Universe customer portrait construction method and device based on big data

Abstract

The invention relates to a global customer portrait construction method and device based on big data, and the method comprises the steps: obtaining the login information of a user, and recognizing first text data inputted by the user; calling first customer information matched with the first text data from a big data network by using a data crawler; inputting the first customer information into a preset characterization model, and obtaining the weight of each keyword in the first customer information calculated by the characterization model through an attention characterization algorithm; generating first portrait label information corresponding to the first customer information on the basis of each keyword weight; establishing a communication interface with the user, continuously obtaining second text data input by the user through the communication interface and a corresponding confirmation signal, and circularly executing calculation of a representation model and generation of a portrait label through the second text data and confirmation information to generate second portrait label information; and constructing a customer portrait through the obtained first portrait label information and the second portrait label information.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for constructing a global customer profile based on big data. Background Art

[0002] In today's information age, the rapid development of big data technology has brought unprecedented opportunities and challenges to all walks of life. In order to improve market competitiveness and achieve precision marketing and personalized services, enterprises are in urgent need of an efficient and comprehensive customer portrait construction method. Traditional customer portrait construction methods mostly rely on limited data sources and manual analysis, which makes it difficult to fully and real-time reflect customer characteristics and behavior patterns, and it is difficult to meet the current complex and changing market needs. Summary of the invention

[0003] The main purpose of the present invention is to provide a method and device for constructing a global customer profile based on big data, so as to solve the traditional customer profile construction model.

[0004] To achieve the above object, the present invention provides a method for constructing a global customer profile based on big data, comprising the following steps:

[0005] Acquire the user's login information and identify the first text data input by the user;

[0006] Retrieving first customer information matching the first text data from a big data network using a data crawler, where the first customer information includes multiple items;

[0007] Inputting the first customer information into a preset representation model, and obtaining the weight of each keyword in the first customer information after the representation model is calculated by an attention representation algorithm;

[0008] Based on the weights of the respective keywords, generating first portrait tag information corresponding to the first customer information;

[0009] Establishing a communication interface with the user, continuously acquiring second text data and a corresponding confirmation signal input by the user through the communication interface, and cyclically executing calculation of the representation model and generation of the portrait label through the second text data and the confirmation information to generate second portrait label information;

[0010] A customer portrait is constructed using the obtained first portrait label information and second portrait label information.

[0011] Furthermore, the step of using a data crawler to retrieve first customer information matching the first text data from a big data network includes:

[0012] Retrieving a pre-bound big data network address, wherein the big data network address includes a plurality of customer information databases;

[0013] Using a dedicated data crawler to input the first text data into a big data network address to obtain the first customer information;

[0014] The plurality of first customer information are classified and data compressed for input into the characterization model.

[0015] Furthermore, the method of calculating the weight of each keyword in the first customer information by using the representation model includes:

[0016] Using the first processing layer of the representation model to perform customer classification on the first customer information;

[0017] The classified first customer information is calculated by the attention representation algorithm of the first processing layer to obtain attention representations corresponding to each keyword in the first customer information;

[0018] Associating each keyword obtained by the attention representation with corresponding customer information to obtain each keyword weight;

[0019] Among them, the attention representation algorithm includes:

[0020]

[0021] Above To obtain the attention representation value, is a text vector in the first customer information. After the customer classification of the first customer information, the text vectors h of each item after classification are accumulated by exp operation. Q The denominator is obtained by Identify h Q At least one keyword related to the first text data And use exp to calculate at least one keyword To get the numerator, and thus get the attention value

[0022] Furthermore, the attention representation algorithm further includes parallel conversion calculation, after the step of classifying the first customer information by using the first processing layer of the representation model, including:

[0023] Vectorizing the information text content of each customer in the first customer information;

[0024] The vectorized information text is calculated using a parallel conversion calculation algorithm to obtain fitting data of each customer in the first customer information. The parallel conversion calculation algorithm is:

[0025]

[0026] Above is a scalar text scalar δ(h Q ), i.e., a text scalar Keywords is in scalar form; the above To fit the text vector h Q The keywords The fitting data, T is the fitting range, and the weight value W is added to each word in the text scalar in the fitting data. S , by comparing the text vector h Q Determine the text vector h based on the topic Q When there are non-keywords in the text, reduce the weight value W corresponding to the non-keywords S , otherwise increase the keyword weight value W S , when there is a semantic connection between keywords and non-keywords, find the weight value W of the two S If the average value is higher than the original W S , then both are keywords; the above deviation value b S , used to better construct the fitting data, because using This item cannot completely fit the data, so the deviation value b needs to be introduced S , divide keywords and non-keywords into pre-set levels (level 1 / 2 / 3 keywords and level 1 / 2 / 3 non-keywords respectively), when fitting data When the non-keywords in the representation model are regarded as 1 / 2 / 3 level non-keywords at the first level, their corresponding weight values ​​are (1 level non-keyword weight value W S -b S , Level 2 weight value W S -2b S , level 3 weight value W S -3b S ), when fitting the data When the keywords in the representation model are regarded as 1 / 2 / 3 level keywords at the first level, their corresponding weight values ​​are (1 level keyword weight value W S +b S , Level 2 weight value W S +2b S , level 3 weight value W S +3b S ); Through parallel transformation in the first layer of the representation model, the text vector is converted into a text scalar, and then the weight and bias processing is introduced to improve the accuracy and reliability of the attention algorithm for the attention calculation of the text vectors between the customer and the previous customer in the first customer information.

[0027] Furthermore, the training process of the representation model includes:

[0028] Formulate a framework to obtain an initial model having a first layer, a second layer and a last layer, and the first layer, the second layer and the last layer of the initial model are preset with corresponding algorithms;

[0029] Import training data into the initial model;

[0030] Calculate the weights and biases with the minimum cross entropy loss in the training data using the gradient descent method through the loss function;

[0031] The weights and biases are introduced into the algorithms corresponding to each layer of the framework; wherein,

[0032] The formulation framework is to create three layers in the initial model, namely, the first layer, the second layer and the last layer. The first layer presets an attention algorithm, which is used to determine the weight of each keyword in the first customer information for the customer label; the second layer presets an influence algorithm, which is used to determine the mutual influence probability of the keywords between each customer information text in the first customer information; the last layer presets a representation integration, which is used to integrate all the obtained influence probabilities, produce influence probability data, and determine the labels of each customer in a macro-level overview form;

[0033] The above training data is: {(Q i , R i ), y i}, the loss function is: in, Right now

[0034] Q, R are the text vectors between two customers in the first customer information, and the training data Q i , R i is the i-th word in the Q, R text vector;

[0035] The above training data y i Q i , R i The probability of mutual influence between them is used to preset known specific values;

[0036] In the above loss function is the predicted Q i , R i The probability of attitude j, the above attitudes include support and negation, when When the threshold is reached, it is supported, otherwise it is denied; the prediction method is to calculate Q i , R i The influence value of the two words on each other, and the influence value and the pre-obtained attention value are combined to calculate the probability of the attitude through the influence value and the attention value is the predicted support / negation, and yi is the support / negation of the actual situation;

[0037] The above is the vector combination of the two words. From the above, we can see that the loss function has been determined. In v and probability y i The specific value of , then, through the gradient descent method, the weight W with the minimum cross entropy in the loss function is obtained l With deviation b l ;

[0038] The “-” in the above is the symbol for minimizing the loss. is the vector concatenation operator;

[0039] In getting the weight W l With deviation b l Afterwards, the weight W l With deviation b l Import it into the algorithms corresponding to each layer of the framework to achieve the training of the representation model.

[0040] Furthermore, based on the weights of the respective keywords, the step of generating first portrait tag information corresponding to the first customer information includes:

[0041] Using the influence algorithm of the second layer in the representation model, an association operation is performed on the keywords with corresponding weights to obtain a first influence representation and a second influence representation;

[0042] The influence algorithm is:

[0043]

[0044]

[0045] Among them, Q C is the first influence representation, R C is the second influence representation. The first influence representation and the second influence representation are influence values ​​calculated by comparing the corresponding weighted keywords in the first customer information. and are the corresponding influence values ​​respectively. and They are respectively the keywords of influence between customers in the first customer information.

[0046] Further, after the step of obtaining the first influence representation and the second influence representation, the method further includes:

[0047] Using the representation integration algorithm of the tail processing layer in the representation model, the attention representation and the influence representation corresponding to the customer text in the first customer information are subjected to a total integration operation to obtain a total representation label associated with the first text data in the first customer information;

[0048] At the same time, based on the influence probability algorithm under the representation integration algorithm, the sub-representation labels associated with each customer in the first customer information are calculated to obtain the label word influence probability between two customers;

[0049] Wherein, the characterization integration algorithm includes:

[0050]

[0051] Q, R are integrated representations of text vectors between customer vectors in the first customer information, and v is the total representation.

[0052] Furthermore, based on the influence probability algorithm under the representation integration algorithm, the sub-representation tags associated with each customer in the first customer information are calculated to obtain the tag word influence probability between two customers, including:

[0053] The total representation v is introduced into the influence probability algorithm to obtain the influence probability data of the sub-representation tags associated with each customer. The influence probability algorithm is:

[0054] y=softmax(W l v+b l );

[0055] Among them, the weight value W has been determined when pre-training the representation model l With deviation value b l ;

[0056] The above y is the probability of influence of label words between customers, and the above weight value W l With deviation value b l It is pre-imported into the influence probability algorithm, and the total representation v is imported into the influence probability algorithm to obtain its corresponding influence probability y.

[0057] Furthermore, the step of constructing a customer portrait by using the obtained first portrait tag information and second portrait tag information includes:

[0058] Performing a threshold comparison and judgment on the total characterization label value and the sub-characterization label value corresponding to the first portrait label information and the second portrait label information;

[0059] After the judgment, the remaining portrait label information is output to obtain the customer portrait.

[0060] The present invention also proposes a device for constructing a global customer portrait based on big data, comprising:

[0061] An acquisition unit, used to acquire the user's login information and identify the first text data input by the user;

[0062] A network unit, configured to retrieve first customer information matching the first text data from a big data network by using a data crawler, wherein the first customer information includes multiple items;

[0063] A model calculation unit, used to input the first customer information into a preset representation model, and obtain the weight of each keyword in the first customer information after the representation model is calculated by an attention representation algorithm;

[0064] A first label unit, configured to generate first portrait label information corresponding to the first customer information based on each of the keyword weights;

[0065] A second label unit is used to establish a communication interface with the user, continuously obtain second text data and a corresponding confirmation signal input by the user through the communication interface, and cyclically perform calculation of the representation model and generation of the portrait label through the second text data and the confirmation information to generate second portrait label information;

[0066] A portrait unit is used to construct a customer portrait by obtaining the first portrait tag information and the second portrait tag information.

[0067] The method and device for constructing a global customer profile based on big data provided by the present invention have the following beneficial effects:

[0068] (1) The use of advanced technologies such as attention representation algorithm, parallel conversion calculation algorithm and influence algorithm can more accurately calculate and identify customers' keyword weights and tag information, thereby improving the accuracy of customer portraits.

[0069] (2) The representation model is trainable and can be continuously optimized and adjusted based on new data and feedback, further improving the accuracy and reliability of customer portraits. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flowchart of a method for constructing a global customer profile based on big data in one embodiment of the present invention;

[0071] Figure 2 It is a structural block diagram of a device for constructing a global customer profile based on big data in one embodiment of the present invention;

[0072] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0074] Reference Figure 1 , which is a flow chart of a method for constructing a global customer profile based on big data proposed by the present invention, comprising the following steps:

[0075] S100, obtaining user login information and identifying first text data input by the user;

[0076] S200, using a data crawler to retrieve first customer information matching the first text data from a big data network, where the first customer information includes multiple items;

[0077] S300, inputting the first customer information into a preset representation model, and obtaining the weight of each keyword in the first customer information after the representation model is calculated by an attention representation algorithm;

[0078] S400, generating first portrait tag information corresponding to the first customer information based on the weights of the respective keywords;

[0079] S500, establishing a communication interface with the user, continuously acquiring second text data and a corresponding confirmation signal input by the user through the communication interface, and cyclically executing calculation of the representation model and generation of the portrait label through the second text data and the confirmation information to generate second portrait label information;

[0080] S600, constructing a customer portrait through the obtained first portrait label information and second portrait label information.

[0081] In S200, the step of using a data crawler to retrieve first customer information matching the first text data from a big data network includes:

[0082] Retrieving a pre-bound big data network address, wherein the big data network address includes a plurality of customer information databases;

[0083] Using a dedicated data crawler to input the first text data into a big data network address to obtain the first customer information;

[0084] The plurality of first customer information are classified and data compressed for input into the characterization model.

[0085] As mentioned in S300, the method for calculating the weight of each keyword in the first customer information by using the representation model includes:

[0086] Using the first processing layer of the representation model to perform customer classification on the first customer information;

[0087] The classified first customer information is calculated by the attention representation algorithm of the first processing layer to obtain attention representations corresponding to each keyword in the first customer information;

[0088] Associating each keyword obtained by the attention representation with corresponding customer information to obtain each keyword weight;

[0089] Among them, the attention representation algorithm includes:

[0090]

[0091] Above To obtain the attention representation value, is a text vector in the first customer information. After the customer classification of the first customer information, the text vectors h of each item after classification are accumulated by exp operation. Q The denominator is obtained by Identify h Q At least one keyword related to the first text data And use exp to calculate at least one keyword To get the numerator, and thus get the attention value

[0092] For example: when T=10, t=7, the value calculated by exp is 0.763.., so the attention value is 0.763; when T=10, t=10, the value calculated by exp is 1, so the attention value is 1 (that is, all words in the text vector are keywords).

[0093] More specifically, the attention representation algorithm further includes parallel conversion calculation, after the step of classifying the first customer information using the first processing layer of the representation model, including:

[0094] Vectorizing the information text content of each customer in the first customer information;

[0095] The vectorized information text is calculated using a parallel conversion calculation algorithm to obtain fitting data of each customer in the first customer information. The parallel conversion calculation algorithm is:

[0096]

[0097] Above is a scalar text scalar δ(h Q ), i.e., a text scalar Keywords is in scalar form; the above To fit the text vector h Q The keywords The fitting data, T is the fitting range, and the weight value W is added to each word in the text scalar in the fitting data. S , by comparing the text vector h Q Determine the text vector h based on the topic Q When the non-keywords in the text are used, the weight value W corresponding to the non-keywords is reduced. S , otherwise increase the keyword weight value W S , when there is a semantic connection between keywords and non-keywords, find the weight value W of the two S If the average value is higher than the original W S , then both are keywords; the above deviation value b S , used to better construct the fitting data, because using This item cannot completely fit the data, so it is necessary to introduce a deviation value b S , divide keywords and non-keywords into pre-set levels (level 1 / 2 / 3 keywords and level 1 / 2 / 3 non-keywords respectively), when fitting data When the non-keywords in the representation model are regarded as 1 / 2 / 3 level non-keywords at the first level, their corresponding weight values ​​are (1 level non-keyword weight value W S -b S , Level 2 weight value W S -2b S , level 3 weight value W S -3b S ), when fitting the data When the keywords in the representation model are regarded as 1 / 2 / 3 level keywords at the first level, their corresponding weight values ​​are (1 level keyword weight value W S +b S , Level 2 weight value W S +2b S , level 3 weight value W S +3b S ); Through parallel transformation in the first layer of the representation model, the text vector is converted into a text scalar, and then the weight and bias processing is introduced to improve the accuracy and reliability of the attention algorithm for the attention calculation of the text vectors between the customer and the previous customer in the first customer information.

[0098] The training process of the above-mentioned representation model includes:

[0099] Formulate a framework to obtain an initial model having a first layer, a second layer and a last layer, and the first layer, the second layer and the last layer of the initial model are preset with corresponding algorithms;

[0100] Import training data into the initial model;

[0101] Calculate the weights and biases with the minimum cross entropy loss in the training data using the gradient descent method through the loss function;

[0102] The weights and biases are introduced into the algorithms corresponding to each layer of the framework; wherein,

[0103] The formulation framework is to create three layers in the initial model, namely, the first layer, the second layer and the last layer. The first layer presets an attention algorithm, which is used to determine the weight of each keyword in the first customer information for the customer label; the second layer presets an influence algorithm, which is used to determine the mutual influence probability of the keywords between each customer information text in the first customer information; the last layer presets a representation integration, which is used to integrate all the obtained influence probabilities, produce influence probability data, and determine the labels of each customer in a macro-level overview form;

[0104] The above training data is: {(Q i , R i ), y i}, the loss function is: in, Right now

[0105] Q, R are the text vectors between two customers in the first customer information, and the training data Q i , R i is the i-th word in the Q, R text vector;

[0106] The above training data y i Q i , R i The probability of mutual influence between them is used to preset known specific values;

[0107] In the above loss function is the predicted Q i , R i The probability of attitude j, the above attitudes include support and negation, when When the threshold is reached, it is supported, otherwise it is denied; the prediction method is to calculate Q i , R i The influence value of the two words on each other, and the influence value and the pre-obtained attention value are combined to calculate the probability of the attitude through the influence value and the attention value is the prediction support / negation, and y i is the support / negation of the real situation;

[0108] The above is the vector combination of the two words. From the above, we can see that the loss function has been determined. In v and probability y i The specific value of , then, through the gradient descent method, the weight W with the minimum cross entropy in the loss function is obtained l With deviation b l ;

[0109] The “-” in the above is the symbol for minimizing the loss. is the vector concatenation operator;

[0110] In getting the weight W l With deviation b l Afterwards, the weight W l With deviation b l Import it into the algorithms corresponding to each layer of the framework to achieve the training of the representation model.

[0111] Specifically, the step of generating first portrait tag information corresponding to the first customer information based on the weights of the respective keywords includes:

[0112] Using the influence algorithm of the second layer in the representation model, an association operation is performed on the keywords with corresponding weights to obtain a first influence representation and a second influence representation;

[0113] The influence algorithm is:

[0114]

[0115]

[0116] Among them, Q C is the first influence representation, R C is the second influence representation. The first influence representation and the second influence representation are influence values ​​calculated by comparing the corresponding weighted keywords in the first customer information. and are the corresponding influence values ​​respectively. and They are respectively the keywords of influence between customers in the first customer information.

[0117] Hypothetical impact value The first influence representation Among them, [1, t]∈ are the influential keywords related to each customer. By obtaining the influence representation, it is convenient to determine the most influential label information from the influence probability according to the representation.

[0118] In one embodiment, after the steps of obtaining the first influence representation and the second influence representation, the following steps are included:

[0119] Using the representation integration algorithm of the tail processing layer in the representation model, the attention representation and the influence representation corresponding to the customer text in the first customer information are subjected to a total integration operation to obtain a total representation label associated with the first text data in the first customer information;

[0120] At the same time, based on the influence probability algorithm under the representation integration algorithm, the sub-representation labels associated with each customer in the first customer information are calculated to obtain the label word influence probability between two customers;

[0121] Wherein, the characterization integration algorithm includes:

[0122]

[0123] Q, R is the integrated representation of the text vectors between the customer vectors in the first customer information, and v is the total representation. The above Q is for example: attention value but Influence Value The first influence representation in are the keywords in the text vector, The keywords that influence the customers are related to each other, namely Q and R are as follows:

[0124]

[0125]

[0126] Furthermore, v is as follows:

[0127]

[0128] Thus, the total representation v of each first vector text and second vector text is obtained.

[0129] Specifically, based on the influence probability algorithm under the representation integration algorithm, the sub-representation tags associated with each customer in the first customer information are calculated to obtain the tag word influence probability between two customers, including:

[0130] The total representation v is introduced into the influence probability algorithm to obtain the influence probability data of the sub-representation tags associated with each customer. The influence probability algorithm is:

[0131] y=softmax(W l v+b1);

[0132] Among them, the weight value W has been determined when pre-training the representation model l With deviation value b l ;

[0133] The above y is the probability of influence of label words between customers, and the above weight value W l With deviation value b l It is pre-imported into the influence probability algorithm, and the total representation v is imported into the influence probability algorithm to obtain its corresponding influence probability y.

[0134] The impact probability y is the probability value of the first text data affecting all customer information. The higher the probability value, the higher the influence. For example, when the probability data y=1, it is the data with the highest influence, and all text data (i.e., the second and third text data) issued after the first text data are affected by it.

[0135] For example, when the customer text keywords crawled by the first text vector (customer 1 in the first customer information) and the second text vector (customer 2 in the first customer information) are t=7 and t=8 respectively, and when the influencing keyword between the first text vector and the second text vector is t=4, then,

[0136]

[0137] Furthermore, when the weight value W pre-imported into the influence probability algorithm l When the deviation value b1 is 0.2 and 1 respectively.

[0138]

[0139] Therefore, the total number of keywords is 10.2, and the attention value 0.76 and the influence value 0.76 are added thereto, which yields a value of 11.72, when the total number of all keywords in the identified first text data is 30.

[0140] Y=11.72 / 30=39%, thus the influence probability data corresponding to customer 1 and customer 2 are obtained.

[0141] Furthermore, the step of constructing a customer portrait by using the obtained first portrait tag information and second portrait tag information includes:

[0142] Performing a threshold comparison and judgment on the total characterization label value and the sub-characterization label value corresponding to the first portrait label information and the second portrait label information;

[0143] After the judgment, the remaining portrait label information is output to obtain the customer portrait.

[0144] refer to Figure 2 The present invention proposes a device for constructing a global customer profile based on big data, comprising:

[0145] The acquisition unit 1 is used to acquire the user's login information and identify the first text data input by the user;

[0146] A network unit 2 is used to retrieve first customer information matching the first text data from a big data network by using a data crawler, where the first customer information includes multiple items;

[0147] Model calculation unit 3, used for inputting the first customer information into a preset representation model, and obtaining the weight of each keyword in the first customer information after the representation model is calculated by the attention representation algorithm;

[0148] A first label unit 4, used for generating first portrait label information corresponding to the first customer information based on each of the keyword weights;

[0149] The second label unit 5 is used to establish a communication interface with the user, continuously obtain the second text data and the corresponding confirmation signal input by the user through the communication interface, and cyclically perform the calculation of the representation model and the generation of the portrait label through the second text data and the confirmation information to generate the second portrait label information;

[0150] The portrait unit 6 is used to construct a customer portrait by using the obtained first portrait label information and second portrait label information.

[0151] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for constructing a global customer profile based on big data, characterized in that: include: Acquire the user's login information and identify the first text data input by the user; Retrieving first customer information matching the first text data from a big data network using a data crawler, where the first customer information includes multiple items; Inputting the first customer information into a preset representation model, and obtaining the weight of each keyword in the first customer information after the representation model is calculated by an attention representation algorithm; Based on the weights of the respective keywords, generating first portrait tag information corresponding to the first customer information; Establishing a communication interface with the user, continuously acquiring second text data and a corresponding confirmation signal input by the user through the communication interface, and cyclically executing calculation of the representation model and generation of the portrait label through the second text data and the confirmation information to generate second portrait label information; A customer portrait is constructed using the obtained first portrait label information and second portrait label information.

2. The method for constructing a global customer profile based on big data according to claim 1 is characterized in that: The step of using a data crawler to retrieve first customer information matching the first text data from a big data network includes: Retrieving a pre-bound big data network address, wherein the big data network address includes a plurality of customer information databases; Using a dedicated data crawler to input the first text data into a big data network address to obtain the first customer information; The plurality of first customer information are classified and data compressed for input into the characterization model.

3. The method for constructing a global customer profile based on big data according to claim 1 is characterized in that: The method for calculating the weight of each keyword in the first customer information by using the representation model includes: Using the first processing layer of the representation model to perform customer classification on the first customer information; The classified first customer information is calculated by the attention representation algorithm of the first processing layer to obtain attention representations corresponding to each keyword in the first customer information; Associating each keyword obtained by the attention representation with corresponding customer information to obtain each keyword weight; Among them, the attention representation algorithm includes: Above To obtain the attention representation value, is a text vector in the first customer information. After the customer classification of the first customer information, the text vectors h of each item after classification are accumulated by exp operation. Q The denominator is obtained by Identify h Q At least one keyword related to the first text data And use exp to calculate at least one keyword To get the numerator, and thus get the attention value 4. The method for constructing a global customer profile based on big data according to claim 3 is characterized in that: The attention representation algorithm further includes parallel conversion calculation, after the step of classifying the first customer information by using the first processing layer of the representation model, including: Vectorizing the information text content of each customer in the first customer information; The vectorized information text is calculated using a parallel conversion calculation algorithm to obtain fitting data of each customer in the first customer information. The parallel conversion calculation algorithm is: Above is a scalar text scalar δ(h Q ), i.e., a text scalar Keywords is in scalar form; the above To fit the text vector h Q The keywords The fitting data, T is the fitting range, and the weight value W is added to each word in the text scalar in the fitting data. S , by comparing the text vector h Q Determine the text vector h based on the topic Q When the non-keywords in the text are used, the weight value W corresponding to the non-keywords is reduced. S , otherwise increase the keyword weight value W S , when there is a semantic connection between keywords and non-keywords, find the weight value W of the two S If the average value is higher than the original W S , then both are keywords; the above deviation value b S , used to better construct the fitting data, because using This item cannot completely fit the data, so the deviation value b needs to be introduced S , divide keywords and non-keywords into pre-set levels (level 1 / 2 / 3 keywords and level 1 / 2 / 3 non-keywords respectively), when fitting data When the non-keywords in the representation model are regarded as 1 / 2 / 3 level non-keywords at the first level, their corresponding weight values ​​are (1 level non-keyword weight value W S -b S , Level 2 weight value W S -2b S , level 3 weight value W S -3b S ), when fitting the data When the keywords in the representation model are regarded as 1 / 2 / 3 level keywords at the first level, their corresponding weight values ​​are (1 level keyword weight value W S +b S , Level 2 weight value W S +2b S , level 3 weight value W S +3b S ); Through parallel transformation in the first layer of the representation model, the text vector is converted into a text scalar, and then the weight and bias processing is introduced to improve the accuracy and reliability of the attention algorithm for the attention calculation of the text vectors between the customer and the previous customer in the first customer information.

5. The method for constructing a global customer profile based on big data according to claim 4 is characterized in that: The training process of the representation model includes: Formulate a framework to obtain an initial model having a first layer, a second layer and a last layer, and the first layer, the second layer and the last layer of the initial model are preset with corresponding algorithms; Import training data into the initial model; Calculate the weights and biases with the minimum cross entropy loss in the training data using the gradient descent method through the loss function; The weights and biases are introduced into the algorithms corresponding to each layer of the framework; wherein, The formulation framework is to create three layers in the initial model, namely, the first layer, the second layer and the last layer. The first layer presets an attention algorithm, which is used to determine the weight of each keyword in the first customer information for the customer label; the second layer presets an influence algorithm, which is used to determine the mutual influence probability of the keywords between each customer information text in the first customer information; the last layer presets a representation integration, which is used to integrate all the obtained influence probabilities, produce influence probability data, and determine the labels of each customer in a macro-level overview form; The above training data is: {(Q i ,R i ),y i }, the loss function is: in, Right now Q, R are the text vectors between two customers in the first customer information, and the training data Q i ,R i is the i-th word in the Q, R text vector; The above training data y i Q i ,R i The probability of mutual influence between them is used to preset known specific values; In the above loss function is the predicted Q i ,R i The probability of attitude j, the above attitudes include support and negation, when When the threshold is reached, it is supported, otherwise it is denied; the prediction method is to calculate Q i ,R i The influence value of the two words on each other, and the influence value and the pre-obtained attention value are combined to calculate the probability of the attitude through the influence value and the attention value is the prediction support / negation, and y i is the support / negation of the real situation; The above is the vector combination of the two words. From the above, we can see that the loss function has been determined. In v and probability y i The specific value of , then, through the gradient descent method, the weight W with the minimum cross entropy in the loss function is obtained l With deviation b l ; The "-" in the above is to minimize the loss. is the vector concatenation operator; In getting the weight W l With deviation b l Afterwards, the weight W l With deviation b l Import it into the algorithms corresponding to each layer of the framework to achieve the training of the representation model.

6. The method for constructing a global customer profile based on big data according to claim 1, characterized in that: The step of generating first portrait tag information corresponding to the first customer information based on the respective keyword weights includes: Using the influence algorithm of the second layer in the representation model, an association operation is performed on the keywords with corresponding weights to obtain a first influence representation and a second influence representation; The influence algorithm is: Among them, Q C is the first influence representation, R C is the second influence representation. The first influence representation and the second influence representation are influence values ​​calculated by comparing the corresponding weighted keywords in the first customer information. and are the corresponding influence values ​​respectively. and They are respectively the keywords of influence between customers in the first customer information.

7. The method for constructing a global customer profile based on big data according to claim 6 is characterized in that: After the steps of obtaining the first influence representation and the second influence representation, the method further includes: Using the representation integration algorithm of the tail processing layer in the representation model, the attention representation and the influence representation corresponding to the customer text in the first customer information are subjected to a total integration operation to obtain a total representation label associated with the first text data in the first customer information; At the same time, based on the influence probability algorithm under the representation integration algorithm, the sub-representation labels associated with each customer in the first customer information are calculated to obtain the label word influence probability between two customers; Wherein, the characterization integration algorithm includes: Q, R are integrated representations of text vectors between customer vectors in the first customer information, and v is the total representation.

8. The method for constructing a global customer profile based on big data according to claim 7 is characterized in that: The step of calculating the sub-representation labels associated with each customer in the first customer information based on the influence probability algorithm under the representation integration algorithm to obtain the label word influence probability between two customers includes: The total representation v is introduced into the influence probability algorithm to obtain the influence probability data of the sub-representation tags associated with each customer. The influence probability algorithm is: y =softmax(W l v+b l ): Among them, the weight value W has been determined when pre-training the representation model l With deviation value b l ; The above y is the probability of influence of label words between customers, and the above weight value W l With deviation value b l It is pre-imported into the influence probability algorithm, and the total representation v is imported into the influence probability algorithm to obtain its corresponding influence probability y.

9. The method for constructing a global customer profile based on big data according to claim 1, characterized in that: The step of constructing a customer portrait by using the obtained first portrait tag information and second portrait tag information includes: Performing a threshold comparison and judgment on the total characterization label value and the sub-characterization label value corresponding to the first portrait label information and the second portrait label information; After the judgment, the remaining portrait label information is output to obtain the customer portrait.

10. A device for constructing a global customer profile based on big data, characterized in that: include: An acquisition unit, used to acquire the user's login information and identify the first text data input by the user; A network unit, configured to retrieve first customer information matching the first text data from a big data network by using a data crawler, wherein the first customer information includes multiple items; A model calculation unit, used to input the first customer information into a preset representation model, and obtain the weight of each keyword in the first customer information after the representation model is calculated by an attention representation algorithm; A first label unit, configured to generate first portrait label information corresponding to the first customer information based on each of the keyword weights; A second label unit is used to establish a communication interface with the user, continuously obtain second text data and a corresponding confirmation signal input by the user through the communication interface, and cyclically perform calculation of the representation model and generation of the portrait label through the second text data and the confirmation information to generate second portrait label information; The portrait unit is used to construct a customer portrait through the obtained first portrait label information and second portrait label information.

Images