Skip to content

Documentation for Confidence Score Functions

Method: dummy_confidences

This function returns a tensor of ones, which serves as a baseline confidence score for events. It uses the provided ranks and event flags to output a uniform confidence value.

Parameters

  • ranks: A FloatTensor containing search result ranks.
  • events: A Tensor indicating event presence; use 1 for event and 0 for non-event.

Usage

  • Purpose: To generate a constant confidence score for each event, useful for preliminary testing or baseline comparisons.

Example

import torch

ranks = torch.tensor([1.0, 2.0, 3.0])
events = torch.tensor([1, 0, 1])

# Returns tensor([1., 1., 1.])
confidence = dummy_confidences(ranks, events)

Method: calculate_confidences

This function computes confidence scores for events (clicks) and non-events (non-clicks) based on ranking similarity and local context. It uses a sliding window to derive an average rank and click proportion, then combines these to produce a normalized score.

Parameters

  • ranks: FloatTensor, list of ranks from search results.
  • results: Tensor, list of 0 (non-event) or 1 (event).
  • window_size: int, size of the context window (must be > 1, default: 3).

Usage

  • Purpose: Generate confidence scores from search results by analyzing local ranking and click patterns.

Example

import torch

ranks = torch.tensor([1, 2, 3])
results = torch.tensor([0, 1, 0])
conf_scores = calculate_confidences(ranks, results, window_size=3)
print(conf_scores)