Study On Multi-sensor Data Fusion Based On The Hierarchical Interaction Lasso Model

In the process of multi-sensor data fusion, the information from different sensors arecombined into integration. It emphasized on specific procedures that the data transfer andto optimize the results of the information.The purpose of multi-sensor data fusion is toobtain a high quality and useful information. Finally, get an accurate describe.Firstly, based on the interaction theory, we introduced the idea of feature interactiveto regression model for feature extraction and classification.And applies on four UCI datasets and a simulation experiment data. The experiments show that widespread interactiveinformation is an important contribution to data classification,then it should be used.Secondly, we put forwared strong hierarchy and weak hierarchy. And use thesemethods in main features and interactive features extraction. Then add lasso penalty to thismodel to achieve the model coefficients reduction and model efficiency. Then we useconvex relaxation to optimize the cost function, and obtain the optimal solution. KKTconditions and augmented Lagrangian function solved the optimization. The difference isthat we use generalized gradient descent in weak hierarchical, but in a strong hierarchicalmodel, alternating direction multiplier method is applicated.Experiment on multi-sensordatasets prove feature interactions are usefull.Hierarchical constraints decrease the amountof data within the model and greatly reduces the loss of time.Finally, the authors propose an interactive hierarchical logistic regression model. Wedefine the interactive model and hierarchical constraints, and then gives the convexrelaxation condition and use the steepest descent method for solving the model. Finally,three experiments based on multi-sensor data collection activity recognition, multi-sensordata EEG data sets and hepatitis were done. The results show hierarchical lasso methodhave obvious advantages in data processing when data interactive well, and better than thelasso method and all variables lasso method. And the most efficient model in my paper islogistic weak hierarchical model. And strong hierarchical model is the most stable.