fg.json と EasyRec による PAI-Rec 特徴量パイプラインの構築 - PAI-Rec

このトピックでは、特徴生成ファイル fg.json とモデル構成ファイル config の構成例について説明します。

サンプルデータ

以下のサンプルデータには、一般的なデータの型のフィールドが含まれています：

ID 特徴： user_id、item_id
カテゴリ特徴： gender、category
数値特徴： age
ルックアップ特徴： user__kv_category_click_1d
多値特徴： tags
テキスト特徴： description
行動シーケンス特徴： click_10_seq
サンプルラベル： is_click

説明

ルックアップ特徴とは、キーバリュー (KV) ストア、キャッシュ、データベーステーブルなど、外部の事前に計算されたマッピングテーブルから特徴値を取得するメカニズムです。これにより、モデルはトレーニングや予測中にこれらの事前計算された値を迅速にクエリできます。これにより、リアルタイムコンピューティングや複雑なデータ処理ロジックが不要になります。このメソッドは、レコメンデーションシステムやクリック率 (CTR) 予測などのシナリオで一般的です。例としては、プロダクトカテゴリに対するユーザーの過去のクリック数などの行動統計や、アイテムの人気度などが挙げられます。

フィールド名	サンプルデータ 1	サンプルデータ 2	サンプルデータ 3
request_id	101	102	103
user_id	1	2	3
item_id	4	5	10
event_unix_time	1672502400	1672502400	1672502400
is_click	0	1	1
age	25	30	22
gender	Male	Female	Female
user__kv_category_click_1d	Electronics:10 Appliances:1 Accessories:2	Electronics:1 Appliances:5 Accessories:1	Electronics:1 Appliances:2 Accessories:11
category	Electronics	Appliances	Accessories
tags	Tech Computer Portable	Home Appliance Refrigerated	Fashion Glasses Sunscreen
description	Portable high-performance laptop	Large-capacity refrigerator	Fashionable sunglasses
click_10_seq	item__item_id:4#item__category:Electronics#user__ts:21041;item__item_id:5#item__category:Appliances#user__ts:168139;item__item_id:10#item__category:Accessories#user__ts:168284
ds	20230101	20230101	20230101

特徴生成構成ファイル fg.json

以下に、構成済みの fg.json ファイルの例を示します。

{
    "features": [
        {
            "feature_name": "user_id",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "user:user_id",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": false
        },
        {
            "feature_name": "item_id",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "item:item_id",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": false
        },
        {
            "feature_name": "age",
            "feature_type": "raw_feature",
            "value_type": "Double",
            "expression": "user:age",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false
        },
        {
            "feature_name": "gender",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "user:gender",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": false
        },
        {
            "feature_name": "user__kv_category_click_1d",
            "feature_type": "lookup_feature",
            "value_type": "Double",
            "map": "user:user__kv_category_click_1d",
            "key": "item:category",
            "needDiscrete": false,
            "needWeighting": false,
            "needKey": false,
            "default_value": "0",
            "combiner": "mean",
            "need_prefix": false
        },
        {
            "feature_name": "category",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "item:category",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": false
        },
        {
            "feature_name": "tags",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "item:tags",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": true
        },
        {
            "feature_name": "description",
            "feature_type": "id_feature",
            "value_type": "String",
            "expression": "item:description",
            "default_value": "-1024",
            "combiner": "mean",
            "need_prefix": false,
            "is_multi": true
        },
        {
            "sequence_name": "click_10_seq",
            "sequence_column": "click_10_seq",
            "sequence_length": 10,
            "sequence_delim": ";",
            "attribute_delim": "#",
            "sequence_table": "item",
            "sequence_pk": "user:click_10_seq",
            "features": [
                {
                    "feature_name": "item_id",
                    "feature_type": "id_feature",
                    "value_type": "String",
                    "expression": "item:item_id",
                    "default_value": "-1024",
                    "combiner": "mean",
                    "need_prefix": false,
                    "is_multi": false,
                    "group": "click_10_seq_feature"
                },
                {
                    "feature_name": "category",
                    "feature_type": "id_feature",
                    "value_type": "String",
                    "expression": "item:category",
                    "default_value": "-1024",
                    "combiner": "mean",
                    "need_prefix": false,
                    "is_multi": false,
                    "group": "click_10_seq_feature"
                },
                {
                    "feature_name": "ts",
                    "feature_type": "raw_feature",
                    "value_type": "Double",
                    "expression": "user:ts",
                    "default_value": "-1024",
                    "combiner": "mean",
                    "need_prefix": false,
                    "group": "click_10_seq_feature"
                }
            ]
        }
    ],
    "reserves": [
        "request_id",
        "user_id",
        "item_id",
        "is_click"
    ]
}

モデル構成ファイル config

以下に、構成済みのモデル構成ファイル config の例を示します。この例には、コンボ特徴と式特徴が含まれています。詳細については、easy_rec をご参照ください。

train_config {
  optimizer_config {
    use_moving_average: false
    adam_optimizer {
      learning_rate {
        exponential_decay_learning_rate {
          initial_learning_rate: 0.001
          decay_steps: 1
          decay_factor: 0.5
          min_learning_rate: 1e-06
        }
      }
    }
  }
  num_steps: 1
  sync_replicas: true
  save_summary_steps: 100
  log_step_count_steps: 100
}
eval_config {
  metrics_set {
    auc {
    }
  }
}
data_config {
  batch_size: 1024
  label_fields: "is_click"
  shuffle: false
  num_epochs: 10000
  input_type: OdpsRTPInput
  separator: ""
  selected_cols: "is_click,features"
  input_fields {
    input_name: "is_click"
    input_type: INT32
    default_val: "0"
  }
  input_fields {
    input_name: "user_id"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "item_id"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "age"
    input_type: DOUBLE
    default_val: "-1024"
  }
  input_fields {
    input_name: "gender"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "user__kv_category_click_1d"
    input_type: DOUBLE
    default_val: "0"
  }
  input_fields {
    input_name: "category"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "tags"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "description"
    input_type: STRING
    default_val: "-1024"
  }
  input_fields {
    input_name: "click_10_seq__item_id"
    input_type: STRING
  }
  input_fields {
    input_name: "click_10_seq__category"
    input_type: STRING
  }
  input_fields {
    input_name: "click_10_seq__ts"
    input_type: STRING
  }
  pai_worker_queue: true
}
feature_configs {
  input_names: "user_id"
  feature_type: IdFeature
  embedding_dim: 8
  hash_bucket_size: 48000
  separator: ""
  combiner: "mean"
}
feature_configs {
  input_names: "item_id"
  feature_type: IdFeature
  embedding_dim: 8
  hash_bucket_size: 27000
  separator: ""
  combiner: "mean"
}
feature_configs {
  input_names: "age"
  feature_type: RawFeature
  embedding_dim: 4
  separator: ""
  boundaries: 1e-08
  boundaries: 10
  boundaries: 20
  boundaries: 30
  boundaries: 40
  boundaries: 50
  boundaries: 60
}
feature_configs {
  input_names: "gender"
  feature_type: IdFeature
  embedding_dim: 4
  hash_bucket_size: 10
  separator: ""
  combiner: "mean"
}
feature_configs {
  input_names: "user__kv_category_click_1d"
  feature_type: RawFeature
  embedding_dim: 4
  separator: ""
  boundaries: 1e-08
  boundaries: 1.0
  boundaries: 2.0
  boundaries: 3.0
  boundaries: 4.0
  boundaries: 5.0
  boundaries: 6.0
}
feature_configs {
  input_names: "category"
  feature_type: IdFeature
  embedding_dim: 4
  hash_bucket_size: 100
  separator: ""
  combiner: "mean"
}
feature_configs {
  input_names: "tags"
  feature_type: TagFeature
  embedding_dim: 4
  hash_bucket_size: 1000
  separator: ""
  combiner: "mean"
}
feature_configs {
  input_names: "description"
  feature_type: SequenceFeature
  embedding_dim: 4
  hash_bucket_size: 10
  separator: ""
  sequence_combiner {
    text_cnn {
      filter_sizes: 2
      filter_sizes: 3
      filter_sizes: 4
      num_filters: 16
      num_filters: 8
      num_filters: 8
    }
  }
}
feature_configs {
  input_names: "click_10_seq__item_id"
  feature_type: SequenceFeature
  embedding_dim: 8
  hash_bucket_size: 27000
  separator: ";"
  combiner: "mean"
  sub_feature_type: IdFeature
}
feature_configs {
  input_names: "click_10_seq__category"
  feature_type: SequenceFeature
  embedding_dim: 4
  hash_bucket_size: 10000
  separator: ";"
  combiner: "mean"
  sub_feature_type: IdFeature
}
feature_configs {
  input_names: "click_10_seq__ts"
  feature_type: SequenceFeature
  embedding_dim: 4
  separator: ";"
  sub_feature_type: RawFeature
}
feature_configs {
  input_names: "click_10_seq__ts"
  feature_type: SequenceFeature
  embedding_dim: 4
  separator: ";"
  sub_feature_type: RawFeature
}
feature_configs {
  input_names: ["age", "gender"]
  feature_name: "combo_age_gender"
  feature_type: ComboFeature
  embedding_dim: 16
  hash_bucket_size: 1000
}
feature_configs {
  input_names: "age"
  feature_name: "age_satisfy1"
  feature_type: ExprFeature
  expression: "age>=18"
}
model_config {
  model_class: "MultiTower"
  feature_groups {
    group_name: "all"
    feature_names: "user_id"
    feature_names: "item_id"
    feature_names: "age"
    feature_names: "gender"
    feature_names: "user__kv_category_click_1d"
    feature_names: "category"
    feature_names: "tags"
    feature_names: "description"
    feature_names: "combo_age_gender"
    feature_names: "age_satisfy1"
    wide_deep: DEEP
    sequence_features {
      group_name: "click_10_seq"
      seq_att_map {
        key: "item_id"
        key: "category"
        hist_seq: "click_10_seq__item_id"
        hist_seq: "click_10_seq__category"
        hist_seq: "click_10_seq__ts"
      }
      tf_summary: false
      allow_key_search: false
      allow_key_transform: true
    }
  }
  embedding_regularization: 5e-06
  multi_tower {
    towers {
      input: "all"
      dnn {
        hidden_units: 256
        hidden_units: 128
      }
    }
    final_dnn {
      hidden_units: 64
      hidden_units: 32
    }
    l2_regularization: 1e-06
  }
}
export_config {
  multi_placeholder: true
}