Context splits into physical and social dimensions

What it is

Context information divides by source into two classes. Physical context: location, time, light level, noise, temperature, device orientation — states of the environment that sensors can measure directly. Social context: who is present, what activity they are engaged in, how they relate to each other, whether they are interruptible — states about people that almost no sensor measures directly; they must be inferred from physical signals.

The split goes back to the classic definitions of context-aware computing: Dey's "any information that can be used to characterize the situation of an entity", where entities may be people, places, or objects. From the start the two classes travelled different routes: environmental quantities are measured, social quantities are guessed.

Why it happens

The classification matters because the two classes differ fundamentally in three respects:

• Source of error. Physical context errs through the sensor itself — drift, occlusion, miscalibration — with physically bounded error. Social context errs through the inference chain — guessing "in a meeting" from sound and location — where every link adds uncertainty, producing larger and less visible errors.
• Privacy sensitivity. Physical context records the environment (how many lux in the room); social context records people (who is here, doing what, with whom). The latter is behavioural data the moment it is captured, one sensitivity tier up.
• How it can be validated. Physical context can be cross-checked against another sensor; the ground truth of social context exists only in the inferred person's head, so validation must return to the person.

The two classes therefore should not share one trust policy — treating "300 lux" and "user is in a meeting" as equally reliable inputs is a direct recipe for misjudging the social side.

Studying it

• The founding literature supplies the taxonomy: Dey's 2001 "Understanding and Using Context" and the earlier "Towards a Better Understanding of Context and Context-Awareness" by Abowd, Dey and colleagues organize context by entity type; activity-recognition work has followed that skeleton since.
• Sensing the social dimension has a dedicated strand: social signal processing, inferring interaction state from speech prosody, posture, and proxemics. Its evaluation protocols — annotation agreement, cross-subject generalisation — transfer well.
• The standard route to ground truth is experience sampling: probe users at the moment the context holds and use their self-report as the comparison baseline. Scripted lab scenarios annotate cleanly but lack ecological validity; in-situ sampling is noisy; the two are usually combined.

One methodological caution: for social context the ground truth is defined by the person being inferred — "resting or ruminating about work" cannot be adjudicated by an outsider. Self-reports are not noise here; they are the only legitimate truth in this kind of research.

Where it stops holding

• The split is not a hard boundary. "At home" is a physical fact that carries social meaning (home implies uninterruptible); calendars are social agreements made physical. Many dimensions straddle the line — the value of the classification is the question "is this dimension measured or inferred?", not the bucket itself.
• Social-context judgements are culturally modulated. The same behaviour means different things in different cultures (loud talk: argument or warmth), and social dimensions are the first to distort when inference models cross cultures.
• Physical does not mean low-risk. Location is thoroughly physical and thoroughly sensitive — the split describes how data is obtained, not its risk level.

Applying it

• Keep the sensor and data inventory columned by dimension: each column marked measured or inferred, with its confidence — "location (GPS, high)" and "in a meeting (inferred, medium)" should never appear as the same grade of input in the system.
• Pair each class with its own policy: consume physical context directly; gate social context behind confirmation before high-consequence actions, or demote it to low-consequence ones.
• How to check: audit the social dimension periodically — sample random moments, compare the system's activity judgement against the user's self-report, tallied per dimension. Calibrate sensors for the physical side, inference for the social side; keep two separate ledgers.

Related

• Same group: Z2.06.2 Static context like location changes slowly; dynamic context like activity changes fast · Z2.06.3 Personal and shared context have different capture boundaries · Z2.06.4 Dimensions can contradict: location right, activity wrong
• Nearby: Z1.03 Defining context · Z2.02 Uncertainty of inference
• Search terms: context awareness · social context · social signal processing · activity recognition