DATASET_ISSUES.md

# GitHub Trending Projects Dataset - Known Issues & Limitations

## Dataset Overview
- **Total Projects:** 423,098
- **Date Range:** 2013-08-21 to 2025-11-30
- **Unique Repositories:** 14,500
- **Success Rate:** 89.8% (17,127/19,064 URLs)

---

## 🚨 Major Issues

### 1. **Missing Star/Fork Count Data (2013-2019)**
**Severity:** High
**Affected:** 25,150 entries (5.9%)

**Problem:**
- 100% of 2013-2019 data lacks star/fork counts
- Only data from 2020+ has star/fork information
- This is due to HTML structure differences in older Wayback Machine snapshots

**Impact:**
- Cannot compare popularity metrics for pre-2020 projects
- Monthly rankings rely solely on trending score for 2013-2019
- Incomplete analysis for historical trends

**Affected Years:**
```
2013: 100% missing (150 entries)
2014: 100% missing (125 entries)
2015: 100% missing (325 entries)
2016: 100% missing (1,200 entries)
2017: 100% missing (1,550 entries)
2018: 100% missing (4,324 entries)
2019: 100% missing (17,475 entries)
2020+: 0% missing (397,949 entries)
```

**Recommendation:**
- Use weighted trending score only for historical analysis
- Clearly document this limitation when presenting data
- Consider scraping current star counts from GitHub API for historical projects

---

### 2. **Uneven Temporal Distribution**
**Severity:** High
**Affected:** All data

**Problem:**
- Snapshot frequency varies dramatically: 1 to 31 snapshots per month
- Some months have 1 snapshot (25 projects), others have 31 (15,763 projects)
- 31x variance in data density across time periods

**Examples:**
```
Sparse months (1 snapshot):
- 2015-04: 25 projects
- 2015-06: 25 projects
- 2016-11: 25 projects

Dense months (31 snapshots):
- 2019-05: 4,650 projects
- 2020-01: 17,446 projects
- 2020-05: 15,763 projects
```

**Impact:**
- Over-representation of 2019-2020 period
- Monthly scores favor periods with more snapshots
- Difficult to compare across time periods fairly
- Projects appearing in dense months get inflated scores

**Recommendation:**
- Normalize scores by dividing by number of snapshots per month
- Weight monthly rankings by data density
- Consider resampling to create uniform temporal distribution

---

### 3. **Inconsistent Star/Fork Count Timing**
**Severity:** Medium
**Affected:** All entries with star counts (67.8%)

**Problem:**
- Star/fork counts are "maximum ever recorded" across all snapshots
- A 2015 project's star count might be from 2025
- A 2025 project's star count is from 2025
- Not temporally consistent or comparable

**Example Issues:**
```
Project A (trending 2015):
- Trending date: 2015-03-15
- Star count: 100,000 (scraped 2025)
- Had 10 years to accumulate stars

Project B (trending 2025):
- Trending date: 2025-03-15
- Star count: 20,000 (scraped 2025)
- Had 0 years to accumulate stars

Issue: Can't fairly compare popularity
```

**Impact:**
- Older projects appear more popular (survival bias)
- Can't analyze "stars at time of trending"
- Misleading for popularity comparisons across eras

**Recommendation:**
- Document this clearly: "Stars represent current popularity, not popularity when trending"
- Consider using trending score only for cross-era comparisons
- For accurate historical analysis, would need to scrape stars from archived snapshots

---

### 4. **Multiple Appearances Bias**
**Severity:** Medium
**Affected:** Scoring methodology

**Problem:**
- Some projects appear 1,900+ times, others appear once
- Scoring favors projects that "stick around" on trending
- Brief but intense viral projects get undervalued

**Distribution:**
```
1 appearance: 1,129 projects (7.8%)
2-5 appearances: 1,852 projects (12.8%)
6-10 appearances: 3,732 projects (25.7%)
11-50 appearances: 6,005 projects (41.4%)
50+ appearances: 1,782 projects (12.3%)
```

**Most Over-Represented:**
```
1. jwasham/coding-interview-university: 1,948 appearances
2. TheAlgorithms/Python: 1,891 appearances
3. donnemartin/system-design-primer: 1,865 appearances
4. public-apis/public-apis: 1,830 appearances
5. EbookFoundation/free-programming-books: 1,737 appearances
```

**Impact:**
- "Evergreen" educational repos dominate rankings
- Viral new projects undervalued if they trend briefly
- Doesn't distinguish between sustained vs. brief trending

**Recommendation:**
- Create separate rankings: "Most Consistent" vs "Peak Trending"
- Add "peak rank achieved" metric
- Consider decay function for repeated appearances

---

### 5. **Linear Scoring Assumption**
**Severity:** Low-Medium
**Affected:** Monthly rankings

**Problem:**
- Current scoring: Rank 1 = 25 pts, Rank 2 = 24 pts (linear)
- Assumes rank 1→2 has same value as rank 24→25
- In reality, top positions have exponentially more visibility

**Distribution:**
```
Rank 1-5: 90,280 entries (21.3%)
Rank 6-10: 90,178 entries (21.3%)
Rank 11-15: 87,522 entries (20.7%)
Rank 16-20: 79,516 entries (18.8%)
Rank 21-25: 75,602 entries (17.9%)
```

**Impact:**
- Undervalues #1 position
- May not reflect actual visibility/impact differences
- Alternative exponential scoring might be more accurate

**Recommendation:**
- Consider exponential scoring: 2^(25-rank)
- Or logarithmic: log(26-rank)
- A/B test different scoring functions against actual star growth

---

### 6. **Failed Scrapes & Missing Data**
**Severity:** Medium
**Affected:** 1,937 URLs (10.2%)

**Problem:**
- SSL/TLS incompatibility with 2014-2019 Wayback snapshots
- Incomplete Wayback Machine captures
- Connection timeouts and 503 errors

**Impact:**
- Gaps in temporal coverage
- Some dates completely missing
- Potential systematic bias if certain types of snapshots fail more

**Affected Periods:**
```
2014-10-01 to 2014-12-21: Many failures
2016-02-24 to 2016-03-11: Several failures
2019-06-12 to 2019-12-31: Heavy failures (mid-2019 SSL issues)
2024-10-28: 3 failures (503 errors)
```

**Recommendation:**
- Retry failed URLs periodically (Wayback Machine availability changes)
- Use GitHub API to fill gaps where possible
- Document missing date ranges in analysis

---

### 7. **Rank Distribution Skew**
**Severity:** Low
**Affected:** Lower-ranked entries

**Problem:**
- Fewer entries at ranks 21-25 (75,602) vs ranks 1-5 (90,280)
- Suggests some snapshots had <25 projects
- Or extraction issues with lower-ranked items

**Impact:**
- Scoring may overvalue top ranks due to sample size
- Statistical significance varies by rank position

**Recommendation:**
- Filter analysis to top 20 for consistency
- Or normalize scores by rank availability

---

## 📊 Dataset Quality Metrics

### Completeness
```
✅ Temporal Coverage: 89.8% (128/142 months have data)
❌ Star/Fork Data: 67.8% complete (missing all pre-2020)
✅ Rank Data: 100% complete
✅ Repository Names: 100% complete
```

### Consistency
```
❌ Snapshot Frequency: Highly inconsistent (1-31 per month)
❌ Star Count Timing: Not temporally aligned
⚠️  Scoring Methodology: Linear assumption (debatable)
```

### Reliability
```
✅ Scraping Success: 89.8%
❌ Failed URLs: 10.2% (recoverable with retry)
✅ Data Validation: No duplicate entries detected
```

---

## 🔧 Recommended Fixes

### High Priority
1. **Add normalized scores** that account for snapshot frequency
2. **Document star count timing issue** prominently in analysis
3. **Create separate pre-2020 and post-2020 analyses** due to missing data
4. **Retry failed URLs** to improve coverage

### Medium Priority
5. **Test exponential scoring** vs linear for better accuracy
6. **Add "peak rank" metric** to identify viral projects
7. **Separate "evergreen" vs "viral" rankings**
8. **Scrape current GitHub API data** to fill historical gaps

### Low Priority
9. Create confidence intervals for sparse months
10. Add data quality flags per entry
11. Document GitHub trending algorithm changes over time

---

## 📝 Usage Guidelines

### ✅ Good Uses
- Identifying trending patterns in 2020-2025 (complete data)
- Analyzing trending frequency/consistency
- Discovering historically significant projects
- Comparative analysis within same time period

### ⚠️ Use With Caution
- Cross-era popularity comparisons (star count issues)
- Monthly comparisons with very different snapshot counts
- Absolute popularity rankings (use GitHub API instead)
- Historical analysis pre-2020 (missing star/fork data)

### ❌ Not Recommended
- Claiming "most popular project ever" (timing issues)
- Direct star count comparisons across decades
- Precise month-to-month trending velocity analysis (uneven sampling)
- Analysis of projects that trended <5 times (insufficient data)

---

## 📈 Data Quality by Year

| Year | Projects | Star Data | Snapshots | Quality Grade |
|------|----------|-----------|-----------|---------------|
| 2013 | 150      | 0%        | Low       | D (Minimal)   |
| 2014 | 125      | 0%        | Low       | D (Minimal)   |
| 2015 | 325      | 0%        | Low       | D (Minimal)   |
| 2016 | 1,200    | 0%        | Low       | D (Minimal)   |
| 2017 | 1,550    | 0%        | Low       | D (Minimal)   |
| 2018 | 4,324    | 0%        | Medium    | C- (Limited)  |
| 2019 | 17,475   | 0%        | High      | C+ (Incomplete)|
| 2020 | 108,672  | 100%      | High      | A- (Excellent)|
| 2021 | 70,006   | 100%      | High      | A- (Excellent)|
| 2022 | 74,915   | 100%      | High      | A- (Excellent)|
| 2023 | 73,674   | 100%      | High      | A- (Excellent)|
| 2024 | 46,538   | 100%      | High      | A- (Excellent)|
| 2025 | 24,144   | 100%      | Medium    | A- (Excellent)|

---

## 🎯 Conclusion

This dataset is **excellent for 2020-2025 analysis** but has **significant limitations for historical (2013-2019) analysis**. The primary issues are:

1. **Missing star/fork data pre-2020** (structural limitation)
2. **Uneven temporal distribution** (Wayback Machine artifact)
3. **Star count timing inconsistency** (methodology issue)

These issues are **documentable and manageable** but should be clearly communicated in any analysis or visualization using this data.

**Overall Grade: B+**
- A+ for recent data (2020-2025)
- C+ for historical data (2013-2019)
- Excellent for trending patterns, limited for absolute popularity metrics