The Application Of Improved SVM Algorithm In Affective Recognition Using Physiological Signals

Improved Support Vector Machine (SVM) was used to recognize affective states based on physiological signals in this paper. SVM as an effective classifier with solid mathematical foundation has been widely used in the field of affective computing. Kernel function parameter σ, penalty coefficient y and loss parameter ε play a vital role in SVM and affect the result directly. However, how to optimize the choosing of these parameters has not been taken seriously. Two swarm intelligence algorithm were used to optimize these parameters of SVM and the results are good, with the best TPR reaching85.6%and TPR-FPR reaching57.1%. While in the mean time the speed of the algorithm was not affected. Finally, real-time affective state recognition was realized based on the improved SVM. Specific work is as follows:(1) The formulation of reliable affection-evoked experimental projects and affective physiological signals projects. MP150which is made by BIOPAC was used as the main capture hardware to get ECG, GSR and pulse signals. Only four video clips were selected from over100clips to ensure the quality. This paper used record-button, questionnaire and video monitoring to capture the precise affective arouse time. Subject should fill a form to record his/her feeling.(2) The establishment of affective physiological signal database. Over200physically and psychologically healthy freshmen participated in the project. Physiological signals of happy, grief, fear, anger were collected. Calm played as the contrast feeling. The affective-evoked time was precisely locked by the affection button and data was saved in database.(3) The design of two new SVM algorithms which were improved by swarm intelligence algorithm. Two swarm intelligence algorithms (IPSO improved by simulated annealing algorithm and ACO improved by simulated annealing algorithm) were introduced to improve SVM and two new SVM algorithms were made.(4) Select the better SVM by comparison. The two new SVM were tested and compared. Original SVM and SA-SVM were used as contrast. The better SVM was found out based on the identification result.(5) The establishment of real-time affective recognition system. A real-time affective recognition system which work with MP150was developed based on the best SVM and affection adjusting interface was reserved. The identification effect of the system is good with average TPR reaching78.9%and average TPR-FPR reaching49.4%. The foundation of real-time affective adjusting was established by this system.The experimental results demonstrate that:(1) The identification effect of SVM can be obviously improved by swarm intelligence algorithm;(2) SVM improved by particle swarm optimization algorithm has better identification effect;(3) The effect of recognition of human effective depend on time scale and the best span of time window is5seconds.