RSSI-based localization technique in Wireless Sensor Network is aimed at building a mapping between signal and physical spaces. The mapping could be overfitting when the number of paired RSSI and location data is small, and the collection of paired data is difficult, so unpaired data could be useful in improving the performance. This paper proposes the Locality Preserving Semi-Supervised Canonical Correlation Analysis (LPSemiCCA) algorithm for localization in Wireless Sensor Network, which combines PCA and CCA smoothly using a tradeoff parameter to overcome problems like sensitivity to data scale of PCA and incapability of utilizing unpaired data of CCA. The algorithm introduces similarity matrices of paired data and whole data to fit the structure of network and employs unpaired data efficiently. Locality Preserving Projection is also applied to construct the objective function in each domain, so the mapping can be calculated in condition of preserving the inner local structure of data. Experimental results in both simulated and realistic data show a higher localization accuracy of the proposed algorithm compared with LapLS, PPLCA and LapSVR.