Personalized Match Recommendations for Fans: From Collaborative Filtering to Vector Retrieval

Once the World Cup kicks off, content explodes: dozens of matches, a flood of news, countless highlights and analyses. A fan who cares about only a few teams drowns in information irrelevant to them. Personalized recommendation solves this — putting the content each fan is most likely to want in front of them.

This guide builds a match content recommendation system, from classic collaborative filtering to modern vector retrieval, and clarifies the challenges specific to sports. For recommendation-system fundamentals, see building a recommendation engine from scratch: collaborative filtering to neural networks; this article focuses on the specific trade-offs of the World Cup scenario.

Two mainstream approaches

Recommendation systems have two main technical routes, each with strengths:

Modern systems usually combine the two, then unify them with vector retrieval. Let's look at each.

Collaborative filtering: find patterns in behavior

The core of collaborative filtering is a "user × content" interaction matrix (who watched what, liked what), mining patterns from it.

python
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
User-content interaction matrix (rows=users, cols=content, value=interaction strength)
In reality this is a huge sparse matrix
interactions = np.array([
    [5, 3, 0, 1],  # user 1
    [4, 0, 0, 1],  # user 2
    [1, 1, 5, 4],  # user 3
])
Item similarity: which content is often consumed by the same group of people
item_sim = cosine_similarity(interactions.T)
To recommend: similar content to what the user interacted with, weighted and ranked

Collaborative filtering's power is discovering unexpected associations — like "people who watch Team A's matches also often watch Team B's tactical analysis" — without you predefining them. But it has a fatal weakness: cold start.

Sports-specific challenges

Challenge 1: cold start is especially severe

The World Cup is a short-cycle, highly time-sensitive event. New users, new matches, new content pour in constantly, and pure collaborative filtering can't recommend accurately without enough behavioral data. Mitigations:

Challenge 2: extreme time sensitivity

Sports content expires fast. Yesterday's match prediction is useless today; once a team is knocked out, content about it plummets in value. Recommendations must have strong time decay:

python
import math
def time_decay_score(base_score, hours_ago, half_life=12):
    # Sports content has a very short half-life; heat halves in 12 hours
    decay = math.pow(0.5, hours_ago / half_life)
    return base_score * decay

Set the time-decay half-life far shorter than for general content — this is what distinguishes sports recommendation from e-commerce and video recommendation.

The modern approach: vector retrieval

The now-mainstream method is mapping both users and content into the same vector space with embeddings, turning recommendation into "find content vectors nearest the user vector" — essentially vector retrieval.

python
Content vector: embed the content's text (title, tags, team)
User vector: a weighted average of recently-interacted content vectors (recent weighted higher)
from openai import OpenAI
client = OpenAI()
def embed(text):
    return client.embeddings.create(
        model="text-embedding-3-small", input=text
    ).data[0].embedding
User vector = time-weighted average of recently-interacted content vectors
Recall = find content most similar to the user vector in the vector DB

This is the same class of technology batch one's RAG used — vector retrieval. Loading content into a vector DB and recalling by similarity, the methods in the complete semantic-search implementation guide apply directly. For vector-DB selection, see the vector database selection guide.

The upside of vector retrieval: it naturally mitigates cold start (new content can be recalled just by being embedded, no behavioral data needed) and unifies handling of text, users, and content. That's why it's gradually replacing pure collaborative filtering.

Full architecture: recall + ranking

Production recommendation systems are usually two-stage:

In the World Cup scenario, the ranking stage should heavily weight timeliness (matches just ended) and the user's followed teams.

Summary

The key to a match recommendation system is sport-specialization on top of a general recommendation framework: an ultra-short time-sensitivity half-life, cold-start solved by picking teams, and vector retrieval unifying multi-channel signals.

It closes the loop with the rest of this series — recommended content comes from automated generation, recall relies on vector retrieval, and the big picture is in the AI and 2026 World Cup roundup.

From a practice standpoint, start with vector retrieval — it's simple to implement, cold-start friendly, and the highest-ROI entry point.