Skip to content
Tutorials Center
My learning

Topics

Lesson 14 of 38 ยท Collaborative Filtering

User-based Collaborative Filtering

Build user-based collaborative filtering from scratch: find similar users, predict ratings with a mean-centered weighted average, evaluate with RMSE and make top-N lists.

Collaborative filtering (CF) recommends using the behavior of many users and needs no item content at all. The user-based version rests on a human idea: people who agreed in the past will probably agree in the future. To predict what Maya will think of a film, find users whose ratings resemble Maya's and look at what they said.

This lesson builds user-based CF from scratch with NumPy, using the similarity measures you already know. You will predict held-out ratings, measure the error, and turn predictions into a top-N list.

The algorithm

To predict user u's rating for item i:

  1. Compute the similarity between u and every other user (over co-rated items).
  2. Keep the k nearest neighbors who actually rated item i and have positive similarity.
  3. Predict a weighted average of their ratings, mean-centered so generous and harsh raters are comparable:

rฬ‚(u, i) = mean(u) + ฮฃ sim(u, v) ยท (r(v, i) - mean(v)) / ฮฃ |sim(u, v)|

flowchart LR
    A["Target user u"] --> B["Similarity to all users"]
    B --> C["Keep k nearest who rated item i"]
    C --> D["Weighted, mean-centered average"]
    D --> E["Predicted rating"]

Subtracting each neighbor's mean converts their rating into "how much above or below their usual".

A synthetic dataset with a hold-out

We create ratings from hidden tastes so a good method can actually find structure, hide 10% of the known ratings, and keep them for testing. The same recipe is reused in the next lessons.

Python Data and a global-mean baseline

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The prediction function

Python User-based k-NN predictions

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Compare against two baselines: always predicting the global mean, and predicting each user's own mean. If user-based CF cannot beat the user mean, the neighbors are adding only noise. In this run k-NN reaches an RMSE of about 0.83, against 0.88 for the global mean and 0.96 for the user mean. The parameter k controls the bias-variance trade-off: too small and a few odd neighbors dominate, too large and you average in people who are not really similar (see shrinkage and tuning).

Top-N recommendations

Rating prediction is rarely the end goal; you show a list. Predict a score for every item the user has not rated, sort, and take the top N.

Python A top-5 list for one user

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Strengths and weaknesses

  • Strengths: simple, no training, serendipitous (neighbors can suggest unexpected items), easy to explain ("people like you rated this high").
  • Weaknesses: the similarity matrix is users ร— users, which becomes huge; it needs overlapping ratings, so sparse data hurts; users change, so neighbors go stale; and it favors popular items.

User-based CF recipe

Find neighbors Similar rating patterns
Keep top k Only those who rated it
Weighted average Center by each user's mean

Predict, rank unseen items, show the top N.

1 / 2

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Why mean-center neighbors' ratings before averaging?

    A 4 from a harsh rater can mean more than a 4 from a generous one; deviations from each user's mean put them on one scale.
  2. 2. What happens if k is very large?

    More neighbors reduce variance but include weakly related users, so predictions flatten.
  3. 3. Which baselines should user-based CF beat?

    If you cannot beat simple means, the neighbor information is not adding value.
  4. 4. Why is user-based CF hard to scale to millions of users?

    Pairwise similarities between all users are expensive to compute and store.

Summary

  • User-based CF predicts a rating from the mean-centered ratings of the k most similar users who rated the item.
  • Evaluate against global-mean and user-mean baselines on held-out ratings.
  • Turn predictions into a ranked list of unseen items for top-N recommendation.
  • It is simple and explainable but costly at scale and weak on sparse data.

Next: Item-based Collaborative Filtering