A data-driven system that identifies the optimal review timing (Sweet Spot)
to maximize vocabulary retention efficiency.
Hệ thống phân tích dữ liệu học tiếng Anh, mô phỏng hành vi học từ vựng của 100 người dùng với 2000 từ.
Mục tiêu: tìm ra Sweet Spot of Review và xây dựng recommendation system cá nhân hóa dựa trên dữ liệu.
Target role: Data Analyst / Data Scientist Intern
Học từ vựng kém hiệu quả thường do 2 nguyên nhân:
- Review quá sớm → lãng phí thời gian (từ chưa kịp quên)
- Review quá muộn → đã quên nhiều, phải học lại từ đầu
Project này xác định Sweet Spot — khoảng thời gian tối ưu để ôn lại — từ dữ liệu thực tế:
- Giảm review không cần thiết
- Ngăn forgetting trước khi xảy ra
- Cải thiện learning efficiency ~10–15% so với random review schedule
Data Generation → ETL → Feature Engineering → Analysis → Insight → Recommendation
↓ ↓ ↓ ↓ ↓ ↓
raw/*.csv cleaned data 6 features 6 queries 7 charts priority score
adaptive-learning-insights/
├── raw/ # Generated datasets
│ ├── words.csv # 2,000 words (level, frequency)
│ ├── users.csv # 100 users (CEFR level)
│ └── quiz_results.csv # 10,000 quiz results
├── processed/ # ETL output + chart files
│ ├── quiz_results_clean.csv # Cleaned data with 6 derived features
│ ├── retention_curve.csv # Retention data for Sweet Spot chart
│ └── chart*.html # 7 interactive Plotly charts
├── scripts/
│ ├── generate_data.py # Data simulation (SRS-controlled spacing)
│ ├── etl.py # ETL pipeline
│ ├── generate_charts.py # 7 visualizations
│ └── recommendation.py # Recommendation system
└── analysis/
└── analysis.ipynb # 6 analytical queries
| File | Records | Key Fields |
|---|---|---|
words.csv |
2,000 | word_id, level (A1–C2), frequency (high/medium/low) |
users.csv |
100 | user_id, level (A1–C2) |
quiz_results.csv |
10,000 | is_correct, time_spent, attempt_number, days_since_last_seen |
Data được simulate với 4 yếu tố logic rõ ràng:
probability = base_from_level_gap # word level vs user level
- frequency_penalty # rare words harder to remember
- forgetting_penalty # step-based: -5% (>3d), -15% (>7d), -20% (>14d)
+ learning_boost # +3% per attempt, max +15%Controlled spaced repetition spacing (phản ánh SRS thực tế):
| Attempt | Interval |
|---|---|
| 2nd review | 1–3 days |
| 3rd review | 5–7 days |
| 4th review | 10–14 days |
| 5th+ review | 15–30 days |
Input: raw/quiz_results.csv
Output: processed/quiz_results_clean.csv (17 columns)
- Remove duplicates
- Filter invalid
time_spent(< 0 hoặc > 300 giây) - Drop missing values on critical fields
| Feature | Logic |
|---|---|
level_gap |
word_level_score − user_level_score |
is_hard_word |
level_gap > 0 |
is_slow_response |
time_spent > median + 1 std |
is_rare_word |
frequency == 'low' |
forgetting_risk |
days_since_last_seen > 7 |
user_performance_score |
Rolling accuracy (window=10) per user |
| # | Query | Finding |
|---|---|---|
| 1 | Word difficulty by level | Higher-level words (C1–C2) show significantly higher error rates, primarily due to mismatch with user proficiency levels |
| 2 | Hardest words | Top words với error rate 80–91%, phần lớn là C1/C2 level |
| 3 | Learning curve | Accuracy cải thiện qua các lần attempt — learning effect rõ ràng |
| 4 | Time vs accuracy | Correlation −0.70: time spent là signal của struggling, không phải careful thinking |
| 5 | User performance | ~25% users có accuracy < 40% — cần intervention |
| 6 | Retention analysis | Accuracy giảm từ 60.7% → 50.0% sau day 10 → Sweet Spot tại day 9 |
| Chart | Type | Insight |
|---|---|---|
chart1_difficulty_by_level.html |
Bar | Error rate tăng theo CEFR level |
chart2_learning_curve.html |
Line | Accuracy cải thiện qua attempts |
chart3_error_heatmap.html |
Heatmap | User level × Word level error pattern |
chart4_time_vs_accuracy.html |
Scatter | Time spent là signal của difficulty |
chart5_user_distribution.html |
Histogram | Distribution of user accuracy |
chart6_sweet_spot.html ⭐ |
Line + markers | Sweet Spot of Review |
chart7_retention_buckets.html |
Bar | Retention by time bucket (supports chart 6) |
Signature insight của project.
Baseline accuracy (0–3 days): 60.7%
Accuracy at day 10: 50.0%
Drop: −10.7%
→ Review BEFORE day 10 để giữ accuracy trên threshold
Threshold definition (10% drop):
| Drop | Interpretation |
|---|---|
| < 5% | Measurement noise — không đáng kể |
| 10% | Significant → actionable threshold ✅ |
| > 15% | Đã quên quá nhiều — recovery tốn kém hơn |
Retention curve:
result = get_recommendation(user_id, top_n=20)
# Returns:
# result['review_words'] — words to review (priority scored)
# result['new_words'] — new words to learn
# result['final_recommendation'] — combined, normalized 0–10
# result['stats'] — user profile summary# Review score — fully explainable weights
review_score = 0.6 × norm(days_since_last_seen) # main: recency
+ 0.3 × norm(wrong_count) # secondary: error history
+ 0.1 × norm(frequency_penalty) # minor: word rarity
# New word score
new_score = frequency_map: {high: 3.0, medium: 2.0, low: 1.0}review_mask = (
(wrong_count >= 2) | # struggling
(days_since > 9) & (accuracy < 0.7) # forgetting risk + weak accuracy
)| User Accuracy | Review slots | New word slots |
|---|---|---|
| < 50% (weak) | 16 | 4 |
| 50–70% (avg) | 15 | 5 |
| ≥ 70% (strong) | 12 | 8 |
# Both scores normalized on COMBINED scale before ranking
# → apples-to-apples comparison, no scale bias
combined = pd.concat([review_tagged, new_tagged])
combined['priority_score'] = normalize(combined['priority_score'])word level type priority_score
word_634 B2 review 10.00 ← wrong×2 + 797 days unseen
word_1078 C1 review 9.97
word_1838 C1 review 9.78
...
word_22 B2 new 0.00 ← guaranteed explore slot
word_6 B2 new 0.00
# 1. Clone & setup
git clone https://github.com/baorphuc/adaptive-learning-insights
cd adaptive-learning-insights
# 2. Create and activate virtual environment (Recommended)
python -m venv venv
# On Windows:
.\venv\Scripts\activate
# On macOS/Linux:
# source venv/bin/activate
# 3. Install dependencies
pip install -r requirements.txt
# 4. Generate data
python scripts/generate_data.py
# 5. Run ETL
python scripts/etl.py
# 6. Generate all charts
python scripts/generate_charts.py
# → Open processed/chart6_sweet_spot.html in browser
# 7. Run recommendation
python scripts/recommendation.py
# → Output: processed/rec_<user_id>_final.csv| Tool | Usage |
|---|---|
| Python 3.10 | Core language |
| pandas | Data manipulation & feature engineering |
| numpy | Numerical operations, normalization |
| Plotly | Interactive HTML visualizations |
| Jupyter | Exploratory analysis notebooks |
- Proficiency mismatch drives error rates — not inherent word difficulty alone
- Time spent = struggling signal — correlation −0.70 with accuracy
- Sweet Spot at day 9 — accuracy drops 10.7% after day 10 without review
- 25% users need intervention — accuracy below 40% threshold
- Learning effect is real — repeated exposure consistently improves accuracy
- Simulated data — không phải hành vi người dùng thực tế; patterns được thiết kế có chủ đích
- Simplified forgetting model — dùng step-based thresholds thay vì exponential decay (Ebbinghaus)
- Rule-based recommendation — không có personalization từ ML; weights được set thủ công
- No real-time adaptation — system không cập nhật khi user thay đổi hành vi
- Sweet Spot generalized — day 9 là trung bình toàn dataset, thực tế sẽ khác nhau theo từng user và word type
MIT License — free to use for educational and portfolio purposes.