2019-08-29工作进展

昨天工作：

1. 将图片数据中不重要的剔除，剩下7000+w张商品图片，特征提取程序还在进行；
2. 对dssm程序的训练集中负样本不合理的例子进行处理，比如query是连衣裙的话，而title的分词结果中如果有“连衣裙”这个词，那这条数据应该划分到正样本中，这样的例子在负样本中占比10%左右；使用处理后的数据进行模型的训练，inference结果：acc:0.75 auc:0.702 precision:0.93；acc结果是当前最优，但是auc和precision较之前最优结果要差一点；
3. 准备转正答辩ppt

今天计划：
图片特征提取完毕之后进行mvdssm网络的训练

1. 关键词法的结果：

inference_query : hs_dssm_dic_query_inf_7 - | query_id | query |
hs_dssm_dic_query_inf_11 : | id | query_emb |
inference_title : hs_dssm_dic_title_inf_10 - | item_id | title |
hs_dssm_dic_title_inf_14 : | id | title_emb |

hs_tmp_157 : | title_id | query_id |

create table hs_tmp_336 as select c., d.title from
(select a., b.query from hs_tmp_157 a join hs_dssm_dic_query_inf_7 b on a.query_id == b.query_id)c join hs_dssm_dic_title_inf_10 d on c.title_id == d.item_id;

create table hs_tmp_337 as select *, search_kg:alinlp_segment(hs_return_clean(title), "MAINSE", "0", "1") as title_seg, search_kg:alinlp_segment(hs_return_clean(query), "MAINSE", "0", "1") as query_seg from hs_tmp_336;

drop table hs_tmp_338;
yes
create table hs_tmp_338 as select hs_return_direct_score_2(title_id, query_id, query, title, title_seg, query_seg) as (title_id, query_id, query, title, title_seg, query_seg, score) from hs_tmp_337;

create table hs_tmp_339 as select * from hs_tmp_338 where score = 0;