Skip to main content

Cold Start for Users and Items

A warm-ID model cannot look up a representation it has never learned. Cold start is therefore a serving contract, not an exception to hide: detect which entity is unknown, choose only available evidence, and evaluate that cohort separately.

Prerequisites​

Understand the split cohorts in temporal and user-aware data splits, content features, and implicit-feedback learning.

Learning goals​

  • Distinguish new-user, new-item, and system cold start.
  • Define deterministic fallbacks without manufacturing latent vectors.
  • Route cold items through content and controlled exploration.
  • Report warm and cold cohorts separately under the frozen evaluation protocol.

Model contract​

PartThis lesson's contract
ObjectiveProduce an eligible initial list when a user ID, item ID, or both have no training state.
Input representationWarm-ID membership, training popularity, declared item content, onboarding answers, and request context.
Learned statePopularity counts and any feature encoders fitted before the cutoff; the routing policy itself is versioned configuration.
Inference pathDetect cold cohort, apply eligibility, select content/onboarding or popularity fallback, then reserve only a declared exploration share.
Evaluation boundaryUse u5 and i7 from the shared fixture as explicit cold cohorts; do not average them into warm-user results.
Principal failure modesEmpty profiles, metadata gaps, popularity lock-in, unsafe exploration, unfair item exposure, and accidental unknown-ID defaults.

Worked example: route both cold cohorts​

The fallback uses positive training actions only. It does not pretend that u5 has a collaborative vector. The new item i7 has declared metadata, so it can enter a content pool but not a warm-ID factor lookup.

import csv
from collections import Counter
from pathlib import Path

path = Path("tests/fixtures/recommender-systems/interactions.csv")
with path.open(encoding="utf-8", newline="") as handle:
rows = list(csv.DictReader(handle))

positive_types = {"click", "save", "complete", "rating"}
train = [row for row in rows if row["split"] == "train"]
test = [row for row in rows if row["split"] == "test"]
train_users = {row["user_id"] for row in train}
train_items = {row["item_id"] for row in train}
cold_users = sorted({row["user_id"] for row in test} - train_users)
cold_items = sorted({row["item_id"] for row in test} - train_items)
counts = Counter(row["item_id"] for row in train if row["event_type"] in positive_types)
fallback = [item for item, _ in sorted(counts.items(), key=lambda pair: (-pair[1], pair[0]))[:3]]
item_metadata = {"i7": {"topic": "machine-learning", "available": True}}
i7_route = "content_pool" if item_metadata["i7"]["available"] else "ineligible"
print(f"cold_users={cold_users} cold_items={cold_items}")
print(f"u5_fallback={fallback} i7_route={i7_route}")

Output:

cold_users=['u5'] cold_items=['i7']
u5_fallback=['i1', 'i4', 'i2'] i7_route=content_pool

This is a traceable fallback, not evidence that popularity is best. An onboarding selection can initialize a content profile; request context can narrow candidates; a bounded exploration policy can collect evidence for a new item. Each source needs consent, retention, safety, and accessibility review.

Limitations and common mistakes​

  • A global default can disadvantage niche users and new providers.
  • Synthetic zero or mean latent vectors conceal unknown state and may produce arbitrary scores.
  • Item metadata can be missing, stale, gamed, or sensitive.
  • Random exploration must still honor eligibility, safety, duplication, and user-control constraints.
  • Warm, cold-user, and cold-item ranking quality need separate denominators; no universal model ranking follows from their average.

Practice​

Design an onboarding flow with at most three questions. For each answer, name the feature it initializes, the user control it needs, and the fallback used when the user skips it.

Next steps​

When identity is absent or long-term taste is less useful than immediate intent, session-based recommendation uses the current bounded sequence.

Primary sources​

Checked 2026-10-17: Google, content-based filtering and candidate-generation overview; Koren, Bell, and Volinsky, matrix-factorization review.