Similarity search optimization using recently-biased symbolic representation

Dimension reduction is one of the important requirements for a successful representation to improve the efficiency of extracting the attracting trend patterns on the time series. Furthermore, an efficient and accurate similarity searching on a huge time series data set is a crucial problem in data mining preprocessing. Symbolic representations have proven to be a very effective way to reduce the dimensionality of time series without loss of knowledge. However, symbolic representations suffer from another challenges promoted by the possibility of losing some principal patterns due to the impractical utilization of dealing with the whole data with the same weight. The methodology utilized in this paper is proposed to overcome symbolic representation pattern mismatch. Moreover, the data dimensionality is reduced by keeping more detail on recent-pattern data and less detail on older ones using modified sliding window controlled by the corresponding classification error rate. Experimental results were made on the UCR standard dataset comparing with the state of the art techniques. The proposed techniques showed promising results. Furthermore, practical experiments were made on the Egyptian stock market indices EGX 30, EGX 70 and EGX 100. The discovered patterns showed the accuracy and effectiveness of the proposed approach.