EduCoder_Study_RS/matching/subject/multi_recall_combine.py

from tqdm import tqdm
import os
import sys
sys.path.append(os.getcwd())
import warnings
import pickle
from config import logger
from config import subject_item_embedding_recall_dict
from config import subject_itemcf_recall_dict
from config import subject_dssm_recall_dict
from config import subject_dssm_usercf_recall_dict
from config import subject_fm_recall_dict
from config import subject_mind_recall0_dict
from config import subject_mind_recall1_dict
from config import subject_pinsage_recall_dict
from config import subject_youtubednn_recall_dict
from config import subject_youtubednn_usercf_recall_dict
from config import subject_final_recall_items_dict

tqdm.pandas()
warnings.filterwarnings('ignore')


def combine_recall_results(user_multi_recall_dict, weight_dict=None, topk=100, save_results=True):
    """
    多路召回合并，将所有的召回策略得到的物品列表合并起来
    """
    final_recall_items_dict = {}
    
    # 对每一种召回结果按照用户进行归一化，方便后面多种召回结果，相同用户的物品之间权重相加
    def norm_user_recall_items_sim(sorted_item_list):

        if len(sorted_item_list) < 2:
            return sorted_item_list
        
        min_sim = sorted_item_list[-1][1]
        max_sim = sorted_item_list[0][1]
        
        norm_sorted_item_list = []
        for item, score in sorted_item_list:
            if max_sim > 0:
                norm_score = 1.0 * (score - min_sim) / (max_sim - min_sim) if max_sim > min_sim else 1.0
            else:
                norm_score = 0.0
            norm_sorted_item_list.append((item, norm_score))
            
        return norm_sorted_item_list
    
    logger.info('开始多路召回合并...')

    for method, user_recall_items in tqdm(user_multi_recall_dict.items()):

        # 在计算最终召回结果的时候，也可以为每一种召回结果设置一个权重
        if weight_dict == None:
            recall_method_weight = 1
        else:
            recall_method_weight = weight_dict[method]
        
        for user_id, sorted_item_list in user_recall_items.items(): 
            # 进行归一化
            user_recall_items[user_id] = norm_user_recall_items_sim(sorted_item_list)
        
        for user_id, sorted_item_list in user_recall_items.items():
            final_recall_items_dict.setdefault(user_id, {})
            
            for item, score in sorted_item_list:
                final_recall_items_dict[user_id].setdefault(item, 0)
                final_recall_items_dict[user_id][item] += recall_method_weight * score  
    
    final_recall_items_dict_rank = {}

    # 多路召回时可以控制最终的召回数量
    for user, recall_item_dict in final_recall_items_dict.items():
        final_recall_items_dict_rank[user] = sorted(recall_item_dict.items(), 
            key = lambda x: x[1], reverse = True)[:topk]

    if save_results: 
        logger.info('保存多路召回合并的结果')
        pickle.dump(final_recall_items_dict_rank, open(subject_final_recall_items_dict, 'wb'))

    return final_recall_items_dict_rank


if __name__ == '__main__':    

    # 定义多路召回的字典，将多路召回的结果都保存在这个字典当中
    user_multi_recall_dict =  {'itemcf_recall': {},
                               'item_embedding_recall': {},
                               'dssm_recall':{},
                               'dssm_usercf_recall':{},
                               'fm_recall':{},
                               'mind_recall0':{},
                               'mind_recall1':{},
                               'pinsage_recall':{},
                               'youtubednn_recall': {},
                               'youtubednn_usercf_recall': {}
                               }                                         

    logger.info('加载itemcf召回结果')
    user_multi_recall_dict['itemcf_recall'] = pickle.load(open(subject_itemcf_recall_dict, 'rb'))

    logger.info('加载item embedding召回结果')        
    user_multi_recall_dict['item_embedding_recall'] = pickle.load(open(subject_item_embedding_recall_dict, 'rb'))

    logger.info('加载dssm召回结果')
    user_multi_recall_dict['dssm_recall'] = pickle.load(open(subject_dssm_recall_dict, 'rb'))

    logger.info('加载dssm usercf召回结果')         
    user_multi_recall_dict['dssm_usercf_recall'] = pickle.load(open(subject_dssm_usercf_recall_dict, 'rb'))    
    
    logger.info('加载fm召回结果')
    user_multi_recall_dict['fm_recall'] = pickle.load(open(subject_fm_recall_dict, 'rb'))

    logger.info('加载mind 用户兴趣1召回结果')        
    user_multi_recall_dict['mind_recall0'] = pickle.load(open(subject_mind_recall0_dict, 'rb'))

    logger.info('加载mind 用户兴趣1召回结果')
    user_multi_recall_dict['mind_recall1'] = pickle.load(open(subject_mind_recall1_dict, 'rb'))

    logger.info('加载pinsage召回结果')         
    user_multi_recall_dict['youtubednn_usercf_recall'] = pickle.load(open(subject_pinsage_recall_dict, 'rb'))    

    logger.info('加载YoutubeDNN召回结果')
    user_multi_recall_dict['youtubednn_recall'] = pickle.load(open(subject_youtubednn_recall_dict, 'rb'))

    logger.info('加载youtube usercf召回结果')         
    user_multi_recall_dict['youtubednn_usercf_recall'] = pickle.load(open(subject_youtubednn_usercf_recall_dict, 'rb'))    

    # 每路召回的权重给不同的值，根据前面召回的情况调整参数的值
    weight_dict = {'itemcf_recall': 0.7,
                   'item_embedding_recall': 1.5,
                   'dssm_recall':2.8,
                   'dssm_usercf_recall':3.9,
                   'fm_recall':2.1,
                   'mind_recall0':1.9,
                   'mind_recall1':1.9,
                   'pinsage_recall':8,
                   'youtubednn_recall': 2.6,
                   'youtubednn_usercf_recall': 3.9
                   }    

    # 最终合并之后每个用户召回300个物品进行排序
    final_recall_items_dict_rank = combine_recall_results(user_multi_recall_dict, weight_dict, topk=300)