With the advent of large-scale real-world anomaly datasets having annotations at the video-level; weakly supervised video anomaly detection (W-VAD) techniques took centre stage in VAD research. Recent W-VAD approaches broadly rely on MIL (Multiple Instance learning) and self-training. MIL typically organises video snippets from a video footage in positive and negative bags and trains simultaneously on both to effectively learn discriminative features between normal and anomalous snippets. However, existing approaches struggle in the presence of hard anomalies: where the difference between normal and anomalous events is very subtle. E.g. transition from normal to anomalous action or vice versa. To this end, we propose WAVE-NET trained on I3D features from both normal and abnormal snippets simultaneously. I3D features are further refined by a transformer-based autoencoder. Instead of static k-top ranking samples we have adopted a variable sampling strategy that considers event (snippet) density within a video footage. The proposed WAVE-NET is evaluated on two prominent large-scale datasets namely: XD-Violence and UCF-Crime. Experimental results establish the superiority of WAVE-NET on both datasets by a significant margin.