{"body":[{"id":"46956778243995648","sort":"1","type":1,"role":"46955279031971266","content":"周啟源覺得舍友謝晨皓說得有道理。作為一個隊長，自己不能太專制，把所有事情都留給自己干。"},{"id":"46956778780852629","sort":"1025","type":1,"role":"46955279031971266","content":"何況，他覺得，一個小隊，實際上，也是成員與成員之間的“開源”。既然自己已成寫完第一問，也有了初步的論文作品，也應該開誠布公，和他們討論一番。"},{"id":"46956780911559921","sort":"2049","type":1,"role":"46955279031971266","content":"于是，三人約好在某個教室，討論題目的解法。如果可以，周啟源還打算安排任務。"},{"id":"46956782270527667","sort":"3073","type":1,"role":"46955279031971266","content":"場景：教學樓的某一教室中"},{"id":"46956783881131257","sort":"4097","type":1,"role":"46955279031971266","content":"周啟源將帶來了兩瓶核桃奶飲料，放在隊友許曉勝、曾喜緣的桌子上。"},{"id":"46956787370798601","sort":"5121","type":1,"role":"46956532088681638","content":"這是我們?nèi)齻€人第一次見面哈。"},{"id":"46956798628307543","sort":"6145","type":1,"role":"46956791665755282","content":"好貼心哦，不過，你還是坐中間吧，這樣方便一點。"},{"id":"46956799970488564","sort":"7169","type":1,"role":"46955279031971266","content":"本來許曉勝坐最右邊，曾喜緣坐中間，周啟源坐在最里面。這么一說，曾喜緣、周啟源兩人覺得有理，就彼此對調(diào)了座位。"},{"id":"46956803997008642","sort":"8193","type":1,"role":"46956532088681638","content":"（打開電腦，展示論文）第一問，根據(jù)污水特征，建立出水指標預測模型，已經(jīng)寫完了。不過，我覺得預測的效果，似乎不太好。"},{"id":"46956807235019886","sort":"9217","type":1,"role":"46956795692292232","content":"你是用什么方法做的呢？"},{"id":"46956809382509375","sort":"10241","type":1,"role":"46956532088681638","content":"我用的 BP 神經(jīng)網(wǎng)絡。評價指標是 R 方、訓練集、測試集按 7:3 劃分。你們看， R 方的值才 0.6 那樣。"},{"id":"46956813660686609","sort":"11265","type":1,"role":"46955279031971266","content":"注：類似于統(tǒng)計學的一元線性回歸，BP 神經(jīng)網(wǎng)絡也是根據(jù)輸入，預測輸出的。R 方是根據(jù)實際值、預測值算出來，評價模型的擬合優(yōu)度指標。越接近于一，則模型預測結果越準確。"},{"id":"46956814734435124","sort":"12289","type":1,"role":"46955279031971266","content":"注：訓練集是用于訓練模型的參數(shù)，測試集用于評價模型。"},{"id":"46956819029401857","sort":"13313","type":1,"role":"46956795692292232","content":"（小聲嘀咕）BP 神經(jīng)網(wǎng)絡年代很久遠了。"},{"id":"46956821176886774","sort":"14337","type":1,"role":"46956791665755282","content":"是的...吧？"},{"id":"46956823324371290","sort":"15361","type":1,"role":"46956532088681638","content":"那有其他改進的意見嗎？"},{"id":"46956824951757738","sort":"16385","type":1,"role":"46956795692292232","content":"我覺得可以用 SVG。"},{"id":"46956827082469441","sort":"17409","type":1,"role":"46956532088681638","content":"SVG？呃...不好意思，我從來沒聽過 SVG 耶..."},{"id":"46956828961508724","sort":"18433","type":1,"role":"46956795692292232","content":"就是那個，emmm，支持...向量..."},{"id":"46956832451179821","sort":"19457","type":1,"role":"46956532088681638","content":"（心想）支持向量機，簡稱 SVM。在特定情況下，也叫支持向量機回歸，即 SVR。可能是曾喜緣記錯了。"},{"id":"46956834615435077","sort":"20481","type":1,"role":"46956532088681638","content":"（心想）不過，為了避免團隊傷了和氣，自己也沒必要去糾正他的錯誤。然而，在使用 BP 神經(jīng)網(wǎng)絡之前，周啟源已經(jīng)嘗試了，包括 SVM 在內(nèi)的許多機器學習模型?？上ВЧ疾皇呛芎?。"},{"id":"46956836746135850","sort":"21505","type":1,"role":"46956791665755282","content":"我們先不說這個了，看看論文吧。"},{"id":"46956838625184607","sort":"22529","type":1,"role":"46955279031971266","content":"三人大致瀏覽了 15 分鐘"},{"id":"46956841577971740","sort":"23553","type":1,"role":"46956532088681638","content":"（自謙）你們也看到了，在語言表達上，我不是一個能手......"},{"id":"46956845621289354","sort":"24577","type":1,"role":"46956791665755282","content":"（打斷，興奮，提高音量）沒事呀。有我們呀，我們來改！"},{"id":"46956848020432156","sort":"25601","type":1,"role":"46956532088681638","content":"我有一個想法，就是之所以 BP 神經(jīng)網(wǎng)絡效果不好，是否因為數(shù)據(jù)帶有時序性呢？也就是說，當前的污水處理的出水，與上一次污水的入水、出水有關呢？"},{"id":"46956850184688212","sort":"26625","type":1,"role":"46956791665755282","content":"是的...吧？"},{"id":"46956852600609335","sort":"27649","type":1,"role":"46955279031971266","content":"曾喜緣目視前方，冷靜思考中"},{"id":"46956857952550755","sort":"28673","type":1,"role":"46956532088681638","content":"（興奮，望向曾喜緣）那么，我可以嘗試，用一些時序預測模型，或許會更好哦。"},{"id":"46956859848360492","sort":"29697","type":1,"role":"46956791665755282","content":"是的...吧？"},{"id":"46956864663432577","sort":"30721","type":1,"role":"46956795692292232","content":"（手捂著額頭）先不說這個，你數(shù)據(jù)預處理方面，是怎么做的？"},{"id":"46956867347790922","sort":"31745","type":1,"role":"46955279031971266","content":"注：機器學習中，為了提高模型的精確性，很有必要進行數(shù)據(jù)的預處理。如將取值范圍弄成 0 到 1，即歸一化；或者均值，方差弄成 0 和 1，即標準化；當然，還有缺失數(shù)據(jù)的處理、以及自變量的降維等"},{"id":"46956869763695514","sort":"32769","type":1,"role":"46956532088681638","content":"因為缺失數(shù)據(jù)比較少，我就直接刪除了。你們覺得，要不要結合整體數(shù)據(jù)，算出他的分布，再填補進去？這樣做，雖然會讓樣本增多，但準確性就降下來了。"},{"id":"46956872179616565","sort":"33793","type":1,"role":"46956791665755282","content":"是的...吧？"},{"id":"46956874058676286","sort":"34817","type":1,"role":"46956532088681638","content":"由于自變量個數(shù)太多，我就用 PCA 降了維。"},{"id":"46956877565104443","sort":"35841","type":1,"role":"46956795692292232","content":"（小聲嘀咕）PCA 是一個很老舊的方法了..."},{"id":"46956880786338373","sort":"36865","type":1,"role":"46956532088681638","content":"（誠懇）那有什么更好的方法沒？"},{"id":"46956882917040381","sort":"37889","type":1,"role":"46955279031971266","content":"曾喜緣不語，把頭擺向一邊，冷靜思考中。"},{"id":"46956885081308963","sort":"38913","type":1,"role":"46955279031971266","content":"會議似乎是周啟源的一言堂，三人繼續(xù)討論，“一致”認為第一問應該使用，時序模型進行預測。至于第二問，尚未有商討結果?，F(xiàn)在已經(jīng)是下午 5 點了，三人都略有疲憊。"},{"id":"46956887212009365","sort":"39937","type":1,"role":"46956795692292232","content":"你想得很好，但我總覺得缺了點什么。"},{"id":"46956890164793705","sort":"40961","type":1,"role":"46956791665755282","content":"是的...吧？可能還需要再改進一下。"},{"id":"46956892597495822","sort":"41985","type":1,"role":"46956532088681638","content":"嗯，我再試一試，看看時序神經(jīng)網(wǎng)絡的效果怎么樣。如果可以，你們寫一下第二問吧。我自己也看看，能不能把第二問順帶解決了。"},{"id":"46956896070382125","sort":"43009","type":1,"role":"46956791665755282","content":"那么，時間不早了，我先走了。"},{"id":"46956898486290903","sort":"44033","type":1,"role":"46956795692292232","content":"我也是，我晚上還由課。曉勝，我們一起去吃飯吧。"},{"id":"46956900382124123","sort":"45057","type":1,"role":"46956532088681638","content":"嗯，那你們慢走。"},{"id":"46956902261173050","sort":"46081","type":1,"role":"46955279031971266","content":"許曉勝、曾喜緣下，只剩周啟源一人，獨自打代碼。約莫 6 點左右，周啟源也去飯?zhí)贸燥埩?}],"sentencecount":46}