rows, sampled proportionally by Category | Preserve category distribution while reducing computational load |
| **Rating = 0 → NaN** | Replaced zero ratings with NaN | Zero means unrated, not bad — imputing would distort analysis |
| **Rating not imputed** | Kept 54.9% of Rating as NaN, created `has_rating` instead | Missing rating is a meaningful signal, not random noise |
| **Size_MB winsorized** | Capped at 65.8 MB (IQR upper fence) | Extreme sizes are edge cases that distort correlations |
| **Installs kept as-is** | No capping on extreme install counts | Viral outliers ARE the story — removing them defeats the purpose |
| **log_installs** | Applied log(1 + Installs) transformation | Raw installs follow a power-law — log scale is needed for analysis |
| **monetization_model** | Combined Free + Ad Supported + In App Purchases into 5 categories | Three boolean columns carry more meaning as a single business model feature |
```

## 👤 Author

**Yonathan Levy**
Econ & Entrepreneurship with Data Science Specialization
Reichman Uni
Class of 2028