Recommending Items in Social Tagging Systems Using Tag and Time Information

Social tagging systems are noisy, ever-changing places. Making recommendations in these environments is deceptively hard—collaborative filtering alone can only get you so far, and just counting tags isn’t enough. This paper grew out of the frustration with standard item recommenders ignoring how people actually use tags and when. Human memory fades, but not linearly or exponentially—a detail most algorithms miss.

Drawing on human memory theory, specifically the Base-Level Learning (BLL) equation, we thought: what if an item’s recommendation value reflected not just what tags a user assigned, but how often and how recently? That led to CIRTT, a two-step system. First, find candidate items via user-based CF like everyone else. Second, re-rank them by combining item similarity with this BLL-forgetting curve over tag usage. It’s a more psychologically plausible model of what users might actually want, not just what they did.

We tested CIRTT—open sourced on GitHub—against established baselines and newer tag/time-aware methods on BibSonomy, CiteULike, and MovieLens. It repeatedly emerged at the top for nDCG, MAP, and recall. Especially telling: the power-law decay from BLL beat out linear or exponential recency models. The system’s better accuracy wasn’t a subtle statistical blip; across three very different datasets, the effect was clear. There’s a hint here that real-world forgetting has something to offer recommender models that rigid math can’t capture.

What did we learn? Item recommenders that take human memory seriously—actually modeling how recent, frequent interaction shapes recall—can beat those that just see users as numbers in a matrix. Next stop: using memory-based recency in other panel elements, maybe even the user similarity calculation. The real world is messy, but a little cognitive science helps recommender systems get messy in the right ways.

@inproceedings{lacic2014recommending,
  title = {Recommending items in social tagging systems using tag and time information},
  author = {Laci\'{c}, E. and Kowald, D. and Seitlinger, P. and Trattner, C. and Parra, D.},
  year = {2014},
  booktitle = {Proceedings of the 1st International Workshop on Social Personalisation at the 25th ACM Conference on Hypertext and Social Media},
  publisher = {CEUR-WS},
}