Research map / Label-efficient and robust learning / Self-supervised and contrastive
Scalable Early Childhood Reading Performance Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the ECRI longitudinal dataset of early reading assessments from 6,916 first-graders in 44 US schools and proposes MaskMLP, an MLP pre-trained with a cosine embedding loss between original and masked inputs, for predicting intervention response.
Problem
Predicting which early-elementary students will make sufficient reading progress requires large longitudinal educational data, which is not publicly available; partially observed (missing) tabular data further complicates modeling.
Approach
A simple self-supervised pre-training of an MLP that aligns embeddings of the original and randomly masked inputs (cosine loss), without imputing missing values, which is then fine-tuned for binary progress classification.
Key results
- MaskMLP reaches 0.6919 accuracy and 0.6899 AUC on word identification (student split), and 0.7704 accuracy on the intervention subset in the school split, best among the compared methods.
- MaskMLP improves over MLP baseline with statistically significant t-test (t=3.876, p=0.0090) across five folds on the word identification task.
- Feature removal shows the baseline word identification score is the most influential input feature.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Identification, school split, all students | AUC: 0.6607 -> 0.6693 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Identification, school split, intervention group | AUC: 0.7394 -> 0.7633 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Attack, school split, all students | AUC: 0.6321 -> 0.6545 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Attack, school split, intervention group | AUC: 0.6403 -> 0.6869 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Identification, student split, all students | AUC: 0.6873 -> 0.6899 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Identification, student split, intervention group | AUC: 0.7198 -> 0.7537 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Attack, student split, all students | AUC: 0.6202 -> 0.6485 | Helped |
| Use masked-input self-supervised pre-training rather than missing-data indicators | MLP on ECRI · Word Attack, student split, intervention group | AUC: 0.6851 -> 0.7333 | Helped |
| Add MSE to cosine pre-training loss (ablation) | MaskMLP · Word Identification, school split | AUC: 0.6693 -> 0.6715 | Helped |
| Replace cosine loss with MSE pre-training loss (ablation) | MaskMLP · Word Identification, school split | AUC: 0.6693 -> 0.6432 | Hurt |
| Replace cosine loss with MSE pre-training loss (ablation) | MaskMLP · Word Identification, student split | AUC: 0.6899 -> 0.6882 | No clear effect |
| Replace cosine loss with Contrastive pre-training loss (ablation) | MaskMLP · Word Identification, school split | AUC: 0.6693 -> 0.6627 | Hurt |
Surprising findings
- Reported benefits of VIME's reconstruction loss and SCARF's contrastive/auxiliary losses did not carry over to this domain; a simple MLP baseline with masked pre-training outperformed them.
- Model performance is lower for at-risk students and for schools with higher free/reduced-price lunch rates.
Limitations
- Model generalizability remains limited across schools and student profiles, especially for at-risk students.
- Performance is lower on students with lower performance gains, reflecting data imbalance.
- Results rely on five-fold cross-validation without error bars across random seeds.
Experimental setup
Datasets: ECRI dataset
Benchmarks: ECRI dataset
Metrics: accuracy AUC
Baselines: Logistic Regression XGBoost LightGBM TabNet GRAPE TabNet (Pretrain) VIME SCARF MLP (Zeros) MLP (Mean) MLP (Indicator)
Builds on: VIME SCARF TabNet
Closest papers by meaning
- EduLLM: Leveraging Large Language Models and Framelet-Based Signed Hypergraph Neural Networks for Student Performance Prediction (ICML 2025 · similarity 0.82)
- CurBench: Curriculum Learning Benchmark (ICML 2024 · similarity 0.82)
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs (NeurIPS 2026 · similarity 0.80)
- A Toolbox for Modelling Engagement with Educational Videos (AAAI 2024 · similarity 0.80)
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.80)
- Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data (NeurIPS 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.