Real validation result: restaurant behavior predicts hotel taste at 0.92–0.95 cosine similarity
Hotel reviews were held out entirely. The engine predicted each reviewer's hotel taste vector from restaurant behavior alone, then measured cosine similarity against the held-out truth. Run on Railway PostgreSQL — not synthetic browser data.
Synthetic Reviewers
800
across restaurant + hotel
Total Reviews
4,408
LLM-generated, two domains
Items Catalogued
450
restaurants + hotels
Cross-Domain Cosine
0.92–0.95
vs held-out hotel ground truth
Hotel Reviews Held Out
100%
zero hotel data used in prediction
Honesty disclosure: this dataset is synthetic — generated with Claude before approaching any partner. The architecture is identical to what would run on real FunNow or Eztable data. The point is not the number — it's that the pipeline works end-to-end, built before asking anyone for data.
Interactive Methodology Demonstration
The validated result above was run on Railway. This interactive simulation demonstrates the same trust propagation methodology on a smaller synthetic dataset — upload your own CSV to re-run the engine on custom data. Expected columns: user_id, reviewer_id, restaurant_id, food, ambiance, service, value, noise.
Methodology demo · synthetic data · 100 simulated users
Section 01
Model Comparison Overview
Three models were tested head-to-head. Each model was given the same training data and asked to predict a user's satisfaction rating for restaurants they hadn't rated. Predictions were compared against held-out ground truth ratings.
Section 02
Prediction Error by Model
Mean Absolute Error (MAE) measures the average gap between predicted and actual ratings (on a 1–10 scale). Lower is better. A MAE of 1.0 means predictions are off by 1 point on average.
MAE — lower is better
MAE by Trust Degree
* Shows Kornerway MAE broken down by whether the prediction came from a direct (1°) or indirect (2°, 3°+) trust connection.
MAE by Attribute
* Shows which attributes Kornerway predicts most and least accurately.
Section 03
Prediction Correlation & Statistical Significance
Pearson correlation (r) measures how well predicted ratings track actual ratings. A value of 1.0 is a perfect match. p-value confirms whether the result is statistically significant (p < 0.05 = significant).
Pearson r — higher is better
Section 04
Does Indirect Trust Still Work?
The core hypothesis: even reviewers reached through 2° or 3° connections should predict user satisfaction better than random. This section isolates predictions made exclusively from indirect trust chains.
Prediction accuracy degradation by trust degree
* Each bar shows the MAE for predictions sourced exclusively from reviewers at that degree of separation. The dotted line shows the random baseline.
Section 05
Attribute-Specific Trust Performance
Kornerway computes separate trust scores per attribute (food, ambiance, service, value, noise). This section tests whether attribute-specific trust outperforms global trust scores for each dimension.
* Global trust = single cosine similarity score across all attributes. Attribute trust = per-dimension similarity score. Improvement = reduction in MAE.
Section 06
Confidence Intervals (Bootstrap, n=1000)
Bootstrap resampling tests whether results are stable or could be due to sampling luck. 1,000 resamples of the test set were run. The 95% confidence interval shows the range within which the true MAE falls with 95% certainty.
