EduCoder_Study_RS/shixun_partner_recommend.py

import numpy as np
import faiss
import pickle
from datetime import datetime
from utils import merge_dict
from config import logger
from config import embedding_dim
from config import shixun_user_w2v_emb_dict
from config import shixuns_user_word2vec_faiss_model_path
from config import shixun_dssm_user_emb_dict
from config import shixun_dssm_user_faiss_model_path
from config import shixun_dssm_user_embedding_index_dict
from config import shixun_user_embedding_index_dict
from config import shixun_user_item_time_dict_data
from config import test_user_id


logger.info('加载学习伙伴召回模型')
user_word2vec_faiss_model = faiss.read_index(shixuns_user_word2vec_faiss_model_path)
dssm_user_faiss_model = faiss.read_index(shixun_dssm_user_faiss_model_path)

logger.info('加载学习伙伴召回字典')
user_w2v_emb_dict = pickle.load(open(shixun_user_w2v_emb_dict, 'rb'))
user_embedding_index_dict = pickle.load(open(shixun_user_embedding_index_dict, 'rb'))
dssm_user_emb_dict = pickle.load(open(shixun_dssm_user_emb_dict, 'rb'))
dssm_user_embedding_index_dict = pickle.load(open(shixun_dssm_user_embedding_index_dict, 'rb'))
user_item_time_dict = pickle.load(open(shixun_user_item_time_dict_data, 'rb'))


def shixun_partner_recommend(user_id, topk=20, method='1'):
    """
    先通过dssm user embedding进行推荐
    不够topk个再通过用户embedding进行推荐
    """
    start_time = datetime.now()

    logger.info(f"本次需要进行学习伙伴推荐的用户ID: {user_id}")

    recommend_results = {}
    recommend_results.clear()      

    recommend_results_dssm = partner_recommend_by_dssm_user_embedding(user_id, topk, False)

    if len(recommend_results_dssm) < topk:
        recommend_results_w2v_embedding = partner_recommend_by_user_w2v_embedding(
            user_id, topk - len(recommend_results_dssm), False)
        recommend_results = merge_dict(recommend_results_dssm, recommend_results_w2v_embedding)
    else:
        recommend_results = recommend_results_dssm   

    if method == '1':
        recommend_results = [key for key, value in recommend_results.items()] 

    # 计算耗时毫秒
    end_time = datetime.utcnow()
    cost_time_millisecond = round(float((end_time - start_time).microseconds / 1000.0), 3)    

    logger.info(f"本次推荐总耗时: {cost_time_millisecond} 毫秒")
    return recommend_results   


def partner_recommend_by_user_w2v_embedding(user_id, topk=20, verbose=True):
    """
    通过word2vec user embedding的相似度推荐学习伙伴
    """
    start_time = datetime.now()
    if verbose:
        logger.info(f"本次需要进行学习伙伴推荐的用户ID: {user_id}")

    recommend_results = {}
    recommend_results.clear()    

    if (user_id not in user_w2v_emb_dict) or (user_id not in user_item_time_dict):
        return recommend_results

    # 取出用户向量
    user_embs = user_w2v_emb_dict[user_id]

    # reshape为二维
    user_embs = user_embs.reshape(-1, embedding_dim)

    # 找topk个相似的向量
    D, I = user_word2vec_faiss_model.search(np.ascontiguousarray(user_embs), topk)

    top_k_index = list(I.ravel())
    top_k_user = {}
    top_k_user.clear()

    # 还原相似向量索引对应的用户
    for index in top_k_index:        
        user_id, item_id_list = user_embedding_index_dict[index][0]
        top_k_user[user_id] = [int(item_id) for item_id in item_id_list]

    # 计算耗时毫秒
    end_time = datetime.utcnow()
    cost_time_millisecond = round(float((end_time - start_time).microseconds / 1000.0), 3) 

    if verbose:
        logger.info(f"本次推荐总耗时: {cost_time_millisecond} 毫秒")
    return top_k_user


def partner_recommend_by_dssm_user_embedding(user_id, topk=20, verbose=True):
    """
    通过dssm user embedding的相似度推荐学习伙伴
    """
    start_time = datetime.now()
    if verbose:
        logger.info(f"本次需要进行学习伙伴推荐的用户ID: {user_id}")

    recommend_results = {}
    recommend_results.clear()    

    if (user_id not in dssm_user_emb_dict) or (user_id not in user_item_time_dict):
        return recommend_results

     # 取出用户向量
    user_embs = dssm_user_emb_dict[user_id]

    # reshape为二维
    user_embs = user_embs.reshape(-1, embedding_dim)

    # 找topk个相似的向量
    D, I = dssm_user_faiss_model.search(np.ascontiguousarray(user_embs), topk + 1)

    top_k_index = list(I.ravel())
    top_k_user = {}
    top_k_user.clear()

    # 还原相似向量索引对应的用户
    for index in top_k_index:        
        cur_user_id = dssm_user_embedding_index_dict[index]

        # 过滤第一个最相似的用户是自己
        if cur_user_id != user_id:
            item_id_list = [int(item_id) for item_id, _ in user_item_time_dict[cur_user_id]]
            top_k_user[cur_user_id] = item_id_list

   # 计算耗时毫秒
    end_time = datetime.utcnow()
    cost_time_millisecond = round(float((end_time - start_time).microseconds / 1000.0), 3)    

    if verbose:
        logger.info(f"本次推荐总耗时: {cost_time_millisecond} 毫秒")
    return top_k_user        


if __name__ == '__main__':
    user_id = test_user_id
    item_id_list = [int(item_id) for item_id, time in user_item_time_dict[user_id]]

    print(str(user_id) + ':')
    print(item_id_list)
    
    print('*' * 100)
    
    recommend_results = shixun_partner_recommend(user_id, topk=20, method='0')    
    for sim_user_id, sim_item_id_list in recommend_results.items():
        print(str(sim_user_id) + ':')
        print(sim_item_id_list)