In plain words
SPAR is a recommendation framework that chooses an item by predicting a sequence of item codes. As each choice narrows the options, it looks up the parts of a person's past activity that fit that point in the sequence. It also arranges the codes to reduce the dominance of branches with many past interactions.
In offline tests on Youku data, SPAR found more next items in its recommendations on average than the strongest tested alternative, OneRec. A separate seven-day Youku homepage test reported higher play conversion and watch time with SPAR than with the production Base model.
The research
When a recommendation system builds an item choice one step at a time, what should it do when useful clues about a person change as the choices narrow, and popular paths already carry much more past traffic?
A fixed summary may miss evidence needed for a later choice, while popularity can crowd out less frequent but relevant targets.
SPAR pairs history retrieval tailored to each generated item-code prefix with a method that limits past-traffic imbalance among sibling choices.
Researchers evaluated SPAR on Youku production data by measuring Hit Ratio, the fraction of evaluated next items found in the top K recommendations. Across K values of 10, 20, 50, and 100, SPAR's result was 2.66% higher on average than OneRec, the strongest reported baseline. This supports improved offline retrieval in that setting.
A seven-day Youku homepage test also reported higher play conversion and watch-time measures for SPAR than for the production Base model. The offline comparison and online test provide evidence in their respective settings.
The retrieval comparison is offline and uses Youku production data; the online engagement results come from a seven-day test of Youku's homepage. The evidence does not establish how SPAR performs on other platforms or whether RQ-PIB balances future item exposure.
Analysis
Deep
- Original paper authors
- Chunrui Zhang, Gaoyun Lin, Chang Zhou, Zheng Hu, Qiran Song
- Submitted to arXiv
- License
- arXiv non-exclusive