Wifi based indoor localization using grossly labeled data for smartphone users

Abstract

Localization in indoor areas deals with the serious problem of collecting data over a broad experimental region while maintaining the location points. Sensor values may be collected during normal movements and work schedules of the people, but those data will be either unlabelled or grossly labeled. This paper addresses the challenge of providing a localization solution for unlabelled or grossly labeled indoor data by a two-phased semi-supervised learning approach. In the first phase, a Rank-Based Iterative Clustering method (RICM) method is proposed that processes the entire dataset iteratively, generating a final cluster at the end of each iteration. The experiments were conducted in a realistic indoor localization dataset. In the first phase, distinct temporary sets of clusters are produced by each clustering algorithm and their performances are evaluated by computing different clustering scores based on the respective temporary set of clusters obtained. Finally, the algorithms are sorted according to their rank. At each iteration, an inner join is performed among each possible pair of clusters obtained from those rank-wise sorted algorithms. Finally, an improved set of clusters is received and the cluster containing maximum data samples is kept in the final cluster set. These samples are removed from the primary dataset and in subsequent iterations, the remaining data samples are re-clustered. This way, a new final cluster is obtained at each iteration and the procedure is repeated until all the final clusters are obtained. In the second phase, classification is performed using random test data and the obtained set of clusters as training data and 97% accuracy is obtained for different supervised classification algorithms. External and internal validation scores were utilized to evaluate the clustering techniques.