<p>Time-series classification is important in various domains and tasks, and has attracted meaningful research over the past decades. Recent years have brought meaningful advancements in using deep neural networks (DNNs) architectures to classify time-series. However, challenges due to measurement errors, missing values, and irregular sampling are still a concern. To minimize their impact on the classification performance, various standardization methods are commonly applied to the raw continuous data as a preprocessing stage to the DNN. Instead, we suggest employing temporal abstraction, wherein the raw time-series is converted into a symbolic representation of time points. The transformed data can then be utilized as input for the DNNs. Specifically, we explore the impact of combining temporal abstraction with convolution-based sequence models and recurrent neural networks. To assess the effectiveness of these methods, we conducted evaluations on a total of 128 univariate and 13 multivariate time-series datasets. Through our framework, which incorporates the temporal abstraction process, we significantly enhanced the performance of various state-of-the-art DNNs used for time-series classification tasks. Our evaluation shows that our methods are significantly superior in classification prediction across all seven evaluation metrics for univariate time-series datasets, outperforming in terms of AUC-ROC for multivariate time-series datasets, compared to predictions using standardized raw data.</p>