Embedding Vector Placement Probability for Position Bias Estimation

The patent introduces embedding vectors as an intermediary representation between items and their placement positions. Instead of directly observing item-position assignments, the system converts items into embedding vectors and learns the distribution of these vectors across potential positions. This intermediary representation allows the system to account for fixed placement patterns while still estimating position bias accurately, resolving the contradiction between simple fixed placement and accurate bias estimation.

The patent transforms the problem from directly analyzing item placement positions to analyzing the distribution of embedding vectors across positions. By changing the parameter space from discrete item IDs to continuous embedding vector distributions, the system can model fixed placement patterns while estimating position bias. This parameter transformation enables accurate bias estimation even when items are consistently placed at fixed positions.

The patent introduces embedding vectors as an intermediary layer between click data and ranking construction. Instead of directly using click counts at fixed positions, the system first converts items to embedding vectors, then models the distribution of these vectors across potential positions to estimate position bias. This intermediary step allows the system to correct for position bias while still efficiently constructing rankings from click data.

The patent replaces the mechanical approach of directly counting clicks at fixed positions with a probabilistic model based on embedding vector distributions. Instead of simple click aggregation, the system uses learned distributions to estimate what clicks would look like without position bias, then adjusts the ranking accordingly. This substitution enables accurate ranking construction that accounts for position bias while maintaining computational efficiency.