{
  "id": 74546,
  "title": "How to Mining Category Information?",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/74546",
  "author_name": "FISH",
  "post_date": "2018-12-13T10:45:26.809000",
  "votes": 11,
  "comment_count": 39,
  "views": 0,
  "content": "<p>Each user has multiple behavior data, and if only basic nunique and count features are constructed, a lot of information may be lost. Here I try Count Vectorizer for category features, and it can improve three thousand points. Feeling is still not enough, I want to know what better way to mine category information.\nIn addition, looking for Chinese teammates</p>",
  "messages": [
    {
      "id": 438243,
      "postDate": "2018-12-13T10:45:26.810Z",
      "content": "<p>Each user has multiple behavior data, and if only basic nunique and count features are constructed, a lot of information may be lost. Here I try Count Vectorizer for category features, and it can improve three thousand points. Feeling is still not enough, I want to know what better way to mine category information.\nIn addition, looking for Chinese teammates</p>",
      "rawMarkdown": "Each user has multiple behavior data, and if only basic nunique and count features are constructed, a lot of information may be lost. Here I try Count Vectorizer for category features, and it can improve three thousand points. Feeling is still not enough, I want to know what better way to mine category information.\nIn addition, looking for Chinese teammates",
      "votes": 11
    },
    {
      "id": 441552,
      "postDate": "2018-12-18T19:34:19.133Z",
      "content": "<p>Luobinli, I would like to share this article related to FFM and its implementation.</p>\n\n<p><a href=\"https://www.analyticsvidhya.com/blog/2018/01/factorization-machines/\">https://www.analyticsvidhya.com/blog/2018/01/factorization-machines/</a></p>",
      "rawMarkdown": "Luobinli, I would like to share this article related to FFM and its implementation.\n\nhttps://www.analyticsvidhya.com/blog/2018/01/factorization-machines/",
      "votes": 8
    },
    {
      "id": 438677,
      "postDate": "2018-12-14T02:40:38.650Z",
      "content": "<p>Maybe you can use the FFM on category feature,it is help me to improve 5 thousand points</p>",
      "rawMarkdown": "Maybe you can use the FFM on category feature,it is help me to improve 5 thousand points",
      "votes": 5,
      "replies": [
        {
          "id": 439348,
          "postDate": "2018-12-15T08:14:33.723Z",
          "content": "<p>You mean make a feature crosses? thank you very much！</p>",
          "rawMarkdown": "You mean make a feature crosses? thank you very much！"
        },
        {
          "id": 439856,
          "postDate": "2018-12-16T14:25:26.633Z",
          "content": "<p>I used it on binary task</p>",
          "rawMarkdown": "I used it on binary task",
          "votes": 1
        },
        {
          "id": 441313,
          "postDate": "2018-12-18T14:35:56.577Z",
          "content": "<p>hi, what did you use as a target for the FFM binary classification?</p>",
          "rawMarkdown": "hi, what did you use as a target for the FFM binary classification?",
          "votes": 1
        },
        {
          "id": 443649,
          "postDate": "2018-12-22T03:10:57.447Z",
          "content": "<p>But how to transfer this regression problem to binary task... I am stuck.</p>",
          "rawMarkdown": "But how to transfer this regression problem to binary task... I am stuck.",
          "votes": 1
        }
      ]
    },
    {
      "id": 448793,
      "postDate": "2019-01-02T05:15:32.100Z",
      "content": "<p>are u using count vectorize in category-1,category-2 and category-3 variables? how r u doing that. any reference would be helpful..</p>",
      "rawMarkdown": "are u using count vectorize in category-1,category-2 and category-3 variables? how r u doing that. any reference would be helpful..",
      "votes": 2
    },
    {
      "id": 438350,
      "postDate": "2018-12-13T14:22:37.970Z",
      "content": "<p>还是中文聊天比较爽，哈哈。。。</p>",
      "rawMarkdown": "还是中文聊天比较爽，哈哈。。。",
      "replies": [
        {
          "id": 440809,
          "postDate": "2018-12-18T01:46:18.003Z",
          "content": "<p>哇 看到中文 好亲切</p>",
          "rawMarkdown": "哇 看到中文 好亲切"
        }
      ]
    },
    {
      "id": 441303,
      "postDate": "2018-12-18T14:20:10.010Z",
      "content": "<p>活捉鱼佬。。。。</p>",
      "rawMarkdown": "活捉鱼佬。。。。"
    },
    {
      "id": 439948,
      "postDate": "2018-12-16T17:56:41.897Z",
      "content": "<p>I think if you one-hot all the categorical features in the historical or new merchants data frames, and after this operation, you aggregate by card id, you can keep a lot of the information, I'm testing this on my model.</p>",
      "rawMarkdown": "I think if you one-hot all the categorical features in the historical or new merchants data frames, and after this operation, you aggregate by card id, you can keep a lot of the information, I'm testing this on my model.",
      "replies": [
        {
          "id": 440061,
          "postDate": "2018-12-17T01:08:19.757Z",
          "content": "<p>Maybe you can just do this one feature by one feature, because the memory problem.</p>",
          "rawMarkdown": "Maybe you can just do this one feature by one feature, because the memory problem."
        },
        {
          "id": 440080,
          "postDate": "2018-12-17T02:15:47.277Z",
          "content": "<p>I do one-hot/aggregate for all raw features(except merchant_id and date related) and get 1484 features. xgboost without parameter tuning get LB score == 3.82, not good enough. May be I should add some statistical features.</p>",
          "rawMarkdown": "I do one-hot/aggregate for all raw features(except merchant_id and date related) and get 1484 features. xgboost without parameter tuning get LB score == 3.82, not good enough. May be I should add some statistical features.",
          "votes": 2
        }
      ]
    },
    {
      "id": 439855,
      "postDate": "2018-12-16T14:22:20.073Z",
      "content": "<p>求当队友</p>",
      "rawMarkdown": "求当队友"
    },
    {
      "id": 439809,
      "postDate": "2018-12-16T12:12:06.640Z",
      "content": "<p>讯飞大佬好</p>",
      "rawMarkdown": "讯飞大佬好",
      "replies": [
        {
          "id": 439817,
          "postDate": "2018-12-16T12:49:08.207Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 439641,
      "postDate": "2018-12-16T01:37:24.800Z",
      "content": "<p>我在想能不能用word2vec对id做向量化提取这样减少损耗</p>",
      "rawMarkdown": "我在想能不能用word2vec对id做向量化提取这样减少损耗",
      "replies": [
        {
          "id": 439654,
          "postDate": "2018-12-16T02:29:19.093Z",
          "content": "<p>可以试试， 感觉先对ID做一些分析比较好，看看有没有什么规律，</p>",
          "rawMarkdown": "可以试试， 感觉先对ID做一些分析比较好，看看有没有什么规律，"
        },
        {
          "id": 439667,
          "postDate": "2018-12-16T03:03:34.517Z",
          "content": "<p>我先试试看看有没有效果</p>",
          "rawMarkdown": "我先试试看看有没有效果"
        },
        {
          "id": 443655,
          "postDate": "2018-12-22T03:30:53.687Z",
          "content": "<p>请问训练word2vec的话，语料怎么构造？？</p>",
          "rawMarkdown": "请问训练word2vec的话，语料怎么构造？？"
        }
      ]
    },
    {
      "id": 439351,
      "postDate": "2018-12-15T08:18:26.877Z",
      "content": "<p>Make Count Vectorizer for feature with many different value(such as city_id) will generate too much features. I just have a memory problem.</p>",
      "rawMarkdown": "Make Count Vectorizer for feature with many different value(such as city_id) will generate too much features. I just have a memory problem.",
      "replies": [
        {
          "id": 450114,
          "postDate": "2019-01-04T10:28:43.653Z",
          "content": "<p>同学，我想请问下，你认为将部分特征进行FFM比较好，还是将整体特征都进行FFM，如果是后者，数据量会不会太大</p>",
          "rawMarkdown": "同学，我想请问下，你认为将部分特征进行FFM比较好，还是将整体特征都进行FFM，如果是后者，数据量会不会太大"
        }
      ]
    },
    {
      "id": 439332,
      "postDate": "2018-12-15T07:03:24.257Z",
      "content": "<p>我做了异常值预测，但是过拟合了，交上去10点多。组队的话，能带一波吗？</p>",
      "rawMarkdown": "我做了异常值预测，但是过拟合了，交上去10点多。组队的话，能带一波吗？",
      "replies": [
        {
          "id": 439633,
          "postDate": "2018-12-16T00:55:44.923Z",
          "content": "<p>不会差那么远吧，我只做异常值预测（相当于二分类），LB score == 3.92，不是很理想，但还是有点作用的。</p>\n\n<p>做异常值预测时，属于类别严重不均衡的分类问题，哎，我不太会处理类别不均衡问题，导致预测效果一般。</p>",
          "rawMarkdown": "不会差那么远吧，我只做异常值预测（相当于二分类），LB score == 3.92，不是很理想，但还是有点作用的。\n\n做异常值预测时，属于类别严重不均衡的分类问题，哎，我不太会处理类别不均衡问题，导致预测效果一般。"
        }
      ]
    },
    {
      "id": 438407,
      "postDate": "2018-12-13T15:55:20.313Z",
      "content": "<p>准备做关键词进行提取吗？</p>",
      "rawMarkdown": "准备做关键词进行提取吗？"
    },
    {
      "id": 438404,
      "postDate": "2018-12-13T15:46:29.357Z",
      "content": "<p>老哥你用的什么方法，准确率这么高，牛逼了</p>",
      "rawMarkdown": "老哥你用的什么方法，准确率这么高，牛逼了",
      "replies": [
        {
          "id": 438630,
          "postDate": "2018-12-14T00:48:32.133Z",
          "content": "<p>是啊，果然厉害！感觉你俩都是高手啊！</p>",
          "rawMarkdown": "是啊，果然厉害！感觉你俩都是高手啊！"
        },
        {
          "id": 439642,
          "postDate": "2018-12-16T01:38:54.783Z",
          "content": "<p>高手二字愧不敢当，就是在这玩玩，哈哈哈</p>",
          "rawMarkdown": "高手二字愧不敢当，就是在这玩玩，哈哈哈"
        }
      ]
    },
    {
      "id": 438351,
      "postDate": "2018-12-13T14:24:22.903Z",
      "content": "<p>我觉得能准确预测出异常值(target=-33.219281)应该是突破点</p>",
      "rawMarkdown": "我觉得能准确预测出异常值(target=-33.219281)应该是突破点",
      "replies": [
        {
          "id": 438372,
          "postDate": "2018-12-13T14:52:11.933Z",
          "content": "<p>嗯呢，在考虑去预测异常值，预测的概率值作为特征，或者作为修正方法</p>",
          "rawMarkdown": "嗯呢，在考虑去预测异常值，预测的概率值作为特征，或者作为修正方法",
          "votes": 1
        }
      ]
    },
    {
      "id": 440852,
      "postDate": "2018-12-18T03:01:58.623Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 440853,
          "postDate": "2018-12-18T03:04:18.627Z",
          "content": "<p>what does you mean? thank you !</p>",
          "rawMarkdown": "what does you mean? thank you !"
        }
      ]
    },
    {
      "id": 438247,
      "postDate": "2018-12-13T10:55:34.963Z",
      "rawMarkdown": "",
      "isDeleted": true,
      "replies": [
        {
          "id": 438306,
          "postDate": "2018-12-13T13:08:52.330Z",
          "content": "<p>你这中文回答让我猝不及防，你这样就可以啊</p>",
          "rawMarkdown": "你这中文回答让我猝不及防，你这样就可以啊",
          "votes": 1
        },
        {
          "id": 438330,
          "postDate": "2018-12-13T13:39:02.560Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 438371,
          "postDate": "2018-12-13T14:51:02.057Z",
          "content": "<p>对类别特征进行词频统计，保留某个特征所有类别信息</p>",
          "rawMarkdown": "对类别特征进行词频统计，保留某个特征所有类别信息"
        },
        {
          "id": 438676,
          "postDate": "2018-12-14T02:35:01.143Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 439837,
          "postDate": "2018-12-16T13:57:22.933Z",
          "content": "<p>我对  ['category_2','category_3','state_id','subsector_id']进行了Count Vectorizer,似乎没有效果。你的Cunnt Vectorizer 有效果？</p>",
          "rawMarkdown": "我对  ['category_2','category_3','state_id','subsector_id']进行了Count Vectorizer,似乎没有效果。你的Cunnt Vectorizer 有效果？"
        },
        {
          "id": 439926,
          "postDate": "2018-12-16T16:54:39.870Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 441552,
      "author_name": "Ashish Gupta",
      "author_url": "",
      "post_date": "2018-12-18T19:34:19.133000",
      "content": "<p>Luobinli, I would like to share this article related to FFM and its implementation.</p>\n\n<p><a href=\"https://www.analyticsvidhya.com/blog/2018/01/factorization-machines/\">https://www.analyticsvidhya.com/blog/2018/01/factorization-machines/</a></p>",
      "votes": 8,
      "replies": []
    },
    {
      "id": 438677,
      "author_name": "Angle",
      "author_url": "",
      "post_date": "2018-12-14T02:40:38.650000",
      "content": "<p>Maybe you can use the FFM on category feature,it is help me to improve 5 thousand points</p>",
      "votes": 5,
      "replies": [
        {
          "id": 439348,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2018-12-15T08:14:33.723000",
          "content": "<p>You mean make a feature crosses? thank you very much！</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 439856,
          "author_name": "Angle",
          "author_url": "",
          "post_date": "2018-12-16T14:25:26.633000",
          "content": "<p>I used it on binary task</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 441313,
          "author_name": "ykskks",
          "author_url": "",
          "post_date": "2018-12-18T14:35:56.577000",
          "content": "<p>hi, what did you use as a target for the FFM binary classification?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 443649,
          "author_name": "Nicotine1026",
          "author_url": "",
          "post_date": "2018-12-22T03:10:57.447000",
          "content": "<p>But how to transfer this regression problem to binary task... I am stuck.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 448793,
      "author_name": "SubikashPal",
      "author_url": "",
      "post_date": "2019-01-02T05:15:32.100000",
      "content": "<p>are u using count vectorize in category-1,category-2 and category-3 variables? how r u doing that. any reference would be helpful..</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 438350,
      "author_name": "Zhuang Jia",
      "author_url": "",
      "post_date": "2018-12-13T14:22:37.970000",
      "content": "<p>还是中文聊天比较爽，哈哈。。。</p>",
      "votes": 0,
      "replies": [
        {
          "id": 440809,
          "author_name": "whoami",
          "author_url": "",
          "post_date": "2018-12-18T01:46:18.003000",
          "content": "<p>哇 看到中文 好亲切</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 441303,
      "author_name": "FlyingFater",
      "author_url": "",
      "post_date": "2018-12-18T14:20:10.010000",
      "content": "<p>活捉鱼佬。。。。</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 439948,
      "author_name": "C4rl05/V",
      "author_url": "",
      "post_date": "2018-12-16T17:56:41.897000",
      "content": "<p>I think if you one-hot all the categorical features in the historical or new merchants data frames, and after this operation, you aggregate by card id, you can keep a lot of the information, I'm testing this on my model.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 440061,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2018-12-17T01:08:19.757000",
          "content": "<p>Maybe you can just do this one feature by one feature, because the memory problem.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 440080,
          "author_name": "Zhuang Jia",
          "author_url": "",
          "post_date": "2018-12-17T02:15:47.277000",
          "content": "<p>I do one-hot/aggregate for all raw features(except merchant_id and date related) and get 1484 features. xgboost without parameter tuning get LB score == 3.82, not good enough. May be I should add some statistical features.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 439855,
      "author_name": "Jianfei Wang",
      "author_url": "",
      "post_date": "2018-12-16T14:22:20.073000",
      "content": "<p>求当队友</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 439809,
      "author_name": "yuyi",
      "author_url": "",
      "post_date": "2018-12-16T12:12:06.640000",
      "content": "<p>讯飞大佬好</p>",
      "votes": 0,
      "replies": [
        {
          "id": 439817,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-12-16T12:49:08.207000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 439641,
      "author_name": "归来仍是少年",
      "author_url": "",
      "post_date": "2018-12-16T01:37:24.800000",
      "content": "<p>我在想能不能用word2vec对id做向量化提取这样减少损耗</p>",
      "votes": 0,
      "replies": [
        {
          "id": 439654,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2018-12-16T02:29:19.093000",
          "content": "<p>可以试试， 感觉先对ID做一些分析比较好，看看有没有什么规律，</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 439667,
          "author_name": "归来仍是少年",
          "author_url": "",
          "post_date": "2018-12-16T03:03:34.517000",
          "content": "<p>我先试试看看有没有效果</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 443655,
          "author_name": "Jacksonler",
          "author_url": "",
          "post_date": "2018-12-22T03:30:53.687000",
          "content": "<p>请问训练word2vec的话，语料怎么构造？？</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 439351,
      "author_name": "Jianfei Wang",
      "author_url": "",
      "post_date": "2018-12-15T08:18:26.877000",
      "content": "<p>Make Count Vectorizer for feature with many different value(such as city_id) will generate too much features. I just have a memory problem.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 450114,
          "author_name": "Grant",
          "author_url": "",
          "post_date": "2019-01-04T10:28:43.653000",
          "content": "<p>同学，我想请问下，你认为将部分特征进行FFM比较好，还是将整体特征都进行FFM，如果是后者，数据量会不会太大</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 439332,
      "author_name": "MissSloane",
      "author_url": "",
      "post_date": "2018-12-15T07:03:24.257000",
      "content": "<p>我做了异常值预测，但是过拟合了，交上去10点多。组队的话，能带一波吗？</p>",
      "votes": 0,
      "replies": [
        {
          "id": 439633,
          "author_name": "Zhuang Jia",
          "author_url": "",
          "post_date": "2018-12-16T00:55:44.923000",
          "content": "<p>不会差那么远吧，我只做异常值预测（相当于二分类），LB score == 3.92，不是很理想，但还是有点作用的。</p>\n\n<p>做异常值预测时，属于类别严重不均衡的分类问题，哎，我不太会处理类别不均衡问题，导致预测效果一般。</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 438407,
      "author_name": "归来仍是少年",
      "author_url": "",
      "post_date": "2018-12-13T15:55:20.313000",
      "content": "<p>准备做关键词进行提取吗？</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 438404,
      "author_name": "归来仍是少年",
      "author_url": "",
      "post_date": "2018-12-13T15:46:29.357000",
      "content": "<p>老哥你用的什么方法，准确率这么高，牛逼了</p>",
      "votes": 0,
      "replies": [
        {
          "id": 438630,
          "author_name": "Zhuang Jia",
          "author_url": "",
          "post_date": "2018-12-14T00:48:32.133000",
          "content": "<p>是啊，果然厉害！感觉你俩都是高手啊！</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 439642,
          "author_name": "归来仍是少年",
          "author_url": "",
          "post_date": "2018-12-16T01:38:54.783000",
          "content": "<p>高手二字愧不敢当，就是在这玩玩，哈哈哈</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 438351,
      "author_name": "Zhuang Jia",
      "author_url": "",
      "post_date": "2018-12-13T14:24:22.903000",
      "content": "<p>我觉得能准确预测出异常值(target=-33.219281)应该是突破点</p>",
      "votes": 0,
      "replies": [
        {
          "id": 438372,
          "author_name": "FISH",
          "author_url": "",
          "post_date": "2018-12-13T14:52:11.933000",
          "content": "<p>嗯呢，在考虑去预测异常值，预测的概率值作为特征，或者作为修正方法</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 440852,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-12-18T03:01:58.623000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 440853,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2018-12-18T03:04:18.627000",
          "content": "<p>what does you mean? thank you !</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 438247,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-12-13T10:55:34.963000",
      "content": "",
      "votes": 0,
      "replies": [
        {
          "id": 438306,
          "author_name": "FISH",
          "author_url": "",
          "post_date": "2018-12-13T13:08:52.330000",
          "content": "<p>你这中文回答让我猝不及防，你这样就可以啊</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 438330,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-12-13T13:39:02.560000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 438371,
          "author_name": "FISH",
          "author_url": "",
          "post_date": "2018-12-13T14:51:02.057000",
          "content": "<p>对类别特征进行词频统计，保留某个特征所有类别信息</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 438676,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-12-14T02:35:01.143000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 439837,
          "author_name": "Jianfei Wang",
          "author_url": "",
          "post_date": "2018-12-16T13:57:22.933000",
          "content": "<p>我对  ['category_2','category_3','state_id','subsector_id']进行了Count Vectorizer,似乎没有效果。你的Cunnt Vectorizer 有效果？</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 439926,
          "author_name": "",
          "author_url": "",
          "post_date": "2018-12-16T16:54:39.870000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "438243": "Each user has multiple behavior data, and if only basic nunique and count features are constructed, a lot of information may be lost. Here I try Count Vectorizer for category features, and it can improve three thousand points. Feeling is still not enough, I want to know what better way to mine category information.\nIn addition, looking for Chinese teammates",
    "441552": "Luobinli, I would like to share this article related to FFM and its implementation.\n\nhttps://www.analyticsvidhya.com/blog/2018/01/factorization-machines/",
    "438677": "Maybe you can use the FFM on category feature,it is help me to improve 5 thousand points",
    "448793": "are u using count vectorize in category-1,category-2 and category-3 variables? how r u doing that. any reference would be helpful..",
    "438350": "还是中文聊天比较爽，哈哈。。。",
    "441303": "活捉鱼佬。。。。",
    "439948": "I think if you one-hot all the categorical features in the historical or new merchants data frames, and after this operation, you aggregate by card id, you can keep a lot of the information, I'm testing this on my model.",
    "439855": "求当队友",
    "439809": "讯飞大佬好",
    "439641": "我在想能不能用word2vec对id做向量化提取这样减少损耗",
    "439351": "Make Count Vectorizer for feature with many different value(such as city_id) will generate too much features. I just have a memory problem.",
    "439332": "我做了异常值预测，但是过拟合了，交上去10点多。组队的话，能带一波吗？",
    "438407": "准备做关键词进行提取吗？",
    "438404": "老哥你用的什么方法，准确率这么高，牛逼了",
    "438351": "我觉得能准确预测出异常值(target=-33.219281)应该是突破点",
    "440852": "",
    "438247": ""
  }
}