add DSSM and YoutubeDNN

README.md

@@ -85,10 +85,6 @@ In [example](https://github.com/ZiyaoGeng/Recommender-System-with-TF2.0/tree/rec
 </table>
 
 
-
-
-
-
 ### Ranking
 
 <table style="text-align:center;margin:auto">
@@ -149,7 +145,7 @@ In [example](https://github.com/ZiyaoGeng/Recommender-System-with-TF2.0/tree/rec
 
 ## Discussion
 
-1. If you have any suggestions or questions about the project, you can leave a comment on `Issue` or email `[email protected]`.
+1. If you have any suggestions or questions about the project, you can leave a comment on `Issue`.
 2. wechat：
 
 <div align=center><img src="https://cdn.jsdelivr.net/gh/BlackSpaceGZY/cdn/img/weixin.jpg" width="20%"/></div>

README_CN.md

@@ -219,13 +219,14 @@ for file in split_file_list[:-1]:
   </tr>
   <tr><td>BPR</td><td>0.5768</td><td>0.2392</td><td>0.3016</td><td>0.3708</td><td>0.2108</td><td>0.2485</td><td>0.7728</td><td>0.4220</td><td>0.5054</td></tr>
   <tr><td>NCF</td><td>0.5711</td><td>0.2112</td><td>0.2950</td><td>0.5448</td><td>0.2831</td><td>0.3451</td><td>0.7768</td><td>0.4273</td><td>0.5103</td></tr>
+  <tr><td>DSSM</td><td>0.5410</td><td>0.2016</td><td>0.2807</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr>
+  <tr><td>YoutubeDNN</td><td>0.6358</td><td>0.3042</td><td>0.3825</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr>
   <tr><td>SASRec</td><td>0.8103</td><td>0.4812</td><td>0.5605</td><td>0.5230</td><td>0.2781</td><td>0.3355</td><td>0.8606</td><td>0.5669</td><td>0.6374</td></tr>
 </table>
 
 
 
 
-
 ### Ranking
 
 <table style="text-align:center;margin:auto">
@@ -289,4 +290,4 @@ for file in split_file_list[:-1]:
 
 ## 讨论
 
-对于项目有任何建议或问题，可以在`Issue`留言，或者发邮件至`[email protected]`。
+对于项目有任何建议或问题，可以在`Issue`留言。

example/m_dssm_demo.py

@@ -0,0 +1,88 @@
+"""
+Created on Apr 1, 2022
+train DSSM demo
+@author: Ziyao Geng([email protected])
+"""
+import os
+from absl import flags, app
+from time import time
+from tensorflow.keras.optimizers import Adam
+
+from reclearn.models.matching import DSSM
+from reclearn.data.datasets import movielens as ml
+from reclearn.evaluator import eval_pos_neg
+
+FLAGS = flags.FLAGS
+
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+# Setting training parameters
+flags.DEFINE_string("file_path", "data/ml-1m/ratings.dat", "file path.")
+flags.DEFINE_string("train_path", "data/ml-1m/ml_train.txt", "train path.")
+flags.DEFINE_string("val_path", "data/ml-1m/ml_val.txt", "val path.")
+flags.DEFINE_string("test_path", "data/ml-1m/ml_test.txt", "test path.")
+flags.DEFINE_string("meta_path", "data/ml-1m/ml_meta.txt", "meta path.")
+flags.DEFINE_integer("embed_dim", 64, "The size of embedding dimension.")
+flags.DEFINE_float("embed_reg", 0.0, "The value of embedding regularization.")
+flags.DEFINE_list("user_mlp", [128], "A list of user MLP hidden units.")
+flags.DEFINE_list("item_mlp", [128], "A list of item MLP hidden units")
+flags.DEFINE_string("activation", "relu", "Activation Name.")
+flags.DEFINE_float("dnn_dropout", 0., "Float between 0 and 1. Dropout of user and item MLP layer.")
+flags.DEFINE_boolean("use_l2norm", False, "Whether user embedding, item embedding should be normalized or not.")
+flags.DEFINE_string("loss_name", "binary_cross_entropy_loss", "Loss Name.")
+flags.DEFINE_float("gamma", 0.5, "If hinge_loss is selected as the loss function, you can specify the margin.")
+flags.DEFINE_float("learning_rate", 0.001, "Learning rate.")
+flags.DEFINE_integer("neg_num", 2, "The number of negative sample for each positive sample.")
+flags.DEFINE_integer("epochs", 10, "train steps.")
+flags.DEFINE_integer("batch_size", 512, "Batch Size.")
+flags.DEFINE_integer("test_neg_num", 100, "The number of test negative samples.")
+flags.DEFINE_integer("k", 10, "recall k items at test stage.")
+
+
+def main(argv):
+    # TODO: 1. Split Data
+    if FLAGS.train_path == "None":
+        train_path, val_path, test_path, meta_path = ml.split_data(file_path=file_path)
+    else:
+        train_path, val_path, test_path, meta_path = FLAGS.train_path, FLAGS.val_path, FLAGS.test_path, FLAGS.meta_path
+    with open(meta_path) as f:
+        max_user_num, max_item_num = [int(x) for x in f.readline().strip('\n').split('\t')]
+    # TODO: 2. Load Data
+    train_data = ml.load_data(train_path, FLAGS.neg_num, max_item_num)
+    val_data = ml.load_data(val_path, FLAGS.neg_num, max_item_num)
+    test_data = ml.load_data(test_path, FLAGS.test_neg_num, max_item_num)
+    # TODO: 3. Set Model Hyper Parameters.
+    model_params = {
+        'user_num': max_user_num + 1,
+        'item_num': max_item_num + 1,
+        'embed_dim': FLAGS.embed_dim,
+        'user_mlp': FLAGS.user_mlp,
+        'item_mlp': FLAGS.item_mlp,
+        'activation': FLAGS.activation,
+        'dnn_dropout': FLAGS.dnn_dropout,
+        'use_l2norm': FLAGS.use_l2norm,
+        'loss_name': FLAGS.loss_name,
+        'gamma': FLAGS.gamma,
+        'embed_reg': FLAGS.embed_reg
+    }
+    # TODO: 4. Build Model
+    model = DSSM(**model_params)
+    model.compile(optimizer=Adam(learning_rate=FLAGS.learning_rate))
+    # TODO: 5. Fit Model
+    for epoch in range(1, FLAGS.epochs + 1):
+        t1 = time()
+        model.fit(
+            x=train_data,
+            epochs=1,
+            validation_data=val_data,
+            batch_size=FLAGS.batch_size
+        )
+        t2 = time()
+        eval_dict = eval_pos_neg(model, test_data, ['hr', 'mrr', 'ndcg'], FLAGS.k, FLAGS.batch_size)
+        print('Iteration %d Fit [%.1f s], Evaluate [%.1f s]: HR = %.4f, MRR = %.4f, NDCG = %.4f'
+              % (epoch, t2 - t1, time() - t2, eval_dict['hr'], eval_dict['mrr'], eval_dict['ndcg']))
+
+
+if __name__ == '__main__':
+    app.run(main)

example/m_youtubednn_demo.py

@@ -0,0 +1,84 @@
+"""
+Created on Apr 5, 2022
+train YoutubeDNN demo
+@author: Ziyao Geng([email protected])
+"""
+import os
+from absl import flags, app
+from time import time
+from tensorflow.keras.optimizers import Adam
+
+from reclearn.models.matching import YoutubeDNN
+from reclearn.data.datasets import movielens as ml
+from reclearn.evaluator import eval_pos_neg
+
+FLAGS = flags.FLAGS
+
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+# Setting training parameters
+flags.DEFINE_string("file_path", "data/ml-1m/ratings.dat", "file path.")
+flags.DEFINE_string("train_path", "data/ml-1m/ml_seq_train.txt", "train path.")
+flags.DEFINE_string("val_path", "data/ml-1m/ml_seq_val.txt", "val path.")
+flags.DEFINE_string("test_path", "data/ml-1m/ml_seq_test.txt", "test path.")
+flags.DEFINE_string("meta_path", "data/ml-1m/ml_seq_meta.txt", "meta path.")
+flags.DEFINE_integer("embed_dim", 64, "The size of embedding dimension.")
+flags.DEFINE_float("embed_reg", 0.0, "The value of embedding regularization.")
+flags.DEFINE_list("user_mlp", [128, 256, 64], "A list of user MLP hidden units.")
+flags.DEFINE_string("activation", "relu", "Activation Name.")
+flags.DEFINE_float("dnn_dropout", 0., "Float between 0 and 1. Dropout of user and item MLP layer.")
+flags.DEFINE_boolean("use_l2norm", False, "Whether user embedding, item embedding should be normalized or not.")
+flags.DEFINE_float("learning_rate", 0.001, "Learning rate.")
+flags.DEFINE_integer("neg_num", 2, "The number of negative sample for each positive sample.")
+flags.DEFINE_integer("seq_len", 200, "The length of user's behavior sequence.")
+flags.DEFINE_integer("epochs", 10, "train steps.")
+flags.DEFINE_integer("batch_size", 512, "Batch Size.")
+flags.DEFINE_integer("test_neg_num", 100, "The number of test negative samples.")
+flags.DEFINE_integer("k", 10, "recall k items at test stage.")
+
+
+def main(argv):
+    # TODO: 1. Split Data
+    if FLAGS.train_path == "None":
+        train_path, val_path, test_path, meta_path = ml.split_seq_data(file_path=FLAGS.file_path)
+    else:
+        train_path, val_path, test_path, meta_path = FLAGS.train_path, FLAGS.val_path, FLAGS.test_path, FLAGS.meta_path
+    with open(meta_path) as f:
+        _, max_item_num = [int(x) for x in f.readline().strip('\n').split('\t')]
+    # TODO: 2. Load Sequence Data
+    train_data = ml.load_seq_data(train_path, "train", FLAGS.seq_len, 0, max_item_num)
+    val_data = ml.load_seq_data(val_path, "val", FLAGS.seq_len, FLAGS.neg_num, max_item_num)
+    test_data = ml.load_seq_data(test_path, "test", FLAGS.seq_len, FLAGS.test_neg_num, max_item_num)
+    # TODO: 3. Set Model Hyper Parameters.
+    model_params = {
+        'item_num': max_item_num + 1,
+        'embed_dim': FLAGS.embed_dim,
+        'user_mlp': FLAGS.user_mlp,
+        'activation': FLAGS.activation,
+        'dnn_dropout': FLAGS.dnn_dropout,
+        'neg_num': FLAGS.neg_num,
+        'batch_size': FLAGS.batch_size,
+        'use_l2norm': FLAGS.use_l2norm,
+        'embed_reg': FLAGS.embed_reg
+    }
+    # TODO: 4. Build Model
+    model = YoutubeDNN(**model_params)
+    model.compile(optimizer=Adam(learning_rate=FLAGS.learning_rate))
+    # TODO: 5. Fit Model
+    for epoch in range(1, FLAGS.epochs + 1):
+        t1 = time()
+        model.fit(
+            x=train_data,
+            epochs=1,
+            validation_data=val_data,
+            batch_size=FLAGS.batch_size
+        )
+        t2 = time()
+        eval_dict = eval_pos_neg(model, test_data, ['hr', 'mrr', 'ndcg'], FLAGS.k, FLAGS.batch_size)
+        print('Iteration %d Fit [%.1f s], Evaluate [%.1f s]: HR = %.4f, MRR = %.4f, NDCG = %.4f'
+              % (epoch, t2 - t1, time() - t2, eval_dict['hr'], eval_dict['mrr'], eval_dict['ndcg']))
+
+
+if __name__ == '__main__':
+    app.run(main)

reclearn/models/losses.py

@@ -21,7 +21,7 @@ def get_loss(pos_scores, neg_scores, loss_name, gamma=None):
     elif loss_name == 'hinge_loss':
         loss = hinge_loss(pos_scores, neg_scores, gamma)
     else:
-        loss = binary_entropy_loss(pos_scores, neg_scores)
+        loss = binary_cross_entropy_loss(pos_scores, neg_scores)
     return loss
 
 
@@ -48,8 +48,8 @@ def hinge_loss(pos_scores, neg_scores, gamma=0.5):
     return loss
 
 
-def binary_entropy_loss(pos_scores, neg_scores):
-    """binary entropy loss.
+def binary_cross_entropy_loss(pos_scores, neg_scores):
+    """binary cross entropy loss.
     Args:
         :param pos_scores: A tensor with shape of [batch_size, neg_num].
         :param neg_scores: A tensor with shape of [batch_size, neg_num].

reclearn/models/matching/__init__.py

@@ -2,11 +2,13 @@
 from reclearn.models.matching.poprec import PopRec
 from reclearn.models.matching.bpr import BPR
 from reclearn.models.matching.ncf import NCF
+from reclearn.models.matching.dssm import DSSM
+from reclearn.models.matching.youtubednn import YoutubeDNN
 from reclearn.models.matching.gru4rec import GRU4Rec
 from reclearn.models.matching.sasrec import SASRec
 from reclearn.models.matching.attrec import AttRec
 from reclearn.models.matching.caser import Caser
 from reclearn.models.matching.fissa import FISSA
 
 
-__all__ = ['PopRec', 'BPR', 'NCF', 'GRU4Rec', 'SASRec', 'AttRec', 'Caser', 'FISSA']
+__all__ = ['PopRec', 'BPR', 'NCF', 'DSSM', 'YoutubeDNN', 'GRU4Rec', 'SASRec', 'AttRec', 'Caser', 'FISSA']

reclearn/models/matching/dssm.py

@@ -0,0 +1,90 @@
+"""
+Created on Mar 31, 2022
+Reference: "Learning Deep Structured Semantic Models for Web Search using Clickthrough Data", CIKM, 2013
+@author: Ziyao Geng([email protected])
+"""
+import tensorflow as tf
+from tensorflow.keras import Model
+from tensorflow.keras.layers import Embedding, Input
+from tensorflow.keras.regularizers import l2
+from reclearn.layers import MLP
+from reclearn.models.losses import get_loss
+
+
+class DSSM(Model):
+    def __init__(self, user_num, item_num, embed_dim, user_mlp, item_mlp, activation='relu',
+                 dnn_dropout=0., use_l2norm=False, loss_name="binary_cross_entropy_loss",
+                 gamma=0.5, embed_reg=0., seed=None):
+        """DSSM: The two-tower matching model commonly used in industry.
+        Args:
+            :param user_num: An integer type. The largest user index + 1.
+            :param item_num: An integer type. The largest item index + 1.
+            :param embed_dim: An integer type. Embedding dimension of user vector and item vector.
+            :param user_mlp: A list of user MLP hidden units such as [128, 64, 32].
+            :param item_mlp: A list of item MLP hidden units such as [128, 64, 32] and
+            the last unit must be equal to the user's.
+            :param activation: A string. Activation function name of user and item MLP layer.
+            :param dnn_dropout: Float between 0 and 1. Dropout of user and item MLP layer.
+            :param use_l2norm: A boolean. Whether user embedding, item embedding should be normalized or not.
+            :param loss_name: A string. You can specify the current point-loss function 'binary_cross_entropy_loss' or
+            pair-loss function as 'bpr_loss'、'hinge_loss'.
+            :param gamma: A scalar. If hinge_loss is selected as the loss function, you can specify the margin.
+            :param embed_reg: A float type. The regularizer of embedding.
+            :param seed: A Python integer to use as random seed.
+        :return:
+        """
+        super(DSSM, self).__init__()
+        if user_mlp[-1] != item_mlp[-1]:
+            raise ValueError("The last value of user_mlp must be equal to item_mlp's.")
+        # user embedding
+        self.user_embedding_table = Embedding(input_dim=user_num,
+                                              input_length=1,
+                                              output_dim=embed_dim,
+                                              embeddings_initializer='random_normal',
+                                              embeddings_regularizer=l2(embed_reg))
+        # item embedding
+        self.item_embedding_table = Embedding(input_dim=item_num,
+                                              input_length=1,
+                                              output_dim=embed_dim,
+                                              embeddings_initializer='random_normal',
+                                              embeddings_regularizer=l2(embed_reg))
+        # user_mlp_layer
+        self.user_mlp_layer = MLP(user_mlp, activation, dnn_dropout)
+        # item_mlp_layer
+        self.item_mlp_layer = MLP(item_mlp, activation, dnn_dropout)
+        self.use_l2norm = use_l2norm
+        self.loss_name = loss_name
+        self.gamma = gamma
+        # seed
+        tf.random.set_seed(seed)
+
+    def call(self, inputs):
+        # user info
+        user_info = self.user_embedding_table(inputs['user'])  # (None, embed_dim)
+        # item info
+        pos_info = self.item_embedding_table(inputs['pos_item'])  # (None, embed_dim)
+        neg_info = self.item_embedding_table(inputs['neg_item'])  # (None, neg_num, embed_dim)
+        # mlp
+        user_info = self.user_mlp_layer(user_info)
+        pos_info = self.item_mlp_layer(pos_info)
+        neg_info = self.item_mlp_layer(neg_info)
+        # norm
+        if self.use_l2norm:
+            user_info = tf.math.l2_normalize(user_info, axis=-1)
+            pos_info = tf.math.l2_normalize(pos_info, axis=-1)
+            neg_info = tf.math.l2_normalize(neg_info, axis=-1)
+        # calculate similar scores.
+        pos_scores = tf.reduce_sum(tf.multiply(user_info, pos_info), axis=-1, keepdims=True)  # (None, 1)
+        neg_scores = tf.reduce_sum(tf.multiply(tf.expand_dims(user_info, axis=1), neg_info), axis=-1)  # (None, neg_num)
+        # add loss
+        self.add_loss(get_loss(pos_scores, neg_scores, self.loss_name, self.gamma))
+        logits = tf.concat([pos_scores, neg_scores], axis=-1)
+        return logits
+
+    def summary(self):
+        inputs = {
+            'user': Input(shape=(), dtype=tf.int32),
+            'pos_item': Input(shape=(), dtype=tf.int32),
+            'neg_item': Input(shape=(1,), dtype=tf.int32)  # suppose neg_num=1
+        }
+        Model(inputs=inputs, outputs=self.call(inputs)).summary()