wenzheli/geographconv

Semi-supervised User Geolocation via Graph Convolutional Networks

Semi-supervised User Geolocation via Graph Convolutional Networks

Introduction

This is a Theano implementation of Semi-supervised User Geolocation via Graph Convolutional Networks paper, published in ACL 2018. It contains 3 geolocation models: 1) gcn, 2) deep cca, and 3) concat network and text. The input data contains Twitter users with their tweets concatenated as a single document, and the @-mentions in their tweets is used to build a graph between users.

The model uses the node features and the graph between them jointly to predict a location for users. For more information about the models refer to the paper. It contains a Theano-based Graph Convolutional Network model which can be used standalone for other experiments.

Geolocation Datasets

Datasets are GEOTEXT a.k.a CMU (a small Twitter geolocation dataset) and TwitterUS a.k.a NA (a bigger Twitter geolocation dataset) both covering continental U.S. which can be downloaded from here

Preprocessed Data

If you want to use the preprocessed data e.g., X, A in your own model download the pickle files from here (1 dump.pkl file for each dataset). A is the normalised Laplacian matrix, and X is the node features (BoW) partitioned into train, dev, and test sets.

Then load the file like this:

import numpy as np
import cPickle
import scipy as sp

def load_obj(filename, serializer=cPickle):
    with gzip.open(filename, 'rb') as fin:
        obj = serializer.load(fin)
    return obj


data = load_obj('dump.pkl')
A, X_train, Y_train, X_dev, Y_dev, X_test, Y_test, U_train, U_dev, U_test, classLatMedian, classLonMedian, userLocation = data
#A is the normalised laplacian matrix as A_hat in Kipf et al. (2016).
#The X_? and Y_? should be concatenated to be feed to GCN.
X = sp.sparse.vstack([X_train, X_dev, X_test])
if len(Y_train.shape) == 1:
    Y = np.hstack((Y_train, Y_dev, Y_test))
else:
    Y = np.vstack((Y_train, Y_dev, Y_test))
print(A.shape, X.shape, Y.shape)
#get train/dev/test indices in X, Y, and A.
train_indices = np.asarray(range(0, X_train.shape[0])).astype(dtypeint)
dev_indices = np.asarray(range(X_train.shape[0], X_train.shape[0] + X_dev.shape[0])).astype(dtypeint)
test_indices = np.asarray(range(X_train.shape[0] + X_dev.shape[0], X_train.shape[0] + X_dev.shape[0] + X_test.shape[0])).astype(dtypeint)

Then build your model and make predictions on X_test to get y_pred. Then use the following function to evaluate the geolocation performance:

from haversine import haversine
import logging
def geo_eval(y_true, y_pred, U_eval, classLatMedian, classLonMedian, userLocation):
    assert len(y_pred) == len(U_eval), "#preds: %d, #users: %d" %(len(y_pred), len(U_eval))
    distances = []
    latlon_pred = []
    latlon_true = []
    for i in range(0, len(y_pred)):
        user = U_eval[i]
        location = userLocation[user].split(',')
        lat, lon = float(location[0]), float(location[1])
        latlon_true.append([lat, lon])
        prediction = str(y_pred[i])
        lat_pred, lon_pred = classLatMedian[prediction], classLonMedian[prediction]
        latlon_pred.append([lat_pred, lon_pred])  
        distance = haversine((lat, lon), (lat_pred, lon_pred))
        distances.append(distance)

    acc_at_161 = 100 * len([d for d in distances if d < 161]) / float(len(distances))

    logging.info( "Mean: " + str(int(np.mean(distances))) + " Median: " + str(int(np.median(distances))) + " Acc@161: " + str(int(acc_at_161)))
        
    return np.mean(distances), np.median(distances), acc_at_161, distances, latlon_true, latlon_pred

mean, median, acc, _, _, _ = geo_eval(y_true, y_pred, U_eval, classLatMedian, classLonMedian, userLocation)

Quick Start

1. Download the datasets and place them in ''./data/cmu'' and ''./data/na''

.
├── data
│   ├── cmu
│   │   ├── user_info.dev.gz
│   │   ├── user_info.test.gz
│   │   ├── user_info.train.gz
│   │
│   └── na
│   	├── user_info.dev.gz
│   	├── user_info.test.gz
│       ├── user_info.train.gz
│  
├── data.py
├── data.pyc
├── deepcca.py
├── experiments
│   ├── cmu-concat-fractions.sh
│   ├── cmu-dcca-fractions.sh
│   ├── cmu-gcn-fractions.sh
│   ├── cmu_layers_highway.sh
│   ├── cmu_layers_nohighway.sh
│   ├── feature_report.sh
│   ├── save_model.sh
│   └── tunebucket.sh
├── gcnmain.py
├── gcnmodel.py
├── gcnmodel.pyc
├── kdtree.py
├── kdtree.pyc
├── mlp.py
├── README.md
├── requirements.txt
└── utils.py

2. Create a new environment:

conda create --name geo python=2.7

Activate the environment:

conda activate geo

Install Libraries:

pip install -r requirements.txt

Upgrade Theano and Lasagne:

pip install --upgrade https://github.com/Theano/Theano/archive/master.zip

pip install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip

2. To run the experiments, look at the experiments directory.

Note 1: The default parameters are not suitable for running the experiments.

Note 2: By changing the seed (e.g. using -seed 1 in command line) the results might slightly change, they might be slightly better or worse than the reported in the paper, but they shouldn't be very different.

For the GCN model:

CMU (runtime: 3min):

THEANO_FLAGS='device=cuda0,floatX=float32' nice -n 9 python -u gcnmain.py -hid 300 300 300 -bucket 50 -batch 500 -d ./data/cmu/ -enc latin1 -mindf 10 -reg 0.0 -dropout 0.5 -cel 5  -highway

NA (runtime: 15h):

THEANO_FLAGS='device=cpu,floatX=float32' python -u gcnmain.py -hid 600 600 600 -bucket 2400 -batch 500 -d ~/data/na/ -mindf 10 -reg 0.0 -dropout 0.5 -cel 15 -highway

WORLD (runtime: 2.5days):

THEANO_FLAGS='device=cpu,floatX=float32' python -u gcnmain.py -hid 900 900 900 -bucket 2400 -batch 500 -d ~/data/world/ -mindf 10 -reg 0.0 -dropout 0.5 -cel 5 -highway

Citation

@InProceedings{P18-1187,
  author = 	"Rahimi, Afshin
		and Cohn, Trevor
		and Baldwin, Timothy",
  title = 	"Semi-supervised User Geolocation via Graph Convolutional Networks",
  booktitle = 	"Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
  year = 	"2018",
  publisher = 	"Association for Computational Linguistics",
  pages = 	"2009--2019",
  location = 	"Melbourne, Australia",
  url = 	"http://aclweb.org/anthology/P18-1187"
}

Contact

Afshin Rahimi [email protected]