【推荐系统】DIN源码分析（1）

源码分析（1）"/>

【推荐系统】DIN源码分析（1）

数据样本构造

关于训练集和测试集合的划分（更为详细的介绍可以看这里：DIN论文官方实现解析）

Amazon Dataset2. Amazon Dataset contains product reviews and metadata from Amazon, which is used as benchmark dataset[13,18, 23]. We conduct experiments on a subset named Electronics, which contains 192,403 users, 63,001 goods, 801 categories and 1,689,188 samples. User behaviors in this dataset are rich, with more than 5 reviews for each users and goods. Features include
goods_id, cate_id, user reviewed goods_id_list and cate_id_list. Let all behaviors of a user be (b1,b2, . . . ,bk , . . . ,bn), the task is to predict the (k+1)-th reviewed goods by making use of the first k reviewed goods. Training dataset is generated with k = 1, 2, . . . ,n-2 for each user. In the test set, we predict the last one given the first n - 1 reviewed goods. For all models, we use SGD as the optimizer with exponential decay, in which learning rate starts at 1 and decay rate is set to 0.1. The mini-batch size is set to be 32.

样本分析

userid = 4
设 用户的历史行为：[1,2,4,5]
随机初始的itemid为 0.
train_set_pos 1 17 [(0, [1], 2, 1), (0, [1], 0, 0), (0, [1, 2], 4, 1), (0, [1, 2], 0, 0), (1, [1], 2, 1), (1, [1], 0, 0), (1, [1, 2], 4, 1), (1, [1, 2], 0, 0), (2, [1], 2, 1), (2, [1], 0, 0), (2, [1, 2], 4, 1), (2, [1, 2], 0, 0), (3, [1], 2, 1), (3, [1], 0, 0), (3, [1, 2], 4, 1), (3, [1, 2], 0, 0),
 (4, [1], 2, 1)] # 生产一条正样本，假设用户在已点击1的情况下，如果预测点击为2，则是正样本
train_set_neg 1 18 [(0, [1], 2, 1), (0, [1], 0, 0), (0, [1, 2], 4, 1), (0, [1, 2], 0, 0), (1, [1], 2, 1), (1, [1], 0, 0), (1, [1, 2], 4, 1), (1, [1, 2], 0, 0), (2, [1], 2, 1), (2, [1], 0, 0), (2, [1, 2], 4, 1), (2, [1, 2], 0, 0), (3, [1], 2, 1), (3, [1], 0, 0), (3, [1, 2], 4, 1), (3, [1, 2], 0, 0),
 (4, [1], 2, 1), (4, [1], 0, 0)] # 生产一条负样本，假设用户在已点击1的情况下，如果预测点击为0，则是负样本
train_set_pos 2 19 [(0, [1], 2, 1), (0, [1], 0, 0), (0, [1, 2], 4, 1), (0, [1, 2], 0, 0), (1, [1], 2, 1), (1, [1], 0, 0), (1, [1, 2], 4, 1), (1, [1, 2], 0, 0), (2, [1], 2, 1), (2, [1], 0, 0), (2, [1, 2], 4, 1), (2, [1, 2], 0, 0), (3, [1], 2, 1), (3, [1], 0, 0), (3, [1, 2], 4, 1), (3, [1, 2], 0, 0),
 (4, [1], 2, 1), (4, [1], 0, 0), (4, [1, 2], 4, 1)] # 生产一条正样本，假设用户在已点击[1,2]的情况下，如果预测点击为4，则是正样本
train_set_neg 2 20 [(0, [1], 2, 1), (0, [1], 0, 0), (0, [1, 2], 4, 1), (0, [1, 2], 0, 0), (1, [1], 2, 1), (1, [1], 0, 0), (1, [1, 2], 4, 1), (1, [1, 2], 0, 0), (2, [1], 2, 1), (2, [1], 0, 0), (2, [1, 2], 4, 1), (2, [1, 2], 0, 0), (3, [1], 2, 1), (3, [1], 0, 0), (3, [1, 2], 4, 1), (3, [1, 2], 0, 0),
 (4, [1], 2, 1), (4, [1], 0, 0), (4, [1, 2], 4, 1), (4, [1, 2], 0, 0)] # 生产一条负样本，假设用户在已点击[1,2]的情况下，如果预测点击为0，则是负样本
 
 
 # 每一个user，用前n-1个item，去预测第n个item.
 # 这里每一个用户的历史行为都是[1,2,4,5], 都用前3次行为，去预测第4次行为，这里第4次真实的行为是点击了5, 0是没有点击的
 test_set 3 5 [(0, [1, 2, 4], (5, 0)), (1, [1, 2, 4], (5, 0)), (2, [1, 2, 4], (5, 0)), (3, [1, 2, 4], (5, 0)),
 (4, [1, 2, 4], (5, 0))] # 测试集，(5, 0) # (pos_id, neg_id)

推荐模型之用户行为序列处理 - billlee的文章 - 知乎（很实用的一些方法）

DIN网络结构

class Model(object):def __init__(self, user_count, item_count, cate_count, cate_list, predict_batch_size, predict_ads_num):# shape: [B],  user id (B: batch size)self.u = tf.placeholder(tf.int32, [None,])# shape: [B]  i: 正样本的itemself.i = tf.placeholder(tf.int32, [None,])# shape: [B]  j: 负样本的itemself.j = tf.placeholder(tf.int32, [None,])# shape: [B], y: labelself.y = tf.placeholder(tf.float32, [None,])# shape: [B, T] # 用户行为特征(User Behavior)中的item序列, T为序列长度self.hist_i = tf.placeholder(tf.int32, [None, None])# shape: [B]; sl: sequence length, User Behavior中序列的真实序列长度self.sl = tf.placeholder(tf.int32, [None,])# learning rateself.lr = tf.placeholder(tf.float64, [])hidden_units = 128# shape: [U, H], user_id的embedding weight. U是user_id的hash bucket sizeuser_emb_w = tf.get_variable("user_emb_w", [user_count, hidden_units])# shape: [I, H//2], item_id的embedding weight. I是item_id的hash bucket size# [I, H//2]item_emb_w = tf.get_variable("item_emb_w", [item_count, hidden_units // 2])# shape: [I], biasitem_b = tf.get_variable("item_b", [item_count],initializer=tf.constant_initializer(0.0))# shape: [C, H//2], cate_id的embedding weight.cate_emb_w = tf.get_variable("cate_emb_w", [cate_count, hidden_units // 2])# shape: [C, H//2]cate_list = tf.convert_to_tensor(cate_list, dtype=tf.int64)# 从cate_list中取出正样本的cateic = tf.gather(cate_list, self.i)# 正样本的embedding，正样本包括item和catei_emb = tf.concat(values = [tf.nn.embedding_lookup(item_emb_w, self.i),tf.nn.embedding_lookup(cate_emb_w, ic),], axis=1)i_b = tf.gather(item_b, self.i)# 从cate_list中取出负样本的catejc = tf.gather(cate_list, self.j)# 负样本的embedding，负样本包括item和catej_emb = tf.concat([tf.nn.embedding_lookup(item_emb_w, self.j),tf.nn.embedding_lookup(cate_emb_w, jc),], axis=1)# 偏置bj_b = tf.gather(item_b, self.j)# 用户行为序列(User Behavior)中的cate序列hc = tf.gather(cate_list, self.hist_i)# 用户行为序列(User Behavior)的embedding，包括item序列和cate序列h_emb = tf.concat([tf.nn.embedding_lookup(item_emb_w, self.hist_i),tf.nn.embedding_lookup(cate_emb_w, hc),], axis=2)# attention操作# 返回用户行为的每个商品的兴趣分布hist_i = attention(i_emb, h_emb, self.sl)#-- attention end ---hist_i = tf.layers.batch_normalization(inputs = hist_i)hist_i = tf.reshape(hist_i, [-1, hidden_units], name='hist_bn')hist_i = tf.layers.dense(hist_i, hidden_units, name='hist_fcn')u_emb_i = hist_ihist_j = attention(j_emb, h_emb, self.sl)#-- attention end ---

参考：

DIN(Deep Interest Network):核心思想+源码阅读注释

=distribute.pc_relevant.none-task-blog-baidujs_baidulandingword-7&spm=1001.2101.3001.4242

DIN算法代码详细解读 - 小2小M的文章 - 知乎

关于构造正负样本的trick：都说数据是上限，推荐系统ctr模型中，构造正负样本有哪些实用的trick？ - 知乎