Active learning (AL) is a popular machine learning algorithm designed to iteratively query and label informative data points from an unlabeled dataset, reducing the need for extensive labeled data. While AL improves labeling efficiency, a major challenge lies in accurately assessing the true value of unlabeled data. In this study, we propose a pioneering AL approach employing Conditional Variational Deep Embedding (CVaDE), aimed at discerning the most informative samples for labeling from both implicit and explicit perspectives. First, CVaDE implicitly learns a latent data representation by co-training a task learner and a discriminator, enabling the estimation of an uncertainty score for each sample. Second, it explicitly performs clustering on the latent representations to compute a representativeness score for each sample. Additionally, we introduce a diversity score that quantifies the average dissimilarity of each sample to the candidate set, aiding in identifying the most valuable samples. Through the combination of uncertainty, representativeness, and diversity scores, CVaDE selects the most informative samples for labeling. Our empirical evaluation on multiple benchmark datasets demonstrates notable performance improvements compared to existing baselines, highlighting the efficacy of our proposed approach.