def build_model(self, user_ids, item_vec, cate_ids, labels,
neighbor_user_ids, keep_prob, phase):
with tf.variable_scope('feature_representation'):
# user
neighbor_user_emb = tf.nn.embedding_lookup(self.user_embedding,
neighbor_user_ids)
# item
item_emb = dense(item_vec, self.emb_dim, ['w1'], keep_prob)
cate_emb = tf.nn.embedding_lookup(self.item_category_embedding,
cate_ids)
item_emb = item_emb + cate_emb
with tf.variable_scope('neighborhood_level_interest_network'):
avg_neighbor_emb = tf.reduce_mean(neighbor_user_emb,
axis=1,
keep_dims=True)
neighbor_emb = neighbor_attention(neighbor_user_emb, item_emb,
avg_neighbor_emb, 1.0, 'relu',
False)
# ef: early fusion
# if: intermediate fusion
# lf: late fusion
with tf.variable_scope('interaction'):
if self.interaction == 'inner':
logits = tf.reduce_sum(neighbor_emb * item_emb, axis=1)
else:
if self.interaction == 'product':
fusion_input = neighbor_emb * item_emb
else:
raise ValueError('invalid fusion way: {}'.format(
self.interaction))
with tf.variable_scope('aggregation_layer'):
logits = dnn(fusion_input, self.agg_layers, keep_prob)
inference_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,
labels=labels))
acc = self.compute_acc(logits, self.labels_ph)
if phase == 'train':
emb_l2_loss = tf.nn.l2_loss(neighbor_user_emb) + tf.nn.l2_loss(
cate_emb)
w_l2_loss = tf.add_n([
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'w' in v.name
])
l2_w_loss = w_l2_loss * self.w_reg
l2_emb_loss = emb_l2_loss * self.emb_reg
return inference_loss, l2_w_loss, l2_emb_loss, acc, logits
else:
return inference_loss, acc, logits
def build_model(self, user_ids, item_ids, cate_ids, labels, keep_prob,
phase):
with tf.variable_scope('feature_representation'):
# user
user_click_cate_distribution = tf.nn.embedding_lookup(
self.user_click_cate_distribution, user_ids)
# item
cate_emb = tf.reduce_mean(tf.nn.embedding_lookup(
self.item_category_embedding, cate_ids),
axis=1)
visual_feat = tf.nn.embedding_lookup(self.visual_feature, item_ids)
visual_emb = dense(visual_feat, self.emb_dim, ['w1'], keep_prob)
item_emb = visual_emb + cate_emb
with tf.variable_scope('category_level_interest'):
category_interest = tf.matmul(user_click_cate_distribution,
self.category_embedding)
with tf.variable_scope('aggregation'):
if self.interaction == 'inner':
fusion_input = category_interest * item_emb
logits = tf.reduce_sum(fusion_input, axis=1)
else:
if self.interaction == 'product':
fusion_input = category_interest * item_emb
else:
raise ValueError('invalid fusion way: {}'.format(
self.interaction))
with tf.variable_scope('aggregation_layer'):
logits = dnn(fusion_input, self.agg_layers, keep_prob)
inference_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,
labels=labels))
acc = self.compute_acc(logits, self.labels_ph)
if phase == 'train':
emb_l2_loss = tf.nn.l2_loss(cate_emb) + tf.nn.l2_loss(
self.category_embedding)
w_l2_loss_lst = [
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'w' in v.name
]
if len(w_l2_loss_lst) != 0:
w_l2_loss = tf.add_n(w_l2_loss_lst)
reg_loss = w_l2_loss * self.w_reg + emb_l2_loss * self.emb_reg
else:
reg_loss = emb_l2_loss * self.emb_reg
return inference_loss, reg_loss, acc, logits
else:
return inference_loss, acc, logits
def build_model(self, item_vec, cate_ids, att_iids, att_cids, intra_mask,
inter_mask, labels, keep_prob):
with tf.variable_scope('item_embedding'):
att_item_vec = self.get_train_cover_image_feature(att_iids)
att_cate_emb = self.get_cate_emb(att_cids)
item_vec = dense(item_vec, self.item_dim, ['w1'], 1.0)
att_item_vec = dense(att_item_vec,
self.item_dim, ['w1'],
1.0,
reuse=True)
cate_emb = self.get_cate_emb(cate_ids)
item_emb = tf.concat([item_vec, cate_emb], axis=-1)
with tf.variable_scope('temporal_hierarchical_attention'):
user_profiles = temporal_hierarchical_attention(
att_cate_emb, att_item_vec, intra_mask, inter_mask,
self.num_heads, keep_prob)
with tf.variable_scope('micro_video_click_through_prediction'):
user_profile = vanilla_attention(user_profiles, item_emb,
inter_mask, keep_prob)
y = dnn(tf.concat([user_profile, item_emb], axis=-1),
self.fusion_layers, keep_prob)
logits = y
# regularization
emb_l2_loss = tf.add_n([
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'embedding' in v.name
])
w_l2_loss = tf.add_n([
tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'w' in v.name
])
l2_norm = emb_l2_loss + w_l2_loss
loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=logits, labels=labels)) + l2_norm * self.reg
acc = self.compute_acc(logits, self.labels_ph)
return loss, acc, logits
def build_model(self, user_ids, item_ids, cate_ids, labels, hist_iids,
hist_length, keep_prob, att_dropout, phase):
with tf.variable_scope('feature_representation'):
# user
# item
visual_feat = tf.nn.embedding_lookup(self.visual_feature, item_ids)
cate_emb = tf.reduce_mean(tf.nn.embedding_lookup(
self.item_category_embedding, cate_ids),
axis=1)
visual_emb = dense(visual_feat, self.emb_dim, ['w1'], 1.0)
item_emb = visual_emb + cate_emb
hist_visual_feat = tf.nn.embedding_lookup(self.visual_feature,
hist_iids)
hist_visual_emb = dense(hist_visual_feat, self.emb_dim, ['w2'],
keep_prob)
hist_item_emb = hist_visual_emb
mask = tf.sequence_mask(hist_length, self.hist_length)
with tf.variable_scope('item_interest_network'):
avg_hist_item_emb = tf.reduce_sum(hist_item_emb * tf.cast(tf.expand_dims(mask, dim=2), tf.float32), axis=1) / \
tf.expand_dims(tf.cast(hist_length, tf.float32), axis=1)
item_interest = item_interest_attention(
hist_item_emb, hist_item_emb, item_emb, avg_hist_item_emb,
mask, 1.0, 'sigmoid', True, att_dropout)
with tf.variable_scope('aggregation'):
if self.interaction == 'inner':
fusion_input = item_interest * item_emb
logits = tf.reduce_sum(fusion_input, axis=1)
else:
if self.interaction == 'product':
fusion_input = item_interest * item_emb
else:
raise ValueError('invalid fusion way: {}'.format(
self.interaction))
with tf.variable_scope('aggregation_layer'):
logits = dnn(fusion_input, self.agg_layers, keep_prob)
inference_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,
labels=labels))
acc = self.compute_acc(logits, self.labels_ph)
if phase == 'train':
emb_l2_loss = tf.nn.l2_loss(cate_emb)
w_l2_loss_lst = [
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'w' in v.name
]
if len(w_l2_loss_lst) != 0:
w_l2_loss = tf.add_n(w_l2_loss_lst)
reg_loss = w_l2_loss * self.w_reg + emb_l2_loss * self.emb_reg
else:
reg_loss = emb_l2_loss * self.emb_reg
return inference_loss, reg_loss, acc, logits
else:
return inference_loss, acc, logits