diff --git a/moiveRe/DeepFMcopy.py b/moiveRe/DeepFMcopy.py new file mode 100644 index 0000000..acf9e8d --- /dev/null +++ b/moiveRe/DeepFMcopy.py @@ -0,0 +1,77 @@ +import django +import os +import torch + +os.environ.setdefault("DJANGO_SETTINGS_MODULE", "moiveRe.settings") +django.setup() + + + + + + + + + +import numpy as np +import pandas as pd +from sklearn.model_selection import train_test_split +from sklearn.preprocessing import LabelEncoder, MinMaxScaler +import torch +from deepctr_torch.models import DeepFM +from deepctr_torch.inputs import SparseFeat, DenseFeat, get_feature_names + +# 生成示例数据 +np.random.seed(2024) +data_size = 10000 + +# 假设有 26 个离散特征 C1, C2, ..., C26 +sparse_features = ['C' + str(i) for i in range(1,4)] +# 假设有 13 个连续特征 I1, I2, ..., I13 +dense_features = ['I' + str(i) for i in range(1,3)] + +data = {feat: np.random.choice([0, 1, 2, 3, 4], data_size) for feat in sparse_features} +data.update({feat: np.random.rand(data_size) for feat in dense_features}) +data['label'] = np.random.randint(0, 2, data_size) + +# 转换为 DataFrame +df = pd.DataFrame(data) +print(df) + + +#预处理部分: +# 对离散特征进行标签编码,转为数值的意思 +for feat in sparse_features: + lbe = LabelEncoder() + df[feat] = lbe.fit_transform(df[feat]) +# 对连续特征进行归一化 +mms = MinMaxScaler() +df[dense_features] = mms.fit_transform(df[dense_features]) + +train, test = train_test_split(df, test_size=0.2) + +# deepctr模型输入 +fixlen_feature_columns = [SparseFeat(feat, df[feat].nunique(), embedding_dim=4) for feat in sparse_features] + [DenseFeat(feat, 1,) for feat in dense_features] + + + + +# DeepFM 模型，开导！ +model = DeepFM(fixlen_feature_columns, fixlen_feature_columns, task='binary') # 重复传入相同的列~ 这里因为是根据CTR来对召回进行排序所以就是binary任务 + +# 编译模型 +model.compile("adam", "binary_crossentropy", + metrics=["binary_crossentropy"], ) #优化器选择balabala，这里直接看库里给的文档 + +# 训练模型 +train_model_input = {name: train[name].values for name in get_feature_names(fixlen_feature_columns)} +test_model_input = {name: test[name].values for name in get_feature_names(fixlen_feature_columns)} +label_distribution = train['label'].value_counts() +print('label_dis:',label_distribution) +history = model.fit(train_model_input, train['label'].values, + batch_size=50, epochs=4, verbose=2, validation_split=0, ) #fit开始train + + + + + diff --git a/moiveRe/DeepFm.py b/moiveRe/DeepFm.py index f89a2ae..e69de29 100644 --- a/moiveRe/DeepFm.py +++ b/moiveRe/DeepFm.py @@ -1,50 +0,0 @@ -import numpy as np -import pandas as pd -import torch -import torch.nn as nn -import torch.optim as optim -from sklearn.model_selection import train_test_split -from deepctr.models import DeepFM -from deepctr.feature_column import SparseFeat, get_feature_names - -# 假设有一个包含广告特征和点击情况的数据集 -data = { - 'user_id': [1, 2, 3, 4, 5], - 'ad_id': [101, 102, 103, 104, 105], - 'age': [25, 30, 35, 40, 45], - 'gender': ['M', 'F', 'M', 'F', 'M'], - 'click': [1, 0, 1, 0, 1] -} -df = pd.DataFrame(data) - -# 定义特征列 -sparse_features = ['user_id', 'ad_id', 'gender'] -dense_features = ['age'] - -# 将特征列转换为SparseFeat和DenseFeat类型 -feature_columns = [SparseFeat(feat, vocabulary_size=df[feat].nunique(), embedding_dim=4) - for feat in sparse_features] + [DenseFeat(feat, 1) for feat in dense_features] - -# 划分训练集和测试集 -train, test = train_test_split(df, test_size=0.2) - -# 创建模型 -model = DeepFM(feature_columns, feature_columns, task='binary') - -# 编译模型 -model.compile("adam", "binary_crossentropy", metrics=['binary_crossentropy']) - -# 将数据转换为模型所需的输入格式 -train_model_input = {name: train[name] for name in sparse_features + dense_features} -test_model_input = {name: test[name] for name in sparse_features + dense_features} -train_target = train['click'].values -test_target = test['click'].values - -# 训练模型 -history = model.fit(train_model_input, train_target, batch_size=256, epochs=10, verbose=2, validation_split=0.2) - -# 预测 -preds = model.predict(test_model_input, batch_size=256) - -# 输出预测结果 -print("预测点击率：", preds) diff --git a/moiveRe/UserExperienceUpdate.py b/moiveRe/UserExperienceUpdate.py new file mode 100644 index 0000000..701fbd7 --- /dev/null +++ b/moiveRe/UserExperienceUpdate.py @@ -0,0 +1,26 @@ + + +import django +from django.conf import settings +import os + +os.environ.setdefault("DJANGO_SETTINGS_MODULE", "moiveRe.settings") # 替换 "yourproject.settings" 为你的项目的 settings 模块路径 +django.setup() + +from moiveReApp.models import UserExperience + + +# 兴趣标签集合 +experience_tags = [ + 'NewtoGame','Middle','Experienced' +] + +# 将标签添加到数据库中 +for tag_name in experience_tags: + tag, created = UserExperience.objects.get_or_create(name=tag_name) + + # 如果标签是新创建的，则打印一条消息 + if created: + print(f'Created tag: {tag_name}') + else: + print(f'Tag already exists: {tag_name}') \ No newline at end of file diff --git a/moiveRe/moiveReApp/models.py b/moiveRe/moiveReApp/models.py index 79c5322..a4dfab5 100644 --- a/moiveRe/moiveReApp/models.py +++ b/moiveRe/moiveReApp/models.py @@ -68,10 +68,19 @@ class Userinterests(models.Model): def __str__(self): return self.name +class UserExperience(models.Model): + name = models.CharField(max_length=100, unique=True) + + def __str__(self): + return self.name + + + class UserProfile(models.Model): # 与用户关联的一对一字段，这里假设你使用了Django内置的User模型 user = models.OneToOneField(User, on_delete=models.CASCADE) + userExperience = models.ManyToManyField(UserExperience, related_name='user_experience') interests = models.ManyToManyField(Userinterests, related_name='interested_users') liked_items = models.ManyToManyField(Question, blank=True, related_name='liked_users') click_items = models.ManyToManyField(Question, blank=True, related_name='clicked_users') diff --git a/moiveRe/moiveReApp/templates/moiveReApp/deepfmrec.html b/moiveRe/moiveReApp/templates/moiveReApp/deepfmrec.html new file mode 100644 index 0000000..048f2d5 --- /dev/null +++ b/moiveRe/moiveReApp/templates/moiveReApp/deepfmrec.html @@ -0,0 +1,229 @@ + + + + + + + Skyrim Mod - 随机Mod推荐 + + + + + +
+

Skyrim Mod

+ {% if user.is_authenticated %} +
Welcome, {{ user.username }}!
+ {% endif %} +
+ +

综合推荐

+ + + + + + + diff --git a/moiveRe/moiveReApp/templates/moiveReApp/index.html b/moiveRe/moiveReApp/templates/moiveReApp/index.html index 4c61a13..b38ad6e 100644 --- a/moiveRe/moiveReApp/templates/moiveReApp/index.html +++ b/moiveRe/moiveReApp/templates/moiveReApp/index.html @@ -187,8 +187,8 @@

Skyrim Mod

Welcome, {{ user.username }}!
{% endif %}
- - UserCF Sort + + 个性❤推荐 diff --git a/moiveRe/moiveReApp/urls.py b/moiveRe/moiveReApp/urls.py index 5c945f5..9f5f56d 100644 --- a/moiveRe/moiveReApp/urls.py +++ b/moiveRe/moiveReApp/urls.py @@ -7,7 +7,7 @@ from django.urls import include, path from django.views.static import serve from .views import question_detail -from .views import register, user_login,random_questions,like_question,usercf,itemcf,bertcall +from .views import register, user_login,random_questions,like_question,usercf,itemcf,bertcall,deepfmrec app_name = "moiveReApp" urlpatterns = [ @@ -18,6 +18,7 @@ path('usercf/', usercf, name='usercf'), path('itemcf/', itemcf, name='itemcf'), path('bertcall/', bertcall, name='bertcall'), + path('deepfmrec/', deepfmrec, name='deepfmrec'), path('like//', like_question, name='like_question'), path('/ratings/', question_detail, name='question_detail'), diff --git a/moiveRe/moiveReApp/views.py b/moiveRe/moiveReApp/views.py index 8449ffd..0c5bfe9 100644 --- a/moiveRe/moiveReApp/views.py +++ b/moiveRe/moiveReApp/views.py @@ -14,6 +14,13 @@ from sklearn.metrics.pairwise import cosine_similarity import math from collections import Counter +import torch +from transformers import DistilBertTokenizer, DistilBertModel +import pandas as pd +from sklearn.preprocessing import LabelEncoder, MinMaxScaler +import torch +from deepctr_torch.models import DeepFM +from deepctr_torch.inputs import SparseFeat, DenseFeat, get_feature_names class IndexView(generic.ListView): template_name = "moiveReApp/index.html" @@ -416,10 +423,10 @@ def itemcf(request): #print(user_similarity_matrix) # 生成推荐列表 recommendations = itemcf_.generate_recommendations(current_user_id, item_similarity_matrix, user_item_matrix, user_dict, item_dict, k=30) - # 获取推荐的电影信息 + # 获取推荐 recommended_movies = Question.objects.filter(id__in=recommendations[:8]) #只展示前ki个 - # 在模板中渲染推荐结果 + return render(request, 'moiveReApp/itemcf.html', {'recommended_movies': recommended_movies}) @@ -449,4 +456,351 @@ def bertcall(request): recommended_movies = Question.objects.filter(id__in=recommendations[:8]) #只展示前ki个 # 在模板中渲染推荐结果 - return render(request, 'moiveReApp/bertcall.html', {'recommended_movies': recommended_movies}) \ No newline at end of file + return render(request, 'moiveReApp/bertcall.html', {'recommended_movies': recommended_movies}) + + + + + + + +#多路召回 + DeepFM 推荐 + + + +class itemcf_deepfm(): + + def build_user_item_matrix(question_list): + """ + 根据物品-用户倒排表构建用户-物品交互矩阵 + + 参数: + - question_list: QuerySet，包含所有问题的列表 + + 返回: + - user_item_matrix: 2D numpy数组，表示用户-物品交互矩阵 + - item_dict: 字典，将物品ID映射到索引 + - user_dict: 字典，将用户ID映射到索引 + """ + + # 获取所有用户的ID和物品的ID + all_users = set() + all_items = set() + + for question in question_list: + item_id = question.id + liked_users = question.liked_by.all() + + all_users.update(user.id for user in liked_users) + all_items.add(item_id) + + # 创建物品-用户的倒排表 + item_user_dict = {} + for question in question_list: + item_id = question.id + liked_users = set(user.id for user in question.liked_by.all()) + item_user_dict[item_id] = liked_users + + # 创建物品相似性矩阵 + num_items = len(all_items) + num_users = len(all_users) + + # 建立物品ID到索引的映射 + item_dict = {item_id: index for index, item_id in enumerate(all_items)} + + # 建立用户ID到索引的映射 + user_dict = {user_id: index for index, user_id in enumerate(all_users)} + + # 初始化用户-物品交互矩阵 + user_item_matrix = np.zeros((num_users, num_items)) + + for item_id, liked_users in item_user_dict.items(): + item_index = item_dict[item_id] + + for user_id in liked_users: + user_index = user_dict[user_id] + user_item_matrix[user_index, item_index] = 1 # 表示用户喜欢该物品 + + return user_item_matrix, item_dict, user_dict #转置一下就是倒排。。 + + def calculate_user_similarity(user_item_matrix): + # 计算物品相似度矩阵 + item_similarity_matrix = cosine_similarity(user_item_matrix.T) #转置才是物品-用户表 + np.fill_diagonal(item_similarity_matrix, 0) + return item_similarity_matrix + + def generate_recommendations(target_user, item_similarity_matrix, user_item_matrix, user_dict, item_dict, k): + + + # 获取目标用户的索引 + target_user_index = user_dict.get(target_user) + if target_user_index is None: + raise ValueError(f"目标用户 '{target_user}' 不存在。") + + # 获取目标用户(未)交互的物品 + target_user_interactions = user_item_matrix[target_user_index] + + # 初始化推荐列表 + recommendations = [] + + + # 遍历每个物品，计算目标用户对未交互物品的感兴趣程度 + for item_id, item_index in item_dict.items(): + if not target_user_interactions[item_index]: + # 目标用户未交互过该物品 + interest_score = 0 + # 获取与目标物品最相似的前K个物品的索引 + similar_item_indices = np.argsort(item_similarity_matrix[item_index])[::-1][:k] + + # 计算目标用户对该物品的感兴趣程度 + for similar_item_index in similar_item_indices: + if user_item_matrix[target_user_index, similar_item_index] == 1: + # 用户u交互过该物品,此即为N(u)∩S（j,k） + similarity_score = item_similarity_matrix[item_index, similar_item_index] + interest_score += similarity_score + + recommendations.append((item_id, interest_score)) + + # 根据感兴趣程度排序推荐列表 + recommendations.sort(key=lambda x: x[1], reverse=True) + + + # 返回排序后的物品ID列表 + return recommendations + +class hotitem(): + + + def recall_hot_item(question_list): + sorted_question_list = sorted(question_list, key=lambda q: q.likes, reverse=True) + recommendations = [(question.id, question.likes) for question in sorted_question_list] + return recommendations + +class newitem(): + @staticmethod + def recall_new_item(question_list): + sorted_question_list = sorted(question_list, key=lambda q: q.pub_date, reverse=True) + recommendations = [(question.id, 1) for question in sorted_question_list] #得分为1 不带偏见的推荐新物品~ + return recommendations + +class U2TAG2I(): + + + + def compute_interest_score(user_profile, questions): + interest_scores = {} + user_interests = set(user_profile.interests.all()) + user_interests_names = [interest.name for interest in user_interests] + #print(user_interests_names) + for question in questions: + if question.category in user_interests_names: + interest_scores[question.id] = 0.3 + else: + interest_scores[question.id] = 0 + + return interest_scores + + def compute_similarity(user_profile, questions): + tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased') + model = DistilBertModel.from_pretrained('distilbert-base-uncased') + device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + model.to(device) + + user_interests_str = ' '.join([interest.name for interest in user_profile.interests.all()]) + + question_details = [question.detail for question in questions] + num_questions = len(question_details) + question_ids = [question.id for question in questions] + question_score = [] + #print(question_ids) + question_details.append(user_interests_str) + texts = question_details + + + + inputs = tokenizer(texts, return_tensors='pt', padding=True, truncation=True) + + # 将张量移动到GPU上 + input_ids = inputs['input_ids'].to(device) + attention_mask = inputs['attention_mask'].to(device) + # 使用DistilBERT模型进行推理 + with torch.no_grad(): + outputs = model(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state + + # 将文本表示向量转移到CPU上 + text_embeddings = outputs.mean(dim=1).cpu().numpy() + + # 计算相似度 + for i in range(num_questions): + similarity = cosine_similarity([text_embeddings[i]], [text_embeddings[-1]])[0][0] + question_score.append(similarity*0.7) + + # 使用 zip() 函数合并两个列表成元组列表 + combined = list(zip(question_ids, question_score)) + # 将元组列表转换为字典 + interest_score = dict(combined) + + return interest_score + + def generate_recommendations(score1,score2): + result_dict = {} + for key in score1: + if key in score2: + result_dict[key] = score1[key] + score2[key] + + sorted_items = sorted(result_dict.items(), key=lambda x: x[1], reverse=True) + top_k = sorted_items + + return top_k + + + +def deepfmrec(request): + current_user_id = request.user.id # 假设你使用了Django的认证系统 + question_list = Question.objects.all() + user_profile = UserProfile.get_user_profile(current_user_id) + + #召回路 + #1.协同过滤itemcf + user_item_matrix, item_dict, user_dict = itemcf_deepfm.build_user_item_matrix(question_list) + item_similarity_matrix = itemcf_deepfm.calculate_user_similarity(user_item_matrix) + r1 = itemcf_deepfm.generate_recommendations(current_user_id, item_similarity_matrix, user_item_matrix, user_dict, item_dict, k=30) + liked_items = list(user_profile.liked_items.values_list('id', flat=True)) + + #2.热门物品 + r2 = hotitem.recall_hot_item(question_list) + r2 = [item for item in r2 if item[0] not in liked_items] + + #3.新出物品 + r3 = newitem.recall_new_item(question_list) + r3 = [item for item in r3 if item[0] not in liked_items] + + #4.U2TAG2L + a = U2TAG2I.compute_interest_score(user_profile, questions=question_list) + b = U2TAG2I.compute_similarity(user_profile, questions=question_list) + + r4 = U2TAG2I.generate_recommendations(a, b) + r4 = [item for item in r4 if item[0] not in liked_items] + + #汇总4路召回 + + # 获取每个类别前 10 个物品的 ID，去除重复项 + top_10_ids = list( + set([item[0] for item in r1[:10]] + [item[0] for item in r2[:10]] + [item[0] for item in r3[:10]] + [item[0] for item in + r4[:10]])) + # print(top_10_ids) + # 根据 top_10_ids 从 r1, r2, r3, r4 中找到对应的得分 + result = [] + for item_id in top_10_ids: + score1 = next((item[1] for item in r1 if item[0] == item_id), None) + score2 = next((item[1] for item in r2 if item[0] == item_id), None) + score3 = next((item[1] for item in r3 if item[0] == item_id), None) + score4 = next((item[1] for item in r4 if item[0] == item_id), None) + result.append((item_id, score1, score2, score3, score4)) + df_origin = pd.DataFrame(result, columns=['id', 'feature1', 'feature2', 'feature3', 'feature4']) + df = pd.DataFrame(result, columns=['id', 'feature1', 'feature2', 'feature3', 'feature4']) + + #进入模型预测: + + rec_id = df['id'] + df.drop(columns=['id'], inplace=True) #id不是特征 + # 进行预测任务，不需要标签列 + # 对特征进行预处理，包括离散特征进行标签编码，连续特征进行归一化 + + # 对离散特征进行标签编码 + sparse_features = ['feature2', 'feature3'] # 假设这些是离散特征的列名 + for feat in sparse_features: + lbe = LabelEncoder() + df[feat] = lbe.fit_transform(df[feat]) + + # 对连续特征进行归一化 + dense_features = ['feature1', 'feature4'] # 假设这些是连续特征的列名 + mms = MinMaxScaler() + df[dense_features] = mms.fit_transform(df[dense_features]) + # deepctr模型输入 + fixlen_feature_columns = [SparseFeat(feat, df[feat].nunique(), embedding_dim=4) for feat in sparse_features] + [ + DenseFeat(feat, 1, ) for feat in dense_features] + + # DeepFM 模型 + model = DeepFM(fixlen_feature_columns, fixlen_feature_columns, task='binary') + # 编译模型 + model.compile("adam", "binary_crossentropy", metrics=["binary_crossentropy"]) + # 获取模型输入 + model_input = {name: df[name].values for name in get_feature_names(fixlen_feature_columns)} + # 进行预测 + pred_ans = model.predict(model_input, batch_size=1) + pred_list = [(id, score) for id, score in zip(rec_id, pred_ans.flatten())] + # 假设 pred_list 是一个包含预测结果的列表，每个元素是一个元组 (id, score) + pred_list_sorted = sorted(pred_list, key=lambda x: x[1], reverse=True) + + # 输出排序后的预测结果列表 + #print(pred_list_sorted) + + # 获取推荐 + recommendations =[item_id for item_id, _ in pred_list_sorted] + recommended_all = Question.objects.filter(id__in=recommendations) + ki=8 + recommended_movies = Question.objects.filter(id__in=recommendations[:ki]) # 只展示前ki个 + print('前者：',recommended_movies) + + + #根据用户反馈实时训练更新=================================================================== + # 获取当前用户喜欢或点击的物品的ID列表 + liked_or_clicked_items_ids = list(user_profile.liked_items.values_list('id', flat=True)) + liked_or_clicked_items_ids.extend(list(user_profile.click_items.values_list('id', flat=True))) + liked_or_clicked_items_ids = set(liked_or_clicked_items_ids) + + #与推荐的物品取交集 + + # 将被当前用户点击或喜欢的物品的ID放入列表中 + like_or_click_items = [movie.id for movie in recommended_all if movie.id in liked_or_clicked_items_ids] + recommend_but_not_click_or_like = [movie.id for movie in recommended_all[ki:] if movie.id not in liked_or_clicked_items_ids] + # 获取 like_or_click_items 对应的行号 + row_indices_true = df_origin[df_origin['id'].isin(like_or_click_items)].index + row_indices_false =df_origin[df_origin['id'].isin(recommend_but_not_click_or_like)].index + + train_data_true= df.loc[row_indices_true, ['feature1', 'feature2', 'feature3', 'feature4']] + train_data_false = df.loc[row_indices_false, ['feature1', 'feature2', 'feature3', 'feature4']] + train_data_false['label'] = 0 + train_data_true['label'] = 1 + + train_data = pd.concat([train_data_true, train_data_false], ignore_index=True) + + #print(train_data) + train_threshold=1 + # 检查 train_data 中的标签分布情况 + label_distribution = train_data['label'].value_counts() + #print('label_dis:',label_distribution) + + if len(train_data) >= train_threshold: + #训练输入 + fixlen_feature_columns = [SparseFeat(feat, train_data[feat].nunique(), embedding_dim=4) for feat in sparse_features] + [ + DenseFeat(feat, 1, ) for feat in dense_features] + + train_model_input = {name: train_data[name].values for name in get_feature_names(fixlen_feature_columns)} + #print('input:',train_model_input) + + history = model.fit(train_model_input, train_data['label'].values, + batch_size=len(train_data), epochs=4, verbose=2, validation_split=0.0, ) + + #训练后预测 + model_input = {name: df[name].values for name in get_feature_names(fixlen_feature_columns)} + pred_ans = model.predict(model_input, batch_size=1) + pred_list = [(id, score) for id, score in zip(rec_id, pred_ans.flatten())] + pred_list_sorted = sorted(pred_list, key=lambda x: x[1], reverse=True) + recommendations = [item_id for item_id, _ in pred_list_sorted] + recommended_movies1 = [] + + for question_id in recommendations: + q =Question.objects.filter(id=question_id) + recommended_movies1.append(q) + from itertools import chain + recommended_movies = list(chain.from_iterable(recommended_movies1)) + + + return render(request, 'moiveReApp/deepfmrec.html', {'recommended_movies': recommended_movies}) + return render(request, 'moiveReApp/deepfmrec.html', {'recommended_movies': recommended_movies}) + + + + diff --git a/moiveRe/recommendationrecall.py b/moiveRe/recommendationrecall.py index 22028df..4e1ccde 100644 --- a/moiveRe/recommendationrecall.py +++ b/moiveRe/recommendationrecall.py @@ -20,7 +20,7 @@ from django.contrib.auth.models import User -user = User.objects.get(username='gly233') +user = User.objects.get(username='gege') user_profile = UserProfile.get_user_profile(user) question_list = Question.objects.all() @@ -129,28 +129,34 @@ def generate_recommendations(target_user, item_similarity_matrix, user_item_matr user_item_matrix, item_dict, user_dict = itemcf.build_user_item_matrix(question_list) item_similarity_matrix = itemcf.calculate_user_similarity(user_item_matrix) -r1 = itemcf.generate_recommendations(user.id, item_similarity_matrix, user_item_matrix, user_dict, item_dict, k=30)[:10] +r1 = itemcf.generate_recommendations(user.id, item_similarity_matrix, user_item_matrix, user_dict, item_dict, k=30) + +# 获取用户已经喜欢的物品的列表 +liked_items = list(user_profile.liked_items.values_list('id', flat=True)) +print(liked_items) +#print(len(liked_items),liked_items) class hotitem(): - def recall_hot_item(question_list,k): + def recall_hot_item(question_list): sorted_question_list = sorted(question_list, key=lambda q: q.likes, reverse=True) - recommendations = [question.id for question in sorted_question_list] - return recommendations[:k] + recommendations = [(question.id, question.likes) for question in sorted_question_list] + return recommendations -r2 = hotitem.recall_hot_item(question_list,10) +r2 = hotitem.recall_hot_item(question_list) +r2 = [item for item in r2 if item[0] not in liked_items] class newitem(): - - def recall_new_item(question_list,k): + @staticmethod + def recall_new_item(question_list): sorted_question_list = sorted(question_list, key=lambda q: q.pub_date, reverse=True) - recommendations = [question.id for question in sorted_question_list] - return recommendations[:k] - -r3 = newitem.recall_new_item(question_list,10) + recommendations = [(question.id, 1) for question in sorted_question_list] #得分为1 不带偏见的推荐新物品~ + return recommendations +r3 = newitem.recall_new_item(question_list) +r3 = [item for item in r3 if item[0] not in liked_items] class U2TAG2I(): @@ -211,14 +217,15 @@ def compute_similarity(user_profile, questions): return interest_score - def generate_recommendations(score1,score2,k): + def generate_recommendations(score1,score2): result_dict = {} for key in score1: if key in score2: result_dict[key] = score1[key] + score2[key] sorted_items = sorted(result_dict.items(), key=lambda x: x[1], reverse=True) - top_k = sorted_items[:k] + top_k = sorted_items + return top_k @@ -226,15 +233,36 @@ def generate_recommendations(score1,score2,k): a=U2TAG2I.compute_interest_score(user_profile,questions=question_list) b=U2TAG2I.compute_similarity(user_profile,questions=question_list) -r4 = U2TAG2I.generate_recommendations(a,b,10) - +r4 = U2TAG2I.generate_recommendations(a,b) +r4 = [item for item in r4 if item[0] not in liked_items] +''' print(r1) print(r2) print(r3) print(r4) - +''' + +# 获取每个类别前 10 个物品的 ID，去除重复项 +top_10_ids = list(set([item[0] for item in r1[:10]] + [item[0] for item in r2[:10]] + [item[0] for item in r3[:10]] + [item[0] for item in r4[:10]])) +#print(top_10_ids) +# 根据 top_10_ids 从 r1, r2, r3, r4 中找到对应的得分 +result = [] +for item_id in top_10_ids: + score1 = next((item[1] for item in r1 if item[0] == item_id), None) + score2 = next((item[1] for item in r2 if item[0] == item_id), None) + score3 = next((item[1] for item in r3 if item[0] == item_id), None) + score4 = next((item[1] for item in r4 if item[0] == item_id), None) + result.append((item_id, score1, score2, score3, score4)) + + +import pandas as pd +df = pd.DataFrame(result, columns=['id', 'feature1', 'feature2', 'feature3', 'feature4']) + +# 输出 DataFrame +print(df) +df.to_csv('test_data.csv', index=False) diff --git a/moiveRe/sortbyDeepFM.py b/moiveRe/sortbyDeepFM.py new file mode 100644 index 0000000..d589cef --- /dev/null +++ b/moiveRe/sortbyDeepFM.py @@ -0,0 +1,72 @@ +import django +import os +import torch + +os.environ.setdefault("DJANGO_SETTINGS_MODULE", "moiveRe.settings") +django.setup() + + + + + + + + + +import pandas as pd +import numpy as np +from sklearn.preprocessing import LabelEncoder, MinMaxScaler +import torch +from deepctr_torch.models import DeepFM +from deepctr_torch.inputs import SparseFeat, DenseFeat, get_feature_names + +# 假设你已经准备好了包含特征的 DataFrame，其中 'id' 列是不参与预测的特征列，其余列是特征 +# 你可以通过删除 'id' 列来获得特征数据 +df = pd.read_csv('test_data.csv') # 假设你的数据集保存为 CSV 文件 +rec_id = df['id'] +# 假设 'id' 列是索引，不是特征，因此将其从 DataFrame 中移除 +df.drop(columns=['id'], inplace=True) +# 进行预测任务，不需要标签列 +# 你需要对特征进行预处理，包括离散特征进行标签编码，连续特征进行归一化 + +# 对离散特征进行标签编码 +sparse_features = ['feature2','feature3'] # 假设这些是离散特征的列名 +for feat in sparse_features: + lbe = LabelEncoder() + df[feat] = lbe.fit_transform(df[feat]) + +# 对连续特征进行归一化 +dense_features = ['feature1','feature4'] # 假设这些是连续特征的列名 +mms = MinMaxScaler() +df[dense_features] = mms.fit_transform(df[dense_features]) +print(df) +# deepctr模型输入 +fixlen_feature_columns = [SparseFeat(feat, df[feat].nunique(), embedding_dim=4) for feat in sparse_features] + [DenseFeat(feat, 1,) for feat in dense_features] + +# DeepFM 模型 +model = DeepFM(fixlen_feature_columns, fixlen_feature_columns, task='binary') + +# 编译模型 +model.compile("adam", "binary_crossentropy", metrics=["binary_crossentropy"]) + +# 获取模型输入 +model_input = {name: df[name].values for name in get_feature_names(fixlen_feature_columns)} + +# 进行预测 +pred_ans = model.predict(model_input, batch_size=1) +pred_list = [(id, score) for id, score in zip(rec_id, pred_ans.flatten())] + +# 假设 pred_list 是一个包含预测结果的列表，每个元素是一个元组 (id, score) +pred_list_sorted = sorted(pred_list, key=lambda x: x[1], reverse=True) + +# 输出排序后的预测结果列表 +print(pred_list_sorted) + + + + + + + + +