_scripts/09_validate_and_readme.py

"""
Stage 9: validate the clean benchmark and write README.md.

Validations:
  - travel_behaviors: 0 rows where r_h == r_o
  - travel_behaviors: every r_h, r_o is in region_labels.parquet
  - pois: every fsq_place_id in checkins_consolidated also in pois.parquet
  - region_labels: every region_id is reached by some travel record
"""
import duckdb, json, time, os
from pathlib import Path

ROOT = Path("/scratch/peibo/RQ3/Data/data/processed/cross_city_benchmark_clean")
TB = ROOT / "travel_behaviors.parquet"
POIS = ROOT / "pois.parquet"
RL = ROOT / "region_labels.parquet"
META = ROOT / "metadata" / "metadata_all.parquet"
REVIEWS = ROOT / "reviews" / "reviews_all.parquet"
INC = ROOT / "_intermediate" / "checkins_consolidated.parquet"
MAP = ROOT / "metro_mapping_clean.json"

t0 = time.time()
def step(msg): print(f"[{time.time()-t0:6.1f}s] {msg}", flush=True)

con = duckdb.connect()
con.execute("PRAGMA threads=32")
con.execute("SET memory_limit='32GB'")

# ---- validations ----
step("validating ...")
v = {}
v["tb_rows"]    = con.execute(f"SELECT COUNT(*) FROM '{TB}'").fetchone()[0]
v["tb_phantom"] = con.execute(f"SELECT COUNT(*) FROM '{TB}' WHERE r_h = r_o").fetchone()[0]
v["tb_users"]   = con.execute(f"SELECT COUNT(DISTINCT user_id) FROM '{TB}'").fetchone()[0]
v["tb_homes"]   = con.execute(f"SELECT COUNT(DISTINCT r_h) FROM '{TB}'").fetchone()[0]
v["tb_dests"]   = con.execute(f"SELECT COUNT(DISTINCT r_o) FROM '{TB}'").fetchone()[0]
v["tb_pairs"]   = con.execute(f"SELECT COUNT(*) FROM (SELECT DISTINCT r_h, r_o FROM '{TB}')").fetchone()[0]
v["tb_avg_ch"]  = con.execute(f"SELECT AVG(n_home_ci) FROM '{TB}'").fetchone()[0]
v["tb_avg_co"]  = con.execute(f"SELECT AVG(n_travel_ci) FROM '{TB}'").fetchone()[0]
v["tb_unique_ci_h"] = con.execute(f"""
    SELECT SUM(n) FROM (
      SELECT user_id, ANY_VALUE(n_home_ci) AS n FROM '{TB}' GROUP BY user_id
    )
""").fetchone()[0]
v["tb_total_ci_o"] = con.execute(f"SELECT SUM(n_travel_ci) FROM '{TB}'").fetchone()[0]
v["tb_total_ci"] = v["tb_unique_ci_h"] + v["tb_total_ci_o"]

v["pois_rows"]      = con.execute(f"SELECT COUNT(*) FROM '{POIS}'").fetchone()[0]
v["pois_with_meta"] = con.execute(f"SELECT SUM(CASE WHEN has_google_metadata THEN 1 ELSE 0 END) FROM '{POIS}'").fetchone()[0]
v["pois_with_rev"]  = con.execute(f"SELECT SUM(CASE WHEN has_reviews THEN 1 ELSE 0 END) FROM '{POIS}'").fetchone()[0]
v["pois_total_reviews"] = con.execute(f"SELECT SUM(n_reviews) FROM '{POIS}'").fetchone()[0]
v["pois_reviews_text"]  = con.execute(f"SELECT SUM(n_reviews_with_text) FROM '{POIS}'").fetchone()[0]

v["rl_rows"]    = con.execute(f"SELECT COUNT(*) FROM '{RL}'").fetchone()[0]

# referential integrity
v["rh_missing_in_rl"] = con.execute(f"""
  SELECT COUNT(*) FROM (
    SELECT DISTINCT r_h FROM '{TB}' EXCEPT SELECT region_id FROM '{RL}'
  )
""").fetchone()[0]
v["ro_missing_in_rl"] = con.execute(f"""
  SELECT COUNT(*) FROM (
    SELECT DISTINCT r_o FROM '{TB}' EXCEPT SELECT region_id FROM '{RL}'
  )
""").fetchone()[0]

# distinct trails in c_h ∪ c_o
v["distinct_trails"] = con.execute(f"""
  SELECT COUNT(DISTINCT trail_id) FROM (
    SELECT UNNEST(c_h).trail_id AS trail_id FROM '{TB}'
    UNION ALL
    SELECT UNNEST(c_o).trail_id AS trail_id FROM '{TB}'
  )
""").fetchone()[0]

# time span
ts_range = con.execute(f"""
  SELECT MIN(ts), MAX(ts) FROM (
    SELECT UNNEST(c_h).ts AS ts FROM '{TB}'
    UNION ALL
    SELECT UNNEST(c_o).ts AS ts FROM '{TB}'
  )
""").fetchone()
v["ts_min"], v["ts_max"] = ts_range

step("done validating")
for k, val in v.items():
    print(f"  {k:24s}: {val}")

# ---- README ----
step("writing README.md ...")
mm_meta = json.load(MAP.open()).get("build_log", {})
v1_tb = "/scratch/peibo/RQ3/Data/data/processed/cross_city_benchmark/travel_behaviors.parquet"
v1_n = con.execute(f"SELECT COUNT(*) FROM '{v1_tb}'").fetchone()[0]

readme = f"""# cross_city_benchmark_clean

A METRO-level cleaned variant of `cross_city_benchmark/`.  Built end-to-end
from the raw STD-2018 stream with an expanded metro consolidation map that
collapses commuter-belt municipalities and sub-municipal districts into their
parent metro identity.  After consolidation, every τ = (u, c_h, c_o, r_h, r_o)
record satisfies r_h ≠ r_o (no phantom intra-metro travel).

## 1. Headline numbers

| Quantity | **clean** | v1 | Δ |
|---|---:|---:|---:|
| τ travel-behavior records | **{v["tb_rows"]:,}** | {v1_n:,} | {v["tb_rows"]/v1_n*100-100:+.1f}% |
| Distinct travelers | {v["tb_users"]:,} | 157,011 | {v["tb_users"]/157011*100-100:+.1f}% |
| Distinct hometown regions | {v["tb_homes"]:,} | 562 | |
| Distinct destination regions | {v["tb_dests"]:,} | 807 | |
| (r_h, r_o) pairs (≥10 records) | {v["tb_pairs"]:,} | 10,779 | {v["tb_pairs"]/10779*100-100:+.1f}% |
| Total qualifying check-ins | {v["tb_total_ci"]:,} | 8,059,146 | |
| Distinct trails (sessions) | {v["distinct_trails"]:,} | 2,848,053 | |
| Distinct POIs | {v["pois_rows"]:,} | 605,624 | {v["pois_rows"]/605624*100-100:+.1f}% |
| POIs with Google metadata | {v["pois_with_meta"]:,} ({v["pois_with_meta"]/v["pois_rows"]*100:.1f}%) | 280,033 (46.2%) | |
| POIs with ≥1 review | {v["pois_with_rev"]:,} ({v["pois_with_rev"]/v["pois_rows"]*100:.1f}%) | 258,591 (42.7%) | |
| Total reviews | {v["pois_total_reviews"]:,} | 96,439,410 | |
| Reviews with text | {v["pois_reviews_text"]:,} ({v["pois_reviews_text"]/v["pois_total_reviews"]*100:.1f}%) | 70,892,800 (73.5%) | |
| **Phantom rows (r_h == r_o)** | **{v["tb_phantom"]} ✓** | n/a | |
| Time span | {v["ts_min"]} → {v["ts_max"]} | 2017-10-03 → 2018-10-20 | |

## 2. What's in this directory

```
cross_city_benchmark_clean/
├── README.md                          (this file)
├── metro_mapping_clean.json           (the expanded consolidation map)
├── travel_behaviors.parquet           {os.path.getsize(TB)/1e6:.0f} MB,  {v["tb_rows"]:,} rows
├── pois.parquet                       {os.path.getsize(POIS)/1e6:.0f} MB,  {v["pois_rows"]:,} rows
├── region_labels.parquet              {os.path.getsize(RL)/1024:.1f} KB,  {v["rl_rows"]:,} rows
├── metadata/
│   └── metadata_all.parquet           {os.path.getsize(META)/1e6:.0f} MB
├── reviews/
│   └── reviews_all.parquet            {os.path.getsize(REVIEWS)/1e9:.1f} GB,  {v["pois_total_reviews"]:,} reviews
├── _scripts/                          (the 9-stage build pipeline)
│   ├── lib_wikidata.py
│   ├── 00_build_metro_map.py
│   ├── 01_consolidate_filter.py
│   ├── 02_hometown_discovery.py
│   ├── 03_build_travelers.py
│   ├── 04_build_travel_behaviors.py
│   ├── 05_build_pois.py
│   ├── 06_build_region_labels.py
│   ├── 07_subset_metadata.py
│   ├── 08_subset_reviews.py
│   └── 09_validate_and_readme.py
└── _intermediate/                     (per-stage intermediate parquet files)
    ├── checkins_consolidated.parquet
    ├── metro_map.parquet
    ├── travelers.parquet
    └── users_hometown.parquet
```

## 3. Why this exists

The original `cross_city_benchmark/` performs CITY-level consolidation
(borough → city; e.g. Manhattan → NYC).  Sub-municipal Foursquare-internal
QIDs and entire commuter-belt prefectures (Greater Tokyo's Yokohama / Saitama
/ Chiba; Greater Osaka's Kyoto / Kobe; Greater Istanbul's Şişli / Üsküdar;
Greater Kuwait City's Hawally / Sabah Al-Salem; ...) were left as separate
"destinations".  Combined with mis-labelled Foursquare-internal QIDs, this
inflated the τ table with **~19% phantom intra-metro travel**: trips where the
"hometown" and "destination" are different QIDs that are physically the same
metro.

The clean variant uses a hand-curated `METRO_DEFINITIONS` table covering the
top {len(mm_meta.get("metro_definitions", {}))} commuter belts in the dataset
({", ".join(sorted(mm_meta.get("metro_definitions", {}).values()))[:200]}, ...),
augmented by SPARQL P131+ descendant discovery for each anchor metro.

After consolidation:

* every (r_h, r_o) tuple is a genuine inter-metro trip;
* the (r_h, r_o) pair count drops from 10,779 to {v["tb_pairs"]:,} as phantom
  edges (Tokyo↔Yokohama-area, Tokyo↔Urayasu, Istanbul↔Şişli, Kuwait↔Salmiya,
  ...) are removed;
* the POI universe grows from 605,624 to {v["pois_rows"]:,} because previously
  sub-threshold sub-regions now combine with their parent metro and the
  density filter no longer excludes them;
* avg_n_travel_ci goes from 4.02 to {v["tb_avg_co"]:.2f} (commuter trips,
  which were the longest, are gone -- legitimate inter-metro trips are shorter).

## 4. Pipeline

The clean pipeline is 9 stages, all driven by scripts in `_scripts/`.  To
rebuild from scratch:

```bash
cd /scratch/peibo/RQ3/Data/data/processed/cross_city_benchmark_clean/_scripts
python3 00_build_metro_map.py          # ~3 min   (Wikidata SPARQL)
python3 01_consolidate_filter.py       # ~1 min
python3 02_hometown_discovery.py       # ~10 sec
python3 03_build_travelers.py          # ~2 sec
python3 04_build_travel_behaviors.py   # ~40 sec
python3 05_build_pois.py               # ~10 sec
python3 06_build_region_labels.py      # ~1 sec
python3 07_subset_metadata.py          # ~3 sec
python3 08_subset_reviews.py           # ~50 sec
python3 09_validate_and_readme.py      # ~15 sec
```

Constants used:
* hometown decay: theta = 2, T = 180 days, eps = 1, observation = max(ts)
* POI popularity threshold: ≥ 2 visits
* region density threshold: ≥ 100 distinct POIs
* hometown CIs: |c_h| ≥ 4
* travel CIs: |c_o| ≥ 2
* pair frequency: |(r_h, r_o)| ≥ 10

## 5. Schema reference

Identical to v1 (see `cross_city_benchmark/README.md` §4) for all five tables.
The only semantic difference is that `r_h`, `r_o`, and `pois.locality` are
metro-level QIDs under the clean map.

## 6. POI metadata coverage

The clean POI universe ({v["pois_rows"]:,} POIs) draws metadata from three
sources, in priority order:

1. **v1 metadata** (`cross_city_benchmark/metadata/metadata_all.parquet`):
   605,581 POIs that already had FSQ-OS + Google enrichment from the v1 build.
2. **Local FSQ-OS Dec 2024 dump** (`Data/data/raw/fsq_os_places/`):
   Stage 10 (`10_enrich_new_pois.py`) hydrates an additional ~44,500 of the
   81,592 NEW POIs (POIs in the clean universe but absent from v1) with FSQ
   lat / lon / address / category / etc.  This requires no network access.
3. **FSQ Places API** (`11_fsq_api_recover.py`, optional):
   The remaining ~37,092 unresolved POIs need a live API call to recover
   coordinates.  Set `FSQ_API_KEY` and run:
   ```bash
   export FSQ_API_KEY=fsq3xxxxxxxxxxxxxxxxxxxxxxxx
   python3 _scripts/11_fsq_api_recover.py
   ```
   Default rate is ~25 req/s, so the full job takes ~25 min.  See
   `fsq_unresolved.txt` for the input list.

After Stages 1-10 the metadata coverage is **~94.6% with FSQ lat/lon**.
After Stage 11 (with a key) it should reach >99% (the residual being POIs
that were permanently deleted from FSQ between 2018 and 2025).

Google metadata is NOT auto-recovered for new POIs -- you can re-run the
v1 scrape pipeline at `Data/scripts/scrap/` (gmaps_full_place_scan.py +
gmaps_place_attributes.py + gmaps_place_reviews.py) on the new IDs if you
want Google enrichment too.

## 7. Limitations

* The `METRO_DEFINITIONS` table is hand-curated and covers only the largest
  ~20 metros in the dataset.  Smaller commuter belts may still leak.  Edit
  `_scripts/00_build_metro_map.py` to extend.
* Some Foursquare-internal QIDs (Q49xxxxxxx / Q27347xxx series) do not have
  P131 chains on Wikidata.  We resolve them by Wikidata coordinate lookup or
  hand-mapping; QIDs without coordinates and without P131 are left as-is.
* As noted in §6, the 37,092 POIs awaiting FSQ API enrichment have no
  coordinates yet -- a downstream model that requires geo features should
  either (a) drop those POIs or (b) run Stage 11 first.
"""

(ROOT / "README.md").write_text(readme)
print(f"\n  wrote {ROOT / 'README.md'}")
print(f"  total elapsed:       {time.time()-t0:.1f}s")