Detail plot of the used data set. The topmost attributes of the global decision tree model were used as axis. 

Context in source publication

Context 1

... is understood that the pain face is not the usual face but a distortion of the neutral face. Taking this into account we averaged the distances, angles and ratios of the non-pain data records and added for every record the ratios of those attribute to their means. As the deformation of the face is potentially highly individual we also added the ratios of those attributes to their means of the considered subject’s non-pain data. Of course all mentioned relational measures are implied in the point data alone. However they would be inaccessible for some classifiers, e.g. Decision Trees (see Sect. 3.1). Hence we decided to include those measures in an explicit form. Based on the data set with 534 entries and 178 attributes (including image id, person number and class) classifiers were trained. In the following the used classifier are briefly described. For further reading see [7] or [8](Support Vector Machines). Decision Trees classify an example according to a set of tree structured if-then- rules. Each inner node of the tree denotes an attribute and branches according to its values. The Decision Tree classifier is the only symbolic classifier we used. We considered using it as the constructed decision model is human under- standable and thus has an explaining element. For learning this trees we used ID3Numerical, a modification of Quinlan’s ID3 [9] which can handle numerical attributes. Naive Bayes is a statistical classifier that is based on Bayes’ Law. It estimates the probability P ( c | x ). For this purpose it assumes that all attributes are statistical independent ( thus naive ). Support Vector Machines try to find a hyperplane – probably in a higher space – that separates the classes. For classification only the data records closest to the hyperplane are used – the Support Vectors . k -Nearest Neighbours ( k NN) considers each data record as points in IR . An unseen instance is now classified searching the k nearest points in the training example space and assigning their mode class. All training examples are kept in the feature space. Therefor k NN is called a lazy learner . The Perceptron considers each data record as a vector of real-valued inputs ( x 1 , . . . , x n ). Those are weighted with ( w 0 , . . . , w n ). For classification it calcu- n lates the linear combination i =0 w i x i , where x 0 is always 1. If the result is greater than 0 it returns positive , negative otherwise. Neural Networks use a graph of sigmoid units to calculate the desired class. Sigmoid units are similar to perceptrons except that they use a sigmoidal output. This output is serving as input for the next unit. In this study we used a linear and acyclic network. Classification by Regression trains a regression model for each class. The class with the higher predicted value is then selected. As base regression model we used linear regression. Linear regression tries to find the linear function that predicts the examples best. For each classifier the learning generally consisted of two phases. First its optimal parameters were obtained then its performance was measured in an cross- validation. As k NN suffers from curse of dimensionality, the attributes were weighted for this classifier before parameter optimization. Cross-Validation was done using 10 partitions. The partitions were drawn according to single data records by means of stratified sampling 4 . Parameters were optimized via a systematic grid search. Parameter combinations were systematically evaluated using 10-folded cross-validations. The combination which performed best was selected. For the Neural Network no systematic approach was chosen. Taking the long training times and the size of the parameter space into account we decided on an evolutionary technique. Attribute Weighting was performed using forward weighting. Initially every attribute was assigned a weight of 0. Each attribute was then independently weighted using a linear search. Attribute Selection was used to overcome Naive Bayes’ assumptions that attributes are statistical independent. Carried out as forward selection, an initial population is created – one individual per attribute. Then further attributes are added to the best ones as long as performance increases. This proceedings were carried out with the whole data set (global) and once for each subject – using only its data (individual classifiers). We ran the experi- ments with RapidMiner 5 on a Fujitsu Siemens Computers LIFEBOOK T Series (Intel c Core TM 2 Duo P8400, 2 GB) and a Fujitsu Siemens Computers Esprimo (Intel R Pentium R 4 3.00 GHz, 1 GB). The results of the different classifiers are shown in Tab. 1. Obviously most classifiers are suitable for this task. Apparently Support Vector Machines and k -nearest Neighbours perform best. Unfortunately it is not possible to build an exact ranking. Due to the number of subjects, no significance tests were made. Decision Tree, Support Vector Machine, Regression, Naive Bayes and k -nearest Neighbours perform similar as adults classifying clinical pain videos[10]. No detailed statement about the Neural Network can be made. The global parameter optimization crashed after 10 days. Since no interim results were available performance estimation was done with manually chosen parameters. A repetition of the parameter optimization was not possible due to time limitations. The detail plot in Fig. 3 shows that the data is probably not linear separable. This explains the bad performance of the Perceptron classifier. It performs even worse than guessing 6 . For most classifiers the individual approach seems to deliver better results. But— due to the small number of subjects—also here no detailed comparison is possible. We showed that many classifiers are suitable for predicting pain by facial expression. Global and individual classifiers perform nearly equally. But due to the low subject number and since we only selected those subjects that showed prototypical facial pain displays[11] it is by now questionable if these findings can be generalized. Currently we are able to decide if a shown face is prototypical for pain or not. However, the true question is whether the person whose face is depicted is experiencing pain. Therefore we will deal with predicting the self-reported VAS values in further work. Additionally we will tackle the issue of mixing up different emotions, first trying to distinguish between pain and disgust. Therefore we will use more subjects for the initial psychological study. Further studies should also address a broader variety of subjects—for example regarding age or gender. Acknowledgements We thank Simone Burkhardt and Ina Schulz for the data collection and picture extraction. The study was supported by the Dr. Werner ...