Curation: re-screen of every AI (UNVERIFIED) record

curation/records/2024_lusaka_mobility_prediction_hackathon.md

@@ -3,10 +3,8 @@ unique_name: 2024_lusaka_mobility_prediction_hackathon
 name: 2024-lusaka-mobility-prediction-hackathon
 checked_by:
 - AI (UNVERIFIED)
-suggestion: 'No'
-decision_markers:
-- Time-series (Forecasting)
-- Missing source information
+suggestion: TBD -> Yes
+priority: 3
 tags:
 - AI-Filled (Verify)
 collections:
@@ -16,10 +14,12 @@ year: '2024'
 domain: technology & internet
 required_split:
 - Temporal (NON-IID)
-problem_type: TBD
+problem_type: Regression
 source_links:
-- https://zindi.africa/competitions/practice-challenge-lusaka/data
 - https://zindi.africa/competitions/2024-lusaka-mobility-prediction-hackathon/data
+- https://zindi.africa/competitions/2024-lusaka-mobility-prediction-hackathon
+- '# a different Yango dataset (average road speeds, Sept 2023), not this task''s data'
+- https://zindi.africa/competitions/practice-challenge-lusaka/data
 source_row: 934
 needs_review:
 - ai_unverified
@@ -30,4 +30,13 @@ type_adapter_id: curation-record-v1
 
 **⚠️ AI-FILLED — UNVERIFIED. The suggestion, metadata, and notes below were drafted by an AI assistant from public knowledge of this competition/dataset, NOT from inspecting the data. A human must verify everything before relying on it.**
 
+**Screening (AI, 2026-10-22):** `TBD -> Yes`, priority 3: a per-trip ride-time regression (Yango's ETA), not demand forecasting; the competition page refutes the old `No` with `Time-series (Forecasting)` and `Missing source information`.
+- Source: Zindi one-day hackathon "2024 Lusaka Mobility Prediction Hackathon Powered" by Yango, 26 Oct 2024 (https://zindi.africa/competitions/2024-lusaka-mobility-prediction-hackathon, read through Zindi's page API); as published: "a sampled dataset of trips ordered via the Yango App from January 2024 in Lusaka": Train.csv (15.6 MB, with the target), Test.csv (6.4 MB, no target), Lusaka_weather.csv (49 KB, joined by time), VariableDefinitions.csv. Coordinates geomasked "by at least X meters and no more than Y meters in a random direction"; "The straight-line distance is calculated based on the original points"; "excessively short trips" removed. Rows and columns not counted (the files need a login).
+- Use case: "Yango uses a wide range of data in order to predict travel and arrival times of its partners' drivers and vehicles" (Description): the platform quotes a ride time when a trip is ordered; fit direct if every column is known at ordering; regime temporal (one month of trips; quoting future trips), as for `nyc_taxi_trip_duration`.
+- Siblings and duplicates: the same Yango task in other cities: `2025_accra_mobility_prediction_hackathon_powered` (Jan 2025, Train.csv 7.9 MB) and `2025_maputo_mobility_prediction_hackathon_powered_by_yango` (2024 trips, 9.0 MB), plus a 2026 Yaounde edition without a record: different data, the same task and publisher, so take one city (this one has the largest train file). The road-speed hackathons (`accra_mobility_prediction_hackathon`, practice-challenge-lusaka, the cities in `mobility_prediction`) are another task. Ride-time precedents: `nyc_taxi_trip_duration` and `pkdd_15_taxi_trip_time_prediction_ii` (`Yes`), `maps_router_eta` (`Yes`, TabReD), `yassir_eta_prediction_challenge` (`TBD -> Yes`). External collections: none found.
+- Papers and pages: no paper; the competition's Description, Data and Evaluation tabs; no public solution or column list found (web search 2026-10-22).
+- Metric: RMSE on the ride time ("The evaluation metric for this competition is Root Mean Squared Error", Evaluation tab); a competition's leaderboard metric (MET-4); inputs none.
+- Risks for /add-dataset: the column list is unread: check that no column is known only after the trip (driven distance, drop-off time) (LEAK); the target's unit; only Train.csv is labelled; the second link (practice-challenge-lusaka) is a different dataset (road speeds, Sept 2023); access: login, and the competition "is only open to Zambian users residing in Zambia" (no criterion, EV-4).
+- Priority: gap 0 (technology & internet, regression and temporal are not thin), use case + (Yango's ETA, named by the source), source 0 (a competition page, no codebook read), shape + (a month of trips with time and weather), risk 0.
+
 A Zindi 'mobility prediction' hackathon for Lusaka with no comments in the record. Mobility-prediction challenges (often ride-hailing/transport demand) typically forecast trips/demand over space and time, which is excluded forecasting rather than a fixed IID/temporal-split tabular task. The record carries essentially no descriptive content, so the verdict is unverified. A human must confirm the exact target and whether it is a one-shot prediction or a multi-step forecast; if it is genuine forecasting it stays No, but if it is a fixed per-trip prediction it could be revisited.

curation/records/2025_accra_mobility_prediction_hackathon_powered.md

@@ -3,10 +3,8 @@ unique_name: 2025_accra_mobility_prediction_hackathon_powered
 name: 2025-accra-mobility-prediction-hackathon-powered
 checked_by:
 - AI (UNVERIFIED)
-suggestion: 'No'
-decision_markers:
-- Time-series (Forecasting)
-- Missing source information
+suggestion: TBD -> 2nd Tier
+priority: 2
 tags:
 - AI-Filled (Verify)
 collections:
@@ -16,9 +14,10 @@ year: '2025'
 domain: technology & internet
 required_split:
 - Temporal (NON-IID)
-problem_type: TBD
+problem_type: Regression
 source_links:
 - https://zindi.africa/competitions/2025-accra-mobility-prediction-hackathon-powered/data
+- https://zindi.africa/competitions/2025-accra-mobility-prediction-hackathon-powered
 source_row: 946
 needs_review:
 - ai_unverified
@@ -29,4 +28,13 @@ type_adapter_id: curation-record-v1
 
 **⚠️ AI-FILLED — UNVERIFIED. The suggestion, metadata, and notes below were drafted by an AI assistant from public knowledge of this competition/dataset, NOT from inspecting the data. A human must verify everything before relying on it.**
 
+**Screening (AI, 2026-10-22):** `TBD -> 2nd Tier`, priority 2: a per-trip ride-time regression (Yango's ETA), not forecasting; the same task as `2024_lusaka_mobility_prediction_hackathon` in another city, with a smaller train file.
+- Source: Zindi one-day hackathon "2025 Accra Mobility Prediction Hackathon Powered" by Yango, 5 Jul 2025 (https://zindi.africa/competitions/2025-accra-mobility-prediction-hackathon-powered, read through Zindi's page API); as published: "a sampled dataset of trips ordered via the Yango App from January 2025 in Ghana": Train.csv (7.9 MB, with the target), Test.csv (3.3 MB), Accra_weather.csv, VariableDefinitions.csv; coordinates geomasked, straight-line distance from the original points, "excessively short trips" removed. Rows and columns not counted (login).
+- Use case: "Yango uses a wide range of data in order to predict travel and arrival times" (Description): a ride time quoted at ordering; fit direct if every column is known then; regime temporal (one month of trips).
+- Siblings and duplicates: `2024_lusaka_mobility_prediction_hackathon` (Jan 2024, 15.6 MB) and `2025_maputo_mobility_prediction_hackathon_powered_by_yango` are the same task and publisher on other cities' data: not copies, but one city is enough (this is the smallest of the three). Not a duplicate of `accra_mobility_prediction_hackathon` (road speeds per segment, Sept 2024), as the old note guessed. Precedents: `nyc_taxi_trip_duration`, `pkdd_15_taxi_trip_time_prediction_ii`, `maps_router_eta` (`Yes`).
+- Papers and pages: no paper; the competition's tabs; no public solution found.
+- Metric: RMSE ("The evaluation metric for this competition is Root Mean Squared Error", Evaluation tab); MET-4; inputs none.
+- Risks for /add-dataset: the column list is unread (post-trip columns, LEAK); the target's unit; access: login, "only open to Ghanaian users residing in Ghana" (no criterion, EV-4).
+- Priority: gap 0, use case + (Yango's ETA), source 0, shape + (a month of trips with time and weather), risk - (a sibling of the Lusaka task; keep one city).
+
 Zindi mobility-prediction hackathon for Accra (the Yango-family ride-hailing series). These challenges generally forecast transport demand/mobility over time and space, which falls under excluded time-series forecasting and is RecSys/operations-adjacent. The record has no comments or descriptive fields to confirm the precise target. A human must check whether the task is a fixed predictive label per row (possibly viable) versus a demand-over-horizon forecast (excluded).

curation/records/2025_maputo_mobility_prediction_hackathon_powered_by_yango.md

@@ -3,10 +3,8 @@ unique_name: 2025_maputo_mobility_prediction_hackathon_powered_by_yango
 name: 2025-maputo-mobility-prediction-hackathon-powered-by-yango
 checked_by:
 - AI (UNVERIFIED)
-suggestion: 'No'
-decision_markers:
-- Time-series (Forecasting)
-- Missing source information
+suggestion: TBD -> 2nd Tier
+priority: 2
 tags:
 - AI-Filled (Verify)
 collections:
@@ -16,9 +14,10 @@ year: '2025'
 domain: technology & internet
 required_split:
 - Temporal (NON-IID)
-problem_type: TBD
+problem_type: Regression
 source_links:
 - https://zindi.africa/competitions/2025-maputo-mobility-prediction-hackathon-powered-by-yango/data
+- https://zindi.africa/competitions/2025-maputo-mobility-prediction-hackathon-powered-by-yango
 source_row: 949
 needs_review:
 - ai_unverified
@@ -29,4 +28,13 @@ type_adapter_id: curation-record-v1
 
 **⚠️ AI-FILLED — UNVERIFIED. The suggestion, metadata, and notes below were drafted by an AI assistant from public knowledge of this competition/dataset, NOT from inspecting the data. A human must verify everything before relying on it.**
 
+**Screening (AI, 2026-10-22):** `TBD -> 2nd Tier`, priority 2: a per-trip ride-time regression (Yango's ETA), not forecasting; the same task as `2024_lusaka_mobility_prediction_hackathon` in another city.
+- Source: Zindi one-day hackathon "2025 Maputo Mobility Prediction Hackathon Powered" by Yango, 11 Oct 2025 (https://zindi.africa/competitions/2025-maputo-mobility-prediction-hackathon-powered-by-yango, read through Zindi's page API); as published: "a sampled dataset of trips via the Yango platform from March 2024 in Maputo" (Description), but "from April 2024 in Mozambique" (Data tab, which ends in a stray "xxx"): the month is to check. Train.csv (9.0 MB, with the target), Test.csv (3.7 MB), Maputo_weather.csv, VariableDefinitions.csv; coordinates geomasked, straight-line distance from the original points, short trips removed. Rows and columns not counted (login).
+- Use case: "Yango uses a wide range of data in order to predict travel and arrival times" (Description): a ride time quoted at ordering; fit direct if every column is known then; regime temporal (one month of trips).
+- Siblings and duplicates: `2024_lusaka_mobility_prediction_hackathon` (15.6 MB) and `2025_accra_mobility_prediction_hackathon_powered` are the same task and publisher on other cities' data: one city is enough. An older Maputo hackathon (maputo-mobility-prediction-hackathon, listed in `mobility_prediction`) is the road-speed task. Precedents: `nyc_taxi_trip_duration`, `pkdd_15_taxi_trip_time_prediction_ii`, `maps_router_eta` (`Yes`).
+- Papers and pages: no paper; the competition's tabs; no public solution found.
+- Metric: RMSE ("The evaluation metric for this competition is Root Mean Squared Error", Evaluation tab); MET-4; inputs none.
+- Risks for /add-dataset: the column list is unread (post-trip columns, LEAK); the data month (March or April 2024); access: login, "only open to Mozambique users residing in Mozambique" (no criterion, EV-4).
+- Priority: gap 0, use case + (Yango's ETA), source 0 (the page contradicts itself on the month), shape + (a month of trips with time and weather), risk - (a sibling of the Lusaka task; keep one city).
+
 Zindi mobility-prediction hackathon for Maputo, explicitly powered by Yango (a ride-hailing platform), so the task is almost certainly ride/trip demand forecasting. That is excluded forecasting and operations/RecSys-adjacent rather than a fixed tabular classification/regression task. The record contains no descriptive content. A human must verify the actual target and split regime before any reconsideration.

curation/records/2030_vision_flood_prediction_in_malawi.md

@@ -4,20 +4,20 @@ name: 2030-vision-flood-prediction-in-malawi
 checked_by:
 - AI (UNVERIFIED)
 suggestion: TBD -> 2nd Tier
-decision_markers:
-- Missing source information
-- Time-series (Forecasting)
+priority: 2
 tags:
 - AI-Filled (Verify)
 collections:
 - New (BeyondArena)
 original_source: Zindi
+year: '2019'
 domain: environmental science & climate
 required_split:
 - '?'
-problem_type: TBD
+problem_type: Regression
 source_links:
 - https://zindi.africa/competitions/2030-vision-flood-prediction-in-malawi/data
+- https://zindi.africa/competitions/2030-vision-flood-prediction-in-malawi
 source_row: 915
 needs_review:
 - ai_unverified
@@ -28,4 +28,14 @@ type_adapter_id: curation-record-v1
 
 **⚠️ AI-FILLED — UNVERIFIED. The suggestion, metadata, and notes below were drafted by an AI assistant from public knowledge of this competition/dataset, NOT from inspecting the data. A human must verify everything before relying on it.**
 
+**Screening (AI, 2026-10-22):** `TBD -> 2nd Tier`, priority 2: the flooded fraction of each ~1 km cell in southern Malawi for a flood event, from elevation, land cover and weekly rainfall; a fixed regression, not a series forecast, but only one event is labelled in the public train file.
+- Source: Zindi, "UNICEF Arm 2030 Vision #1: Flood Prediction in Malawi", Dec 2019 to May 2020, sponsored by Arm and UNICEF (2030Vision) (https://zindi.africa/competitions/2030-vision-flood-prediction-in-malawi, read through Zindi's page API); as published: one Train.csv (7.1 MB) of ~1 km x 1 km cells (0.01 degree, X and Y centres) with "the fraction (percentage) of that rectangle that was flooded in 2015", mean elevation, dominant land cover (MODIS LC_Type1), and weekly precipitation "for 18 weeks beginning 2 months before the flooding" for both the 2015 and 2019 events; the test is the same cells' extent for cyclone Idai, March 2019 ("target_2019"), scored on Zindi. Rows not counted.
+- Use case: "a machine learning model that helps predict the location and extent of floods in southern Malawi" (Description), for humanitarian preparedness (UNICEF; Zindi blog "Zindi solutions will help Malawi prepare better for flooding"); actor and decision not named, fit approximate: the rainfall covers the flood weeks, so it estimates extent at the event rather than ahead. Regime: by event (train 2015, test 2019) with strong spatial autocorrelation; with only 2015 labelled, our split would be spatially grouped within one event, which does not mirror a next flood.
+- Siblings and duplicates: none among the records (searched malawi, flood); external collections: none found. A participant's repo (kmezhoud, learn_by_example, Flood_prediction_Malawi) claims 2015 and 2019 datasets: whether the 2019 labels are public is to check.
+- Papers and pages: no paper; the competition's tabs, discussions 644, 646, 1241, 1363; a Kaggle U-Net baseline ("zindi-flood-in-malawi-unet-baseline") treats the cells as a raster.
+- Metric: RMSE on the fraction ("The error metric for this competition is the Root Mean Squared Error", Evaluation tab); MET-4; inputs none.
+- Risks for /add-dataset: the target is derived from satellite flood maps and the cells form a grid where a U-Net is a baseline: judge SEL-12 on what wins (the features are tabular); the 2019 labels; the split (one event, spatial blocks); many cells at 0 (a hurdle shape, per a moderator).
+- Markers: the old `Missing source information` is refuted (Zindi, UNICEF and Arm, a documented page); `Time-series (Forecasting)` dropped: the target is an event's extent per cell, not future values of a series, though Zindi tags it "Forecast".
+- Priority: gap + (a spatially grouped regime is thin), use case 0 (approximate, actor unnamed), source 0, shape + (thousands of cells, 20+ features, spatial structure), risk - (one labelled event, raster structure).
+
 A Zindi flood-prediction challenge for Malawi with no descriptive comments. Flood prediction can be either a spatio-temporal forecast (excluded) or a fixed per-location/per-event risk classification on tabular hydrological/geospatial features (potentially viable). Without details the modality (it may rely on satellite/remote-sensing rasters rather than tabular features) and the target are unknown. A human must inspect whether the data is genuinely tabular and whether the task is a fixed prediction vs. a time-horizon forecast.

curation/records/absa_corporate_client_activity_forecasting_challenge.md

@@ -28,6 +28,15 @@ type_adapter_id: curation-record-v1
 
 **⚠️ AI-FILLED — UNVERIFIED. The suggestion, metadata, and notes below were drafted by an AI assistant from public knowledge of this competition/dataset, NOT from inspecting the data. A human must verify everything before relying on it.**
 
+**Screening (AI, 2026-10-22):** `No` (`Time-series (Forecasting)`, SEL-10): the source asks for a forecast of each corporate client's activity series, the 6-hour windows of the day after seven weeks of raw event logs; no fixed per-row task table is published.
+- Source: Zindi, "Absa Corporate Client Activity Forecasting Challenge", 1-27 Nov 2022, Absa CIB (https://zindi.africa/competitions/absa-corporate-client-activity-forecasting-challenge, read through Zindi's page API); as published: Train.csv (58 MB) of events (`eventdatetime`, `event`, `userid`, `dayofweek`, `useremaildomain`, `userrole`, `companyprofileid`, `country`) for "4 200 customers, across five countries over a period of seven weeks"; Test.csv (0.7 MB) lists user x 6-hour slots of one day, e.g. `User2504_Wed_3_6`.
+- Deciding words: subtitle "Can you forecast if a corporate banking client will be active in the following week?"; Data tab "The training set is 7 weeks of data, and the last day of the final week will be the test set, where the target is whether a client performed event 14 during that day at 6-hour intervals"; Zindi's problem type "Forecast".
+- Use case: Absa CIB anticipating platform usage ("help Absa anticipate their clients' needs, allowing for better planning, delivery and customer service"); the prediction is the next values of each client's event series over a one-day horizon.
+- Why not SEL-11: the label is built from the same event series the features come from, one horizon day is shared by all clients, and the older note records that the winning solution was an LSTM; a per-client churn-like reframing would be our task, not the source's (SEL-8).
+- Siblings and duplicates: `absa_customer_income_prediction_challenge` is another dataset; none else.
+- Metric: F1 on hard labels ("The error metric for this competition is the F1 score", Evaluation tab), MET-4.
+- Reject check (AI, 2026-10-22): stands. SEL-10 (Must). Data tab: "The training set is 7 weeks of data, and the last day of the final week will be the test set, where the target is whether a client performed event 14 during that day at 6-hour intervals". Zindi type_of_problem ["Forecast"]. The target event is in the same raw event log as the features. All clients share one horizon day, no row table is published, and the older note says the winner used an LSTM. A TP-10 per-client reframing would be our own task (SEL-8). https://zindi.africa/competitions/absa-corporate-client-activity-forecasting-challenge
+
 Predict whether a corporate client performs a specific action within a 6-hour window across a day, using seven weeks of prior event history. This is a sequential/temporal forecasting task over future windows (the winning solution used an LSTM), which is excluded time-series forecasting rather than a fixed IID tabular task. The temporal tag and the 'forecasting' framing both reinforce this. A human could double-check whether an aggregated per-client tabular reformulation exists, but as posed it should be rejected.
 
 ---