WSPI: a wavelet-structured popularity index for stable content ranking via trend–noise separation

Abstract Data popularity estimation ranks items by their share of demand to guide caching. Scores built from raw counts in a short window move with transient spikes; a long window is stable but reacts late and is less accurate. Our main novelty is to score popularity from the trend that a wavelet transform extracts from each item’s demand, not from its raw counts. DWT+AF and DTCWT+AF weight this trend by recency, with the discrete and dual-tree complex wavelet transforms; WSPI multiplies the DTCWT trend by a bounded factor from the spread of energy over the scales. We compare these three training-free methods with six conventional baselines on YouTube, NYC Yellow Taxi (three granularities) and MovieLens (two), over 131,443 windows, with block-bootstrap confidence intervals. With the baselines at their default 7-slot window, a wavelet-based method has the highest ranking stability (RSI@10) in all six scenarios, and WSPI has 2.0–2.8 times lower rank displacement under injected spikes than the best conventional baseline in each scenario, excluding PFRF, whose NDCG@10 is below 0.6. With an equal 64-slot window, a moving average is more stable and WSPI more accurate. WSPI offers an accuracy–stability trade-off and reacts later than short-window baselines to genuine surges.