Add HRDR model (#518)

* Add HRDR model

* refactor code

* refactor code to use save and load function

Author: lthoang

README.md

@@ -112,6 +112,7 @@ The recommender models supported by Cornac are listed below. Why don't you join
 |      | [Causal Inference for Visual Debiasing in Visually-Aware Recommendation (CausalRec)](cornac/models/causalrec), [paper](https://arxiv.org/abs/2107.02390) | [requirements.txt](cornac/models/causalrec/requirements.txt) | [causalrec_clothing.py](examples/causalrec_clothing.py)
 |      | [Explainable Recommendation with Comparative Constraints on Product Aspects (ComparER)](cornac/models/comparer), [paper](https://dl.acm.org/doi/pdf/10.1145/3437963.3441754) | N/A | [PreferredAI/ComparER](https://github.com/PreferredAI/ComparER)
 | 2020 | [Adversarial Training Towards Robust Multimedia Recommender System (AMR)](cornac/models/amr), [paper](https://ieeexplore.ieee.org/document/8618394) | [requirements.txt](cornac/models/amr/requirements.txt) | [amr_clothing.py](examples/amr_clothing.py)
+|      | [Hybrid neural recommendation with joint deep representation learning of ratings and reviews (HRDR)](cornac/models/hrdr), [paper](https://www.sciencedirect.com/science/article/abs/pii/S0925231219313207) | [requirements.txt](cornac/models/hrdr/requirements.txt) | [hrdr_example.py](examples/hrdr_example.py)
 | 2019 | [Embarrassingly Shallow Autoencoders for Sparse Data (EASEᴿ)](cornac/models/ease), [paper](https://arxiv.org/pdf/1905.03375.pdf) | N/A | [ease_movielens.py](examples/ease_movielens.py)
 | 2018 | [Collaborative Context Poisson Factorization (C2PF)](cornac/models/c2pf), [paper](https://www.ijcai.org/proceedings/2018/0370.pdf) | N/A | [c2pf_exp.py](examples/c2pf_example.py)
 |      | [Multi-Task Explainable Recommendation (MTER)](cornac/models/mter), [paper](https://arxiv.org/pdf/1806.03568.pdf) | N/A | [mter_exp.py](examples/mter_example.py)

cornac/models/__init__.py

@@ -36,6 +36,7 @@
 from .global_avg import GlobalAvg
 from .hft import HFT
 from .hpf import HPF
+from .hrdr import HRDR
 from .ibpr import IBPR
 from .knn import ItemKNN
 from .knn import UserKNN

cornac/models/hrdr/__init__.py

@@ -0,0 +1 @@
+from .recom_hrdr import HRDR

cornac/models/hrdr/hrdr.py

@@ -0,0 +1,239 @@
+import numpy as np
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import layers, initializers
+from tensorflow.python.keras.preprocessing.sequence import pad_sequences
+
+from ...utils import get_rng
+from ...utils.init_utils import uniform
+from ..narre.narre import TextProcessor, AddGlobalBias
+
+
+def get_data(batch_ids, train_set, max_text_length, by="user", max_num_review=32):
+    batch_reviews, batch_num_reviews = [], []
+    review_group = (
+        train_set.review_text.user_review
+        if by == "user"
+        else train_set.review_text.item_review
+    )
+    for idx in batch_ids:
+        review_ids = []
+        for inc, (jdx, review_idx) in enumerate(review_group[idx].items()):
+            if max_num_review is not None and inc == max_num_review:
+                break
+            review_ids.append(review_idx)
+        reviews = train_set.review_text.batch_seq(
+            review_ids, max_length=max_text_length
+        )
+        batch_reviews.append(reviews)
+        batch_num_reviews.append(len(reviews))
+    batch_reviews = pad_sequences(batch_reviews, maxlen=max_num_review, padding="post")
+    batch_num_reviews = np.array(batch_num_reviews).astype(np.int32)
+    batch_ratings = (
+        np.zeros((len(batch_ids), train_set.num_items), dtype=np.float32)
+        if by == "user"
+        else np.zeros((len(batch_ids), train_set.num_users), dtype=np.float32)
+    )
+    rating_group = train_set.user_data if by == "user" else train_set.item_data
+    for batch_inc, idx in enumerate(batch_ids):
+        jds, ratings = rating_group[idx]
+        for jdx, rating in zip(jds, ratings):
+            batch_ratings[batch_inc, jdx] = rating
+    return batch_reviews, batch_num_reviews, batch_ratings
+
+class Model(keras.Model):
+    def __init__(self, n_users, n_items, n_vocab, global_mean, embedding_matrix,
+                 n_factors=32, embedding_size=100, id_embedding_size=32,
+                 attention_size=16, kernel_sizes=[3], n_filters=64,
+                 n_user_mlp_factors=128, n_item_mlp_factors=128,
+                 dropout_rate=0.5, max_text_length=50):
+        super().__init__()
+        self.l_user_review_embedding = layers.Embedding(n_vocab, embedding_size, embeddings_initializer=embedding_matrix, mask_zero=True, name="user_review_embedding")
+        self.l_item_review_embedding = layers.Embedding(n_vocab, embedding_size, embeddings_initializer=embedding_matrix, mask_zero=True, name="item_review_embedding")
+        self.l_user_embedding = layers.Embedding(n_users, id_embedding_size, embeddings_initializer="uniform", name="user_embedding")
+        self.l_item_embedding = layers.Embedding(n_items, id_embedding_size, embeddings_initializer="uniform", name="item_embedding")
+        self.user_bias = layers.Embedding(n_users, 1, embeddings_initializer=tf.initializers.Constant(0.1), name="user_bias")
+        self.item_bias = layers.Embedding(n_items, 1, embeddings_initializer=tf.initializers.Constant(0.1), name="item_bias")
+        self.user_text_processor = TextProcessor(max_text_length, filters=n_filters, kernel_sizes=kernel_sizes, dropout_rate=dropout_rate, name='user_text_processor')
+        self.item_text_processor = TextProcessor(max_text_length, filters=n_filters, kernel_sizes=kernel_sizes, dropout_rate=dropout_rate, name='item_text_processor')
+
+        self.l_user_mlp = keras.models.Sequential([
+            layers.Dense(n_user_mlp_factors, input_dim=n_items, activation="relu"),
+            layers.Dense(n_user_mlp_factors // 2, activation="relu"),
+            layers.Dense(n_filters, activation="relu"),
+            layers.BatchNormalization(),
+        ])
+        self.l_item_mlp = keras.models.Sequential([
+            layers.Dense(n_item_mlp_factors, input_dim=n_users, activation="relu"),
+            layers.Dense(n_item_mlp_factors // 2, activation="relu"),
+            layers.Dense(n_filters, activation="relu"),
+            layers.BatchNormalization(),
+        ])
+        self.a_user = keras.models.Sequential([
+            layers.Dense(attention_size, activation="relu", use_bias=True),
+            layers.Dense(1, activation=None, use_bias=True)
+        ])
+        self.user_attention = layers.Softmax(axis=1, name="user_attention")
+        self.a_item = keras.models.Sequential([
+            layers.Dense(attention_size, activation="relu", use_bias=True),
+            layers.Dense(1, activation=None, use_bias=True)
+        ])
+        self.item_attention = layers.Softmax(axis=1, name="item_attention")
+        self.ou_dropout = layers.Dropout(rate=dropout_rate)
+        self.oi_dropout = layers.Dropout(rate=dropout_rate)
+        self.ou = layers.Dense(n_factors, use_bias=True, name="ou")
+        self.oi = layers.Dense(n_factors, use_bias=True, name="oi")
+        self.W1 = layers.Dense(1, activation=None, use_bias=False, name="W1")
+        self.add_global_bias = AddGlobalBias(init_value=global_mean, name="global_bias")
+
+    def call(self, inputs, training=False):
+        i_user_id, i_item_id, i_user_rating, i_user_review, i_user_num_reviews, i_item_rating, i_item_review, i_item_num_reviews = inputs
+        user_review_h = self.user_text_processor(self.l_user_review_embedding(i_user_review), training=training)
+        item_review_h = self.item_text_processor(self.l_item_review_embedding(i_item_review), training=training)
+        user_rating_h = self.l_user_mlp(i_user_rating)
+        item_rating_h = self.l_item_mlp(i_item_rating)
+        a_user = self.a_user(
+            tf.multiply(
+                user_review_h,
+                tf.expand_dims(user_rating_h, 1)
+            )
+        )
+        a_user_masking = tf.expand_dims(tf.sequence_mask(tf.reshape(i_user_num_reviews, [-1]), maxlen=i_user_review.shape[1]), -1)
+        user_attention = self.user_attention(a_user, a_user_masking)
+        a_item = self.a_item(
+            tf.multiply(
+                item_review_h,
+                tf.expand_dims(item_rating_h, 1)
+            )
+        )
+        a_item_masking = tf.expand_dims(tf.sequence_mask(tf.reshape(i_item_num_reviews, [-1]), maxlen=i_item_review.shape[1]), -1)
+        item_attention = self.item_attention(a_item, a_item_masking)
+        ou = tf.multiply(user_attention, user_review_h)
+        ou = tf.reduce_sum(ou, 1)
+        if training:
+            ou = self.ou_dropout(ou, training=training)
+        ou = self.ou(ou)
+        oi = tf.multiply(item_attention, item_review_h)
+        oi = tf.reduce_sum(oi, 1)
+        if training:
+            oi = self.oi_dropout(oi, training=training)
+        oi = self.oi(oi)
+        pu = tf.concat([
+            user_rating_h,
+            ou,
+            self.l_user_embedding(i_user_id)
+        ], axis=-1)
+        qi = tf.concat([
+            item_rating_h,
+            oi,
+            self.l_item_embedding(i_item_id)
+        ], axis=-1)
+        h0 = tf.multiply(pu, qi)
+        r = self.add_global_bias(
+            tf.add_n([
+                self.W1(h0),
+                self.user_bias(i_user_id),
+                self.item_bias(i_item_id)
+            ])
+        )
+        return r
+
+class HRDRModel:
+    def __init__(self, n_users, n_items, vocab, global_mean,
+                 n_factors=32, embedding_size=100, id_embedding_size=32,
+                 attention_size=16, kernel_sizes=[3], n_filters=64,
+                 n_user_mlp_factors=128, n_item_mlp_factors=128,
+                 dropout_rate=0.5, max_text_length=50, max_num_review=32,
+                 pretrained_word_embeddings=None, verbose=False, seed=None):
+        self.n_users = n_users
+        self.n_items = n_items
+        self.n_vocab = vocab.size
+        self.global_mean = global_mean
+        self.n_factors = n_factors
+        self.embedding_size = embedding_size
+        self.id_embedding_size = id_embedding_size
+        self.attention_size = attention_size
+        self.kernel_sizes = kernel_sizes
+        self.n_filters = n_filters
+        self.n_user_mlp_factors = n_user_mlp_factors
+        self.n_item_mlp_factors = n_item_mlp_factors
+        self.dropout_rate = dropout_rate
+        self.max_text_length = max_text_length
+        self.max_num_review = max_num_review
+        self.verbose = verbose
+        if seed is not None:
+            self.rng = get_rng(seed)
+            tf.random.set_seed(seed)
+
+        embedding_matrix = uniform(shape=(self.n_vocab, self.embedding_size), low=-0.5, high=0.5, random_state=self.rng)
+        embedding_matrix[:4, :] = np.zeros((4, self.embedding_size))
+        if pretrained_word_embeddings is not None:
+            oov_count = 0
+            for word, idx in vocab.tok2idx.items():
+                embedding_vector = pretrained_word_embeddings.get(word)
+                if embedding_vector is not None:
+                    embedding_matrix[idx] = embedding_vector
+                else:
+                    oov_count += 1
+            if self.verbose:
+                print("Number of OOV words: %d" % oov_count)
+
+        embedding_matrix = initializers.Constant(embedding_matrix)
+        self.graph = Model(
+            self.n_users, self.n_items, self.n_vocab, self.global_mean, embedding_matrix,
+            self.n_factors, self.embedding_size, self.id_embedding_size,
+            self.attention_size, self.kernel_sizes, self.n_filters,
+            self.n_user_mlp_factors, self.n_item_mlp_factors,
+            self.dropout_rate, self.max_text_length
+        )
+
+    def get_weights(self, train_set, batch_size=64):
+        P = np.zeros((self.n_users, self.n_filters + self.n_factors + self.id_embedding_size))
+        Q = np.zeros((self.n_items, self.n_filters + self.n_factors + self.id_embedding_size))
+        A = np.zeros((self.n_items, self.max_num_review))
+        for batch_users in train_set.user_iter(batch_size, shuffle=False):
+            i_user_review, i_user_num_reviews, i_user_rating = get_data(batch_users, train_set, self.max_text_length, by='user', max_num_review=self.max_num_review)
+            user_review_embedding = self.graph.l_user_review_embedding(i_user_review)
+            user_review_h = self.graph.user_text_processor(user_review_embedding, training=False)
+            user_rating_h = self.graph.l_user_mlp(i_user_rating)
+            a_user = self.graph.a_user(
+                tf.multiply(
+                    user_review_h,
+                    tf.expand_dims(user_rating_h, 1)
+                )
+            )
+            a_user_masking = tf.expand_dims(tf.sequence_mask(tf.reshape(i_user_num_reviews, [-1]), maxlen=i_user_review.shape[1]), -1)
+            user_attention = self.graph.user_attention(a_user, a_user_masking)
+            ou = self.graph.ou(tf.reduce_sum(tf.multiply(user_attention, user_review_h), 1))
+            pu = tf.concat([
+                user_rating_h,
+                ou,
+                self.graph.l_user_embedding(batch_users)
+            ], axis=-1)
+            P[batch_users] = pu.numpy()
+        for batch_items in train_set.item_iter(batch_size, shuffle=False):
+            i_item_review, i_item_num_reviews, i_item_rating = get_data(batch_items, train_set, self.max_text_length, by='item', max_num_review=self.max_num_review)
+            item_review_embedding = self.graph.l_item_review_embedding(i_item_review)
+            item_review_h = self.graph.item_text_processor(item_review_embedding, training=False)
+            item_rating_h = self.graph.l_item_mlp(i_item_rating)
+            a_item = self.graph.a_item(
+                tf.multiply(
+                    item_review_h,
+                    tf.expand_dims(item_rating_h, 1)
+                )
+            )
+            a_item_masking = tf.expand_dims(tf.sequence_mask(tf.reshape(i_item_num_reviews, [-1]), maxlen=i_item_review.shape[1]), -1)
+            item_attention = self.graph.item_attention(a_item, a_item_masking)
+            oi = self.graph.oi(tf.reduce_sum(tf.multiply(item_attention, item_review_h), 1))
+            qi = tf.concat([
+                item_rating_h,
+                oi,
+                self.graph.l_item_embedding(batch_items)
+            ], axis=-1)
+            Q[batch_items] = qi.numpy()
+            A[batch_items, :item_attention.shape[1]] = item_attention.numpy().reshape(item_attention.shape[:2])
+        W1 = self.graph.W1.get_weights()[0]
+        bu = self.graph.user_bias.get_weights()[0]
+        bi = self.graph.item_bias.get_weights()[0]
+        mu = self.graph.add_global_bias.get_weights()[0][0]
+        return P, Q, W1, bu, bi, mu, A