Context is the set of external circumstances that can change a recommendation’s relevance without changing the user’s tastes or the item. Device, time of day, location, and season are contextual signals. A recommendation can therefore be appropriate for a person in one situation and poorly timed in another. section 1.3.3
Table 1.1 Contextual differences in one recommendation
|
User
|
Item
|
Device
|
Time
|
Relevance
|
|
Alice
|
Thorgal (epic, multi-volume)
|
Mobile,
|
lunch break12:30 PM
|
Low — too long for the format and moment
|
|
Alice
|
Thorgal (epic, multi-volume)
|
Laptop
|
weekend10:00 AM
|
High — right format, plenty of time
|
|
Alice
|
Thorgal (epic, multi-volume)
|
Laptop
|
weekend11:30 PM
|
Medium — right format, but fatigue matters
|
The table holds the user and comic constant while changing device and time. A multi-volume comic rates poorly during a mobile lunch break, highly on a weekend laptop, and only moderately late at night. section 1.3.3
Contextual data differs from content metadata, which describes items, and behavioral data, which records interactions. A context-based recommender relies on context alone; a context-aware recommender adds it to content-based recommendations or collaborative filtering. The latter is the more common approach. section 1.3 section 1.3.3
A practical pattern is reranking: generate candidates from item similarity, then use current context to change their order. For example, the same ten comic candidates can receive an afternoon ranking that favors adventure and an evening ranking that favors humor. The contextual layer changes priority, not the candidate set. section 1.3.3
Before adding contextual signals, account for their operational costs:
- Capture the signals needed at recommendation time; additional features increase data dimensionality. section 1.3.3
- Do not assume every user has a clear contextual pattern. section 1.3.3
- Treat location and device data as privacy-sensitive inputs. section 1.3.3





























