Results appear here
Load sample data or enter your own dataset, then calculate similarities.
Analysis results
ReadyCosine similarity compares vector direction. It works well with sparse ratings. Larger positive values indicate closer item behaviour.
Adjusted cosine removes each user’s average rating. Pearson removes each item’s average. Both reduce rating-scale bias.
Jaccard, Dice, and overlap compare shared presence. They suit clicks, purchases, tags, and other binary interactions.
Distance similarities convert distance into similarity using 1 ÷ (1 + distance). Smaller original distances therefore produce larger scores.
- Choose an input mode and delimiter.
- Paste data, upload a file, or load a sample.
- Select a similarity method and preprocessing settings.
- Choose target items and optional recommendation user.
- Calculate, inspect diagnostics, then export your results.
| Mode | First column | Remaining columns | Example |
|---|---|---|---|
| Ratings | User | Items | U1,5,4,,2 |
| Binary | User | Items | U1,1,0,1,0 |
| Features | Item | Numeric features | Movie A,1,0.7,45 |
| Text | Item | Description | Movie A,space exploration adventure |
What is item similarity?
It measures how closely two items behave or appear. Data may come from users, attributes, or text. Higher scores usually mean stronger similarity.
Which method should I choose?
Use cosine for sparse numeric vectors. Use adjusted cosine for rating data. Use Jaccard or Dice for binary interactions.
Why is adjusted cosine useful?
Users apply rating scales differently. Adjusted cosine subtracts each user’s mean first. This reduces generous or strict user bias.
What does shrinkage do?
Shrinkage reduces scores based on tiny overlaps. Strong evidence receives less reduction. This improves ranking stability in sparse data.
Can the calculator predict recommendations?
Yes, for rating and interaction matrices. It combines similar neighbours already known by a user. Predictions depend on available overlap.
How are text items compared?
The calculator tokenizes descriptions and builds TF-IDF vectors. Similarity then compares weighted term patterns. Common uninformative words receive less influence.
What causes an undefined score?
Two items may lack enough shared values. Constant vectors can also block correlation. Diagnostics explain the detected problem.
Does zero mean missing?
That depends on your input mode. Blank cells are treated as missing by default. Binary zero represents no recorded interaction.
Are saved projects uploaded anywhere?
No server upload is required for browser storage. Projects remain in local browser storage. Clearing site data removes them.