Geo ML @ MediaEval Placing Task 2015

Geo ML @ MediaEval Placing Task 2015MediaEval 2015, Wurzen, Germany

Nghia Duong-Trung, Martin Wistuba, Lucas Rego Drumond, LarsSchmidt-Thieme

Information Systems and Machine Learning Lab (ISMLL)University of Hildesheim, Germany

September 14-15, 2015

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 1 / 8

Geo ML @ MediaEval Placing Task 2015

Outline

Three-step approach

Experimental results

Discussions and outlook

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 2 / 8

Geo ML @ MediaEval Placing Task 2015 Three-step approach

Outline

Three-step approach

Experimental results

Discussions and outlook

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 3 / 8

Geo ML @ MediaEval Placing Task 2015 Three-step approach

Notations

I m, v , n geotagged multimedia items in the training and test data,number of features respectively

I Each item is annotated with a geo-location y ∈ R2, y = (y lat , y lon)where y lat ∈ R is the latitude and y lon ∈ R is the longitude

I X train ∈ Rm×n, and the respective labels Y train ∈ Rm×2

I find a model f : Rn → R2 such that for some test data X test ∈ Rv×n,the error

∑vi=1 d(f (X test

i ),Y testi ) is minimal

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 3 / 8

Geo ML @ MediaEval Placing Task 2015 Three-step approach

Three-step approach

I Find a model f : Rn → R2

I Three steps:

I find a cluster assignment vector c ∈ Cm, where the i-th element cicontains the cluster assigned to the i-th instance based on itsgeo-location yi

I learn a model on X train and c in order to map the test instances tothose clusters

I predicted geo-location ytesti is the geo-location of the nearest neighborin the same cluster

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 4 / 8

Geo ML @ MediaEval Placing Task 2015 Three-step approach

Three-step approach

Algorithm 1 Three-step approach

Require: X train, X test , Y train, cost s, number of clusters k1: {Step 1: k-means clustering}2: c← Kmeans(Y train, k)3: {Step 2: Linear SVM}4: g ← LinearSVM(X train, s, c)5: {Step 3: k-nearest neighbor}6: for i = 1 . . . v do7: ci ← g(X test

i )8: X ,Y ← rows of X train,Y train belonging to cluster ci9: yi ← 1NNRegression(X ,Y )

10: end for11: return yi

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 5 / 8

Geo ML @ MediaEval Placing Task 2015 Experimental results

Outline

Three-step approach

Experimental results

Discussions and outlook

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 6 / 8

Geo ML @ MediaEval Placing Task 2015 Experimental results

Experimental results

I Implementation achieves a median error of 352.47 km to the test set.

distance # items percentage

0.001 km 504 0.05 %

0.01 km 1051 0.11 %

1 km 11849 1.25 %

10 km 287807 30.03 %

100 km 418831 44.09 %

1000 km 566791 59.67 %

10000 km 911364 95.94 %

40000 km 949889 100.00 %

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 6 / 8

Geo ML @ MediaEval Placing Task 2015 Discussions and outlook

Outline

Three-step approach

Experimental results

Discussions and outlook

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 7 / 8

Geo ML @ MediaEval Placing Task 2015 Discussions and outlook

Discussion and outlook

I Conceptual to understand and simple to implement

I Easily replace and integrate new techniques in the workflow withoutredesigning the complete system

I Outlook

I Only apply K-means on Y train without any textual knowledge orlanguage models

I Do not apply feature ranking, dimension reduction techniques. (alreadydid in another following up paper)

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 7 / 8

Geo ML @ MediaEval Placing Task 2015 Discussions and outlook

Thank you for your attention.

Duong-Trung et al., Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany 8 / 8