Static context like location changes slowly; dynamic context like activity changes fast

What it is

A second axis for typing context: rate of change. Slow context — where home is, the floor plan, household composition, the broad rhythm of the day — holds for months, so one measurement lasts. Fast context — current activity, who is present, the conversation underway — turns over in minutes, so yesterday's judgement is worthless almost immediately. "Location" itself straddles the axis: where home is sits at the slow end; where the person is right now sits at the fast end.

The use of this typing is setting sampling and update policy: slow dimensions get infrequent sampling with long-lived cached results; fast dimensions get high-frequency, low-latency treatment. Rate-policy mismatch is the common hidden source of both wasted sensing and stale decisions in ambient systems.

Why it happens

Rate dictates policy because two cost curves run in opposite directions:

• High-frequency sampling of a slow dimension is pure waste — confirming the floor plan every hour spends energy and bandwidth on "nothing changed". The right shape is low-frequency polling plus event-triggered remeasurement (re-survey on moving house).
• Low-frequency sampling or long caching of a fast dimension spends stale data as if it were current — "watching TV" five minutes ago is worthless for deciding whether to interrupt now. Fast dimensions must follow the change itself: event-driven, short windows, consume at the moment of judgement.

The classic trade-off in activity recognition is this logic in engineering form: lengthen the recognition window and accuracy rises while the conclusion ages; shorten it and conclusions stay fresh but error rates climb. Where to place the window is set by how much latency that dimension's rate of change tolerates.

Studying it

• The activity-recognition literature measures the rate-latency trade-off directly: train and recognise on the same sensor data at several window lengths, plot the accuracy-latency curve, and read the viable window off its knee.
• Context-aware system evaluations commonly build a stability profile per dimension: measure the distribution of change intervals (how long between changes); the mode of that distribution decides polling versus event-driven handling.
• A general validation move: compare the system's sampling policy against measured change intervals and flag the two mismatch classes — "sampling far faster than change" (waste) and "cache far longer than change interval" (staleness).

One methodological caution: measure change intervals in real homes, not scripted labs — scripted activity transitions are choreographed and systematically fast, inflating every dimension's apparent dynamism.

Where it stops holding

• Fast/slow is not intrinsic to a dimension but relative to the application. "Person at home" is a second-by-second fast variable for a security alert and an hour-scale slow variable for content recommendation. Rate grading must be done against "which action does this judgement feed"; divorced from use, fast and slow mean nothing.
• Slow dimensions jump too. Moving house, separation, renovation flip the slowest dimensions overnight. Purely time-driven policies for slow dimensions (refresh annually) miss jumps and need event triggers layered on top.
• Power constraints invert the policy: battery-powered sensors are forced into infrequent sampling even when they monitor a fast dimension — that mismatch is a hardware-budget compromise rather than a design error. Once identified, either change the power source or accept the staleness.

Applying it

• Tag every context dimension with one of three update policies: event-driven (fast), low-frequency polling (slow), event-triggered resurvey (jump insurance for slow). A dimension without a tag does not enter an automation condition.
• Fast dimensions go subscribe-and-push: push on change, silence otherwise; slow dimensions go lookup: local cache, re-verify only on expiry.
• How to check: export each dimension's "last confirmed at" distribution, overlay it on the measured change intervals, and read the two together — sampling hugging the change rhythm passes; the gaps on either side are the waste and the staleness.

Related

• Same group: Z2.06.1 Context splits into physical and social dimensions · Z2.06.3 Personal and shared context have different capture boundaries · Z2.06.4 Dimensions can contradict: location right, activity wrong
• Nearby: Z2.09 Timeliness and expiry of context
• Search terms: context dynamics · activity recognition window · sampling strategy · context-aware computing