DecoupledTemporalEncoding/model.py at main · AlibabaResearch/DecoupledTemporalEncoding · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
# -*- coding: utf-8 -*-
import tensorflow as tf

from layers import (
    masked_average_pooling,
    build_pairwise_time_gaps_from_seconds,
    single_feature_embedding,
    macro_temporal_encoding_personalized,
    causal_attention
)


def build_graph(config):
    seq_len = config.seq_len
    d_model = config.d_model
    num_heads = config.num_heads

    labels_ph = tf.placeholder(tf.float32, shape=[None, 1], name="labels")

    # sequence inputs
    item_ids_ph = tf.placeholder(tf.int32, shape=[None, seq_len], name="item_ids")
    seq_mask_ph = tf.placeholder(tf.bool, shape=[None, seq_len], name="seq_mask")
    hour_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="hour_0_23")
    weekday_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="weekday_0_6")
    delta_t_hours_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="delta_t_hours")
    seq_time_seconds_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="seq_time_seconds")
    prev_day_segment_stat_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="prev_day_segment_stat")
    hist30_segment_avg_stat_ph = tf.placeholder(tf.float32, shape=[None, seq_len], name="hist30_segment_avg_stat")

    # user side features
    user_active_degree_ph = tf.placeholder(tf.int32, shape=[None], name="user_active_degree")
    is_lowactive_period_ph = tf.placeholder(tf.int32, shape=[None], name="is_lowactive_period")
    is_live_streamer_ph = tf.placeholder(tf.int32, shape=[None], name="is_live_streamer")
    is_video_author_ph = tf.placeholder(tf.int32, shape=[None], name="is_video_author")
    follow_user_num_range_ph = tf.placeholder(tf.int32, shape=[None], name="follow_user_num_range")
    fans_user_num_range_ph = tf.placeholder(tf.int32, shape=[None], name="fans_user_num_range")
    register_days_range_ph = tf.placeholder(tf.int32, shape=[None], name="register_days_range")
    onehot_feat1_ph = tf.placeholder(tf.int32, shape=[None], name="onehot_feat1")
    onehot_feat2_ph = tf.placeholder(tf.int32, shape=[None], name="onehot_feat2")
    onehot_feat3_ph = tf.placeholder(tf.int32, shape=[None], name="onehot_feat3")

    # target item side features
    target_video_id_ph = tf.placeholder(tf.int32, shape=[None], name="target_video_id")
    target_author_id_ph = tf.placeholder(tf.int32, shape=[None], name="target_author_id")
    target_video_type_ph = tf.placeholder(tf.int32, shape=[None], name="target_video_type")
    target_music_type_ph = tf.placeholder(tf.int32, shape=[None], name="target_music_type")

    # sequence item embedding
    with tf.variable_scope("sequence_embedding", reuse=tf.AUTO_REUSE):
        item_table = tf.get_variable(
            "item_table",
            shape=[config.video_vocab_size, d_model],
            dtype=tf.float32,
            initializer=tf.contrib.layers.xavier_initializer()
        )
        item_emb = tf.nn.embedding_lookup(item_table, item_ids_ph)  # [N, L, D]

    # user context for macro gating
    user_context = masked_average_pooling(item_emb, seq_mask_ph)  # [N, D]

    # macro-temporal module
    macro_temporal_emb, macro_scalar, macro_beta = macro_temporal_encoding_personalized(
        hour_0_23=hour_ph,
        weekday_0_6=weekday_ph,
        delta_t_hours=delta_t_hours_ph,
        seq_mask=seq_mask_ph,
        d_model=d_model,
        user_context=user_context,
        burst_value=prev_day_segment_stat_ph,
        avg_brust_value=hist30_segment_avg_stat_ph,
        scope="macro_temporal"
    )

    seq_input = item_emb + macro_temporal_emb  # [N, L, D]

    # micro-sequential module inputs
    pairwise_time_gaps = build_pairwise_time_gaps_from_seconds(seq_time_seconds_ph)  # [N, L, L]

    # decoder-only backbone: single masked self-attention block
    decoder_output, att_vec = causal_attention(
        queries=seq_input,
        keys=seq_input,
        num_units=d_model,
        num_output_units=d_model,
        num_heads=num_heads,
        scope="decoder_block_self_attention",
        reuse=tf.AUTO_REUSE,
        query_masks=seq_mask_ph,
        key_masks=seq_mask_ph,
        pairwise_time_gaps=pairwise_time_gaps,
        use_micro_bias=True,
        micro_gamma=config.micro_gamma,
        micro_tau_init=config.micro_tau_init,
        micro_alpha_init=config.micro_alpha_init
    )  # [N, L, D]

    # index 0 = most recent behavior
    target_repr = decoder_output[:, 0, :]  # [N, D]

    # user side encoder
    with tf.variable_scope("user_side_encoder", reuse=tf.AUTO_REUSE):
        u1 = single_feature_embedding(user_active_degree_ph, config.user_active_degree_vocab_size, 4, "user_active_degree_emb")
        u2 = single_feature_embedding(is_lowactive_period_ph, 2, 2, "is_lowactive_period_emb")
        u3 = single_feature_embedding(is_live_streamer_ph, 2, 2, "is_live_streamer_emb")
        u4 = single_feature_embedding(is_video_author_ph, 2, 2, "is_video_author_emb")
        u5 = single_feature_embedding(follow_user_num_range_ph, config.range_vocab_size, 4, "follow_user_num_range_emb")
        u6 = single_feature_embedding(fans_user_num_range_ph, config.range_vocab_size, 4, "fans_user_num_range_emb")
        u7 = single_feature_embedding(register_days_range_ph, config.range_vocab_size, 4, "register_days_range_emb")
        u8 = single_feature_embedding(onehot_feat1_ph, config.onehot_feat1_vocab_size, 4, "onehot_feat1_emb")
        u9 = single_feature_embedding(onehot_feat2_ph, config.onehot_feat2_vocab_size, 4, "onehot_feat2_emb")
        u10 = single_feature_embedding(onehot_feat3_ph, config.onehot_feat3_vocab_size, 8, "onehot_feat3_emb")

        user_side_concat = tf.concat([u1, u2, u3, u4, u5, u6, u7, u8, u9, u10], axis=-1)
        user_side_repr = tf.layers.dense(user_side_concat, d_model, activation=tf.nn.relu, name="user_side_proj")

    # target item side encoder
    with tf.variable_scope("target_item_side_encoder", reuse=tf.AUTO_REUSE):
        t1 = single_feature_embedding(target_video_id_ph, config.video_vocab_size, 16, "target_video_id_emb")
        t2 = single_feature_embedding(target_author_id_ph, config.author_vocab_size, 8, "target_author_id_emb")
        t3 = single_feature_embedding(target_video_type_ph, config.video_type_vocab_size, 2, "target_video_type_emb")
        t4 = single_feature_embedding(target_music_type_ph, config.music_type_vocab_size, 4, "target_music_type_emb")

        target_side_concat = tf.concat([t1, t2, t3, t4], axis=-1)
        target_side_repr = tf.layers.dense(target_side_concat, d_model, activation=tf.nn.relu, name="target_side_proj")

    # final fusion + task tower
    final_repr = tf.concat([target_repr, user_side_repr, target_side_repr], axis=-1)
    with tf.variable_scope("task_tower", reuse=tf.AUTO_REUSE):
        h1 = tf.layers.dense(final_repr, 16, activation=tf.nn.relu, name="fc16")
        h2 = tf.layers.dense(h1, 8, activation=tf.nn.relu, name="fc8")
        logits = tf.layers.dense(h2, 1, activation=None, name="fc1")

    probs = tf.sigmoid(logits)
    loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=labels_ph, logits=logits))

    optimizer = tf.train.AdamOptimizer(learning_rate=config.learning_rate)
    train_op = optimizer.minimize(loss)

    return {
        "labels_ph": labels_ph,
        "item_ids_ph": item_ids_ph,
        "seq_mask_ph": seq_mask_ph,
        "hour_ph": hour_ph,
        "weekday_ph": weekday_ph,
        "delta_t_hours_ph": delta_t_hours_ph,
        "seq_time_seconds_ph": seq_time_seconds_ph,

        "user_active_degree_ph": user_active_degree_ph,
        "is_lowactive_period_ph": is_lowactive_period_ph,
        "is_live_streamer_ph": is_live_streamer_ph,
        "is_video_author_ph": is_video_author_ph,
        "follow_user_num_range_ph": follow_user_num_range_ph,
        "fans_user_num_range_ph": fans_user_num_range_ph,
        "register_days_range_ph": register_days_range_ph,
        "onehot_feat1_ph": onehot_feat1_ph,
        "onehot_feat2_ph": onehot_feat2_ph,
        "onehot_feat3_ph": onehot_feat3_ph,

        "target_video_id_ph": target_video_id_ph,
        "target_author_id_ph": target_author_id_ph,
        "target_video_type_ph": target_video_type_ph,
        "target_music_type_ph": target_music_type_ph,

        "loss": loss,
        "train_op": train_op,
        "probs": probs,
        "macro_scalar": macro_scalar,
        "macro_beta": macro_beta,
        "att_vec": att_vec
    }