제1회 국민대학교 AI빅데이터 분석 경진대회

알고리즘 | 정형 | 추천시스템 | Recall

  • moneyIcon Prize : 400 만원
  • 1,171명 마감

 

Recall@5 평가산식 코드

2023.10.16 09:45 2,334 Views
import pandas as pd
import numpy as np


def recall5(answer_df, submission_df):
    """
    Calculate recall@5 for given dataframes.
    
    Parameters:
    - answer_df: DataFrame containing the ground truth
    - submission_df: DataFrame containing the predictions
    
    Returns:
    - recall: Recall@5 value
    """
    
    primary_col = answer_df.columns[0]
    secondary_col = answer_df.columns[1]
    
    # Check if each primary_col entry has exactly 5 secondary_col predictions
    prediction_counts = submission_df.groupby(primary_col).size()
    if not all(prediction_counts == 5):
        raise ValueError(f"Each {primary_col} should have exactly 5 {secondary_col} predictions.")


    # Check for NULL values in the predicted secondary_col
    if submission_df[secondary_col].isnull().any():
        raise ValueError(f"Predicted {secondary_col} contains NULL values.")
    
    # Check for duplicates in the predicted secondary_col for each primary_col
    duplicated_preds = submission_df.groupby(primary_col).apply(lambda x: x[secondary_col].duplicated().any())
    if duplicated_preds.any():
        raise ValueError(f"Predicted {secondary_col} contains duplicates for some {primary_col}.")


    # Filter the submission dataframe based on the primary_col present in the answer dataframe
    submission_df = submission_df[submission_df[primary_col].isin(answer_df[primary_col])]
    
    # For each primary_col, get the top 5 predicted secondary_col values
    top_5_preds = submission_df.groupby(primary_col).apply(lambda x: x[secondary_col].head(5).tolist()).to_dict()
    
    # Convert the answer_df to a dictionary for easier lookup
    true_dict = answer_df.groupby(primary_col).apply(lambda x: x[secondary_col].tolist()).to_dict()
    
    
    individual_recalls = []
    for key, val in true_dict.items():
        if key in top_5_preds:
            correct_matches = len(set(true_dict[key]) & set(top_5_preds[key]))
            individual_recall = correct_matches / min(len(val), 5) # 공정한 평가를 가능하게 위하여 분모(k)를 'min(len(val), 5)' 로 설정함 
            individual_recalls.append(individual_recall)


    recall = np.mean(individual_recalls)
    
    return recall
로그인이 필요합니다
0 / 1000
jazzy
2023.12.03 20:39

평가지표로 Recall@K 를 선정한 이유를 알 수 있을까요?