Neural Embeddings for Populated GeoNames Locations

This dataset contains embeddings and associated files for all Geonames populated locations with population greater than 0.

These are expressed as DeepWalk embeddings for populated places in Geonames.
Specific files and file groups are described below. All embeddings were generated using the gensim package in Python.
Geonames is a large, comprehensive knowledge base of geographical locations, both populated and unpopulated, and at different administrative levels.

FILES:

populated-place-embeddings-{1-7}.jl.gz are compressed JSON
lines files where each line is a JSON containing a single
key-value pair. For example:
{"": [-0.00016037290333770216,
-0.16331960260868073, 0.14559058845043182, 0.014049071818590164,
0.016628244891762733, -0.1897198110818863, 0.14188186824321747,
0.11444853991270065, 0.018866628408432007, -0.04606882855296135,
-0.15467466413974762, 0.18563739955425262, -0.053200043737888336,
-0.16911742091178894, -0.05657876282930374, -0.00039542425656691194,
0.06017935276031494, 0.019657278433442116, 0.014765074476599693,
0.1783478558063507, -0.0005339447525329888, -0.10394854098558426,
-0.17260663211345673, 0.022552259266376495, 0.060257066041231155,
0.298672616481781, 0.06802289932966232, -0.03377196565270424,
0.13092969357967377, 0.007119285874068737, -0.011325502768158913,
-0.03717223182320595, 0.10660370439291, 0.09025651216506958,
0.023348767310380936, -0.1824614256620407, 0.13780122995376587,
-0.03130345419049263, -0.10430310666561127, 0.02684781886637211,
0.06622970849275589, -0.11856741458177567, -0.053687382489442825,
0.14932486414909363, 0.03837030380964279, 0.0744427815079689,
0.08571769297122955, 0.007302495185285807, 0.11334840208292007,
0.0385720431804657, -0.15567876398563385, -0.134855255484581,
0.10208329558372498, 0.07373132556676865, -0.08734219521284103,
0.0397501066327095, -0.07437381893396378, 0.019372688606381416,
-0.03451831266283989, 0.015789508819580078, 0.05273417755961418,
-0.06393982470035553, -0.09630796313285828, -0.047562047839164734,
-0.07217235118150711, -0.05011208727955818, 0.08369360864162445,
0.08336202055215836, 0.05594024807214737, 0.12139936536550522,
0.051332831382751465, 0.027806641533970833, 0.09555482864379883,
-0.14043797552585602, -0.003694443963468075, 0.009580886922776699,
0.19412516057491302, 0.05366421863436699, 0.10408374667167664,
0.12925657629966736, -0.07961612194776535, -0.07847319543361664,
-0.23122484982013702, -0.1003454178571701, 0.1332612931728363,
-0.00953035056591034, 0.030777478590607643, 0.054927386343479156,
-0.17683790624141693, 0.0793348103761673, 0.06144792586565018,
-0.05212269350886345, 0.07661330699920654, 0.08598912507295609,
-0.026763606816530228, 0.0698607936501503, 0.16680659353733063,
0.07281855493783951, -0.0883372575044632, 0.015735004097223282]}
is the 100-dimensional embedding of Presidio, which has Geonames ID
3521133. This way, each location is canonically
identified but is also human-readable.

To combine all the files into a single jl file use the cat command (after decompressing) in a UNIX-like terminal

The first 100 lines of populated-place-embeddings-1.jl.gz are in
samples-100-locations.jsonl (renamed from .jl to prevent Github from
erroneously concluding this is a Julia project!) and can be viewed in a
browser
mapped_places.txt.gz: To efficiently read things in memory and
process, we mapped each of the mnemonic Geonames keys to an integer.
E.g.
287145
we use tab to separate the two items. 100 sample lines were written out
to samples-100-mapped_places.txt
Note that mapped_places.txt contains keys that you will not find in the
embedding files. These are places that are not
populated anymore i.e. Geonames shows them as having population 0 (or
missing a population feature).
random-walk-corpus.txt.gz: contains the random walk corpus that was used by DeepWalk for the embeddings. Each line
contains a random walk sequence, space separated. Each item in the sequence is the mapped int of the mnemonic Geonames key.
We initiated 5 samples per node, up to a maximum depth of
weighted_adjacency_list.tsv.gz: contains the directed weighted network that was used to generate the random walks. The
items in each line are tab-delimited and are of the form:
[node-out] [node-in1] [weight1] [node-in2] [weight2]...
where node-out is the node from which the edges are outgoing (to node-in1, node-in2...). The weights represent a probability
distribution that is inversely related to the distance between the places represented by the nodes. For more details,
see the paper.

samples-100-weighted_adjacency_list.tsv contains the first 100 lines of (11) for easy viewing.