Context inference inherently errs

What it is

Context is not read off the world; it is inferred — a probabilistic judgement from limited, noisy sensor evidence to "what the user is doing". The claim here: the residual error of that inference cannot be eliminated. It is a structural property of context systems, not an engineering deficiency.

The reason lies in the form of the problem: the true context state (the user is asleep) is never directly observed; the system holds only indirect evidence (stillness, low heart rate, lights off). Any inference from incomplete evidence retains uncertainty — independent of sensor quality. A better sensor compresses the error but cannot zero it, because the information is simply not in the data (a motionless person may be asleep or reading; the evidence does not distinguish the two).

Why it happens

Error sources are layered, and each layer contributes an irreducible part:

• Ambiguity at the evidence layer. Distinct context states produce identical evidence (sleep and quiet reading look the same to PIR plus heart rate) — a class separability problem. Where the evidence carries no discriminating information, no classifier can help. This sets the floor of irreducible error.
• Bias at the prior layer. Inference models carry priors (most people sleep after eleven); individuals who deviate (night-shift nurses, insomniacs) suffer systematic misjudgement — error is not random but biased toward majority patterns, with minority routines absorbing most of it.
• Drift at the distribution layer. Models learn from historical data while life changes: a new job, a newborn, the seasons shift behavioural distributions, and the gap between training and deployment distributions widens over time — context distributions are non-stationary.

Stacked, the consequence: an activity recogniser reporting "95% accurate" may deliver far less for atypical users, at atypical hours, months after deployment. Error can be compressed, monitored, and absorbed by design — but not eliminated, so downstream behaviour must treat "the judgement may be wrong" as a standing input.

Studying it

• Confusion matrices and stratified reporting: report false positives and negatives stratified by user group, time of day and scenario rather than one aggregate accuracy — the error distribution predicts field performance better than the mean.
• Non-stationarity studies: longitudinal deployments tracking a fixed model's accuracy decay over months to quantify drift; contrasts between frozen models and adaptive/online-learning ones that keep learning after deployment.
• Uncertainty quantification: have models output confidence rather than bare labels (probabilities, ensemble variance) and study the calibration between stated confidence and observed error rate — is a judgement offered at 80% confidence right 80% of the time? Well-calibrated confidence can be consumed directly by downstream rules.

One methodological caution: clean lab data (protocol-executed actions, single occupant) improves all three error layers at once and yields optimistic numbers; in-the-wild datasets with real noise and real routines are the basis for design decisions.

Where it stops holding

• Error is not evenly shared. Misjudgement concentrates on atypical users and scenarios (shift workers, multi-generation households, overlapping occupants in small flats); aggregate accuracy conceals this structural unfairness — evaluate the worst quantile of error, not its mean.
• Fusion has a ceiling. Gains are limited when signals are highly correlated (PIR and a camera both triggered by the same cat); genuinely independent evidence sources reduce uncertainty, and independence is precisely what is privacy-costly (more kinds of capture).
• The kind of error determines the kind of harm. Misses (asleep, judged absent) and false alarms (absent, judged present) have entirely different downstream costs. Eliminating error is impossible, but choosing which error to absorb is the controllable design variable.

Applying it

• Pass confidence together with the inferred label to the behaviour layer, and let rules tier on it: high confidence — act; medium — suggest; low — log only. Thresholds come from calibration data, not intuition.
• For each class of context judgement, state explicitly which error is acceptable: security errs toward false alarms, convenience errs toward misses. Error preference is a product decision and belongs in writing.
• Give every deployed model a review cadence: sample-check inference against self-report monthly; when agreement drops below threshold, trigger retraining or fall back to manual rules. Never train once and run forever.
• How to check: keep an error ledger of "judgement → reality → consequence", grouped by user and scenario; list the three most error-concentrated scenarios and answer for each — swap the proxy, add redundancy, or make it conservative? A scenario with no answer is one that should not be automated.

Related

• Same group: Z1.03.1 Context comprises location, time, activity and co-presence · Z1.03.2 Sensors yield only proxy measures of context
• Nearby: Z2.02 Inference uncertainty · Z2.03 The asymmetric costs of false alarms and misses
• Search terms: activity recognition · calibration · concept drift · uncertainty