通過本篇文檔，使用者可以更好的體驗和瞭解OpenSearch行業演算法版推出的定製排序模型功能。
操作步驟
-
在特徵管理中建立如下欄位特徵：（以system_item表為例，如果需要的特徵不在system_item表中，可以先將外部MaxCompute表註冊進來）如title欄位分別建立原值（custom_title）， 分詞後產生lookup特徵（custom_title_match），分詞後統計詞數量（custom_title_len），其他欄位類似，可根據業務需要增加。以ctr中使用的欄位為例：
類似地，description 欄位建立 custom_description、custom_desc_match、custom_desc_len；brand_name 欄位建立 custom_tags、custom_tags_match；category_name 欄位建立 custom_category。
-
結合system_internal表中的內建特徵，以及上步中建立的欄位特徵，下面進行特徵產生（特徵工程）。
以目前CTR常用的特徵產生為例：可以通過OpenApi（建立演算法資源）大量註冊：
其中ResourceType選擇feature_generator，Data填以下內容：（注意每個input features中以custom_開頭的特徵需要提前準備好，如缺失請在第一步中添加）
[
{
"input": {
"features": [
{
"type": "user",
"name": "system_raw_q_ultra"
},
{
"type": "item",
"name": "system_item_id"
}
]
},
"generator": "combo",
"output": "comb_q_nid"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_user_id"
},
{
"type": "item",
"name": "system_item_id"
}
]
},
"generator": "combo",
"output": "comb_uid_nid"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_user_id"
},
{
"type": "item",
"name": "custom_tags"
}
]
},
"generator": "combo",
"output": "comb_uid_tags"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_raw_q_ultra"
},
{
"type": "item",
"name": "custom_tags"
}
]
},
"generator": "combo",
"output": "comb_q_tags"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_exp_time"
}
]
},
"generator": "id",
"output": "exp_time"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "id",
"output": "terms2"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_raw_q_ultra"
}
]
},
"generator": "id",
"output": "raw_q_ultra"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_user_id"
}
]
},
"generator": "id",
"output": "user_id"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_item_id"
}
]
},
"generator": "id",
"output": "item_id"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_description"
}
]
},
"generator": "id",
"output": "description"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_desc_len"
}
]
},
"generator": "id",
"output": "desc_len"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_title"
}
]
},
"generator": "id",
"output": "title"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_title_len"
}
]
},
"generator": "id",
"output": "title_len"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_category"
}
]
},
"generator": "id",
"output": "category"
},
{
"input": {
"features": [
{
"type": "item",
"name": "custom_tags"
}
]
},
"generator": "id",
"output": "tags"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ctr_30"
}
]
},
"generator": "id",
"output": "all_nid_ctr_30"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ctr_7"
}
]
},
"generator": "id",
"output": "all_nid_ctr_7"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ctr_1"
}
]
},
"generator": "id",
"output": "all_nid_ctr_1"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_pv_30"
}
]
},
"generator": "id",
"output": "all_nid_pv_30"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_pv_7"
}
]
},
"generator": "id",
"output": "all_nid_pv_7"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_pv_1"
}
]
},
"generator": "id",
"output": "all_nid_pv_1"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ipv_30"
}
]
},
"generator": "id",
"output": "all_nid_ipv_30"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ipv_7"
}
]
},
"generator": "id",
"output": "all_nid_ipv_7"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_all_nid_ipv_1"
}
]
},
"generator": "id",
"output": "all_nid_ipv_1"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "custom_title_match"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "term_title_match"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "custom_desc_match"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "term_desc_match"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "custom_tags_match"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "term_tags_match"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "system_qterm_match_decay"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "term_os_kw_match"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_query_cnt"
}
]
},
"generator": "id",
"output": "opensearch_query_cnt"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_qterm_cnt"
}
]
},
"generator": "id",
"output": "opensearch_qterm_cnt"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "system_query_ctr_decay"
},
{
"role": "key",
"type": "user",
"name": "system_raw_q_ultra"
}
]
},
"generator": "lookup",
"output": "os_q_ctr_decay"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "system_qterm_ctr_decay"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "os_term_ctr_decay"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "system_query_ctr_decay"
},
{
"role": "key",
"type": "user",
"name": "system_raw_q_ultra"
}
]
},
"generator": "lookup",
"output": "os_q_ctr_decay_nokey"
},
{
"input": {
"features": [
{
"role": "map",
"type": "item",
"name": "system_qterm_ctr_decay"
},
{
"role": "key",
"type": "user",
"name": "system_terms2"
}
]
},
"generator": "lookup",
"output": "os_term_ctr_decay_nokey"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_query_seq_decay"
}
]
},
"generator": "id",
"output": "os_q_seq_decay"
},
{
"input": {
"features": [
{
"type": "item",
"name": "system_qterm_seq_decay"
}
]
},
"generator": "id",
"output": "os_term_seq_decay"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "system_qterm_seq_decay"
}
],
"method": "query_common_ratio"
},
"generator": "overlap",
"output": "os_qterm_q_common_ratio"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "system_qterm_seq_decay"
}
],
"method": "title_common_ratio"
},
"generator": "overlap",
"output": "os_qterm_title_common_ratio"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "custom_title"
}
],
"method": "query_common_ratio"
},
"generator": "overlap",
"output": "title_q_common_ratio"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "custom_title"
}
],
"method": "title_common_ratio"
},
"generator": "overlap",
"output": "title_title_common_ratio"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "custom_description"
}
],
"method": "query_common_ratio"
},
"generator": "overlap",
"output": "desc_q_common_ratio"
},
{
"input": {
"features": [
{
"role": "query",
"type": "user",
"name": "system_terms2"
},
{
"role": "title",
"type": "item",
"name": "custom_description"
}
],
"method": "title_common_ratio"
},
"generator": "overlap",
"output": "desc_title_common_ratio"
},
{
"input": {
"features": [
{
"type": "user",
"name": "system_term_seq_length"
}
],
"dimension": 1
},
"generator": "raw",
"output": "term_seq_length"
}
]建立完成後，可以在對應頁面上進行編輯。
至此特徵準備完成，具體如何使用特徵需要在模型代碼中進行指定。
以快速入門為基礎，修改需要使用的特徵列表。 一般用embedding_colums。
在定製排序模型配置頁面，填寫模型描述名稱（如 model1），並在 Python 檔案中定義 self.embedding_columns 變數，將上述特徵配置中的輸出欄位（如 comb_q_nid、comb_uid_nid、comb_uid_tags 等）作為 embedding 特徵列表傳入模型。
-
在自訂排序模型中指定需要使用的特徵描述和模型描述
在左側導覽列選擇定製排序模型，進入建立排序模型頁面，配置以下參數：選擇目標應用，輸入模型名稱（例如 rank_model，長度1-30個字元，以字母開頭，可包含大小寫字母、數字和底線），模型類型選擇自訂排序模型，按需開啟定時訓練開關，選擇已建立的特徵描述（例如 fg1）和模型描述（例如 model1）。