Staypoint Detection in the MultiMNO Pipeline

Eurostat’s MultiMNO reference implementation detects staypoints—periods when a subscriber is stationary in a consistent location—by combining Present Population estimation with the Continuous Time Segmentation module. Together these stages transform raw Mobile Network Operator (MNO) events into longitudinal insights that underpin official statistics on residence, mobility, and tourism. Telcofy reuses the same logic as documented in the open-source repository on GitHub.

Input data and prerequisite layers

Staypoint detection builds upon the silver layer generated earlier in the pipeline:

• SilverEventFlaggedDataObject: cleaned usage events annotated with semantic quality flags and network metadata.
• SilverCellConnectionProbabilitiesDataObject: grid-to-cell probabilities derived from radio coverage and calibration routines.
• SilverCellIntersectionGroupsDataObject: for each day, the list of neighbouring cells that form a consistent spatial footprint.
• EventCacheDataObject: cached “first events of the following day” that allow segments to connect across date boundaries.

These datasets are materialised by upstream components such as Event Cleaning, Semantic Cleaning, and Cell Footprint Estimation (see Methodological Framework for the full overview).

Present Population: probabilistic occupancy counts

The Present Population component (present_population_estimation.py) calculates occupancy per grid tile and timestamp. Although its main output is population estimates, the method establishes the temporal window that later feeds staypoint logic:

1. Event windowing – for each query timestamp, the code keeps events within a configurable tolerance window (tolerance_period_s) and selects the event closest to the time point per device (Window spec sorted by time distance, see calculate_devices_per_cell).
2. Device-to-grid weighting – device counts per cell are merged with cell-to-grid probabilities and iterated through a Bayesian normalisation loop until differences fall below min_difference_threshold.
3. Persisted cache – intermediate population distributions are written to pp_cache to track convergence and supply stable inputs to downstream modules.

This step ensures that when a device remains within one cell (or cluster of overlapping cells) across multiple time points, the occupancy signal remains coherent, a critical prerequisite for consistent staypoint labelling.

Continuous Time Segmentation: stay, move, abroad, unknown

The Continuous Time Segmentation component (continuous_time_segmentation.py) converts daily batches of events into time segments tagged with behavioural states. Key stages include:

1. Daily slicing with cross-day continuity
   For each date in [data_period_start, data_period_end], events are filtered by error flags and domains, then unioned with the first event of the next day (via EventCacheDataObject). This ensures that a stay spanning midnight is not split prematurely.

2. Context enrichment
   Events are joined with intersection groups to recover all overlapping cells, enabling spatial buffering when multiple towers cover the same location. If segments from the previous day exist, they are joined back to preserve ending state and cell set.

3. Segmentation parameters
   Configuration knobs (defined in continuous_time_segmentation.ini) drive the classification logic:

   • min_time_stay_s: minimum dwell time for a STAY segment (default 900s).
   • max_time_missing_stay_s, max_time_missing_move_s, max_time_missing_abroad_s: largest gap tolerated between events before inserting an UNKNOWN state.
   • pad_time_s: slight temporal expansion applied to isolated segments between UNKNOWN blocks.
   • domains_to_include: inbound, domestic, outbound domains processed; outbound-only runs still depend on domestic events to close segments.
   • local_mcc: identifies the home country, used to classify ABROAD segments.

4. User-level UDF
   Events per user are grouped and fed to a Pandas UDF (segmentation_return_schema). The UDF tracks sequences of events, extends segments while devices remain within overlapping cells, and creates new segments when:

   • the device crosses domain boundaries (e.g., domestic → outbound),
   • the time gap exceeds the relevant max_time_missing_* threshold,
   • or overlapping cells no longer intersect.

5. State assignment
   Each segment receives one of four enumerated states (SegmentStates):

   • STAY – dwell longer than min_time_stay.
   • MOVE – transitions with sufficient evidence of displacement.
   • ABROAD – activity outside the local MCC.
   • UNKNOWN – gaps or ambiguous signals beyond configured tolerances.

6. Persistence with continuity
   Segments are written to the silver layer (time_segments_silver) with is_last markers indicating the trailing segment per user/day. Subsequent days reuse these markers to carry context forward.

Outputs and downstream usage

• Staypoint table (time_segments_silver): each record contains start/end timestamps, last observed event, list of contributing cells, state, and a hashed time_segment_id. This dataset powers Midterm/Longterm Permanence calculations, usual environment labelling, and tourism metrics.
• Population snapshots: the present_population_silver table is refreshed for each analysed timestamp and feeds daily dashboards as well as the iterative permanence scores.

Telcofy extends these outputs with add-on analytics (e.g., persistence scoring, anomaly flagging) while preserving the upstream logic to remain interoperable with Eurostat deliverables.

Telcofy alignment checklist

• ✅ Reuse the Continuous Time Segmentation configuration structure so auditors can compare parameters directly with the MultiMNO defaults.
• ✅ Preserve the four-state taxonomy and is_last markers, even when enriching segments with Telcofy-specific metadata.
• ✅ Document any changes to tolerance windows or dwell thresholds and justify them against ESS quality guidance.
• ✅ Propagate population caches and time segments into Telcofy’s compliance artifacts so EU stakeholders can trace staypoint derivations end-to-end.