CIS635 Semester Project - Data Fusion Methods for Estimating Funding Impacts on Educational Outcomes
Does how you aggregate school funding data change what machine learning can predict? We test three competing formulations — per-pupil, aggregate sum, and percentage allocation — using a multi-source district-level data fusion pipeline adapted from Chango et al. (2021).
Alysha Nadeem · Ishra Naznin · MD Imran Khaled · Kenaniah Williams
Grand Valley State University — CIS 635: Knowledge Discovery and Data Mining — April 2026
Which aggregation of school district funding data — per-pupil, aggregate total, or percentage allocation — best predicts student academic achievement, and to what extent does funding explain achievement when socioeconomic context is included?
Presentation YouTube Video Link here. Code Demo YouTube Video Link here.
| Metric | Value |
|---|---|
| Best accuracy | 68.94% (Experiment 1, Sum formulation) |
| Best AUC | 0.89 (Logistic regression, numerical) |
| Sample size | 12,100 districts (77.8% match rate) |
| Strongest predictor | Free/reduced lunch eligibility (r = 0.84) |
| Per-pupil expenditure correlation | r = 0.17 (weak direct relationship) |
| Source | Dataset | Description |
|---|---|---|
| A | SEDA 6.0 | District-level achievement scores, grades 3–8, pooled 2008–2019. Empirical Bayes adjusted composite score used as target variable. |
| B | NCES F-33 (FY 2018–19) | District revenue and expenditure data. Tested in three aggregation variants: per-pupil, aggregate sum, and percentage allocation. |
| C | SEDA Covariates | SES composite, poverty rate, free/reduced lunch eligibility, racial composition, urbanicity. 12 demographic variables. |
The final merged dataset joins SEDA achievement data with F-33 finance data on district LEAID, then left-joins demographic covariates on the same key. Rows with missing values, division-by-zero artifacts, non-positive expenditure, fewer than 50 enrolled students, or per-pupil expenditure above $50,000 were removed.
| Source | Type of Data | Fusion Role |
|---|---|---|
| SEDA 6.0 (Achievement) | Standardized test score estimates | Outcome variable |
| NCES F-33 (Finance) | School financial records | Primary predictors |
| SEDA Covariates | Socioeconomic & demographic variables | Control variables |
- Per-pupil: All revenue and expenditure figures divided by enrollment. Conventional policy normalization. Variables:
exppp,instpp,suppp,cfpp,fedrevpp,locrevpp. - Aggregate sum: Raw total dollar amounts, not divided by enrollment. Preserves absolute scale of district resources. Variables:
expsum,instsum,supsum,cfsum,fedrevsum,locrevsum. - Percentage allocation: Each category as a fraction of its parent total. Captures spending prioritization. Variables:
instpct,suppct,cfmargin,fedrevpct,locrevpct.
Adapted from the four-experiment progressive fusion framework in Chango et al. (2021). Each experiment was applied across all three funding formulations and two data representations (numerical and discretized), evaluated with 10-fold stratified cross-validation.
| Experiment | Strategy | Avg Accuracy (Sum) | Top Classifier |
|---|---|---|---|
| 1 | Merge all attributes | 66.68% | Logistic regression |
| 2 | Global feature selection (SelectKBest) | 64.87% | Logistic regression |
| 3 | Decision-level ensemble (weighted soft vote) | 64.75% | Logistic regression |
| 4 | Ensemble + per-source feature selection | 63.97% | Logistic regression |
Naive merging (Experiment 1) consistently outperformed more complex fusion strategies across all three funding formulations, in contrast to Chango et al. (2021) where feature selection and ensembles improved results.
Mapped from Weka (original paper) to scikit-learn equivalents:
| Original (Weka) | scikit-learn Equivalent |
|---|---|
| J48 | DecisionTreeClassifier (Gini, max_depth=10) |
| REPTree | DecisionTreeClassifier (min_impurity_decrease=0.01) |
| RandomTree | DecisionTreeClassifier (random splitter, sqrt features) |
| JRip | LogisticRegression (max_iter=2000) |
| NNge | KNeighborsClassifier (k=5) |
| PART | DecisionTreeClassifier (entropy, max_depth=8) |
- Numerical: Min-Max normalization to [0, 1].
- Discretized: Equal-width binning into three categories (Low=1, Medium=2, High=3) after normalization.
- Target variable: Tertile (quantile) binning into Low / Medium / High achievement districts, producing approximately balanced classes (~33.3% each). Equal-width binning was not used for the target as it produced extreme class imbalance (~0.3% Low) given the near-normal distribution of achievement scores.
Weighted soft voting across three per-source classifiers with weights A=2, B=1, C=2. Source B (finance) received lower weight as the primary research variable under investigation, mirroring the weighting logic in Chango et al. Equal-weight ensembles (1:1:1) were tested for comparison and produced modestly lower performance.
- Socioeconomic variables dominate feature selection in every formulation. No finance variable from Source B survives global selection when competing against demographic predictors.
- Aggregate sum funding preserves the strongest predictive signal, suggesting that per-pupil normalization removes meaningful variation in absolute district resources.
- Federal revenue per pupil is a negative predictor of achievement, acting as a poverty proxy that reflects Title I targeting of high-need districts.
- Local revenue per pupil is a positive predictor, consistent with property-tax-based wealth disparities across districts.
- Differences between funding formulations are under 0.6% average accuracy across all experiments. The choice of aggregation does not meaningfully alter model performance.
- Discretized data provides a marginal advantage in Experiment 3 but not consistently elsewhere, unlike the original paper where discretization uniformly helped.
| Experiment | Per-Pupil Num Acc | Per-Pupil Disc Acc | Sum Num Acc | Sum Disc Acc | Pct Num Acc | Pct Disc Acc |
|---|---|---|---|---|---|---|
| 1: Merge all | 66.50% | 66.39% | 66.68% | 66.57% | 66.79% | 66.24% |
| 2: Feature selection | 64.87% | 65.40% | 64.87% | 64.97% | 64.87% | 64.97% |
| 3: Ensembles | 65.15% | 64.82% | 64.27% | 64.35% | 64.75% | 65.36% |
| 4: Ensembles + selection | 64.68% | 64.03% | 63.23% | 63.40% | 64.00% | 64.59% |
- Cross-sectional design: finance data from a single fiscal year (2018–19) paired with achievement estimates pooled across 2008–2019.
- Small districts (fewer than 50 students) and extreme per-pupil values (above $50,000) were excluded, which may under-represent rural and very small districts.
- Achievement coarsened into tertiles for class balance; this sacrifices granularity compared to regression approaches.
- Restricted to six white-box classifiers from the original paper for comparability. More powerful models may yield higher accuracy at the cost of interpretability.
- No external validation on a held-out cohort of districts or non-US contexts.
| Column | Meaning |
|---|---|
cs_mn_avg_eb |
Primary outcome: Empirical Bayes average achievement |
cs_mn_mth |
Math achievement |
cs_mn_lrn |
Learning rate across cohorts |
cs_mn_tav |
Achievement trend over time |
| Column | Meaning |
|---|---|
V33 |
Student enrollment |
TCURELSC |
Total current education expenditures |
TCURINST |
Instructional expenditures |
TLOCREV |
Local revenue (property taxes) |
TFEDREV |
Federal revenue |
TSTREV |
State revenue |
TOTALEXP |
Total expenditures |
| Column | Meaning |
|---|---|
exppp |
Per-pupil current expenditure |
instpp |
Per-pupil instructional expenditure |
locrevpp |
Per-pupil local revenue (wealth proxy) |
fedrevpp |
Per-pupil federal revenue |
| Column | Meaning |
|---|---|
perfrl |
% Free/reduced lunch (poverty proxy) |
sesavgall |
Composite SES index |
lninc50avgall |
Log median household income |
povertyavgall |
Poverty rate |
totenrl |
Enrollment |
perwht, perblk, perhsp |
Racial composition |
urban, suburb, town, rural |
Locale indicators |
| Column | Meaning |
|---|---|
sedalea / LEAID |
District merge key |
sedaleaname |
SEDA district name |
NAME |
F-33 district name |
stateabb / STABBR |
State abbreviation |
fips |
State FIPS code |
YEAR |
Fiscal year |
| Column | Meaning |
|---|---|
multi_comp |
Multi-county district flag |
subcat |
Subcategory type |
subgroup |
Student subgroup |
gap |
Achievement gap estimate |
cellcount |
Observation coverage |
tot_asmts |
Total assessments used |
SEDA variables follow a naming pattern:
| Suffix | Meaning |
|---|---|
_ol |
OLS estimate |
_eb |
Empirical Bayes estimate (preferred) |
_se |
Standard error |
_se_adj |
Adjusted standard error |
Example:
cs_mn_avg_eb→ Primary outcome variable (district-level achievement)
| Column | Meaning |
|---|---|
exppp |
Per-pupil current expenditure |
instpp |
Per-pupil instructional spending |
locrevpp |
Per-pupil local revenue (property wealth proxy) |
fedrevpp |
Per-pupil federal revenue |
- Data collection (SEDA + F-33 + Covariates)
- Early fusion at district level
- Feature engineering (per-pupil metrics)
- Correlation analysis
- Regression modeling
- Interpretation of funding–achievement relationships
This project provides a unified framework for analyzing how:
- School funding
- Socioeconomic conditions
- Demographic composition
jointly influence academic achievement across U.S. districts.
| Method | Description |
|---|---|
| Late Fusion | Combine model outputs from separate data streams |
| Hybrid Fusion | Combine early + late fusion approaches |
| Causal modeling | Improve inference beyond correlation/regression |
All datasets used are publicly available through:
- Stanford Educational Opportunity Project (SEDA)
- National Center for Education Statistics (NCES CCD)
Please refer to original sources for citation requirements.
- Chango, W., Cerezo, R., & Romero, C. (2021). Multi-source and multimodal data fusion for predicting academic performance in blended learning university courses. Computers & Electrical Engineering, 89, 106908.
- Meng, T., Jing, X., Yan, Z., & Pedrycz, W. (2020). A survey on machine learning for data fusion. Information Fusion, 57, 115–129.
- Roy, J. (2011). Impact of school finance reform on resource equalization and academic performance: Evidence from Michigan. Education Finance and Policy, 6(2), 137–167.