Does item-item filtering need personal profiles?

It can work from interaction or rating histories without detailed descriptive profiles, although those histories still require appropriate handling.

Can it recommend an item with no interactions?

Pure collaborative similarity has little evidence for that item, so another signal or fallback is usually needed.

Is the similarity matrix necessarily stored in full?

No. Systems often keep a limited neighbor list or use other techniques to avoid storing every item pair.

AI Glossary / Retrieval and knowledge systems / Intermediate

Item-item collaborative filtering

Definition

Item-item collaborative filtering is a recommendation approach that uses relationships between items inferred from users' interactions or ratings. To recommend an item to a user, it examines items that user has already engaged with and how similar they are to candidate items. The similarity comes from collaborative behavior rather than necessarily from item descriptions.

What it means

Two items can be related because many users interact with both or give them similar ratings. That relationship can help identify another item that a particular user may find useful, even when the items have different wording or surface features.

The approach differs from user-user collaborative filtering, which first looks for similar people. It also differs from content-based similarity, which compares item attributes directly. A practical recommendation system can combine these approaches, but the source of each similarity signal should remain clear.

How it works

Start with user-item observations and define how they should be interpreted. For explicit ratings, a similarity calculation may account for differences in users' rating scales. For implicit interactions, the system must consider exposure and the meaning of missing observations.

Compute or estimate item-to-item similarities and retain useful neighbors. For a target user and candidate item, combine evidence from the user's interactions with neighboring items. One simple explicit-rating formulation takes a similarity-weighted average of ratings, while other variants use centered ratings or different scoring rules.

Precomputing item neighbors can make serving efficient, but similarities need updating as behavior and the catalog change. New items without sufficient interactions require a fallback or additional information.

Three practical examples

1. A film service recommends a candidate because its rating pattern is related to films the viewer has already rated positively.

2. A learning platform uses co-enrollment patterns to identify courses that may interest someone who completed related material.

3. A product-discovery system generates candidates from items frequently associated in interaction histories, then applies availability and suitability checks before presenting recommendations.

Think of it this way

Think of learning which library books tend to appeal to the same readers. When someone enjoys one book, the library can suggest another with a related readership pattern. The relationship is useful evidence, but it does not prove that the books cover the same subject or that every reader will enjoy both.

Worked example

Suppose an invented user has rated item A as five out of five and item B as two out of five. Candidate C has positive similarity 0.8 with A and 0.2 with B. Use a simple weighted average with no rating centering.

The predicted score for C is (0.8 times 5 + 0.2 times 2)/(0.8 + 0.2) = 4.4/1.0 = 4.4. Candidate D instead has similarities 0.3 with A and 0.7 with B, giving (0.3 times 5 + 0.7 times 2)/(0.3 + 0.7) = 2.9.

Under this explicitly stated rule, C ranks above D. The calculation assumes those similarities were already estimated from an appropriate training dataset and that the user's two ratings are valid inputs.

It does not explain how reliable the similarities are, account for user rating biases or prove that the user will assign 4.4 to C. A production method might center ratings, limit neighbors, discount weakly supported relationships or use a different model for implicit interactions. These are illustrative recommendation scores, not guaranteed future ratings.

Illustrative case study

Fictional case: Spruce Courses builds an item-neighbor recommender using historical enrollments. Reviewers notice that a mandatory introductory course is linked with almost everything because nearly every learner takes it.

The team examines popularity effects and the support behind each relationship, then evaluates alternative similarity and filtering choices. It also checks prerequisite information separately so that behavioral similarity does not recommend an unsuitable course. The final evaluation includes new learners and less popular courses, where interaction-based evidence may be particularly limited.

Common misconceptions and mistakes

Item similarity does not necessarily mean content similarity. It can reflect shared audiences, complementary uses, bundling or presentation patterns. Co-occurrence alone therefore does not establish substitutability or causation.

Missing interactions are not direct dislikes, and a relationship based on a handful of users can be unstable. Do not apply a formula designed for explicit ratings unchanged to raw clicks without examining its assumptions. Item-item filtering also does not eliminate cold-start problems: a new item may lack collaborative evidence, and a new user may have no history from which to personalize.

Business use

Use the method as a recommendation baseline or candidate-generation component when there is enough reliable interaction history. Define the behavior being modeled, how similarities are calculated and how many neighbors are retained. Evaluate both prediction or ranking quality and coverage across the catalog.

Measure recommendation usefulness alongside popularity concentration, diversity and response time. Refresh relationships as the catalog changes and provide fallbacks for sparse histories. Explanations such as related to an item you used should accurately reflect the signal; they should not claim that the user explicitly endorsed something inferred only from a click.

Check your understanding

Two products are frequently bought together. Does that prove they are interchangeable alternatives?

Answer: No. They may be complements, such as an item and an accessory. The recommendation relationship must be interpreted according to the interaction data and the intended task.

Related terms

Collaborative filtering
Nearest neighbor
Cosine similarity
user-item matrix
nearest neighbours
adjusted cosine
implicit feedback
cold start and recommendation ranking
comparisons and limits behind item-based recommendations

Learn next

calculate a small item-neighbor score from explicit ratings, then examine how rating centering changes the result. Repeat the reasoning for implicit interactions and identify which assumptions no longer apply before choosing an evaluation method.

Sources

Primary sources: Sarwar, Karypis, Konstan and Riedl, Item-Based Collaborative Filtering Recommendation Algorithms: https://compjournalism.com/files/-dmsl/sarwar_2001.pdf

Hu, Koren and Volinsky, Collaborative Filtering for Implicit Feedback Datasets, including item-oriented prediction and implicit-data limitations: https://yifanhu.net/PUB/cf.pdf

Previous: IT risk management Next: Item matrix