Session-Based Recommendation
Session models answer a short-horizon question: given the ordered actions in the current visit, what is a plausible next item? This differs from estimating a user's durable taste. The sessionization rule is part of the model input and must be versioned.
Prerequisites
Understand timestamps, temporal splits, and the cold-start routing contract.
Learning goals
- Define session boundaries using observable product semantics.
- Preserve event order without crossing the global evaluation cutoff.
- Distinguish transition tables, nearest-session methods, and learned sequence models.
- Provide a fallback for empty or one-item sessions.
Model contract
| Part | This lesson's contract |
|---|---|
| Objective | Predict the next eligible lesson from the current visit's ordered context. |
| Input representation | Per-user timestamped events split after 30 minutes of inactivity, deduplicated to item order within each session. |
| Learned state | Training-only item transitions or sequence-model parameters, plus the timeout, event filter, and vocabulary. |
| Inference path | Assign the request to a session, encode its ordered prefix, retrieve next-item candidates, then fall back if support is absent. |
| Evaluation boundary | Cut data globally first, build session prefixes inside each split, and score next-item predictions without future events. |
| Principal failure modes | Arbitrary timeout, cross-device identity, repeated refreshes, merged journeys, sparse transitions, and popularity-only fallback. |
Worked example: make the boundary observable
The example sessionizes u1 independently within train and test. Multiple event types for the same item collapse inside a session, because the target here is next item, not next event type.
import csv
from datetime import datetime, timedelta
from pathlib import Path
path = Path("tests/fixtures/recommender-systems/interactions.csv")
with path.open(encoding="utf-8", newline="") as handle:
rows = [row for row in csv.DictReader(handle) if row["user_id"] == "u1"]
def timestamp(value):
return datetime.fromisoformat(value.replace("Z", "+00:00"))
def sessions(split):
events = sorted((row for row in rows if row["split"] == split), key=lambda row: row["timestamp"])
result = []
previous = None
for row in events:
current = timestamp(row["timestamp"])
if previous is None or current - previous > timedelta(minutes=30):
result.append([])
if not result[-1] or result[-1][-1] != row["item_id"]:
result[-1].append(row["item_id"])
previous = current
return result
training_sessions = sessions("train")
current_session = sessions("test")[0]
print(f"u1_sessions={training_sessions} boundaries={len(training_sessions)}")
print(f"current_session={current_session} fallback=content_then_popularity")
Output:
u1_sessions=[['i1'], ['i1'], ['i2']] boundaries=3
current_session=['i5'] fallback=content_then_popularity
There is no supported next-item transition in these u1 training sessions, so the honest path is a content-then-popularity fallback. A recurrent or attention model cannot manufacture sequence evidence absent from the data. For next-item evaluation, create prefixes only from events that would have been known at each prediction time.
Limitations and common mistakes
- A 30-minute session boundary is a product assumption, not a natural law.
- Sorting after mixing train and test leaks future sequence context.
- Anonymous sessions can be shared by several people; logged-in sessions can span devices.
- Repeated impressions may reflect rendering rather than intent.
- Large neural session models should beat strong recency, popularity, and nearest-session baselines under the same candidates before their complexity is justified.
Practice
Change the timeout to two days and list which u1 sessions merge. Explain how that changes the target rather than merely adding data.
Next steps
Individual models are usually only sources. Candidate generation combines their nominations and measures whether relevant items survive retrieval.
Primary sources
Checked 2026-10-17: Hidasi et al., “Session-based Recommendations with Recurrent Neural Networks”, and Ludewig and Jannach, “Evaluation of Session-based Recommendation Algorithms”.