Formula Used
Weighted prediction
Predicted rating = Σ(similarity × neighbour rating) ÷ Σ|similarity|
Mean-centred prediction
Prediction = target mean + Σ[similarity × (neighbour rating − neighbour mean)] ÷ Σ|similarity|
Baseline estimate
Baseline = global mean + user bias + item bias
Matrix factorisation
Prediction = global mean + user bias + item bias + user vector · item vector
How to Use
Choose user-based, item-based, or matrix factorisation. Select the similarity and prediction settings. Keep defaults for a quick first calculation.
Paste CSV rating rows into the data field. Enter the target user and target item. Adjust rating limits, neighbours, and filters.
Submit the form to calculate predictions. Review neighbour contributions and recommendation rankings. Export the results for later analysis.
Example Data Format
| User | Item | Rating | Category |
|---|---|---|---|
| Alice | Movie A | 5 | Drama |
| Bob | Movie C | 4 | Action |
| Cara | Movie B | 3 | Comedy |
Frequently Asked Questions
What is collaborative filtering?
It predicts preferences using behaviour from similar users or items. It does not require detailed item descriptions. Quality depends on interaction coverage.
When should user-based filtering be used?
Use it when users have meaningful overlapping ratings. It explains results through similar people. Very large user sets may require optimisation.
When is item-based filtering better?
Item relationships often change more slowly than users. This can improve serving speed. It works well for stable catalogues.
What does Pearson similarity measure?
Pearson correlation compares rating patterns after mean adjustment. It reduces differences in personal rating scales. Scores range from negative one to one.
Why use similarity shrinkage?
Shrinkage reduces confidence in similarities from few overlaps. It prevents unstable neighbours dominating predictions. Larger values produce stronger shrinkage.
What is a cold-start problem?
New users or items have little interaction history. Collaborative methods cannot find reliable neighbours. Fallback averages or popular items help.
What does RMSE show?
RMSE measures typical prediction error with larger penalties. Lower values generally indicate better rating predictions. Compare models on identical test data.
What does NDCG measure?
NDCG evaluates whether relevant items appear near the top. Higher values indicate better ranking quality. It is useful for recommendation lists.
Why use matrix factorisation?
Matrix factorisation learns compact user and item vectors. It can reveal hidden preference dimensions. Tuning factors and regularisation remains important.