CN116308464B

CN116308464B - Target client acquisition system and method

Info

Publication number: CN116308464B
Application number: CN202310526584.5A
Authority: CN
Inventors: 梁少杰; 陈琪钛
Original assignee: Guangzhou Woti Mobile Technology Co ltd
Current assignee: Guangzhou Woti Mobile Technology Co ltd
Priority date: 2023-05-11
Filing date: 2023-05-11
Publication date: 2023-09-08
Anticipated expiration: 2043-05-11
Also published as: CN116308464A

Abstract

The invention provides a target customer acquisition system and method, the technical conception is that recruitment information and company service information play a vital role in target customer locking, secondly, the recruitment information and the service information do not need to be judged and screened manually, but are directly converted into a text data form, and then are processed and refined by a neural network system, and finally, the most critical customer popularization probability information can be acquired, thereby judging whether the inspected customer deserves commercial popularization. In the aspect of target client acquisition, the method can release enterprises from massive data, and the enterprises can judge by themselves by means of the technology, so that the client acquisition efficiency is greatly improved.

Description

Target client acquisition system and method

Technical Field

The invention relates to customer data screening, in particular to a target customer acquisition system and method.

Background

Currently, a large number of manufacturers attempt to accurately acquire target clients using each large ranking list on the internet, however, the ranking lists are open to the general public, and you see, i.e., others see. Ranking enterprises on such a list have already undergone one round of popularization and delivery. Most of these top-level enterprises already have stable suppliers. In other words, the corporation's collaboration company is essentially established already when the corporation is on the top. At this point, the business is lost, and the likelihood of successfully acquiring the target customer and achieving the intent of the collaboration is very low.

Therefore, the Internet resources are fully utilized, the current general guest-obtaining mode is broken, the first machine is occupied in business competition, and the urgent needs of all large manufacturers are met.

Disclosure of Invention

The invention provides a target client acquisition method and a target client acquisition system, which effectively solve the technical problems existing in the prior art.

Specifically, the invention provides a target client acquisition method, which comprises the following steps: crawling information of a target client from the Internet, wherein the information comprises recruitment information and business information; the recruitment information and the service information comprise quantifiable text data and unquantifiable text data, the quantifiable text data is converted into numerical features, the unquantifiable text data is converted into vector forms, the numerical features and the vector forms are combined into input word vectors, and the input word vectors are input into a neural network structure; the neural network structure comprises an input layer, a bidirectional LSTM layer, a CNN network layer, a global max pooling layer, a full connection layer and a Softmax network layer, wherein the input layer inputs the input word vector to the bidirectional LSTM layer, the bidirectional LSTM layer trains the input word vector, and the bidirectional LSTM layer carries out dot product operation on the trained word vector, so that a two-dimensional similarity matrix is constructed; the two-dimensional similarity matrix is accessed into a CNN convolutional neural network in the CNN network layer, so that a feature extraction network is constructed from the two-dimensional similarity matrix, and neural features are extracted; the nerve characteristics are accessed into the global maximum pooling layer, and the global maximum pooling layer eliminates non-key element characteristics in the nerve characteristics; the global maximum pooling layer is connected with the full-connection layer, and the full-connection layer integrates the neural characteristics with non-key element characteristics removed into local information with category differentiation; the full connection layer is then connected to a Softmax network layer that converts the local information integrated by the full connection layer into a probability distribution, thereby deriving the probability of promotion for the target client.

Preferably, the recruitment information comprises recruitment post, recruitment academy and salary treatment; the business information comprises the company industry, the number of people on a company scale, the financing stage, the product type and the product quantity.

Optionally, payroll treatments, product quantities, company-scale population and financing stages are converted into numerical features, recruitment posts, recruitment academia, company industry, academic requirements, product types are converted into vector forms.

Optionally, the recruitment information and the traffic information are represented in discrete form when converted to an input word vector.

Preferably, the bi-directional LSTM layers include a backward LSTM layer and a forward LSTM layer.

More preferably, the backward LSTM layer and the forward LSTM layer acquire forward text sequence information and backward text sequence information of the input word vector at the same time, and fuse the forward text sequence information and the backward text sequence information.

More preferably, the backward LSTM layer and the forward LSTM layer perform training of a bidirectional cyclic neural network on the input word vector to obtain forward text sequence information and backward text sequence information at each moment, and a two-dimensional similarity matrix is constructed based on splicing of the forward text sequence information and the backward text sequence information, wherein each element in the two-dimensional similarity matrix is a dot product result of the forward text sequence information and the backward text sequence information at different moments.

More preferably, the backward LSTM layer and the forward LSTM layer each construct an LSTM network structure of neurons, and a residual network is added in each LSTM network structure.

More preferably, the LSTM network structure is made up of 128 neurons.

The invention also provides a target client acquisition system which adopts the target client acquisition method.

The invention firstly notices that recruitment information and company business information play a vital role in target client locking, but not noticeable before the role, secondly, the recruitment information and the business information do not need to be judged and screened manually, but are directly converted into a text data form, and then processed and refined by a neural network system, and finally the most critical client popularization probability information can be obtained, thereby judging whether the inspected client deserves commercial popularization. In a word, in the aspect of target client acquisition, the method can release enterprises from massive data, and the enterprises can judge by themselves by means of a technical means, so that the client acquisition efficiency is greatly improved. In small terms, the technical means effectively promote the business development of enterprises, and in large terms, the economical and smooth circulation of commodities can be promoted.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following discussion will discuss the embodiments or the drawings required in the description of the prior art, and it is obvious that the technical solutions described in connection with the drawings are only some embodiments of the present invention, and that other embodiments and drawings thereof can be obtained according to the embodiments shown in the drawings without inventive effort for a person skilled in the art.

Fig. 1 shows a flow chart of the operation of an LSTM network in a target client acquisition system according to the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made in detail with reference to the accompanying drawings, wherein it is apparent that the embodiments described are only some, but not all embodiments of the present invention. All other embodiments, which can be made by a person of ordinary skill in the art without the need for inventive faculty, are within the scope of the invention, based on the embodiments described in the present invention.

As noted in the background section, it is not practical to accurately obtain a target client using a list ranking. Therefore, the concept of the invention considers that the flow of the passenger obtaining mode is advanced, and the technical mode is utilized to obtain the target clients which are not necessarily in a list and even still remain in a word girl in reality, wherein the target clients may not have great name, but have enough strength and actual requirements, but are easy to cut in due to relatively low exposure rate, so that the passenger obtaining is successful.

To this end, the present inventive concept contemplates data screening from the cut-in of recruitment information (e.g., recruitment posts, recruitment treatments, etc.) and business information (e.g., financing amounts, company sizes, etc.) on the internet, thereby obtaining a suitable acquisition list. Recruitment information may be obtained from a recruitment website. The recruitment website is used for recruiting the company, the business is in an ascending period, and the company recruits the company, so that on one hand, the lack of relevant personnel for the company to develop relevant business is indicated, and on the other hand, the upstream industrial requirement of the company in the aspect of relevant business is also indicated. Company information often reflects the own strength of the company and determines the purchasing power and strategic direction of the company.

Based on the above-mentioned thought, the target client obtaining method and system of the invention crawls company information on the internet, including recruitment information and business information, such as the number of persons on a company scale, recruitment posts, financing conditions, product types, product numbers, corresponding upper-level times and the like, and takes such company information as input dimension, adopts a multi-layer Bi-directional long-short-term memory neural network (i.e. Bi-LSTM, hereinafter written into Bi-directional LSTM) model as training, finally obtains corresponding results through the operation of the model, and can achieve accurate locking of ideal target clients by taking the results as the acquisition standards.

The specific flow steps of the target client acquisition method according to the present invention will be described in detail hereinafter.

First, crawling an internet website gathers a plurality of key information, which may include recruitment posts, recruitment academies, payroll treatments, and the like, and business key information, which may include company industry, company-scale population, financing stage, product category, product quantity, and the like.

As described above, these key information form text data that ultimately is trained through the bi-directional LSTM model to ultimately form the acquisition criteria. It should be noted, however, that the bi-directional LSTM model does not directly process text data, and that words in the text need to be first converted into vector representations, called Word vectors (Word vectors). Word vectors are vector representations that map words into a low-dimensional continuous vector space.

Specifically, the recruitment key information and the traffic key information include quantifiable text data and unquantifiable text data that are converted into numerical feature and vector forms, respectively. The quantifiable text data of payroll treatment, product quantity, company scale number and financing stage are directly converted into numerical characteristics, and recruitment post, recruitment academy, company industry, academy requirements and product types, which are unquantifiable text data, can be converted into vector forms. The numerical feature and vector form are combined into an input word vector for customizing the input of the neural network model.

Preferably, the recruitment key information and the traffic key information are not directly related to each other in context, nor are they required in word order, so that a discrete representation (one-hot representation) can be used in converting to a word vector. Specifically, a dictionary (vocabolary) needs to be prepared, each word in the text is converted into its serial number ID in the dictionary, and a special word [ UNK ] needs to be set to represent the unknown word. In processing text, if words not in the dictionary are encountered, the text is uniformly processed as [ UNK ].

After completing the transition from the key information to word vectors and numerical features, these input word vectors will be input to a neural network model that classifies based on text data, as shown in fig. 1. Fig. 1 shows a flow chart of the operation of a bi-directional LSTM network in a target client acquisition system according to the present invention.

The neural network model adopts a multi-layer two-way long-short-term memory neural network (i.e. two-way LSTM) structure and is used for classifying recruitment information and business information represented by word vectors and judging whether the company needs to be popularized with advertisements.

The neural network model comprises an input layer, a bidirectional LSTM layer (divided into a backward LSTM layer and a forward LSTM layer), a network layer, a global maximum pooling layer, a full connection layer and a Softmax network layer. Layer by layer detailed description will follow.

The word vectors are input to the bi-directional LSTM layer through the input layer, shown in the figure as the backward LSTM layer and the forward LSTM layer. And training and outputting the word vector by the bidirectional LSTM layer.

The bidirectional LSTM layer is a bidirectional layer, so that forward text sequence information and backward text sequence information of an input word vector can be obtained at the same time, and the forward text sequence information and the backward text sequence information are fused, thereby helping to better capture context and context information in a text sequence when recruitment information and company service information are processed, so that text data can be more comprehensively understood, the processing precision and effect of the text data are improved, and further, the accuracy of locking and popularization prediction of a target client is improved.

The backward LSTM layer and the forward LSTM layer can respectively construct an LSTM network structure (for example, composed of 128 neurons) composed of neurons, and a residual network is added in each LSTM network structure. The formula for the residual network is expressed as follows:

y＝F(x ,Wi)+x

where F (x, wi) is the output of the forward LSTM layer, x is the output of the backward LSTM layer, and y is the overall input of the bi-directional LSTM network.

Then, dot product operation can be performed on the outputs of the two LSTM network structures, thereby constructing a two-dimensional similarity matrix.

In practice, the bidirectional LSTM layer performs training of a bidirectional cyclic neural network on an input word vector to obtain forward text sequence information and backward text sequence information at each moment. These text sequence information can be used to construct a similarity matrix. Based on the splicing operation of the forward text sequence information and the backward text sequence information, a two-dimensional similarity matrix can be constructed, wherein each element in the matrix is the dot product result of the forward text sequence information and the backward text sequence information at different moments.

Then, the two-dimensional similarity matrix is connected into a CNN convolutional neural network, and a feature extraction network is constructed on the basis of the two-dimensional similarity matrix so as to extract neural features.

In the feature extraction network, different features, namely, extraction of nerve features, can be extracted through construction of a two-dimensional similarity matrix, and the features can help to judge which clients are more in line with popularization targets of enterprises, so that the passenger acquisition efficiency is improved, and the popularization cost is reduced.

Thus, in the process of target client acquisition, extracting neural features can help determine which clients have higher popularization probabilities, thereby helping enterprises to more accurately target clients.

Subsequently, the max pooling layer is accessed. The maximum pooling layer is equivalent to a data filter and is mainly used for eliminating non-key element features in the CNN convolutional neural network, so that the key points are highlighted, the calculation speed is improved, and the robustness of the extracted network features is improved.

And the maximum pooling layer is accessed into the full-connection layer. The full-connection layer is mainly used for integrating the neural characteristics extracted after the filtering of the maximum pooling layer for one time, and especially integrating the local information with category differentiation in the maximum pooling layer.

In particular, each neuron of the fully connected layer is connected to all neurons of the previous layer, and thus the characteristic information of the previous layer (i.e., the information of the neural characteristics) can be more comprehensively acquired.

In the target client acquisition process, the full connection layer can be used for further processing the feature information extracted by the previous layers to obtain higher-level feature representation, so that the prediction and judgment of the target client popularization probability can be more accurately carried out. Meanwhile, the full connection layer can be used for integrating relevant features of a target client to obtain more comprehensive and representative feature representation, so that the accuracy and generalization capability of the model are improved.

The full-connection layer is then connected to the Softmax network layer, feature data integrated through the full-connection layer is converted into probability distribution, and finally target client promotion probability is obtained and used for judging whether the company is worth business to contact promotion service.

The target client acquisition system and method provided by the present invention has been described in detail so far. The invention firstly notices that recruitment information and company business information play a vital role in target client locking, but not noticeable before the role, secondly, the recruitment information and the business information do not need to be judged and screened manually, but are directly converted into a text data form, and then processed and refined by a neural network system, and finally the most critical client popularization probability information can be obtained, thereby judging whether the inspected client deserves commercial popularization. In a word, in the aspect of target client acquisition, the method can release enterprises from massive data, and the enterprises can judge by themselves by means of a technical means, so that the client acquisition efficiency is greatly improved. In small terms, the technical means effectively promote the business development of enterprises, and in large terms, the economical and smooth circulation of commodities can be promoted.

The foregoing description of the exemplary embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any modifications, equivalents, and variations which fall within the spirit and scope of the invention are intended to be included in the scope of the invention.

Claims

1. A method for obtaining a target client, the method comprising:

crawling information of a target client from the Internet, wherein the information comprises recruitment information and business information;

the recruitment information and the service information comprise quantifiable text data and unquantifiable text data, the quantifiable text data is converted into numerical type features, the unquantifiable text data is converted into vector form, the numerical type features and the vector form are combined into input word vectors, the input word vectors are input into a neural network structure, and the neural network structure comprises an input layer, a bidirectional LSTM layer, a CNN network layer, a global max pooling layer, a full connection layer and a Softmax network layer;

wherein the input layer inputs the input word vector to a bidirectional LSTM layer for training, the bidirectional LSTM layer comprises a backward LSTM layer and a forward LSTM layer,

the backward LSTM layer and the forward LSTM layer simultaneously acquire the forward text sequence information and the backward text sequence information of the input word vector, and fuse the forward text sequence information and the backward text sequence information,

the backward LSTM layer and the forward LSTM layer perform training of a bidirectional cyclic neural network on the input word vector to obtain forward text sequence information and backward text sequence information at each moment, and construct a two-dimensional similarity matrix based on splicing of the forward text sequence information and the backward text sequence information, wherein each element in the two-dimensional similarity matrix is a dot product result of the forward text sequence information and the backward text sequence information at different moments,

the two-dimensional similarity matrix is accessed into a CNN convolutional neural network in the CNN network layer, so that a feature extraction network is constructed from the two-dimensional similarity matrix, and neural features are extracted;

the nerve characteristics are accessed into the global maximum pooling layer to eliminate non-key element characteristics in the nerve characteristics;

the global maximum pooling layer is connected with the full-connection layer, and the full-connection layer integrates the neural characteristics with non-key element characteristics removed into local information with category differentiation;

the full connection layer is then connected to a Softmax network layer that converts the local information integrated by the full connection layer into a probability distribution, thereby deriving the probability of promotion for the target client.

2. The target customer acquisition method of claim 1, wherein the recruitment information comprises a recruitment post, a recruitment academy, and a payroll treatment; the business information comprises the company industry, the number of people on a company scale, the financing stage, the product type and the product quantity.

3. The method of claim 2, wherein payroll treatments, product quantities, company size numbers and financing stages are converted to numerical features, recruitment posts, recruitment academies, company industries, academies requirements, product categories are converted to vector forms.

4. The target client acquisition method according to claim 1 or 2, characterized in that the recruitment information and the traffic information are represented in discrete form when converted to an input word vector.

5. The target client acquisition method of claim 1 wherein the backward LSTM layer and the forward LSTM layer each construct LSTM network structures of neurons, each LSTM network structure having a residual network added thereto.

6. The target client acquisition method of claim 5, wherein the LSTM network structure is comprised of 128 neurons.

7. A target client acquisition system for performing the target client acquisition method according to any one of claims 1 to 6.