201108130 六、發明說明： 【發明所屬之技術領域】 本發明係關於網路技術，特別是關於一種網站訪客價 値的計算方法及系統。 【先前技術】 目前，很多企業都透過網站來進行資訊展示和產品行 銷，隨著網路技術的發展，網站和訪客之間也可以實現更 多的互動。例如，網站主可以透過網路管理系統看到訪客 的一些資訊，根據這些資訊來判斷訪客的價値，然後進一 步對具有一定價値的訪客做一些業務行爲，例如與其聯繫 ，向其提供更詳細的資訊等等》 可以想像，對於具有大量訪客的網站而言，如果以人 工方式來判斷每個訪客的價値，工作量大且效率難以保證 。爲了實現對網站訪客價値的自動判斷，現有技術中，是 透過把各類“訪客資訊”中的單個或組合設置爲條件，一 旦滿足條件，就認爲該訪客具有一定價値並通知網站主。 這些條件可以包括：來訪次數超過幾次、訪問頁面超過幾 個、是否訪問過某些特定頁面、訪客是否來自某些特定的 省市，等等。 透過對現有技術的硏究，發明人發現上述方法只適用 於處理簡單的條件組合，然而在實際需求中，訪客資訊和 訪客價値之間往往存在著更爲複雜的對應關係，例如：如 果訪客來自北京，則其訪問頁面A的價値大、如果訪客來 -5- 201108130 自上海，則其訪問頁面B的價値大；男性訪客訪問頁面C 的價値大、女性訪客訪問頁面D的價値大。類似或更爲 複雜的情況還有很多，這些對應關係往往是非線性的，甚 至是不確知的（即網站主並沒有意識到自己的主動行爲中 存在著某種對應關係），如果使用條件組合的方法，難以 把這些複雜的對應關係全部納入考慮。因此，現有的自動 判斷訪客價値的方法，其判斷結果往往會與人工判斷的結 果有很大的偏差，無法達到預期效果。 【發明內容】 有鑒於此，本申請案提供了一種網站訪客價値的計算 方法及系統，以解決現有自動判斷訪客價値的方法，其判 斷結果與人工判斷結果存在很大偏差的問題。技術方案如 下： 本申請案提供一種網站訪客價値的計算方法，包括： 初始化網站訪客價値計算模型，所述計算模型爲神經 網路模型，以訪客資訊爲輸入，以訪客價値爲輸出； 使用資料樣本對所述計算模型進行訓練，確定所述計 算模型； 獲得訪客資訊，使用所確定的計算模型，計算所述訪 客的價値 本申請案還提供一種網站訪客價値的計算系統，包括 初始化單元，用於初始化網站訪客價値計算模型，所 -6- 201108130 述計算模型爲神經網路模型，以訪客資訊爲輸入，以訪客 價値爲輸出： 訓練單兀’用於使用資料樣本對所述計算模型進行訓 練，確定所述計算模型； 計算單元’用於獲得訪客資訊，使用所確定的計算模 型，計算所述訪客的價値。 上述技術方案，以神經網路作爲網站訪客價値的計算 模型。透過訓練可以令模型學習網站主的主動行爲，從而 將業務需求中訪客資訊與訪客價値間各種複雜的對應關係 納入模型。訓練完成後，透過使用該模型，即可根據訪客 的各種資訊，自動計算出該訪客的價値，由於模型是根據 人工判斷的結果訓練出來的，因此，其計算的結果與人工 判斷的結果趨於一致。 【實施方式】 爲了使本技術領域的人員更好地理解本申請案所提供 的方案，下面結合附圖和實施方式對本申請案的實施例作 進一步的詳細說明。 圖1所示爲本申請案實施例一種網站訪客價値的計算 方法的流程圖，具體包括以下步驟： S101，初始化網站訪客價値計算模型，所述計算模型 爲神經網路模型，以訪客資訊爲輸入，以訪客價値爲輸出 爲了實現根據訪客資訊自動計算訪客的價値，需要建 201108130 立各類訪客資訊與訪客價値的對應關係。在實際需求中， 各類訪客資訊與訪客價値的關係往往是非線性甚至不確知 的’因此，選擇神經網路作爲計算模型，可以有效地解決 以上問題。 本實施例所採用的神經網路模型如圖2所示，包括輸 入層、隱含層和輸出層。該模型的數學描述如下： Y = F{X) =f2iW2fx{WxX + Bx) + B2) 其中，X爲輸入向量；Y爲輸出向量；201108130 VI. Description of the Invention: [Technical Field of the Invention] The present invention relates to network technology, and more particularly to a method and system for calculating a website visitor price. [Prior Art] At present, many companies use the website for information display and product marketing. With the development of network technology, more interaction between websites and visitors can be achieved. For example, the website owner can see some information of the visitor through the network management system, judge the price of the visitor based on the information, and then further perform some business activities for the visitor with a pricing defect, for example, to contact him and provide more detailed information to him. Etc. It is conceivable that for a website with a large number of visitors, if the price of each visitor is judged manually, the workload is large and the efficiency is difficult to guarantee. In order to realize the automatic judgment of the website visitor price, in the prior art, by setting a single or a combination of various "visitor information" as a condition, once the condition is satisfied, the visitor is considered to have a pricing price and notify the website owner. These conditions can include: more than a few visits, more than a few pages visited, whether certain pages have been visited, whether visitors are from certain provinces, and so on. Through the study of the prior art, the inventors found that the above method is only suitable for dealing with a simple combination of conditions. However, in actual demand, there is often a more complicated correspondence between visitor information and visitor price, for example, if the visitor comes from In Beijing, the price of visit page A is large. If the visitor comes to -5 - 201108130 from Shanghai, the price of visit page B is large; the price of male visitor C is large, and the price of female visitor D is large. There are many similar or more complicated situations. These correspondences are often non-linear or even unsure (that is, the website owner does not realize that there is a corresponding relationship in his active behavior), if a conditional combination is used. Methods, it is difficult to take all these complex correspondences into consideration. Therefore, the existing method of automatically determining the price of the visitor often has a large deviation from the result of the manual judgment and cannot achieve the desired effect. SUMMARY OF THE INVENTION In view of this, the present application provides a method and system for calculating a website visitor price to solve the existing method for automatically determining the visitor price, and the judgment result is greatly deviated from the manual judgment result. The technical solution is as follows: The application provides a method for calculating a website visitor price, comprising: initializing a website visitor price calculation model, wherein the calculation model is a neural network model, with visitor information as an input, and a visitor price as an output; Performing training on the calculation model to determine the calculation model; obtaining visitor information, and calculating the price of the visitor using the determined calculation model. The application further provides a calculation system for website visitor price, including an initialization unit, for Initialize the website visitor price calculation model, -6- 201108130 The calculation model is a neural network model, with the visitor information as input and the visitor price as the output: the training unit 用于 is used to train the calculation model using the data sample, Determining the calculation model; the calculation unit 'for obtaining visitor information, and using the determined calculation model, calculating the price of the visitor. The above technical solution uses a neural network as a calculation model for website visitor price. Through training, the model can learn the active behavior of the website owner, thus incorporating the complex correspondence between the visitor information and the visitor price in the business demand. After the training is completed, by using the model, the price of the visitor can be automatically calculated according to various information of the visitor. Since the model is trained according to the result of manual judgment, the result of the calculation and the result of the manual judgment tend to Consistent. [Embodiment] In order to make those skilled in the art better understand the solutions provided by the present application, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings and embodiments. FIG. 1 is a flowchart of a method for calculating a website visitor price according to an embodiment of the present application, which specifically includes the following steps: S101: Initializing a website visitor price calculation model, the calculation model is a neural network model, and the visitor information is input. In order to realize the automatic calculation of the visitor's price based on the visitor information, it is necessary to establish 201108130 to establish the correspondence between various visitor information and the visitor price. In actual demand, the relationship between various types of visitor information and visitor price is often non-linear or even uncertain. Therefore, the choice of neural network as a computational model can effectively solve the above problems. The neural network model used in this embodiment is shown in Fig. 2, and includes an input layer, an implicit layer, and an output layer. The mathematical description of the model is as follows: Y = F{X) = f2iW2fx{WxX + Bx) + B2) where X is the input vector; Y is the output vector;
Wi爲隱含層權値矩陣，Βι爲隱含層偏置向量，f!爲 隱含層傳遞函數； W2爲輸出層權値矩陣，B2爲輸出層偏置向量，f2爲 輸出層傳遞函數； 計算模型以訪客資訊爲輸入，X爲η維縱向量，其中 每一維的數値表示一類訪客資訊；以訪客價値爲輸出，Υ 爲m維縱向量；假設在模型的隱含層有k個節點，貝lj Wi 爲kxn矩陣，Βι爲k維縱向量，W2爲mxk矩陣’ B2爲取 維縱向量。 對於本實施例而言，所要計算的訪客價値可以用一個 數値來表示，因此輸出向量Y爲1維向量。需要說明的是 ，輸入向量X和輸出向量Y的維數是由具體需求所決定 的，本實施例並不對其構成限定。 在本實施例中，隱含層傳遞函數fi選用神經元的非線 -8 - 201108130 性作用函數Sigmoid函數，即形如f(x)=i/(i+e·*)的函數； 輸出層傳遞函數f2爲線性函數。理論硏究表明，神經網路 在其隱含層中使用Sigmoid傳遞函數，在輸出層中使用線 性傳遞函數，就能夠以任意精確度逼近任何感興趣的函數 〇 對於隱含層的節點數目k而言，較多的隱含層節點可 以加快神經網路的學習速度，但是也會使模型的結構變得 複雜，降低模型的適應能力。節點的數目可以在訓練的過 程中進行調整，例如採用測試樣本與訓練樣本的誤差交叉 評價的試錯法來選擇隱含層節點數。本領域技術人員也可 以根據具體需求’採用其他方法來確定隱含層節點的數目 ，本實施例不對其進行限定。 S1 〇2，使用資料樣本對所述計算模型進行訓練，確定 所述計算模型； 在S 1 0 1中，對訪客價値計算模型進行了初始化，在 本步驟中，將透過訓練對模型中的、W2、B丨、B2等參 數進行確定。 對於某個網站，可以選取一批歷史資料，作爲用於訓 練的樣本。其中，樣本輸入爲訪客資訊，可分爲“訪客屬 性資訊”和“訪客行爲資訊”兩大類。其中，“訪客屬性 資訊”一般爲相對靜態的資訊，包括訪客的性別、年齡、 地域等；而“訪客行爲資訊”則是短期內可能發生動態變 化的資訊，例如訪問網站的次數、所訪問的頁面、訪問停 留時間、訪客是否進行了某些主動行爲（如聊天、發郵件 -9- 201108130 )、進行主動行爲的次數等等。 爲了令神經網路計算模型能夠識別訪客資訊，可以對 上述訪客資訊進行數値化處理，例如：對於時間、次數等 信息’可以直接用其實際數値表示；對於取値爲“是/否 ”的資訊’可以使用數値0/1來表示；而對於訪客地域、 所訪問頁面這類的資訊，可以分別針對每種資訊的特點， 預先對各種可能的取値進行編碼，以“訪客地域”爲例， 可以用001代表北京、002代表上海，等等。 對應每一位元訪客的一組資訊，網站主會根據這些資 訊，對這位元訪客的價値給出判斷，需要說明的是，這個 判斷過程應該是客觀的，即對於相同的資訊，其判斷結果 不會以人的主觀意識而改變。當然，根據具體需求，判斷 的結果的表現形式是多種多樣的：可以是簡單的是/否有 價値’或者是對價値劃分出具體等級，也可以是一些具體 的業務行爲，例如：與訪客聯繫（打電話、網路聊天、發 郵件等），將訪客加入CRM (客戶管理系統）等等，這些 人工判斷的結果即爲樣本的期望輸出。 與樣本輸入相對應，對作爲樣本期望輸出的判斷結果 也應該進行數値化處理，其具體方法與對訪客資訊進行數 値化處理的方法相類似。需要說明的是，如果判斷結果是 以“具體業務行爲”來表示，那麼對於同一位訪客，可能 會進行多種業務行爲，這時，可以根據具體的需求，針對 各種業務行爲的重要程度，對判斷結果進行加權平均處理 。例如：將某位訪客加入C R Μ意味著這名訪客價値很大 -10- 201108130 ’而如果僅與某位訪客進行試探性的聊天，則說明這位訪 客的價値相對較小，那麼在進行加權平均處理的時候，“ 加入CRM”的權重就應該大於“聊天”。 可以想像，上述資料樣本的輸入和期望輸出，即“訪 客資訊”和“訪客價値”之間，是存在著某種對應關係的 。但是客觀業務需求的複雜性導致這種對應關係難以直接 描述。而透過對神經網路進行訓練，神經網路可以將樣本 資料中所包含的對應關係以權値、偏置等形式表現在計算 模型中，從而解決現有技術所存在的問題。 BP (誤差反向傳播）演算法是比較常用的一類用於對 神經網路進行訓練的演算法，常用的BP演算法包括GD ( Gradient Descent ) 、 GDA ( Gradient Descent withWi is the implicit layer weight matrix, Βι is the implicit layer offset vector, f! is the hidden layer transfer function; W2 is the output layer weight matrix, B2 is the output layer offset vector, and f2 is the output layer transfer function; The calculation model takes the visitor information as input, X is the η-dimensional longitudinal quantity, where the number of each dimension represents a type of visitor information; the visitor price is the output, Υ is the m-dimensional longitudinal quantity; assuming that there are k hidden layers in the model Node, Bay lj Wi is a kxn matrix, Βι is a k-dimensional longitudinal quantity, and W2 is a mxk matrix 'B2 is a dimension longitudinal dimension. For the present embodiment, the visitor price to be calculated can be represented by a number, so the output vector Y is a one-dimensional vector. It should be noted that the dimension of the input vector X and the output vector Y is determined by specific requirements, and is not limited in this embodiment. In this embodiment, the implicit layer transfer function fi uses a non-linear -8 - 201108130 sexual action function Sigmoid function of the neuron, that is, a function of the form f(x)=i/(i+e·*); The transfer function f2 is a linear function. Theoretical studies show that the neural network uses the Sigmoid transfer function in its hidden layer and uses the linear transfer function in the output layer to approximate any function of interest with arbitrary precision. The number of nodes for the hidden layer k In terms of more, the hidden layer nodes can speed up the learning of the neural network, but it also complicates the structure of the model and reduces the adaptability of the model. The number of nodes can be adjusted during the training process, for example, using the trial and error method of the cross-evaluation of the test sample and the training sample to select the number of hidden layer nodes. Those skilled in the art can also use other methods to determine the number of hidden layer nodes according to specific needs, which is not limited in this embodiment. S1 〇2, using the data sample to train the calculation model to determine the calculation model; in S 1 0 1 , the guest price calculation model is initialized, and in this step, the training model is Parameters such as W2, B丨, and B2 are determined. For a website, a batch of historical data can be selected as a sample for training. Among them, the sample input is visitor information, which can be divided into two categories: “visitor attribute information” and “visitor behavior information”. Among them, “visitor attribute information” is generally relatively static information, including the gender, age, and region of the visitor; and “visitor behavior information” is information that may change dynamically in a short period of time, such as the number of visits to the website and the visits. The page, the time of the visit, whether the visitor has performed some active behavior (such as chat, email -9- 201108130), the number of active actions, and so on. In order to enable the neural network computing model to recognize the visitor information, the above-mentioned visitor information may be digitized, for example, information such as time, number of times, etc. may be directly represented by its actual number; for the reference is "yes/no" The information 'can be expressed by the number 値0/1; and for the information of the visitor area and the visited page, each possible information can be coded separately for the characteristics of each type of information, in order to "visitor area" For example, you can use 001 for Beijing, 002 for Shanghai, and so on. According to a group of information for each visitor, the website owner will judge the price of the meta-visitor based on the information. It should be noted that the judgment process should be objective, that is, for the same information, the judgment is made. The results will not change with the subjective consciousness of the person. Of course, depending on the specific needs, the outcome of the judgment is manifested in a variety of ways: it can be simple yes/no price 値' or the price 値 is divided into specific grades, or it can be some specific business behavior, for example: contact with the visitor (Call, chat, email, etc.), join visitors to CRM (customer management system), etc. The result of these manual judgments is the expected output of the sample. Corresponding to the sample input, the judgment result as the expected output of the sample should also be digitized, and the specific method is similar to the method of digitizing the visitor information. It should be noted that if the judgment result is expressed by “specific business behavior”, multiple business behaviors may be performed for the same visitor. In this case, the judgment result may be determined according to the specific requirements and the importance degree of various business behaviors. Perform a weighted average process. For example, adding a visitor to CR means that the visitor price is very large -10- 201108130 'and if only a tentative chat with a visitor indicates that the visitor's price is relatively small, then the weight is being weighted. When averaging, the weight of “joining CRM” should be greater than “chat”. It is conceivable that there is a certain correspondence between the input and expected output of the above data sample, that is, "visitor information" and "visit price". However, the complexity of objective business requirements makes this correspondence difficult to describe directly. By training the neural network, the neural network can express the corresponding relationship contained in the sample data in the calculation model in the form of weights and offsets, thereby solving the problems existing in the prior art. The BP (Error Back Propagation) algorithm is a commonly used algorithm for training neural networks. The commonly used BP algorithms include GD (Gradient Descent) and GDA (Gradient Descent with
Adaptive learning rate ) 、GDM ( Gradient Descent with Momentum ) 、GDX ( Gradient Descent with momentum and adaptive learning rate)及 LM ( Levenberg-Marquardt )等，本實施例選用具有較好收斂性能的LM BP演算法 ，能夠達到最佳的實施效果。首先在全體資料樣本中隨機 選定一個訓練集（例如1 /2的樣本量）輸入計算模型，將 計算所得到的樣本輸出値與期望輸出値進行比較，根據比 較結果不斷調整計算模型的未確定參數（如W!、W2、B! 、B2等），使得樣本輸出値儘量接近期望輸出値。模型的 精確度可以用均方誤差來評價’當樣本輸出値與期望輸出 値的誤差滿足一定的精確度要求時（例如均方誤差小於 10% )，即可確定計算模型° -11 - 201108130 在訓練過程中，還可以將訓練集之外的資料樣本作爲 確證集，用來輔助控制精確度。理想的訓練過程中，訓練 集的均方誤差應該隨著訓練次數增加而逐漸減小。確證集 的作用就是每進行一次（或幾次）訓練後即驗證確證集的 均方誤差是否減少，如果是，繼續對模型進行訓練，否則 停止訓練，重新確認模型的參數。引入確證集的好處在於 ，可以自適應得到神經網路最優誤差控制標準，避免選擇 誤差控制標準的困難，還可以避免出現網路學習產生的對 訓練集過適應的現象》 s 1 03，獲得訪客資訊，使用所確定的計算模型，計算 所述訪客的價値。 訓練完成後，使用所確定的計算模型即可自動對訪客 價値進行計算。其中，本領域技術人員可以根據具體需求 ，選擇獲得訪客資訊的具體方法。例如，對於性別、年齡 、地域等“訪客屬性資訊”，可以由該訪客在網站的註冊 資訊中獲得，其中，訪客的地域資訊還可以透過該訪客的 IP位址獲得。而對於“訪客行爲資訊”則可以由網站後台 的管理系統獲得。 在獲得某位元訪客的資訊之後，將這些資訊輸入所確 定的計算模型，就可以得到表示該訪客價値的輸出値。由 於計算模型是根據人工判斷的結果訓練出來的，因此，其 計算結果與人工判斷的結果是趨於一致的。網站主根據該 輸出値’就可以決定是否對該訪客做進一步的業務行爲。 根據實際的業務需求，如果計算模型的輸入僅包括“ -12- 201108130 訪客屬性資訊”這類相對靜態的資訊’那麼模型所 訪客價値是可以一次性計算出來的’並且在業務需 的前提下，可以認爲這個値不會發生變化；如果計 的輸入包括“訪客行爲資訊”這類短期內可能發生 化的資訊，則模型相應所輸出的訪客價値也應該是 化的。針對這種情況，可以由網站管理系統週期性 算模型提供/更新用戶資訊（或者由網站主來人工 供/更新用戶資訊的操作），以實現對訪客價値的 算。 在計算模型中，還可以預先設置一個（或多個 價値閾値，當模型的輸出値大於該閾値時，透過聲 幕視覺變化等方式提醒網站主，從而減少網站主對 無價値”訪客的關注，提高工作效率。 在本申請案的較佳實施方式中，對於當前已確 算模型，還可以進一步對其進行修正。例如，由當 定的模型參數中，可以提取出每種訪客資訊與訪客 相關性，如果發現某些訪客資訊與最終計算出的訪 相關性很小，說明這些資訊對最終結果的影響不大 可以將這些資訊所對應的分量從計算模型的輸入向 除，從而提高模型的計算性能。具體來講，可以預 一個相關性閾値，當檢測某個訪客資訊與訪客價値 性小於這個閾値時，則從刪除對應的分量，或者向 發出提示資訊，讓網站主來決定是否刪除該分量。 此外，在計算模型的使用過程中，還可以週期 輸出的 求不變 算模型 動態變 動態變 地向計 執行提 動態計 )訪客 音、螢 大量“ 定的計 前所確 價値的 客價値 ，那麼 量中刪 先設定 的相關 網站主 性或隨 -13- 201108130 機地對當前所確定的計算模型的實際輸出値與期望輸出値 進行比較，當然，期望輸出値是需要由人工判斷得到的， 如果比較結果表明實際輸出値與期望輸出値的誤差超過了 某個預設的閾値，那麼說明當前的計算模型已經過時，其 原因可能是具體的業務需求方面有了變化，這時可以重新 對計算模型進行訓練，以適應業務需求的變化。可以理解 的是，如果網站主自己能夠意識到業務需求發生了變化， 也可以主動觸發計算模型的重構步驟。 下面將透過一個具體的實施例，對本申請案的網站訪 客價値的計算方法進行說明β假設某企業分別針對初中生 和高中生提供兩類產品，並且將這兩類產品資訊分別發佈 在網站的Α頁面和Β頁面上。網站主在判斷訪客的價値時 ，遵循的原則包括： 1) 如果訪客年齡在10-16歲，並且訪問頁面A超過 3次，則認爲其有價値； 2) 如果訪客年齡在14-20歲，並且訪問頁面B超過3 次，則認爲其有價値； 3) 如果訪客年齡在35歲以上（可以認爲是學生家長 )，並且訪問頁面A或B超過3次，則認爲其有價値； 4) 任意訪客訪問頁面A或B超過1 0次，則認爲其有 價値； 5) 任意年齡大於等於10歲的訪客，如果向網站發送 了站內諮詢消息，則認爲其有價値。 根據上述需求，可以首先確定模型的計算模型的輸入 -14- 201108130 參量包括：訪客年齡、訪問網頁A (或B)的次數、訪客 是否發送過站內諮詢消息三類；計算模型的輸出即訪客價 値。所初始化的計算模型Y = F(X)中，X爲3維向量，Y爲 1維向量。 進一步的’需要對計算模型的輸入和輸出進行數値化 處理。對於“訪客年齡”及“訪問頁面A (或B)的次數 ”，可以直接採用實際數値，對於“訪客是否發送過站內 諮詢消息” ’可以用數値1表示“是”，用數値〇表示“ 否”。同樣，作爲樣本期望輸出的判斷結果“訪客是否有 價値”，也用數値1表示“是”，用數値0表示“否”。 透過訓練確定計算模型 Y = F(X)後，就可以使用所確 定的模型計算訪客的價値了，對本實施例而言，“訪客年 齡”可以由訪客在網站的註冊資訊中獲得，較佳地，還可 以由當前日期與訪客的出生曰期（可以由註冊資訊中獲得 )相減來獲得訪客的年齡。 對於“訪問頁面A (或B )的次數”以及“訪客是否 發送過站內諮詢消息” ’這兩類訪客資訊，可以透過網站 後台的管理系統獲得。由於這兩類資訊很有可能在短期內 發生變化，因此，網站管理系統應該即時或週期性地向計 算模型提供最新的資訊。計算模型根據網站管理系統所提 供的資訊，即時或週期性地計算訪客的價値。 網站管理系統獲得訪客的相關資訊後’以預定的資料 結構發送至計算模型的輸入埠。假設預定的資料結構形式 如下： -15- 201108130 {訪客標識，年齡，訪問頁面A次數，訪| 數’是否發送過站內諮詢消息} 其中，訪客標識不輸入模型參與計算，而 標識該訪客，例如，可以是該訪客在網站所I 或者是網站爲其分配的訪客編號等等。 以下假設網站管理系統以1小時爲週期， 更新訪客新型。訪客1(網站爲其分配的訪客 )首次訪問網站後，在第一個更新時刻，網站 計算模型發送的該訪客的資訊資料爲： {001， 15， 1 ， 1 ， 0} 需要說明的是，我們所期望的模型結果是 (〇和1 )，但是透過訓練得到的計算模型， 都是非離散形式的，可能是[0,1]範圍內的任意 超出該範圍（視對計算模型的精確度要求而定 這個問題，可以將模型所輸出的數値進行標準 如以0.5爲閩値，大於該閾値則認爲訪客有價 閾値則認爲訪客沒有價値。 如果根據人工判斷的原則，認爲該訪客是 的，由於計算模型是根據人工判斷的結果所訓 那麼經判別後的模型輸出値應該和人工判斷的 即輸出一個爲小於〇 . 5的値。 後續每間隔一小時，網站管理系統會將最 訊提供給計算模型。計算模型也會根據網站管 供的資訊，即時更新計算結果。假設在某一更 間頁面B次 是用於唯一 Ξ冊的ID， 向計算模型 編號爲001 管理系統向 離散形式的 其輸出一般 値，或者略 )。爲解決 離散化，例 値，小於該 不具有價値 練出來的， 結果一致， 新的訪客資 理系統所提 新時刻，網 -16- 201108130 站管理系統向計算模 {001， 15， 3 ， 1 可以看出，該資 模型會輸出一個大於 時，計算模型可將結 而言，由於結果僅涉 該訪客的標識回饋至 如果計算結果涉及多 標識一倂回饋至網站 螢幕視覺變化等方式 取進一步的業務行爲 本領域技術人員 計算模型提供多名訪 體實現方式，並行或 爲價値的訪客標識（ 〇 相應於上面的方 種網站訪客價値的計 初始化單元3 1 0 ，所述計算模型爲神 訪客價値爲輸出； 訓練單元 320， 行訓練，確定所述計; 計算單元3 3 0， 型發送的該訪客的資訊資料爲： ，0} 訊可以滿足上述的條件1 )，則計算 0.5的値，即認爲該訪客有價値。此 果回饋至網站管理系統，對本實施例 及“是”或“否”兩種，因此可僅將 網站管理系統，表示該訪客有價値。 個値，則應將具體的計算値也和訪客 管理系統。網站管理系統透過聲音、 提醒網站主，以便網站主對該訪客採 〇 可以理解’網站管理系統每次可以向 客的資訊，而計算模型可以根據其具 串列計算多名訪客的價値。並將有認 以及I十算結果）回饋至網站管理系統 法實施例’本申請案實施例還提供一 算系統，參見圖3所示，包括： ’用於初始化網站訪客價値計算模型 經網路模型，以訪客資訊爲輸入，以 用於使用資料樣本對所述計算模型進 障模型； 用於獲得訪客資訊，使用所確定的計 -17- 201108130 算模型，計算所述訪客的價値。 其中，所述初始化的計算模型可以是以下的形式： Y = f2(F2MWlX + Bi) + B2) X爲輸入向量；Y爲輸出向量； W1爲隱含層權値矩陣，B1爲隱含層偏置向量，f!爲 隱含層傳遞函數； W2爲輸出層權値矩陣，B2爲輸出層偏置向量，；^爲 輸出層傳遞函數； f!爲神經元的非線性作用函數，f2爲線性函數。 其中，根據業務的具體需求，作爲計算模型輸入/輸 出的訪客資訊和訪客價値，可以是數値化的訪客資訊和訪 客價値。 所述訓練單元320，可以使用誤差反向傳播演算法對 所述計算模型進行訓練。當所述計算模型的樣本輸出値與 期望輸出値的誤差滿足精確度要求時，確定所述計算模型 〇 以上所提供的系統，以神經網路作爲網站訪客價値的 計算模型。透過訓練可以令模型學習網站主的主動行爲， 從而將實際需求中訪客資訊和訪客價値的各種複雜的對應 關係納入模型。訓練完成後，透過使用該模型，即可根據 訪客的各種資訊，自動計算出該訪客的價値，由於模型是 根據人工判斷的結果訓練出來的，因此，其計算的結果與 人工判斷的結果趨於一致。 參見圖4所示，本申請案實施例所提供的網站訪客價 -18- 201108130 値的計算系統，還可以進一步包括修正單元340，用於對 當前所確定的計算模型進行修正，並將修正後的計算模型 確定爲新計算模型； 則所述計算單元3 30，用於使用所述修正單元340所 確定的新計算模型，計算訪客的價値。 其中，所述修正單元340可以包括： 相關性獲得子單元’用於由當前所確定的計算模型中 ’獲得訪客資訊類別與訪客價値的相關性； 輸入向量刪除子單元，用於將相關性小於預設閩値的 訪客資訊類別從所述當前所確定的計算模型的輸入向量中 刪除。 應用上述系統，可以將對最終結果的影響不大的訪客 資訊從計算模型的輸入向量中刪除，從而提高模型的計算 性能》 所述修正單元3 4 0還可以包括： 比較子單元，用於週期性或隨機地對當前所確定的計 算模型的實際輸出値與期望輸出値進行比較； 重構子單元，用於當所述比較子單元所得到的誤差大 於預設閾値時，重新對所述計算模型進行訓練❶ 應用上述系統，可以即時對網路誤差進行監控，並根 據監控結果對計算模型進行重構，以適應業務需求的動態 變化。 本領域技術人員可以理解，修正單元340的構成方式 ’也可以是上述兩種方式的結合。 -19- 201108130 對於系統實施例而言，由於其基本相應於方法實施例 ，所以描述得比較簡單，相關之處參見方法實施例的部分 說明即可。以上所描述的系統實施例僅僅是示意性的，其 中所述作爲分離部件說明的單元可以是或者也可以不是物 理上分開的，作爲單元顯示的部件可以是或者也可以不是 物理單元，即可以位於一個地方，或者也可以分佈到多個 網路單元上。可以根據實際的需要選擇其中的部分或者全 部模組來實現本實施例方案的目的。本領域普通技術人員 在不付出創造性勞動的情況下，即可以理解並實施。 爲了描述的方便，描述以上系統時以功能分爲各種單 元分別描述。當然，在實施本發明時可以把各單元的功能 在同一個或多個軟體和/或硬體中實現。 透過以上的實施方式的描述可知，本領域的技術人員 可以清楚地瞭解到本申請案中的實施例可借助軟體加必需 的通用硬體平臺的方式來實現。基於這樣的理解，本申請 案中的實施例的技術方案本質上或者說對現有技術做出貢 獻的部分可以以軟體產品的形式體現出來，該電腦軟體產 品可以儲存在儲存媒體中，如ROM/RAM、磁碟、光碟等 ’包括若干指令用以使得一台電腦設備（可以是個人電腦 ，伺服器，或者網路設備等）執行本申請案中的各個實施 例或者實施例的某些部分所述的方法。 以上所述僅是本發明的具體實施方式，應當指出，對 於本技術領域的普通技術人員來說，在不脫離本發明原理 的前提下，還可以做出若干改進和潤飾，這些改進和潤飾 -20- 201108130 也應視爲本申請案的保護範圍。 【圖式簡單說明】 圖1爲本申請案實施例實現網站訪客價値計算方法的 流程圖； 圖2爲本申請案實施例的神經網路模型結構示意圖： 圖3爲本申請案實施例網站訪客價値計算系統的結構 示意圖； 圖4爲本申請案實施例網站訪客價値計算系統的另〜 種結構示意圖。 【主要元件符號說明】 3 1 〇 :初始化單元 320 :訓練單元 3 3 0 :計算單元 3 4 0 :修正單元 • 21 -Adaptive learning rate ), GDM (Gradient Descent with Momentum ), GDX (Gradient Descent with momentum and adaptive learning rate), and LM ( Levenberg-Marquardt ), etc., in this embodiment, an LM BP algorithm with better convergence performance can be used. The best implementation effect. First, a training set (for example, a sample size of 1 /2) is randomly selected in the entire data sample to input a calculation model, and the obtained sample output 値 is compared with the expected output ,, and the undetermined parameters of the calculation model are continuously adjusted according to the comparison result. (such as W!, W2, B!, B2, etc.), so that the sample output 値 is as close as possible to the desired output 値. The accuracy of the model can be evaluated by the mean square error 'when the error between the sample output 値 and the desired output 满足 meets certain accuracy requirements (eg, the mean square error is less than 10%), the calculation model can be determined ° -11 - 201108130 During the training process, data samples outside the training set can also be used as a confirmation set to assist control accuracy. In the ideal training process, the mean square error of the training set should be gradually reduced as the number of training increases. The role of the corroboration set is to verify that the mean square error of the corroboration set is reduced after each (or several) trainings. If so, continue to train the model, otherwise stop training and reconfirm the parameters of the model. The advantage of introducing a corroboration set is that it can adaptively obtain the neural network optimal error control standard, avoid the difficulty of selecting the error control standard, and avoid the phenomenon that the network learning produces over-adaptation to the training set. s 1 03 Visitor information, using the determined calculation model, to calculate the price of the visitor. After the training is completed, the visitor price is automatically calculated using the determined calculation model. Among them, the person skilled in the art can select a specific method for obtaining visitor information according to specific needs. For example, the “visitor attribute information” for gender, age, geography, etc. can be obtained by the visitor in the registration information of the website, and the geographical information of the visitor can also be obtained through the visitor's IP address. For "visitor behavior information", it can be obtained by the management system in the background of the website. After obtaining the information of a certain bit visitor, the information is input into the determined calculation model, and the output 表示 indicating the visitor price is obtained. Since the calculation model is trained based on the results of manual judgment, the calculation results are consistent with the results of manual judgment. Based on the output, the website owner can decide whether to conduct further business activities for the visitor. According to the actual business needs, if the input of the calculation model only includes the relatively static information such as “ -12- 201108130 Visitor Attribute Information”, then the model visitor price can be calculated at one time, and under the premise of business needs, It can be assumed that this flaw will not change; if the input of the meter includes information such as "visitor behavior information" that may occur in a short period of time, the corresponding visitor price of the model should also be normalized. In this case, the website management system can periodically provide/update user information (or the operation of manually providing/updating user information by the website owner) to realize the calculation of the visitor price. In the calculation model, one (or more price thresholds may be preset), and when the output of the model is greater than the threshold, the website owner is reminded by visual changes of the sound curtain, thereby reducing the attention of the website owner to the priceless visitor. In the preferred embodiment of the present application, the currently validated model can be further modified. For example, from the predetermined model parameters, each visitor information can be extracted and related to the visitor. Sexuality, if some visitor information is found to have little correlation with the final calculated visit, it means that the impact of the information on the final result is small, the component corresponding to the information can be removed from the input of the calculation model, thereby improving the calculation of the model. Performance. Specifically, a correlation threshold may be pre-selected. When detecting a visitor information and a visitor price is less than the threshold, the corresponding component is deleted, or a prompt message is sent to the website owner to decide whether to delete the component. In addition, in the use of the calculation model, it is also possible to calculate the periodic output. Dynamic dynamic change to the implementation of the dynamic meter) visitor tone, firefly a large number of "predetermined price of the customer before the price", then the amount of the relevant website set first deleted or with the current -13- 201108130 The actual output 确定 of the determined calculation model is compared with the expected output ,. Of course, the expected output 値 needs to be manually determined. If the comparison result indicates that the error between the actual output 値 and the desired output 超过 exceeds a certain threshold 値Then, the current calculation model is outdated. The reason may be that the specific business requirements have changed. At this time, the calculation model can be retrained to adapt to changes in business needs. Understandably, if the website owner can realize that the business needs have changed, he can also actively trigger the reconstruction step of the calculation model. In the following, a specific embodiment will be used to explain the calculation method of the website visitor price of the present application. β It is assumed that an enterprise provides two types of products for junior high school students and high school students respectively, and the information of the two types of products is separately posted on the website. On the page and on the page. The principles that website owners follow when judging a visitor's price include: 1) If the visitor is 10-16 years old and visits page A more than 3 times, the price is considered to be 値; 2) If the visitor is 14-20 years old And if you visit page B more than 3 times, you think it is worth 値; 3) If the visitor is over 35 years old (can be considered as the parent of the student) and visits page A or B more than 3 times, it is considered to be valuable 値4) If any visitor visits page A or B more than 10 times, it is considered to be worthwhile; 5) Any visitor who is 10 years old or older, if they send a website consultation message to the website, considers it to be worthwhile. According to the above requirements, you can first determine the input of the model's calculation model.-14- 201108130 Parameters include: visitor age, the number of visits to page A (or B), whether the visitor sent the in-site consultation message; the output of the calculation model is the visitor price値. In the initialized calculation model Y = F(X), X is a 3-dimensional vector and Y is a 1-dimensional vector. Further, the input and output of the computational model need to be digitized. For "Visitor Age" and "Number of Visits to Page A (or B)", you can use the actual number directly. For "Whether the visitor has sent an in-site consultation message", you can use the number 値1 to indicate "Yes". Indicates "No". Similarly, as a result of the judgment of the expected output of the sample, "whether the visitor has a price," the number 値1 indicates "yes" and the number 値0 indicates "no". After the calculation model Y = F(X) is determined through training, the determined model can be used to calculate the price of the visitor. For the present embodiment, the "visitor age" can be obtained by the visitor in the registration information of the website, preferably You can also get the age of the visitor by subtracting the current date from the visitor's birth period (which can be obtained from the registration information). Visitor information such as "Number of Visits to Page A (or B)" and "Whether visitors have sent in-site consultation messages" can be obtained through the back-end management system of the website. Since these two types of information are likely to change in the short term, the website management system should provide the latest information to the calculation model on an immediate or periodic basis. The calculation model calculates the price of the visitor instantly or periodically based on the information provided by the website management system. After the website management system obtains the relevant information of the visitor, it is sent to the input of the calculation model in a predetermined data structure. Assume that the predetermined data structure is as follows: -15- 201108130 {Visitor ID, age, number of visits to page A, visits | number 'whether sent in-site consultation message} Where the visitor ID does not enter the model to participate in the calculation, but identifies the visitor, for example It can be the visitor number assigned by the visitor on the website I or the website, and so on. The following assumes that the website management system updates the new visitors in a one-hour cycle. After visitor 1 (the visitor assigned to the website) visits the website for the first time, at the first update moment, the visitor's information sent by the website calculation model is: {001, 15, 1 , 1 , 0} It should be noted that The model results we expect are (〇 and 1), but the computational models obtained through training are all non-discrete, and may be any outside the range of [0,1] (accuracy requirements for the parallel computing model) To solve this problem, the number of the output of the model can be set to a standard of 0.5, and if it is greater than the threshold, the visitor has a price threshold, and the visitor is considered to have no price. If the principle is based on manual judgment, the visitor is considered to be Since the calculation model is trained according to the result of manual judgment, the model output after the discriminating should be compared with the manual judgment, that is, the output is less than 〇. 5. After every hour, the website management system will report the most. Provided to the calculation model. The calculation model also updates the calculation results on the basis of the information provided by the website. It is assumed that the B-time is the ID for unique registration in a certain page. The calculation model number 001 outputs to the management system in the form of generally discrete Zhi, or omitted). In order to solve the discretization, for example, less than the price that has not been priced, the results are consistent, and the new visitor management system proposes a new moment. Net-16- 201108130 Station Management System to the calculation model {001, 15, 3, 1 It can be seen that the asset model will output a greater than, when the calculation model can be compared, the result is only related to the visitor's logo feedback. If the calculation result involves multiple identifications, feedback to the website screen visual changes, etc. Business Behavior The computing model of the person skilled in the art provides multiple visitor implementations, in parallel or as a price of the visitor identification (〇 corresponding to the above-mentioned website website price 初始化 initialization unit 3 1 0 , the calculation model is the god visitor price 値For the output; the training unit 320, the line training, determine the meter; the computing unit 3 3 0, the type of information sent by the visitor is: , 0} can satisfy the above condition 1), then calculate 0.5, that is, Think that the visitor has a price. This is fed back to the website management system, for the present embodiment and "yes" or "no", so only the website management system can be used to indicate that the visitor has a price. In the case of 値, the specific calculations and the visitor management system should be used. The website management system reminds the website owner through the voice, so that the website owner can understand the visitor's information that the website management system can visit the customer every time, and the calculation model can calculate the price of multiple visitors according to the serial number. And will give back to the website management system method. The embodiment of the present application also provides a calculation system, as shown in Figure 3, including: 'used to initialize the website visitor price calculation model through the network The model is input with the visitor information for using the data sample to the obstacle model of the calculation model; for obtaining the visitor information, and calculating the price of the visitor using the determined calculation model -17-201108130. The initialization calculation model may be in the following form: Y = f2(F2MWlX + Bi) + B2) X is an input vector; Y is an output vector; W1 is an implicit layer weight matrix, and B1 is an implicit layer bias Set vector, f! is the implicit layer transfer function; W2 is the output layer weight matrix, B2 is the output layer offset vector; ^ is the output layer transfer function; f! is the nonlinear action function of the neuron, f2 is linear function. Among them, according to the specific needs of the business, as the visitor information and visitor price of the input/output of the calculation model, it can be a number of visitor information and visitor price. The training unit 320 can train the calculation model using an error back propagation algorithm. When the error of the sample output 値 and the desired output 値 of the calculation model satisfies the accuracy requirement, the calculation model 确定 is determined as the system provided above, and the neural network is used as a calculation model of the website visitor price. Through training, the model can learn the active behavior of the website owner, so as to incorporate various complex correspondences between visitor information and visitor price in actual demand into the model. After the training is completed, by using the model, the price of the visitor can be automatically calculated according to various information of the visitor. Since the model is trained according to the result of manual judgment, the result of the calculation and the result of the manual judgment tend to Consistent. As shown in FIG. 4, the calculation system of the website visitor price -18-201108130 提供 provided by the embodiment of the present application may further include a correction unit 340 for correcting the currently determined calculation model, and The calculation model is determined as a new calculation model; then the calculation unit 303 is configured to calculate the price of the visitor using the new calculation model determined by the modification unit 340. The correction unit 340 may include: a correlation obtaining subunit 'for obtaining a correlation between a visitor information category and a visitor price in the currently determined calculation model; and an input vector deletion subunit for using the correlation to be less than The preset visitor information category is deleted from the input vector of the currently determined calculation model. Applying the above system, the visitor information with little influence on the final result can be deleted from the input vector of the calculation model, thereby improving the calculation performance of the model. The correction unit 300 can further include: a comparison subunit for the period Sexually or randomly comparing the actual output 値 of the currently determined calculation model with the expected output ;; reconstructing the subunit for recalculating the calculation when the error obtained by the comparison subunit is greater than a preset threshold 値Model training ❶ Applying the above system, the network error can be monitored immediately, and the calculation model can be reconstructed according to the monitoring result to adapt to the dynamic changes of business requirements. Those skilled in the art can understand that the configuration of the correction unit 340 can also be a combination of the above two methods. -19- 201108130 For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, ie may be located A place, or it can be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the objectives of the solution of the embodiment. Those of ordinary skill in the art can understand and implement without any creative effort. For the convenience of description, the above system is described by function into various units separately. Of course, the functions of the various units can be implemented in the same or multiple software and/or hardware in the practice of the invention. It will be apparent to those skilled in the art from the above description of the embodiments that the embodiments of the present application can be implemented by means of a software plus a necessary universal hardware platform. Based on such understanding, the technical solutions of the embodiments in the present application may be embodied in the form of a software product in essence or in the form of a software product, which may be stored in a storage medium such as a ROM/ 'RAM, diskette, optical disk, etc.' includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform various embodiments of the present application or portions of the embodiments. The method described. The above is only a specific embodiment of the present invention, and it should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present invention. 20- 201108130 should also be considered as the scope of protection of this application. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flow chart of a method for calculating a website visitor price according to an embodiment of the present application; FIG. 2 is a schematic structural diagram of a neural network model according to an embodiment of the present application: FIG. 3 is a website visitor of an embodiment of the present application. FIG. 4 is a schematic structural diagram of a website price calculation system of the embodiment of the present application. FIG. [Main component symbol description] 3 1 〇 : Initialization unit 320 : Training unit 3 3 0 : Calculation unit 3 4 0 : Correction unit • 21 -