Anomaly detection is crucial in cybersecurity logs; nevertheless, system logs' extensive size and complexity render manual analysis impractical. Traditional supervised methods necessitate extensive labelled datasets, whereas unsupervised methods lack robustness. To address these difficulties, we proposed a novel semi-supervised framework for anomaly detection in cybersecurity logs. It employs a hybrid feature representation, deep learning, traditional models, and ensemble techniques. The framework has many critical layers: hybrid feature representation TF-IDF (sparse feature), SBERT (semantic feature), and statistical features. Anomaly detection employs an Auto-Encoder, a Bi-LSTM module, and two traditional models: an isolation forest and a one-class support vector machine. The outputs of these models are integrated using a two-layer approach: weighted averaging (soft voting) and stacking via a random forest optimizer. Experimental findings on the HDFS dataset demonstrate that this hybrid semi-supervised approach enhances detection accuracy, scalability, and robustness, offering an efficient method for enhancing cybersecurity via log-based anomaly detection.