Proceedings of the Second ACM International Conference on Web Search and Data Mining
Incorporating post-click behaviors into a click model
Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval
LETOR: A benchmark collection for research on learning to rank for information retrieval
Information Retrieval
Position-normalized click prediction in search advertising
Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining
| Hi-index | 0.00 |
Click-through rate (CTR) prediction and relevance ranking are two fundamental problems in web advertising. In this study, we address the problem of modeling the relationship between CTR and relevance for sponsored search. We used normalized relevance scores comparable across all queries to represent relevance when modeling with CTR, instead of directly using human judgment labels or relevance scores valid only within same query. We classified clicks by identifying their relevance quality using dwell time and session information, and compared all clicks versus selective clicks effects when modeling relevance. Our results showed that the cleaned click signal outperforms raw click signal and others we explored, in terms of relevance score fitting. The cleaned clicks include clicks with dwell time greater than 5 seconds and last clicks in session. Besides traditional thoughts that there is no linear relation between click and relevance, we showed that the cleaned click based CTR can be fitted well with the normalized relevance scores using a quadratic regression model. This relevance-click model could help to train ranking models using processed click feedback to complement expensive human editorial relevance labels, or better leverage relevance signals in CTR prediction.