WO2019223145A1 - 电子装置、推销名单推荐方法、系统和计算机可读存储介质 - Google Patents

电子装置、推销名单推荐方法、系统和计算机可读存储介质 Download PDF

Info

Publication number
WO2019223145A1
WO2019223145A1 PCT/CN2018/102111 CN2018102111W WO2019223145A1 WO 2019223145 A1 WO2019223145 A1 WO 2019223145A1 CN 2018102111 W CN2018102111 W CN 2018102111W WO 2019223145 A1 WO2019223145 A1 WO 2019223145A1
Authority
WO
WIPO (PCT)
Prior art keywords
sales
machine learning
learning model
online machine
customer
Prior art date
Application number
PCT/CN2018/102111
Other languages
English (en)
French (fr)
Inventor
黄博
毕野
王建明
吴振宇
肖京
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019223145A1 publication Critical patent/WO2019223145A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Definitions

  • the present application relates to the field of data processing technologies, and in particular, to an electronic device, a sales list recommendation method, a system, and a computer-readable storage medium.
  • the present application provides an electronic device, a sales list recommendation method, and a computer-readable storage medium, which are intended to improve the model prediction effect and timeliness, and reduce labor costs.
  • a first aspect of the present application provides an electronic device including a memory and a processor.
  • the memory stores a sales list recommendation system operable on the processor.
  • the sales list recommendation system is executed by the processor, To achieve the following steps:
  • the feature vector corresponding to each customer to be sold is substituted into a pre-trained online machine learning model for analysis and prediction, and the sales prediction results of each customer to be sold are obtained;
  • the obtained feature vectors are substituted into the online machine learning model, and the FTRL (Follow The Regularized Leader) algorithm is used to quickly and iteratively train each feature vector. Each iteration is solved and the model parameters that minimize the sum of all previous loss functions are updated. .
  • the second aspect of the present application provides a method for recommending a sales list.
  • the method includes the following steps:
  • the feature vector corresponding to each customer to be sold is substituted into a pre-trained online machine learning model for analysis and prediction, and the sales prediction results of each customer to be sold are obtained;
  • Each obtained feature vector is substituted into the online machine learning model, and each feature vector is rapidly and iteratively trained using the FTRL algorithm, and each iteration is solved and the model parameters that minimize the sum of all previous loss functions are updated.
  • a third aspect of the present application provides a computer-readable storage medium that stores a sales list recommendation system that can be executed by at least one processor such that the at least one processor Perform the following steps:
  • the feature vector corresponding to each customer to be sold is substituted into a pre-trained online machine learning model for analysis and prediction, and the sales prediction results of each customer to be sold are obtained;
  • Each obtained feature vector is substituted into the online machine learning model, and each feature vector is rapidly and iteratively trained using the FTRL algorithm, and each iteration is solved and the model parameters that minimize the sum of all previous loss functions are updated.
  • the fourth aspect of the present application provides a sales list recommendation system, including:
  • An acquisition module configured to obtain the customer data of each customer to be promoted after receiving the list of customers to be promoted, and convert the customer data of each customer to be promoted into corresponding feature vectors and store them;
  • the analysis and prediction module is used to substitute the feature vector corresponding to each customer to be sold into a pre-trained online machine learning model for analysis and prediction, and to obtain the sales prediction result of each customer to be sold;
  • a screening module for screening a list of customers to be promoted whose sales prediction results meet preset screening conditions, and sending the list to an agent center for sales processing;
  • the feedback data sorting module is used to obtain the feature vector of the customer corresponding to each sales record in the sales feedback data after receiving the sales feedback data from the agent center, and to obtain the obtained feature vector according to the sales result of each sales record Divided into positive and negative samples;
  • An online training module is used to substitute each obtained feature vector into the online machine learning model, and use FTRL algorithm to perform fast iterative training on each feature vector. Each iteration solves the model parameters that minimize the sum of all previous loss functions and updates .
  • an online machine learning model is used to analyze and predict the sales prediction results of the customers to be sold, and a list of selected customers is selected according to the sales prediction results and sent to the agent center for sales processing, and the agent center sales processing is received in real time.
  • the completed sales feedback data from the positive and negative samples sorted out from the sales feedback data, iteratively trains the online machine learning model to update the model.
  • the proposed solution uses an online machine learning model for recommendation prediction.
  • the model requires less training data. It does not need to wait for enough data to be collected before training. It can receive agent center feedback in real time.
  • the data is updated iteratively to avoid the deterioration of the prediction effect caused by the model not being updated for a long time, to ensure the good prediction effect and timeliness of the model, and the iterative training of the model is automatically iterated online, without manual training offline, human low cost.
  • FIG. 1 is a schematic flowchart of an embodiment of a sales list recommendation method of this application
  • FIG. 2 is a schematic flowchart of initial training of an online machine learning model in a method for recommending a sales list of this application;
  • FIG. 3 is a schematic diagram of an operating environment of a preferred embodiment of a sales list recommendation system of the present application
  • FIG. 4 is a program module diagram of an embodiment of a sales list recommendation system of the present application.
  • This application proposes a sales list recommendation method.
  • a processor in an electronic device implements the sales list recommendation method when executing a sales list recommendation system.
  • FIG. 1 is a schematic flowchart of an embodiment of a sales list recommendation method of the present application.
  • the sales list recommendation method includes:
  • step S10 after receiving the list of customers to be promoted, the customer data of each of the customers to be promoted is obtained, and the customer data of each of the customers to be promoted is converted into corresponding feature vectors and stored;
  • the customer database in the electronic device stores the customer data of each customer.
  • the customer data includes data and labels of many dimensions of the customer, such as gender, age, occupation, income level, consumer purchase records (including purchase method, channel, quantity, amount , Payment methods, etc.).
  • the electronic device After receiving the list of customers to be promoted to be uploaded, the electronic device obtains the customer data of each customer from the customer database according to the customer ID in the list of customers to be promoted, and then converts the customer data of each customer into corresponding feature vectors, and The feature vector corresponding to each customer is stored in the customer database.
  • Step S20 Substituting the feature vector corresponding to each customer to be sold into a pre-trained online machine learning model for analysis and prediction, and obtaining the sales prediction result of each customer to be sold;
  • the electronic device After the electronic device obtains the feature vectors of the customers to be sold, the obtained feature vectors are substituted into the pre-trained online machine learning model in the electronic device.
  • the online machine learning model analyzes and predicts each feature vector to obtain each Sales forecast results of sales customers.
  • the sales prediction result may be: sales success rate (for example, 30%, 80%, etc.), sales difficulty (for example, easy, normal, difficult, etc.), or sales recommendation score.
  • Step S30 Screen a list of customers to be promoted whose sales prediction results meet the preset screening conditions, and send the list to the agent center for sales processing;
  • the electronic device After the electronic device obtains the sales forecast results for each of the customers to be recommended, it uses a preset filter condition to screen each sales forecast result in order to obtain a sales forecast result that satisfies the screening conditions. The customers corresponding to the obtained sales forecast results are then screened. For selected customers, this selected customer list is sent to the agent center for sales processing.
  • step S40 after receiving the sales feedback data of the agent center, the feature vector of the customer corresponding to each sales record in the sales feedback data is obtained, and the obtained feature vector is divided into positive and negative according to the sales result of each sales record.
  • the agent center After the agent center obtains the list of selected customers recommended by the electronic device, it is assigned to the agent to carry out sales. Because the sales through the phone or the Internet usually have different closing times (some may be the same day, some may be the week or month ); When the agent center confirms the sales results of the customers who have been processed (successful or failed), the agent center will periodically or real-time feed back the latest determined sales data to the electronic device, that is, the agent center may feedback one or A batch of sales data to electronic devices. After the electronic device receives the sales feedback data from the agent center, it obtains the customer's feature vector corresponding to each sales record in the sales feedback data from the customer database (the customer's customer data has been characterized before the customer's prediction is made).
  • each feature vector obtained is converted and stored, so it can be directly searched in the customer database according to the customer ID to obtain), each feature vector obtained, according to the sales result of the sales record (successful or failed sales), the corresponding characteristics of the successful sales record
  • the vector is used as a positive sample, and the feature vector corresponding to the sales failure record is used as a negative sample for training and updating the online machine learning model.
  • Step S50 Substituting each obtained feature vector into the online machine learning model, and using FTRL (Follow The Regularized Leader) algorithm to perform fast iterative training on each feature vector, and each iteration solves a model that minimizes the sum of all previous loss functions Parameters and update.
  • FTRL Frellow The Regularized Leader
  • the electronic device substitutes the obtained positive and negative samples (that is, the obtained feature vectors) into the online machine learning model, and quickly iterates each sample through the FTRL algorithm to update the online machine learning model.
  • the model parameters with the smallest sum of all previous loss functions are updated after solving the model parameters and then the next iteration is repeated until all the samples are iterated and the updated online machine learning model is finally obtained.
  • the obtained feature vector is a predicted feature vector of the online machine learning model
  • the obtained feature vector is used to train the online machine learning model after confirming the sample type (that is, confirming whether it is a positive sample or a negative sample). , Correction and learning effect is better.
  • the FTRL algorithm can produce sparse results, and the trained model will be smaller, making online machine learning models more conducive to online storage and real-time prediction.
  • an online machine learning model is used to analyze and predict the sales forecast results of the customers to be marketed, and a list of selected customers is selected based on the sales forecast results and sent to the agent center for sales processing, and the agent center sales are received in real time.
  • After processing the sales feedback data iteratively trains the online machine learning model from the positive and negative samples arranged in the sales feedback data to update the model.
  • the proposed solution uses an online machine learning model for recommendation prediction. The model requires less training data. It does not need to wait for enough data to be collected before training. It can receive agent center feedback in real time.
  • the data is updated iteratively to avoid the deterioration of the prediction effect caused by the model not being updated for a long time, to ensure the good prediction effect and timeliness of the model, and the iterative training of the model is automatically iterated online, without manual training offline, human low cost.
  • FIG. 2 is a schematic flowchart of initial training of an online machine learning model in a recommendation method of a sales list of this application.
  • the initial training process of the online machine learning model is:
  • Step S60 Obtain a preset number of sales samples, obtain customer data of customers corresponding to each sales sample, convert each customer data into corresponding feature vectors, and divide the feature vectors corresponding to each customer data into positives according to the sales results of each sales sample. Negative samples
  • the sales samples are historical sales data and have confirmed sales results, that is, each sales sample already knows the success or failure of the sales.
  • the customer data of each customer is obtained from the customer database.
  • Each customer data is converted into a corresponding feature vector.
  • the sales result is a sales sample with successful sales, and its corresponding feature vector is used as a positive sample; the sales result is a sales sample with failed sales, and its corresponding feature vector is used as a negative sample.
  • Step S70 Initialize the preset model parameters of the online machine learning model, input the positive and negative samples into the online machine learning model, and perform iterative training using the FTRL algorithm, and minimize the sum of all the previous loss functions each iteration. Model parameters to confirm the latest model parameters of the online machine learning model after the iterative training is completed.
  • An electronic machine learning model is established in advance in the electronic device.
  • the model parameters of the online machine learning model are initialized (either directly assigned or randomly initialized).
  • the samples are input into the online machine learning model, and iterate quickly through FTRL calculation.
  • One iteration yields new model parameters, that is, an online machine learning model is updated, and the next iteration is based on the updated online machine learning model.
  • the latest model parameters of the online machine learning model will be finally confirmed, that is, the latest online machine learning model will be obtained for analysis and prediction of the recommended customers.
  • the objective function of the model parameter vector of the online machine learning model is:
  • W is the model parameter vector (w is usually randomly initialized or directly assigned), and ⁇ s represents the learning rate;
  • the first term g 1: t * w is an estimate of the contribution to the loss function, that is, the gradient or cumulative gradient;
  • the third term ⁇ 1 ⁇ w ⁇ 1 is L1 regular.
  • model parameter vector of the online machine learning model is solved as follows:
  • this application also proposes a sales list recommendation system.
  • FIG. 3 is a schematic diagram of an operating environment of a preferred embodiment of the sales list recommendation system 10 of the present application.
  • the sales list recommendation system 10 is installed and operated in the electronic device 1.
  • the electronic device 1 may be a computing device such as a desktop computer, a notebook, a palmtop computer, and a server.
  • the electronic device 1 may include, but is not limited to, a memory 11, a processor 12, and a display 13.
  • FIG. 3 only shows the electronic device 1 with components 11-13, but it should be understood that it is not required to implement all the illustrated components, and more or fewer components may be implemented instead.
  • the memory 11 may be an internal storage unit of the electronic device 1 in some embodiments, such as a hard disk or a memory of the electronic device 1.
  • the memory 11 may also be an external storage device of the electronic device 1 in other embodiments, such as a plug-in hard disk provided on the electronic device 1, a Smart Media Card (SMC), and a Secure Digital (SD). Cards, flash cards, etc.
  • the memory 11 may include both an internal storage unit of the electronic device 1 and an external storage device.
  • the memory 11 is used to store application software installed in the electronic device 1 and various types of data, such as program codes of the sales list recommendation system 10.
  • the memory 11 may also be used to temporarily store data that has been output or is to be output.
  • the processor 12 may be a central processing unit (CPU), a microprocessor or other data processing chip in some embodiments, and is configured to run program codes or process data stored in the memory 11, for example, to perform a sales list recommendation. System 10 and so on.
  • CPU central processing unit
  • microprocessor or other data processing chip in some embodiments, and is configured to run program codes or process data stored in the memory 11, for example, to perform a sales list recommendation. System 10 and so on.
  • the display 13 may be an LED display, a liquid crystal display, a touch-type liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, or the like.
  • the display 13 is used to display information processed in the electronic device 1 and to display a visualized user interface.
  • the components 11-13 of the electronic device 1 communicate with each other through a system bus.
  • FIG. 4 is a program module diagram of the preferred embodiment of the sales list recommendation system 10 of the present application.
  • the sales list recommendation system 10 may be divided into one or more modules, and the one or more modules are stored in the memory 11 and composed of one or more processors (the processor 12 in this embodiment) Implemented to complete this application.
  • the sales list recommendation system 10 may be divided into an acquisition module 101, an analysis prediction module 102, a screening module 103, a feedback data collation module 104, and an online training module 105.
  • the modules referred to in this application refer to a series of computer program instruction segments capable of performing specific functions, which are more suitable than programs for describing the execution process of the sales list recommendation system 10 in the electronic device 1, wherein:
  • the obtaining module 101 is configured to obtain the customer data of each customer to be promoted after receiving the list of customers to be promoted, and convert the customer data of each customer to be promoted into corresponding feature vectors and store them;
  • the customer database in the electronic device stores the customer data of each customer.
  • the customer data includes data and labels of many dimensions of the customer, such as gender, age, occupation, income level, consumer purchase records (including purchase method, channel, quantity, amount , Payment methods, etc.).
  • the electronic device After receiving the list of customers to be promoted to be uploaded, the electronic device obtains the customer data of each customer from the customer database according to the customer ID in the list of customers to be promoted, and then converts the customer data of each customer into corresponding feature vectors, and The feature vector corresponding to each customer is stored in the customer database.
  • the analysis and prediction module 102 is configured to substitute a feature vector corresponding to each customer to be sold into a pre-trained online machine learning model for analysis and prediction, and obtain a sales prediction result of each customer to be sold;
  • the obtained feature vectors are substituted into the pre-trained online machine learning model in the electronic device.
  • the online machine learning model analyzes and predicts each feature vector to obtain the customers to be sold Sales forecast results.
  • the sales prediction result may be: sales success rate (for example, 30%, 80%, etc.), sales difficulty (for example, easy, normal, difficult, etc.), or sales recommendation score.
  • a screening module 103 is configured to screen a list of customers to be promoted whose sales prediction results meet preset screening conditions, and send the list to the agent center for sales processing;
  • the electronic device After the electronic device obtains the sales forecast results for each of the customers to be recommended, it uses a preset filter condition to screen each sales forecast result in order to obtain a sales forecast result that satisfies the screening conditions. The customers corresponding to the obtained sales forecast results are then screened. For selected customers, this selected customer list is sent to the agent center for sales processing.
  • the feedback data sorting module 104 is configured to obtain the feature vector of the customer corresponding to each sales record in the sales feedback data after receiving the sales feedback data of the agent center, and to obtain the obtained features according to the sales results of the sales records.
  • Vector is divided into positive and negative samples;
  • the agent center After the agent center obtains the list of selected customers recommended by the electronic device, it is assigned to the agent to carry out sales. Because the sales through the phone or the Internet usually have different closing times (some may be the same day, some may be the week or month ); When the agent center confirms the sales results of the customers who have been processed (successful or failed), the agent center will periodically or real-time feed back the latest determined sales data to the electronic device, that is, the agent center may feedback one or A batch of sales data to electronic devices. After the electronic device receives the sales feedback data from the agent center, it obtains the customer's feature vector corresponding to each sales record in the sales feedback data from the customer database (the customer's customer data has been characterized before the customer's prediction is made).
  • each feature vector obtained is converted and stored, so it can be directly searched in the customer database according to the customer ID to obtain), each feature vector obtained, according to the sales result of the sales record (successful or failed sales), the corresponding characteristics of the successful sales record
  • the vector is used as a positive sample, and the feature vector corresponding to the sales failure record is used as a negative sample for training and updating the online machine learning model.
  • the online training module 105 is used to substitute each obtained feature vector into the online machine learning model, and use FTRL (Follow The Regularized Leader) algorithm to perform fast iterative training on each feature vector, and each iteration solves all previous loss functions. And minimum model parameters and update.
  • FTRL Frellow The Regularized Leader
  • the electronic device substitutes the obtained positive and negative samples (that is, the obtained feature vectors) into the online machine learning model, and quickly iterates each sample through the FTRL algorithm to update the online machine learning model.
  • the model parameters with the smallest sum of all previous loss functions are updated after solving the model parameters and then the next iteration is repeated until all the samples are iterated and the updated online machine learning model is finally obtained.
  • the obtained feature vector is a predicted feature vector of the online machine learning model
  • the obtained feature vector is used to train the online machine learning model after confirming the sample type (that is, confirming whether it is a positive sample or a negative sample). , Correction and learning effect is better.
  • the FTRL algorithm can produce sparse results, and the trained model will be smaller, making online machine learning models more conducive to online storage and real-time prediction.
  • an online machine learning model is used to analyze and predict the sales forecast results of the customers to be marketed, and a list of selected customers is selected based on the sales forecast results and sent to the agent center for sales processing, and the agent center sales are received in real time.
  • After processing the sales feedback data iteratively trains the online machine learning model from the positive and negative samples arranged in the sales feedback data to update the model.
  • the proposed solution uses an online machine learning model for recommendation prediction. The model requires less training data. It does not need to wait for enough data to be collected before training. It can receive agent center feedback in real time.
  • the data is updated iteratively to avoid the deterioration of the prediction effect caused by the model not being updated for a long time, to ensure the good prediction effect and timeliness of the model, and the iterative training of the model is automatically iterated online, without manual training offline, human low cost.
  • the initial training process of the online machine learning model includes:
  • the sales samples are historical sales data and have confirmed sales results, that is, each sales sample already knows the success or failure of the sales.
  • the customer data of each customer is obtained from the customer database.
  • Each customer data is converted into a corresponding feature vector.
  • the sales result is a sales sample with successful sales, and its corresponding feature vector is used as a positive sample; the sales result is a sales sample with failed sales, and its corresponding feature vector is used as a negative sample.
  • model parameters of the preset online machine learning model input the positive and negative samples into the online machine learning model, and use the FTRL algorithm for iterative training. Each iteration solves to minimize the sum of all previous loss functions. Model parameters to confirm the latest model parameters of the online machine learning model after the iterative training is completed.
  • An electronic machine learning model is established in advance in the electronic device.
  • the model parameters of the online machine learning model are initialized (either directly assigned or randomly initialized).
  • the samples are input into the online machine learning model, and iterate quickly through FTRL calculation.
  • One iteration yields new model parameters, that is, an online machine learning model is updated, and the next iteration is based on the updated online machine learning model.
  • the latest model parameters of the online machine learning model will be finally confirmed, that is, the latest online machine learning model will be obtained for analysis and prediction of the recommended customers.
  • the objective function of the model parameter vector of the online machine learning model is:
  • W is the model parameter vector (w is usually randomly initialized or can be directly assigned), and ⁇ s represents the learning rate;
  • the first term g 1: t * w is an estimate of the contribution to the loss function, that is, the gradient or cumulative gradient;
  • the third term ⁇ 1 ⁇ w ⁇ 1 is L1 regular.
  • model parameter vector of the online machine learning model is solved as follows:
  • the sales list recommendation system of the electronic device in this embodiment preferably adopts spark-based big data technology and redis memory storage technology, which can greatly improve the computing performance and reduce the time cost and labor cost of responding to the business.
  • the present application also proposes a computer-readable storage medium storing a sales list recommendation system, and the sales list recommendation system may be executed by at least one processor to enable the at least one process
  • the device executes the sales list recommendation method in any of the above embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Accounting & Taxation (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Engineering & Computer Science (AREA)
  • Game Theory and Decision Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请公开一种电子装置、推销名单推荐方法、系统和计算机可读存储介质，该方法包括：在接收到待推销客户名单后，获取各个待推销客户的客户数据，分别转换为对应的特征向量并存储；将各个特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；筛选出满足预设筛选条件的待推销客户的名单，并将其下发至坐席中心进行推销处理；接收到坐席中心的推销反馈数据后，获取每一条推销记录对应的客户的特征向量，并将获取的特征向量分成正、负样本；将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，更新模型模型参数。本申请方案提升模型预测效果，降低了人力成本。

Description

电子装置、推销名单推荐方法、系统和计算机可读存储介质
本申请基于巴黎公约申明享有2018年5月23日递交的申请号为CN 2018105023313、名称为“电子装置、推销名单推荐方法和计算机可读存储介质”中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请涉及数据处理技术领域，特别涉及一种电子装置、推销名单推荐方法、系统和计算机可读存储介质。
背景技术
传统的推荐系统通常是在针对业务(比如电话销售)的建模过程中，利用一批来自业务方的已有数据，学习到一个固化的模型；该模型的泛化能力，不仅依赖于精心设计的模型，更需要一次性灌注大量数据来保证。但是在业务实际的推荐过程中，特别是通过电话坐席收集样本的场景下，难以一次性获得大量训练数据；另一方面，模型上线后，一段时间是静态的，即使预测效果变差了，也需要等到下一次收集到足够多的数据后再次更新，影响预测效果，并且一般是通过人工重新训练，周期是周或者月，对人力成本要求高，且模型时效性差。
发明内容
本申请提供一种电子装置、推销名单推荐方法及计算机可读存储介质，旨在提升模型预测效果和时效性，降低人力成本。
本申请第一方面提供一种电子装置，包括存储器和处理器，所述存储器上存储有可在所述处理器上运行的推销名单推荐系统，所述推销名单推荐系统被所述处理器执行时实现如下步骤：
在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL(Follow The Regularized Leader)算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
本申请第二方面提供一种推销名单推荐方法，该方法包括步骤：
在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
本申请第三方面提供一种计算机可读存储介质，所述计算机可读存储介质存储有推销名单推荐系统，所述推销名单推荐系统可被至少一个处理器执行，以使所述至少一个处理器执行如下步骤：
在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的 特征向量分成正、负样本；
将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
本申请第四方面提供一种推销名单推荐系统，包括：
获取模块，用于在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
分析预测模块，用于将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
筛选模块，用于筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
反馈数据整理模块，用于在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
在线训练模块，用于将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
本申请技术方案，通过采用在线机器学习模型分析预测出待推销客户的推销预测结果，根据推销预测结果筛选出精选的客户的名单下发至坐席中心进行推销处理，并实时接收坐席中心推销处理完毕的推销反馈数据，从推销反馈数据中的整理出的正、负样本，对在线机器学习模型进行迭代训练以更新模型。相较于现有的传统推荐模型而言，本申请方案采用在线机器学习模型进行推荐预测，模型对训练数据量的要求小，无需等待收集足够多的数据后再训练，可以实时接收坐席中心反馈的数据以进行迭代更新，避免模型长时间不更新而带来的预测效果变差，保证模型的良好预测效果和时效性，并且模型的迭代训练是自动在线迭代的，无需下线人工训练，人力成本低。
附图说明
图1为本申请推销名单推荐方法一实施例的流程示意图；
图2为本申请推销名单推荐方法中在线机器学习模型的初始训练的流程示意图；
图3为本申请推销名单推荐系统较佳实施例的运行环境示意图；
图4为本申请推销名单推荐系统一实施例的程序模块图。
具体实施方式
以下结合附图对本申请的原理和特征进行描述，所举实例只用于解释本申请，并非用于限定本申请的范围。
本申请提出一种推销名单推荐方法，电子装置中的处理器在执行其推销名单推荐系统时实现该推销名单推荐方法。
如图1所示，图1为本申请推销名单推荐方法一实施例的流程示意图。
本实施例中，该推销名单推荐方法包括：
步骤S10，在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
电子装置中的客户数据库存储有各个客户的客户数据，客户数据包括客户的很多维度的数据和标签，例如，性别、年龄、职业、收入水平、消费购买记录(包含购买方式、渠道、数量、金额、付款方式等)等相关特征数据。电子装置在接收到待上传的待推销客户名单后，根据待推销客户名单中的客户ID从客户数据库中获取到各个客户的客户数据，然后将各个客户的客户数据转换为对应的特征向量，并将各个客户对应的特征向量存储到客户数据库中。
步骤S20，将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
电子装置在得到各个待推销客户的特征向量后，将得到的各个特征向量代入电子装置中预先训练好的在线机器学习模型中，该在线机器学习模型对各个特征向量进行分析预测，得出各个待推销客户的推销预测结果。本实施例中，该推销预测结果可以为：推销成功率(例如，30％、80％等)、推销难度(例如易、一般、难等)或推销推荐分值等。
步骤S30，筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
电子装置在得出各个待推荐客户的推销预测结果后，利用预先设置的筛选条件对各个推销预测结果进行筛选，以筛选得到满足筛选条件的推销预测结果，筛选得到的推销预测结果对应的客户则为精选的客户，将这精选的客 户的名单下发至坐席中心进行推销处理。
步骤S40，在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
坐席中心获取到电子装置推荐的精选的客户的名单后，分配给坐席人员去进行推销，由于通过电话或网络推销通常成交时间不一(有的可能是当天成交、有的可能是周或月)；当坐席中心在确认已推销处理的客户的推销结果(推销成功或推销失败)后，坐席中心会定时或实时将最新确定的推销数据反馈给电子装置，即坐席中心可能一次性反馈一个或一批推销数据到电子装置。电子装置接收到坐席中心的推销反馈数据后，则从客户数据库中获取推销反馈数据中每一条推销记录对应的客户的特征向量(该客户的客户数据在对该客户进行预测前已经进行过特征向量转换并存储了，因此这里可以直接根据客户ID在客户数据库中查找以获取)，获取的各个特征向量，根据推销记录的推销结果(推销成功或推销失败)，将推销成功的推销记录对应的特征向量作为正样本，将推销失败的推销记录对应的特征向量作为负样本，以用于对在线机器学习模型进行训练更新。
步骤S50，将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL(Follow The Regularized Leader)算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
电子装置将获取的正、负样本(即获取的特征向量)代入所述在线机器学习模型中，通过FTRL算法对各个样本快速迭代以对该在线机器学习模型进行更新，每次迭代都求解除让之前所有损失函数之和最小的模型参数，求解出后更新模型参数再进行下一次迭代，如此重复，直至所有样本都迭代完成，最终得到更新后的在线机器学习模型。由于该获取的特征向量为所述在线机器学习模型进行预测过的特征向量，将该获取的特征向量确认样本类型(即确认是正样本还是负样本)后，再用来对在线机器学习模型进行训练，修正和学习效果更佳。另外，FTRL算法能够产生稀疏化的效果，训练出的模型会比较小，使得在线机器学习模型更利于线上存储和实时预测。
本实施例技术方案，通过采用在线机器学习模型分析预测出待推销客户的推销预测结果，根据推销预测结果筛选出精选的客户的名单下发至坐席中 心进行推销处理，并实时接收坐席中心推销处理完毕的推销反馈数据，从推销反馈数据中的整理出的正、负样本，对在线机器学习模型进行迭代训练以更新模型。相较于现有的传统推荐模型而言，本申请方案采用在线机器学习模型进行推荐预测，模型对训练数据量的要求小，无需等待收集足够多的数据后再训练，可以实时接收坐席中心反馈的数据以进行迭代更新，避免模型长时间不更新而带来的预测效果变差，保证模型的良好预测效果和时效性，并且模型的迭代训练是自动在线迭代的，无需下线人工训练，人力成本低。
如图2所示，图2为本申请推销名单推荐方法中在线机器学习模型的初始训练的流程示意图；在本实施中，所述在线机器学习模型的初始训练过程为：
步骤S60，获取预设数量的推销样本，获取各个推销样本对应的客户的客户数据，将各个客户数据转换为对应的特征向量，根据各推销样本的推销结果将各个客户数据对应的特征向量分成正、负样本；
所述推销样本为历史推销数据，都是有确认的推销结果的，即各个推销样本都是已经知道推销成功还是失败，根据各个推销样本所属客户，从客户数据库中获取各个客户的客户数据，将各个客户数据转换为对应的特征向量。将推销结果为推销成功的推销样本，其对应的特征向量作为正样本；推销结果为推销失败的推销样本，其对应的特征向量作为负样本。
步骤S70，初始化预设的在线机器学习模型的模型参数，将所述正、负样本输入所述在线机器学习模型中，采用FTRL算法进行迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数，以在迭代训练完成后确认所述在线机器学习模型的最新模型参数。
电子装置中预先建立了在线机器学习模型，在对该在线机器学习模型迭代训练前，先初始化该在线机器学习模型的模型参数(可以直接赋值，也可以随机初始化)，电子装置将上述正、负样本输入该在线机器学习模型中，通过FTRL算快速迭代，一次迭代得出一次新的模型参数，即对在线机器学习模型进行一次更新，下一次迭代则在更新的在线机器学习模型基础上进行，在对样本数据迭代训练完成后，最终会确认得出在线机器学习模型的最新模型参数，即得到最新的在线机器学习模型，用于对待推荐客户的分析预测。
本实施例中，所述在线机器学习模型的模型参数向量的目标函数为：
Figure PCTCN2018102111-appb-000001
其中，W就是模型参数向量(w通常采用随机初始化，也可以直接赋值)，σ s表示学习速率；
第一项g 1:t*w是对损失函数的贡献的一个估计，也即梯度或累计梯度；
第二项
Figure PCTCN2018102111-appb-000002
这个累加和是限定w在每次迭代中不要变化太大，也限定了新的迭代结果不要离已迭代过的解太远；
第三项λ 1‖w‖ 1为L1正则。
具体的，该在线机器学习模型的模型参数向量的求解方式为：
在第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
Figure PCTCN2018102111-appb-000003
Predict p t＝σ(x t·w)using the w t,i computed above
Observe label y t∈{0,1}
for all i∈I do
g i＝(p t-y t)x i#gradient of loss w.r.t.w i
Figure PCTCN2018102111-appb-000004
z i←z i+g i-σ iw t,i
n i←n i+g i 2
求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1(即t+1次迭代后的模型参数向量)。
此外，本申请还提出一种推销名单推荐系统。
请参阅图3，是本申请推销名单推荐系统10较佳实施例的运行环境示意图。
在本实施例中，推销名单推荐系统10安装并运行于电子装置1中。电子装置1可以是桌上型计算机、笔记本、掌上电脑及服务器等计算设备。该电子装置1可包括，但不仅限于，存储器11、处理器12及显示器13。图3仅示出了具有组件11-13的电子装置1，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
存储器11在一些实施例中可以是电子装置1的内部存储单元，例如该电子装置1的硬盘或内存。存储器11在另一些实施例中也可以是电子装置1的外部存储设备，例如电子装置1上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，存储器11还可以既包括电子装置1的内部存储单元也包括外部存储设备。存储器11用于存储安装于电子装置1的应用软件及各类数据，例如推销名单推荐系统10的程序代码等。存储器11还可以用于暂时地存储已经输出或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)，微处理器或其他数据处理芯片，用于运行存储器11中存储的程序代码或处理数据，例如执行推销名单推荐系统10等。
显示器13在一些实施例中可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。显示器13用于显示在电子装置1中处理的信息以及用于显示可视化的用户界面。电子装置1的部件11-13通过系统总线相互通信。
请参阅图4，是本申请推销名单推荐系统10较佳实施例的程序模块图。在本实施例中，推销名单推荐系统10可以被分割成一个或多个模块，一个或者多个模块被存储于存储器11中，并由一个或多个处理器(本实施例为处理器12)所执行，以完成本申请。例如，在图4中，推销名单推荐系统10可以被分割成获取模块101、分析预测模块102、筛选模块103、反馈数据整理模块104及在线训练模块105。本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段，比程序更适合于描述推销名单推荐系统10在电子装置1中的执行过程，其中：
获取模块101，用于在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
电子装置中的客户数据库存储有各个客户的客户数据，客户数据包括客户的很多维度的数据和标签，例如，性别、年龄、职业、收入水平、消费购买记录(包含购买方式、渠道、数量、金额、付款方式等)等相关特征数据。电子装置在接收到待上传的待推销客户名单后，根据待推销客户名单中的客户ID从客户数据库中获取到各个客户的客户数据，然后将各个客户的客户数据转换为对应的特征向量，并将各个客户对应的特征向量存储到客户数据库中。
分析预测模块102，用于将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
在得到各个待推销客户的特征向量后，将得到的各个特征向量代入电子装置中预先训练好的在线机器学习模型中，该在线机器学习模型对各个特征向量进行分析预测，得出各个待推销客户的推销预测结果。本实施例中，该推销预测结果可以为：推销成功率(例如，30％、80％等)、推销难度(例如易、一般、难等)或推销推荐分值等。
筛选模块103，用于筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
电子装置在得出各个待推荐客户的推销预测结果后，利用预先设置的筛选条件对各个推销预测结果进行筛选，以筛选得到满足筛选条件的推销预测结果，筛选得到的推销预测结果对应的客户则为精选的客户，将这精选的客户的名单下发至坐席中心进行推销处理。
反馈数据整理模块104，用于在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
坐席中心获取到电子装置推荐的精选的客户的名单后，分配给坐席人员去进行推销，由于通过电话或网络推销通常成交时间不一(有的可能是当天成交、有的可能是周或月)；当坐席中心在确认已推销处理的客户的推销结果(推销成功或推销失败)后，坐席中心会定时或实时将最新确定的推销数据反馈给电子装置，即坐席中心可能一次性反馈一个或一批推销数据到电子装置。电子装置接收到坐席中心的推销反馈数据后，则从客户数据库中获取推 销反馈数据中每一条推销记录对应的客户的特征向量(该客户的客户数据在对该客户进行预测前已经进行过特征向量转换并存储了，因此这里可以直接根据客户ID在客户数据库中查找以获取)，获取的各个特征向量，根据推销记录的推销结果(推销成功或推销失败)，将推销成功的推销记录对应的特征向量作为正样本，将推销失败的推销记录对应的特征向量作为负样本，以用于对在线机器学习模型进行训练更新。
在线训练模块105，用于将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL(Follow The Regularized Leader)算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
电子装置将获取的正、负样本(即获取的特征向量)代入所述在线机器学习模型中，通过FTRL算法对各个样本快速迭代以对该在线机器学习模型进行更新，每次迭代都求解除让之前所有损失函数之和最小的模型参数，求解出后更新模型参数再进行下一次迭代，如此重复，直至所有样本都迭代完成，最终得到更新后的在线机器学习模型。由于该获取的特征向量为所述在线机器学习模型进行预测过的特征向量，将该获取的特征向量确认样本类型(即确认是正样本还是负样本)后，再用来对在线机器学习模型进行训练，修正和学习效果更佳。另外，FTRL算法能够产生稀疏化的效果，训练出的模型会比较小，使得在线机器学习模型更利于线上存储和实时预测。
本实施例技术方案，通过采用在线机器学习模型分析预测出待推销客户的推销预测结果，根据推销预测结果筛选出精选的客户的名单下发至坐席中心进行推销处理，并实时接收坐席中心推销处理完毕的推销反馈数据，从推销反馈数据中的整理出的正、负样本，对在线机器学习模型进行迭代训练以更新模型。相较于现有的传统推荐模型而言，本申请方案采用在线机器学习模型进行推荐预测，模型对训练数据量的要求小，无需等待收集足够多的数据后再训练，可以实时接收坐席中心反馈的数据以进行迭代更新，避免模型长时间不更新而带来的预测效果变差，保证模型的良好预测效果和时效性，并且模型的迭代训练是自动在线迭代的，无需下线人工训练，人力成本低。
本实施例中，所述在线机器学习模型的初始训练过程包括：
1、获取预设数量的推销样本，获取各个推销样本对应的客户的客户数据，将各个客户数据转换为对应的特征向量，根据各推销样本的推销结果将各个客户数据对应的特征向量分成正、负样本；
所述推销样本为历史推销数据，都是有确认的推销结果的，即各个推销样本都是已经知道推销成功还是失败，根据各个推销样本所属客户，从客户数据库中获取各个客户的客户数据，将各个客户数据转换为对应的特征向量。将推销结果为推销成功的推销样本，其对应的特征向量作为正样本；推销结果为推销失败的推销样本，其对应的特征向量作为负样本。
2、初始化预设的在线机器学习模型的模型参数，将所述正、负样本输入所述在线机器学习模型中，采用FTRL算法进行迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数，以在迭代训练完成后确认所述在线机器学习模型的最新模型参数。
电子装置中预先建立了在线机器学习模型，在对该在线机器学习模型迭代训练前，先初始化该在线机器学习模型的模型参数(可以直接赋值，也可以随机初始化)，电子装置将上述正、负样本输入该在线机器学习模型中，通过FTRL算快速迭代，一次迭代得出一次新的模型参数，即对在线机器学习模型进行一次更新，下一次迭代则在更新的在线机器学习模型基础上进行，在对样本数据迭代训练完成后，最终会确认得出在线机器学习模型的最新模型参数，即得到最新的在线机器学习模型，用于对待推荐客户的分析预测。
本实施例中，所述在线机器学习模型的模型参数向量的目标函数为：
Figure PCTCN2018102111-appb-000005
其中，W就是模型参数向量(w通常采用随机初始化，也可以直接赋值)，σ s表示学习速率；
第一项g 1:t*w是对损失函数的贡献的一个估计，也即梯度或累计梯度；
第二项
Figure PCTCN2018102111-appb-000006
这个累加和是限定w在每次迭代中不要变化太大，也限定了新的迭代结果不要离已迭代过的解太远；
第三项λ 1‖w‖ 1为L1正则。
具体的，该在线机器学习模型的模型参数向量的求解方式为：
在第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
Figure PCTCN2018102111-appb-000007
Predict p t＝σ(x t·w)using the w t,i computed above
Observe label y t∈{0,1}
for all i∈I do
g i＝(p t-y t)x i #gradient of loss w.r.t.w i
Figure PCTCN2018102111-appb-000008
z i←z i+g i-σ iw t,i
n i←n i+g i 2
求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1(即t+1次迭代后的模型参数向量)。
本实施例的电子装置的推销名单推荐系统优选采用基于spark大数据技术和redis内存存储技术，如此可大大提高了运算性能，缩减了响应业务的时间成本和人力成本。
进一步地，本申请还提出一种计算机可读存储介质，所述计算机可读存储介质存储有推销名单推荐系统，所述推销名单推荐系统可被至少一个处理器执行，以使所述至少一个处理器执行上述任一实施例中的推销名单推荐方法。
以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是在本申请的发明构思下，利用本申请说明书及附图内容所作的等效结构变换，或直接/间接运用在其他相关的技术领域均包括在本申请的专利保护范围内。

Claims (20)

  1. 一种电子装置，其特征在于，所述电子装置包括存储器和处理器，所述存储器上存储有可在所述处理器上运行的推销名单推荐系统，所述推销名单推荐系统被所述处理器执行时实现如下步骤：
    在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
    将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
    筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
    在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
    将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL(Follow The Regularized Leader)算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
  2. 如权利要求1所述的电子装置，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100001
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100002
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  3. 如权利要求2所述的电子装置，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量Xt，引入超参数α、β、λ 1、λ 2，通过以下公式计算：
    Figure PCTCN2018102111-appb-100003
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100004
    #equals
    Figure PCTCN2018102111-appb-100005
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再带回目标函数，则求解出w t+1。
  4. 如权利要求1所述的电子装置，其特征在于，所述在线机器学习模型的初始训练过程为：
    获取预设数量的推销样本，获取各个推销样本对应的客户的客户数据，将各个客户数据转换为对应的特征向量，根据各推销样本的推销结果将各个客户数据对应的特征向量分成正、负样本；
    初始化预设的在线机器学习模型的模型参数，将所述正、负样本输入所述在线机器学习模型中，采用FTRL算法进行迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数，以在迭代训练完成后确认所述在线机器学习模型的最新模型参数。
  5. 如权利要求4所述的电子装置，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100006
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100007
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  6. 如权利要求5所述的电子装置，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量Xt，引入超参数α、β、λ 1、λ 2，通过以下公式计算：
    Figure PCTCN2018102111-appb-100008
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100009
    #equals
    Figure PCTCN2018102111-appb-100010
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再带回目标函数，则求解出w t+1。
  7. 一种推销名单推荐方法，其特征在于，该方法包括步骤：
    在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
    将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
    筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
    在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
    将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和 最小的模型参数并更新。
  8. 如权利要求7所述的推销名单推荐方法，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100011
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100012
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  9. 如权利要求8所述的推销名单推荐方法，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
    Figure PCTCN2018102111-appb-100013
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100014
    #equals
    Figure PCTCN2018102111-appb-100015
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1。
  10. 如权利要求7所述的推销名单推荐方法，其特征在于，所述在线机器学习模型的初始训练过程为：
    获取预设数量的推销样本，获取各个推销样本对应的客户的客户数据， 将各个客户数据转换为对应的特征向量，根据各推销样本的推销结果将各个客户数据对应的特征向量分成正、负样本；
    初始化预设的在线机器学习模型的模型参数，将所述正、负样本输入所述在线机器学习模型中，采用FTRL算法进行迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数，以在迭代训练完成后确认所述在线机器学习模型的最新模型参数。
  11. 如权利要求10所述的推销名单推荐方法，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100016
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100017
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  12. 如权利要求11所述的推销名单推荐方法，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
    Figure PCTCN2018102111-appb-100018
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100019
    #equals
    Figure PCTCN2018102111-appb-100020
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1。
  13. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有推销名单推荐系统，所述推销名单推荐系统可被至少一个处理器执行，以使所述至少一个处理器执行如下步骤：
    在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
    将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
    筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
    在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
    将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数并更新。
  14. 如权利要求13所述的计算机可读存储介质，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100021
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100022
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  15. 如权利要求14所述的计算机可读存储介质，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
    Figure PCTCN2018102111-appb-100023
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100024
    #equals
    Figure PCTCN2018102111-appb-100025
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1。
  16. 如权利要求13所述的计算机可读存储介质，其特征在于，所述在线机器学习模型的初始训练过程为：
    获取预设数量的推销样本，获取各个推销样本对应的客户的客户数据，将各个客户数据转换为对应的特征向量，根据各推销样本的推销结果将各个客户数据对应的特征向量分成正、负样本；
    初始化预设的在线机器学习模型的模型参数，将所述正、负样本输入所述在线机器学习模型中，采用FTRL算法进行迭代训练，每次迭代求解让之前所有损失函数之和最小的模型参数，以在迭代训练完成后确认所述在线机器学习模型的最新模型参数。
  17. 如权利要求16所述的计算机可读存储介质，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100026
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100027
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
  18. 如权利要求17所述的计算机可读存储介质，其特征在于，所述在线机器学习模型的模型参数向量的计算方式为：
    对第t次迭代新加入的特征向量x t，引入超参数α、β、λ 1、λ 2，通过以下计算处理：
    Figure PCTCN2018102111-appb-100028
    Predict p t＝σ(x t·w)using the w t,i computed above
    Observe label y t∈{0,1}
    for all i∈I do
    g i＝(p t-y t)x i #gradient of loss w.r.t.w i
    Figure PCTCN2018102111-appb-100029
    #equals
    Figure PCTCN2018102111-appb-100030
    z i←z i+g i-σ iw t,i
    n i←n i+g i 2
    求解出g、σ和w t，再将求解出的g、σ及w t带回目标函数，则求解出w t+1。
  19. 一种推销名单推荐系统，其特征在于，所述推销名单推荐系统包括：
    获取模块，用于在接收到待推销客户名单后，获取各个待推销客户的客户数据，将各个待推销客户的客户数据分别转换为对应的特征向量并存储；
    分析预测模块，用于将各个待推销客户对应的特征向量代入预先训练好的在线机器学习模型中进行分析预测，得出各个待推销客户的推销预测结果；
    筛选模块，用于筛选出推销预测结果满足预设筛选条件的待推销客户的名单，将所述名单下发至坐席中心进行推销处理；
    反馈数据整理模块，用于在接收到坐席中心的推销反馈数据后，获取所述推销反馈数据中的每一条推销记录对应的客户的特征向量，并根据各推销记录的推销结果将获取的特征向量分成正、负样本；
    在线训练模块，用于将获取的各个特征向量代入所述在线机器学习模型中，采用FTRL算法对各个特征向量进行快速迭代训练，每次迭代求解让之 前所有损失函数之和最小的模型参数并更新。
  20. 如权利要求19所述的推销名单推荐系统，其特征在于，所述在线机器学习模型的模型参数向量的目标函数为：
    Figure PCTCN2018102111-appb-100031
    其中，W就是模型参数向量，σ s表示学习速率；
    第一项g 1:t*w是对损失函数的贡献的一个估计；
    第二项
    Figure PCTCN2018102111-appb-100032
    这个累加和是限定w在每次迭代中不要变化太大；
    第三项λ 1||w|| 1为L1正则。
PCT/CN2018/102111 2018-05-23 2018-08-24 电子装置、推销名单推荐方法、系统和计算机可读存储介质 WO2019223145A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810502331.3A CN108805613A (zh) 2018-05-23 2018-05-23 电子装置、推销名单推荐方法和计算机可读存储介质
CN201810502331.3 2018-05-23

Publications (1)

Publication Number Publication Date
WO2019223145A1 true WO2019223145A1 (zh) 2019-11-28

Family

ID=64092807

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/102111 WO2019223145A1 (zh) 2018-05-23 2018-08-24 电子装置、推销名单推荐方法、系统和计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN108805613A (zh)
WO (1) WO2019223145A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114549048A (zh) * 2022-01-18 2022-05-27 北京健康之家科技有限公司 外呼用户的推荐方法和装置、电子设备

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111797852A (zh) * 2019-04-09 2020-10-20 Oppo广东移动通信有限公司 算法模型的调整方法、装置、存储介质及电子设备
CN110400209A (zh) * 2019-07-03 2019-11-01 上海上湖信息技术有限公司 用户信用的确定方法及装置、存储介质、终端
CN112001434A (zh) * 2020-08-14 2020-11-27 深圳世间乐见科技有限公司 一种颜值预测方法、装置、终端及介质
CN112084408B (zh) * 2020-09-08 2023-11-21 中国平安财产保险股份有限公司 名单数据筛选方法、装置、计算机设备及存储介质
CN115329084A (zh) * 2022-08-26 2022-11-11 苏州大学 基于稀疏线性在线学习的垃圾邮件分类方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105184321A (zh) * 2015-09-10 2015-12-23 北京金山安全软件有限公司 一种针对于ftrl模型的数据处理方法及装置
CN105989374A (zh) * 2015-03-03 2016-10-05 阿里巴巴集团控股有限公司 一种在线训练模型的方法和设备
CN107194532A (zh) * 2017-04-07 2017-09-22 广东精点数据科技股份有限公司 基于大数据的保险业务分析方法
CN107609060A (zh) * 2017-08-28 2018-01-19 百度在线网络技术（北京）有限公司 资源推荐方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105989374A (zh) * 2015-03-03 2016-10-05 阿里巴巴集团控股有限公司 一种在线训练模型的方法和设备
CN105184321A (zh) * 2015-09-10 2015-12-23 北京金山安全软件有限公司 一种针对于ftrl模型的数据处理方法及装置
CN107194532A (zh) * 2017-04-07 2017-09-22 广东精点数据科技股份有限公司 基于大数据的保险业务分析方法
CN107609060A (zh) * 2017-08-28 2018-01-19 百度在线网络技术（北京）有限公司 资源推荐方法及装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114549048A (zh) * 2022-01-18 2022-05-27 北京健康之家科技有限公司 外呼用户的推荐方法和装置、电子设备

Also Published As

Publication number Publication date
CN108805613A (zh) 2018-11-13

Similar Documents

Publication Publication Date Title
WO2019223145A1 (zh) 电子装置、推销名单推荐方法、系统和计算机可读存储介质
US10452992B2 (en) Interactive interfaces for machine learning model evaluations
US9990558B2 (en) Generating image features based on robust feature-learning
US11574201B2 (en) Enhancing evolutionary optimization in uncertain environments by allocating evaluations via multi-armed bandit algorithms
WO2021027256A1 (zh) 处理交互序列数据的方法及装置
CA2953817C (en) Feature processing tradeoff management
US10963810B2 (en) Efficient duplicate detection for machine learning data sets
CN111639516B (zh) 基于机器学习的分析平台
CN103502899B (zh) 动态预测建模平台
US20190251471A1 (en) Machine learning device
CN109948121A (zh) 文章相似度挖掘方法、系统、设备及存储介质
CN110413786B (zh) 基于网页文本分类的数据处理方法、智能终端及存储介质
CN114663198A (zh) 基于用户画像的产品推荐方法、装置、设备及存储介质
WO2019061664A1 (zh) 电子装置、基于用户上网数据的产品推荐方法及存储介质
WO2023004632A1 (zh) 知识图谱的更新方法、装置、电子设备、存储介质及程序
Chen et al. An overview of diffusion models: Applications, guided generation, statistical rates and optimization
KR101700030B1 (ko) 사전 정보를 이용한 영상 물체 탐색 방법 및 이를 수행하는 장치
US11436529B1 (en) Method, apparatus, and computer program product for natural language processing
CN112464042A (zh) 根据关系图卷积网络的任务标签生成方法及相关装置
CN112258368A (zh) 政务事项的处理方法和装置
CN113822390B (zh) 用户画像构建方法、装置、电子设备和存储介质
WO2023121731A1 (en) Machine learning product development life cycle model
US20230012316A1 (en) Automation of leave request process
Cady Data Science: The Executive Summary-A Technical Book for Non-Technical Professionals
Signoretto et al. High level high performance computing for multitask learning of time-varying models

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18920088

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18920088

Country of ref document: EP

Kind code of ref document: A1