TiSASRec

TiSASRec starts from a simple observation about SASRec: treating a user's history as an evenly spaced sequence throws away real information. Someone who buys coffee every morning is a different pattern from someone who buys coffee once a year, even if the order of items looks identical. Published at WSDM 2020 by Jiacheng Li, Yujie Wang, and Julian McAuley, it extends SASRec's self-attention with time-interval-aware positional encodings — the model sees both the absolute position of each item and the actual elapsed time between consecutive interactions.

The paper's experiments showed the time-aware variants beating plain positional encodings and outperforming state-of-the-art sequential models on both sparse and dense datasets. It is a surgical improvement rather than a new paradigm, which is exactly why practitioners like it: drop it in wherever you were using SASRec and get better handling of irregular session gaps. The authors released TensorFlow code and the community ported it to PyTorch.

It remains ID-based, so new items with no history are still a problem — but for logged-in users with bursty, irregular behavior, the time signal is one of the cheapest wins in sequential recsys.

• Paper-reported (WSDM 2020, pp. 322–330): TiSASRec outperformed various state-of-the-art sequential models on both sparse and dense datasets across different evaluation metrics; ablations showed time-interval encodings beating absolute-only positional encodings
• No independent benchmark body re-scores it; reproduced in community ports