TWI722774B - Marketing object decision-making method and marketing system using mobile phone number - Google Patents

Abstract

A marketing object decision-making method includes: inputting an initial data into a first predictive model. The initial data includes a target mobile phone number and multiple feature fields. Each feature field is used to fill in a feature of the door number and has not yet been filled in. Each feature field of the entry number feature is regarded as a missing feature field; the first prediction model generates a prediction result according to the digital composition of the target mobile phone number to fill in each missing feature field as a feature of the door number, so that the initial The data becomes a piece of data to be evaluated without missing feature fields; the data to be evaluated is input into a second prediction model; the second prediction model generates and outputs a decision result based on the correlation between the door number features of the data to be evaluated , The decision result indicates whether the target mobile phone number is a target marketing object suitable for marketing.

Description

Marketing object decision-making method and marketing system using mobile phone number

The invention relates to a marketing object decision-making method, in particular to a marketing object decision-making method involving machine learning. The invention also relates to a marketing system capable of implementing the marketing target decision method.

Since the popularization of smart phones, delivering marketing content to mobile phones through marketing systems has become a more direct and instant marketing method than traditional media.

However, as the population with mobile phones continues to grow, the number of mobile phones is increasing. Therefore, if the marketing system only transmits marketing content to all mobile phone numbers in the database indiscriminately, it will be more expensive. More traffic and longer time to deliver marketing content. Such an approach makes the marketing system inefficient and unable to target the target customer group. As a result, the marketing conversion rate is low and the traffic cost is increased.

Therefore, how to improve the shortcomings of the existing marketing system has become a topic worthy of discussion.

One of the objectives of the present invention is to provide a marketing target decision-making method that can overcome the shortcomings of the prior art.

The marketing target decision-making method of the present invention is implemented by a marketing system. The marketing target decision-making method includes: (A) Obtaining multiple training data, each training data includes a reference mobile phone number composed of multiple numbers, and multiple respective and The reference feature corresponding to the reference mobile phone number, and each reference feature is related to a user or a mobile phone device corresponding to the reference mobile phone number; (B) training a first prediction model and a second prediction model, the The first prediction model is trained by machine learning by analyzing the relationship between the number composition of each reference mobile phone number and each reference feature corresponding to the reference mobile phone number; (C) Data is input into the first predictive model. The initial data includes a target mobile phone number composed of multiple numbers, and a plurality of feature fields corresponding to the target mobile phone number. Each feature field corresponds to the target mobile phone number. It is related to a user or a mobile phone device corresponding to the number, and is used to fill in a door number feature corresponding to the target mobile phone number, and each feature field that has not yet been filled with the door number feature in the feature fields The bit is regarded as a missing feature field; (D) For each missing feature field, the first prediction model generates a prediction result corresponding to the missing feature field according to the digital composition of the target mobile phone number, and then The prediction result is filled in the missing feature field as the door number feature corresponding to the target mobile phone number, so that the initial data becomes a piece of data to be evaluated without missing feature fields; (E) the data to be evaluated is entered into the missing feature field. The second prediction model; (F) The second prediction model generates and outputs a decision result based on the correlation between the door number features of the data to be evaluated, and the decision result indicates whether the target mobile phone number is an appropriate Target marketing objects for accepting marketing.

In some implementation aspects of the marketing target decision method of the present invention, in step (A), each training material also contains a pair of reference marketing results that should refer to the mobile phone number, and the reference marketing result indicates the reference mobile phone number Whether it is a target marketing object suitable for marketing; in step (B), the second predictive model performs machine learning by analyzing the correlation between the reference features of each training data and the reference marketing result By training.

In some implementation aspects of the marketing target decision method of the present invention, in step (A), one of the reference features of each training material is used as a target reference feature, and, in step (B), for For the target reference features of the training data, the machine learning method of the first prediction model includes: for a first part of the numbers of each reference mobile phone number, analyzing the reference mobile phone numbers The relationship between the number composition method of the first part and the target reference features to generate the first correlation data corresponding to the first part and the target reference features; these numbers for each reference mobile phone number A second part of the second part of the reference mobile phone number, the analysis of the relationship between the digital composition of the second part of the reference mobile phone number and the target reference characteristics, to generate a correspondence between the second part and the target reference The second relevance data of the feature; at least the relevance between the digital composition of the reference mobile phone numbers and the target reference characteristics is calculated based on the first relevance data and the second relevance data, so as to complete the comparison These targets refer to machine learning of features.

In some implementation aspects of the marketing target decision method of the present invention, in step (D), the method for generating the prediction result of the first prediction model includes: a first part of the numbers of the target mobile phone number Analyze the digital composition to generate the first speculation data corresponding to the first part. The first speculation data includes a first speculation result and a first confidence value indicating the estimated accuracy of the first speculation result ; Analyze the digital composition of a second part of the numbers of the target mobile phone number to generate a second guess data corresponding to the second part, the second guess data includes a second guess result, And a second confidence value indicating the estimated accuracy of the second prediction result; at least according to the magnitude relationship between the first confidence value and the second confidence value, and at least from the first prediction result and the second confidence value 2. The prediction result is determined from the prediction result.

In some implementation aspects of the marketing target decision-making method of the present invention, the marketing target decision-making method further includes a step (F) after step (F): (H) if the decision result indicates that the target mobile phone number is the target marketing target, Send a marketing message to the mobile phone device corresponding to the target mobile phone number for the mobile phone device corresponding to the target mobile phone number to output the marketing message, if the decision result indicates that the target mobile phone number is not the same The target marketing target does not send the marketing message to the mobile phone device corresponding to the target mobile phone number.

In some implementation aspects of the marketing target decision method of the present invention, in step (A), the reference features of each training data include at least a region reference feature, a gender reference feature, and an operating system reference feature. The region The reference feature indicates the residential area of the user corresponding to the reference mobile phone number, the gender reference feature indicates the gender of the user corresponding to the reference mobile phone number, and the operating system indicates the reference feature Corresponding to the operating system type of the mobile phone device corresponding to the reference mobile phone number; in step (C), the feature fields include at least a region feature field, a gender feature field, and an operating system feature field And, the door number feature used to be filled in the area feature field is an area feature indicating the residential area of the user corresponding to the target mobile phone number, which is used to be filled in the gender feature field The door number feature of is a gender feature indicating the gender of the user corresponding to the target mobile phone number, and the door number feature used to be filled in the operating system feature field is a gender feature indicating the target mobile phone door The operating system characteristics of the operating system type of the mobile device corresponding to the number.

In some implementation aspects of the marketing target decision method of the present invention, in step (A), the training data is used as a plurality of first training data, and the reference mobile phone number of each first training data is used as a first training data. A reference mobile phone number; the marketing target decision-making method includes before step (B): (G) obtaining a plurality of second training data, and each second training data includes a second reference mobile phone door composed of multiple numbers Number, multiple reference features each corresponding to the second reference mobile phone number, and a reference marketing result corresponding to the second reference mobile phone number, each reference feature of each second training material and the second reference mobile phone The door number corresponding to a user or a mobile phone device is related, and the reference marketing result indicates whether the second reference mobile phone number is a target marketing target suitable for marketing; in step (B), the second The prediction model is trained by machine learning by analyzing the correlation between the reference features of each second training data and the reference marketing result.

Another object of the present invention is to provide a marketing system capable of implementing the method for decision-making of marketing objects.

The marketing system of the present invention can perform the following steps: (A) Obtain multiple training data, each training data includes a reference mobile phone number composed of multiple numbers, and a plurality of reference features corresponding to the reference mobile phone number, And each reference feature is related to a user or a mobile phone device corresponding to the reference mobile phone number; (B) training a first prediction model and a second prediction model, the first prediction model is analyzed by each The relationship between the digital composition of the reference mobile phone number and each reference feature corresponding to the reference mobile phone number is trained by machine learning; (C) an initial data is input into the first prediction model, and the initial The data includes a target mobile phone number composed of multiple numbers, and a plurality of feature fields corresponding to the target mobile phone number, and each feature field corresponds to a user or a mobile phone corresponding to the target mobile phone number Device-related, and used to fill in a door number feature corresponding to the target mobile phone number, and each feature field in the feature fields that has not yet been filled in with the entry number feature is regarded as a missing feature field;( D) For each missing feature field, the first prediction model generates a prediction result corresponding to the missing feature field according to the digital composition of the target mobile phone number, and uses the prediction result as a corresponding target mobile phone number Fill in the missing feature field with the door number feature of, so that the initial data becomes a piece of data to be evaluated without missing feature fields; (E) input the data to be evaluated into the second prediction model; (F) the first data The second prediction model generates and outputs a decision result based on the correlation between the door number characteristics of the data to be evaluated, the decision result indicating whether the target mobile phone number is a target marketing object suitable for marketing.

In some implementations of the marketing system of the present invention, in step (A), each training material also includes a pair of reference marketing results that should refer to the mobile phone number, and the reference marketing result indicates whether the reference mobile phone number is A target marketing object suitable for marketing; in step (B), the second predictive model is trained by machine learning by analyzing the correlation between the reference features of each training data and the reference marketing result Become.

In some implementation aspects of the marketing system of the present invention, in step (A), one of the reference features of each training data is used as a target reference feature, and, in step (B), for the reference features For the target reference features of the training data, the machine learning method of the first prediction model includes: for a first part of the numbers of each reference mobile phone number, analyzing the first parts of the reference mobile phone number The relationship between the number composition method and the target reference features to generate a pair of first correlation data corresponding to the first part and the target reference features; for each of the numbers in the reference mobile phone number A second part analyzes the correlation between the digital composition of the second parts of the reference mobile phone numbers and the target reference features to generate a correspondence between the second part and the target reference features The second relevance data; at least based on the first relevance data and the second relevance data, calculate the relevance between the digital composition of the reference mobile phone numbers and the target reference characteristics, so as to complete the Machine learning of target reference features.

In some implementation aspects of the marketing system of the present invention, in step (D), the method for the first prediction model to generate the prediction result includes: a first part of the numbers for the target mobile phone number The method is analyzed to generate the first guess data corresponding to the first part. The first guess data includes a first guess result and a first confidence value indicating the estimated accuracy of the first guess result; The digital composition of a second part of the numbers of the target mobile phone number is analyzed to generate second guess data corresponding to the second part. The second guess data includes a second guess result and a A second confidence value indicating the estimated accuracy of the second prediction result; at least according to the magnitude relationship between the first confidence value and the second confidence value, and at least from the first prediction result and the second prediction The prediction result is determined from the result.

In some implementation aspects of the marketing system of the present invention, the marketing system also executes after step (F): (H) if the decision result indicates that the target mobile phone number is the target marketing target, send a marketing message to The mobile phone device corresponding to the target mobile phone number is used for the mobile phone device corresponding to the target mobile phone number to output the marketing message. If the decision result indicates that the target mobile phone number is not the target marketing object, then Do not send the marketing message to the mobile phone device corresponding to the target mobile phone number.

In some implementation aspects of the marketing system of the present invention, in step (A), the reference features of each training data include at least a region reference feature, a gender reference feature, and an operating system reference feature. The region reference feature Indicate the residential area of the user corresponding to the reference mobile phone number, the gender reference feature indicates the gender of the user corresponding to the reference mobile phone number, and the operating system refers to the feature to indicate the corresponding The operating system type of the mobile phone device corresponding to the reference mobile phone number; in step (C), the feature fields include at least a region feature field, a gender feature field, and an operating system feature field, and , The door number feature used to be filled in the area feature field is an area feature indicating the residential area of the user corresponding to the target mobile phone number, used for the gender feature field filled in The door number feature is a gender feature that indicates the gender of the user corresponding to the target mobile phone number. The door number feature used to be filled in the operating system feature field is a gender feature that indicates the target mobile phone door number. The operating system characteristics corresponding to the operating system type of the mobile device.

In some implementation aspects of the marketing system of the present invention, in step (A), the training data is used as a plurality of first training data, and the reference mobile phone number of each first training data is used as a first reference Mobile phone number; the marketing system also executes before step (B): (G) to obtain multiple second training data, each second training data contains a second reference mobile phone number composed of multiple numbers, multiple Each reference feature corresponding to the second reference mobile phone number, and the reference marketing result corresponding to the second reference mobile phone number, each reference feature of each second training material is associated with the second reference mobile phone number. Corresponding to a user or a mobile phone device, and the reference marketing result indicates whether the second reference mobile phone number is a target marketing target suitable for marketing; in step (B), the second prediction model is It is trained by machine learning by analyzing the correlation between the reference features of each second training data and the reference marketing result.

The effect of the present invention is that the marketing system can predict whether the target mobile phone number is a target marketing target suitable for marketing according to the digital composition of the target mobile phone number. As a result, compared with the prior art for all mobile phones With the method of transmitting marketing content by number, the marketing system of this embodiment can effectively target target marketing targets for more efficient and accurate marketing, thereby reducing the traffic and time cost required for marketing, and increasing the conversion rate of marketing. It is worth mentioning that even if the initial data’s house number features are not complete and there are deficiencies, the first prediction model can also use predictive analysis to fill in the missing parts of the initial data, even if the initial data All feature fields in are all missing feature fields. The first prediction model can still predict the door number feature corresponding to each missing feature field, so as to ensure that the second prediction model can be based on the complete data to be evaluated. "Is it a target marketing target". That is to say, in extreme applications, even if the initial data only contains the target number and does not contain any characteristics of the number, the marketing system can still predict the corresponding All the door number characteristics of the target door number, and then based on the predicted door number characteristics to produce the decision result corresponding to the initial data.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same numbers. In addition, the "electrical connection" mentioned in this patent specification generally refers to a wired electrical connection between multiple electronic devices/devices/components connected through conductive materials, and a radio connection through wireless communication technology for wireless signal transmission. . Moreover, the “electrical connection” mentioned in this patent specification also refers to the “direct electrical connection” formed by the direct connection between two electronic equipment/devices/components, and the “direct electrical connection” between the two electronic equipment/devices/components. "Indirect electrical connection" formed by connecting other electronic equipment/devices/components.

Referring to FIG. 1, a first embodiment of the marketing system 1 of the present invention is implemented as a server, for example, and includes a storage unit 11 and a processing unit 12 electrically connected to the storage unit 11. In this embodiment, the storage unit 11 can be implemented as one or more hard disks. On the other hand, the processing unit 12 can be implemented as a control circuit motherboard and a control circuit motherboard. Processor, but not limited to this.

The storage unit 11 stores a first prediction model M1 and a second prediction model M2 that have not been trained. In this embodiment, the first prediction model M1 and the second prediction model M2 can be implemented as two neural networks (Artificial Neural Network, ANN for short), for example, respectively. However, in other embodiments, the The first prediction model M1 and the second prediction model M2 can also be achieved by using other types of machine learning (Machine Learning) technologies, such as Support Vector Machine (SVM) or Random Forest (Random Forest), and It is not limited to this embodiment.

1 and 2 at the same time, the following exemplarily describes in detail how the marketing system 1 of this embodiment implements the method for decision-making of marketing objects.

First, in step S1, the processing unit 12 obtains a plurality of training data, and stores the training data in the storage unit 11.

In this embodiment, the training data are, for example, multiple historical marketing records for specific types of commodities (such as cosmetics), but it is not limited to this. Moreover, in this embodiment, each training data includes, for example, a reference mobile phone number composed of multiple numbers, a plurality of reference features corresponding to the reference mobile phone number, and a pair of references that should refer to the mobile phone number. Marketing results. Among them, each reference feature is related to a user or a mobile device corresponding to the reference mobile phone number, and the reference marketing result is, for example, "Yes" or "No" to indicate whether the reference mobile phone number is a suitable one. It is the target marketing target who accepts marketing, but it is not limited to this.

More specifically, in this embodiment, each reference mobile phone number is, for example, composed of 10 numbers (for example, "0912345678").

On the other hand, in this embodiment, each training data includes three reference features, and the three reference features are, for example, a region reference feature, a gender reference feature, and an operating system reference feature, but not Not limited to this.

In more detail, the area reference feature indicates, for example, the residential area of the user corresponding to the reference mobile phone number, such as "Northern", "Central", "Southern" or "Eastern". On the other hand, the gender reference feature indicates, for example, the gender of the user corresponding to the reference mobile phone number, such as "male" or "female". In another aspect, the operating system reference feature, for example, indicates the type of operating system installed on the mobile device corresponding to the reference mobile phone number, such as "Android" or "iOS".

For ease of understanding, Table 1 below shows an exemplary training material. [Table 1] Reference mobile phone number Regional reference characteristics Gender reference characteristics Operating system reference features Reference marketing results 0912345678 North male Android Yes

It should be noted that the region reference feature, the gender reference feature, and the operating system reference feature in this embodiment are for ease of understanding, so they are presented exemplarily in an intuitive and vernacular way. It should be understood that in actual In the implementation aspect, the reference features can be implemented as a predefined number or computer code, and are not limited to the aforementioned example aspect.

It is added that in this step S1, the processing unit 12 may first establish an electrical connection with a data storage device not shown in the figure (for example, a hard disk, a flash drive or another computer device). , And obtain the training data from the data storage device, or, the processing unit 12 may also obtain the training data from the cloud, for example, by downloading from the Internet. Since the method for obtaining the training data is not the focus of the present invention, the method for obtaining the training data by the processing unit 12 is not particularly limited in this embodiment.

After the processing unit 12 obtains the training data, the flow proceeds to step S2.

In step S2, the processing unit 12 trains the first prediction model M1 and the second prediction model M2 stored in the storage unit 11 according to the training data. Moreover, in this embodiment, the training method of the first prediction model M1 and the second prediction model M2 is, for example, deep learning based on the training data, but it is not limited to this.

The following describes the training method of the first prediction model M1.

In this embodiment, the first predictive model M1 performs machine learning by analyzing the correlation between the numeric composition of each reference mobile phone number and each reference feature corresponding to the reference mobile phone number (in this In the embodiment, it is trained by deep learning).

In order to further explain the training method of the first prediction model M1, the following Table 2 shows the training data used for several examples. It is supplemented that some numbers in Table 2 are represented by "*" symbols. . [Table 2] Reference mobile phone number Regional reference characteristics Gender reference characteristics Operating system reference features Reference marketing results 0912345678 North male Android no 0910****** Central Female iOS Yes 0956****** North male Android no 0917****** South male Android no … … … … … 0922****** east male iOS no 0935****** South Female iOS Yes 0912****** east Female Android no

With reference to the training data exemplified in Table 2 above, and taking the regional reference features of the training data as an example, the first prediction model M1 of this embodiment can be used in the process of machine learning. Inferential statistics or descriptive statistics are used to perform statistical analysis to summarize the relationship between the digital composition of the reference mobile phone numbers and the reference characteristics of the regions. In this way, the first prediction model M1 can learn which specific digital composition mode of the reference mobile phone number has a greater probability of corresponding to which type of regional reference feature, that is, learn the corresponding regional reference features of each type of regional reference feature. The reference mobile phone number has a greater probability of common features in the number composition. For example: the reference mobile phone number with the first five digits of "09111" corresponds to the regional reference feature with a greater probability (for example, 70%) of "North ", but not limited to this.

With reference to FIG. 3, in more detail, the first prediction model M1, for example, will first use the region reference feature of each training data as a target reference feature among all the reference features of the training data, and for the training data For the target reference features of the data, the machine learning method of the first prediction model M1 includes the following sub-steps.

First, proceed to sub-step S21.

In sub-step S21, for a first part of the numbers of each reference mobile phone number, the first prediction model M1 analyzes the digital composition method of the first parts of the reference mobile phone number and the target reference The correlation between the features (in this case, the reference features of the regions) to generate a pair of first correlation data that should equal the first part and the target reference features.

More specifically, in this embodiment, the first part of each reference mobile phone number is, for example, the first five digits of the reference mobile phone number. For example, for the reference phone number of "0912345678", the first part refers to "09123". In other words, in the sub-step S21 of this embodiment, the first prediction model M1 will perform a statistical analysis on the relationship between the number composition of the first five codes of the reference mobile phone numbers and the reference features of the regions. , And the first relevance data is the statistical analysis result describing the relevance between the five codes before the reference mobile phone numbers and the reference features of the regions.

Next, proceed to sub-step S22.

In sub-step S22, for a second part of the numbers of each reference mobile phone number, the first prediction model M1 analyzes the digital composition of the second parts of the reference mobile phone numbers and the The correlation between the target reference features (in this example, the regional reference features), to generate a second correlation data corresponding to the second part and the target reference features.

More specifically, in this embodiment, the second part of each reference mobile phone number is, for example, the sixth to eight digits of the reference mobile phone number. For example, for the reference phone number of "0912345678", the second part refers to "456". In other words, in the sub-step S22 of this embodiment, the first predictive model M1 will be based on the correlation between the digital composition of the sixth to eighth codes of the reference mobile phone numbers and the reference features of the regions. Statistical analysis, and the second correlation data is the statistical analysis result describing the correlation between the sixth to eighth codes of the reference mobile phone numbers and the reference features of the regions.

Next, proceed to sub-step S23.

In sub-step S23, for a third part of the numbers of each reference mobile phone number, the first prediction model M1 analyzes the digital composition of the third parts of the reference mobile phone numbers and the The correlation between the target reference features (in this example, these regional reference features) is used to generate third correlation data corresponding to the third part and the target reference features.

More specifically, in this embodiment, the third part of each reference mobile phone number is, for example, the last two digits of the reference mobile phone number. For example, for the reference phone number of "0912345678", the third part refers to "78". In other words, in the sub-step S23 of this embodiment, the first prediction model M1 will perform a statistical analysis on the relationship between the digital composition of the last two codes of the reference mobile phone numbers and the reference features of the regions. , And the third correlation data is the statistical analysis result describing the correlation between the last two codes of the reference mobile phone numbers and the reference features of the regions.

Finally, proceed to sub-step S24.

In sub-step S24, the first predictive model M1 merges the first relevance data, the second relevance data, and the third relevance data with each other to generate a comprehensive analysis result to complete the reference to the targets Machine learning of features (in this case, reference features of these regions). Moreover, the comprehensive analysis result indicates the correlation between the digital composition mode of the reference mobile phone numbers and the target reference features.

To sum up, in the sub-steps S21 to S24 of this embodiment, the first prediction model M1 is based on the reference mobile phone numbers and the target reference features (in the foregoing description, reference is made to the regions For example, when performing machine learning, each reference mobile phone number is divided into multiple parts (in this embodiment, the first part, the second part, and the third part). Part), and perform multiple statistical analyses on each part of the reference mobile phone number in sequence and the target reference features. Finally, the results of multiple statistical analyses (in this embodiment are the aforementioned first A relevance data, the second relevance data and the third relevance data) are combined with each other to complete the machine learning for the target reference features. That is to say, the first prediction model M1 of this embodiment is trained by deep learning in a “layered training” manner based on the reference mobile phone numbers and the target reference features.

It is worth mentioning that if k-fold cross-validation (k-fold cross-validation) is used to test the learning results of the first predictive model M1, the reference mobile phone number is divided into more than "no". In this embodiment, the method of machine learning with “layered training” can make the first prediction model M1 have better predictions by directly using the “wholes” of each reference mobile phone number for correlation analysis. Accuracy. Further, by using k-fold cross-validation for testing, compared to splitting each reference mobile phone number in other ways, this embodiment splits each reference mobile phone number into the first part (the first five codes), The second part (the sixth to the eighth code) and the third part (the last two codes) are respectively subjected to correlation analysis, so that the first prediction model M1 has better prediction accuracy after the training is completed. However, it should be understood that even if the reference mobile phone number is split in other ways (for example, split into the first four codes, the fifth to seventh codes, and the last three codes), it is still an implementable aspect of the present invention.

The above is the description of the machine learning of the first prediction model M1 using these regional reference features as the target reference features.

In addition, after the machine learning for the reference features of the regions is completed, the first prediction model M1 will, for example, change the gender reference feature of each training data as the new target reference feature, and based on the reference mobile phone numbers and The gender reference features are subjected to machine learning for the gender reference features in the same hierarchical training method as in the aforementioned sub-step S21 to sub-step S24.

Then, after the machine learning for the gender reference features is completed, the first prediction model M1 will, for example, change the operating system reference feature of each training data as a new target reference feature, and then use the reference mobile phone number And these operating system reference features, the machine learning for the operating system reference features is performed in the same hierarchical training method as in the aforementioned sub-step S21 to sub-step S24.

In this way, the first prediction model M1 completes the machine learning of the correlations between the reference mobile phone numbers and the regional reference features, the gender reference features, and the operating system reference features.

Next, the training method of the second prediction model M2 will be described below.

In this embodiment, the second prediction model M2 performs machine learning (in this embodiment, deep learning) by analyzing the correlation between the reference features of each training data and the reference marketing result. Trained.

With reference to the training data exemplified in Table 2 above, the second predictive model M2 of this embodiment may, for example, also use inferential statistics or descriptive statistics to perform statistical analysis in the process of machine learning, so as to summarize Find out the correlation between the reference features of the training materials (that is, the geographic reference features, the gender reference features, and the operating system reference features) and the reference marketing results. In this way, the second predictive model M2 can learn which reference marketing results corresponding to the specific composition of reference features have a greater probability of being "Yes", that is, learning the training data for which the reference marketing result is "Yes" There is a common feature with greater probability in the composition of the reference features. For example, in the same training data, if the region reference feature, the gender reference feature, and the operating system reference feature are "Northern" and "Female" respectively And "iOS", the reference marketing result of the training data has a higher probability (for example, 70%) to be "Yes", but it is not limited to this.

After completing the training of the first prediction model M1 and the second prediction model M2, the process proceeds to step S3.

Referring to FIGS. 1, 2 and 4 at the same time, in step S3, the processing unit 12 inputs a piece of initial data D (shown in FIG. 4) into the first prediction model M1.

In this embodiment, as shown in FIG. 4, the initial data D includes a target mobile phone number consisting of multiple numbers ("0987654321" as shown in FIG. 4), and a plurality of mobile phone numbers that are connected to the target mobile phone. The feature field corresponding to the target mobile phone number, each feature field is related to a user or a mobile phone device corresponding to the target mobile phone number, and is used to fill in a house number feature corresponding to the target mobile phone number, and, In these feature fields, each feature field that has not been filled in with the entry number feature is regarded as a missing feature field. In other words, if any feature field in the initial data D is a missing feature field, it means that the initial data D itself is not complete.

Specifically, in this embodiment, like the reference mobile phone numbers, the target mobile phone number is, for example, composed of 10 numbers, and the initial data D includes, for example, three characteristic fields, and the The three feature fields are, for example, a region feature field C1, a gender feature field C2, and an operating system feature field C3 shown in FIG. 4, respectively.

In more detail, the area feature field C1 is, for example, used to fill in the door number feature that is used as a area feature f1, and the area feature f1 is, for example, used to indicate the user corresponding to the target mobile phone number. Where you live (e.g. "North", "Central", "South" or "East"). On the other hand, the gender feature field C2 is, for example, used to fill in the door number feature that is used as a gender feature f2, and the gender feature f2 is, for example, used to indicate the user corresponding to the target mobile phone number. Gender (e.g. "male" or "female"). On the other hand, the operating system feature field C3 is, for example, used to fill in a feature of the operating system feature f3, and the operating system feature f3, for example, is used to indicate the target mobile phone number corresponding to the feature The type of operating system installed on the mobile device (such as "Android" or "iOS"), but it is not limited to this.

And, taking the initial data D shown in FIG. 4 as an example, the area feature field C1 has not yet been filled in the area feature f1 and is in a blank state. Therefore, the area feature field C1 will be used as in this embodiment A missing feature field of (here as a first missing feature field Cm1 shown in Figure 4). On the other hand, the gender feature field C2 has been filled in with the gender feature f2 that appears as "female". Therefore, the gender feature field C2 will not be regarded as a missing feature field. On the other hand, the operating system feature field C3 has not yet been filled in the operating system feature f3. Therefore, the operating system feature field C3 will be used as another missing feature field in this embodiment (here as shown in Figure 4). A second missing feature field is shown Cm2).

After inputting the initial data D into the first prediction model M1, the flow proceeds to step S4.

In step S4, for each missing feature field of the initial data D, the first predictive model M1 performs predictive analytics based on the digital composition of the target mobile phone number, and generates a corresponding missing feature. Then, the first prediction model M1 fills the missing feature field with the prediction result as the door number feature corresponding to the target mobile phone number, so that the initial data D becomes a piece without any Data to be evaluated D'in the missing feature field. In other words, the initial data D is used as the input data of the first prediction model M1, and the data to be evaluated D'is used as the output data of the second prediction model M2. In this way, although the gate number feature of the initial data D is missing, the first prediction model M1 can use its machine learning results to fill in the missing parts of the initial data D by means of predictive analysis.

Taking the initial data D shown in FIG. 4 as an example, for the first missing feature field Cm1, the first prediction model M1 will generate a corresponding first missing according to the digital composition of the target mobile phone number The prediction result of the feature field Cm1 (here as a regional prediction result, such as "Northern"), and then the first prediction model M1 fills the prediction result of the region into the first missing feature field Cm1, so that the region The prediction result is added to the initial data D, and is taken as the area feature f1 corresponding to the target mobile phone number.

In conjunction with FIG. 5, in more detail, the method for generating the prediction result of the region by the first prediction model M1 includes the following sub-steps.

First, proceed to sub-step S41.

In sub-step S41, for a first part of the numbers of the target mobile phone number, the first prediction model M1 performs a predictive analysis on the number composition of the first part to generate a first part corresponding to the first part. Speculate information. The first prediction data includes a first prediction result and a first confidence value indicating the prediction accuracy of the first prediction result. For example, the first guess result may be "Northern", and the first confidence value may be "72%", for example.

In more detail, in this embodiment, the first part of the target mobile phone number is, for example, the first five digits of the target mobile phone number. For example, for the reference phone number of "0987654321", the first part refers to "09876". In other words, in the sub-step S41 of this embodiment, the first predictive model M1 performs predictive analysis on the digital composition of the first five digits of the target mobile phone number, thereby inferring the target mobile phone number corresponding to the What are the characteristics of the region?

After the first estimation data is generated, the sub-step S42 is then performed.

In sub-step S42, for a second part of the numbers of the target mobile phone number, the first predictive model M1 performs a predictive analysis on the digital composition of the second part to generate a corresponding second part The second speculation data. The second prediction data includes a second prediction result and a second confidence value indicating the prediction accuracy of the second prediction result. For example, the second guess result may be "Southern", and the second confidence value may be "45%", for example.

In more detail, in this embodiment, the second part of the target mobile phone number is, for example, the sixth to eight digits of the target mobile phone number. For example, for the reference mobile phone number of "0987654321", the second part refers to "543". In other words, in the sub-step S42 of this embodiment, the first predictive model M1 performs predictive analysis on the digital composition of the sixth to eighth digits of the target mobile phone number, thereby inferring the location of the target mobile phone number. What are the corresponding regional characteristics?

After the second estimation data is generated, the sub-step S43 is then performed.

In sub-step S43, for a third part of the numbers of the target mobile phone number, the first prediction model M1 performs a predictive analysis on the number composition of the third part to generate a corresponding third part The third speculation data. The third prediction data includes a third prediction result, and a third confidence value indicating the prediction accuracy of the third prediction result. For example, the third guess result may be “East”, and the third confidence value may be “34%”, for example.

In more detail, in this embodiment, the third part of the target mobile phone number is, for example, the last two digits of the target mobile phone number. For example, for the reference phone number of "0987654321", the third part refers to "21". In other words, in the sub-step S43 of this embodiment, the first predictive model M1 performs predictive analysis on the digital composition of the last two codes of the target mobile phone number, thereby inferring the target mobile phone number corresponding to the What are the characteristics of the region?

After the third estimation data is generated, the sub-step S44 is then performed.

In sub-step S44, the first prediction model M1 uses the first prediction result, the second prediction result, and the relationship between the first confidence value, the second confidence value, and the third confidence value. The prediction result of the region is determined from the third prediction result.

To sum up, in the sub-steps S41 to S44 of this embodiment, the first prediction model M1 firstly focuses on the first part of the target mobile phone number when performing predictive analysis on the target mobile phone number. , The second part and the third part perform three predictive analyses to generate the first inferred data, the second inferred data, and the third inferred data, respectively. Then, the first predictive model M1 is based on the first to The size relationship between the third confidence value, and the prediction result of the region is determined from the first to third prediction results. In this example, the "72%" of the first confidence value is the highest among the first to third confidence values. Therefore, the first prediction model M1 in this example will, for example, predict the first The result (that is, the "Northern" shown in Figure 4) is used as the regional prediction result corresponding to the target mobile phone number.

It is worth noting that, in response to the hierarchical training method described in the aforementioned sub-steps S21 to S24, for the target mobile phone number, the first prediction model M1 is also for the first to the first to the first to the target mobile phone number. The three parts analyze and predict separately in the way of "stratified prediction", and determine the prediction result according to the magnitude relationship of the first to third confidence values. Through the test of k-fold cross-validation, the aforementioned "stratified prediction" The method can enable the first prediction model M1 to have better prediction accuracy.

Next, for the second missing feature field Cm2 shown in FIG. 4, the first prediction model M1 will generate another prediction corresponding to the second missing feature field Cm2 according to the digital composition of the target mobile phone number Result (here as an operating system prediction result, such as "iOS"), and then the first prediction model M1 fills the operating system prediction result into the second missing feature field Cm2, so that the operating system prediction result is It is added to the initial data D, and is taken as the operating system feature f3 corresponding to the target mobile phone number. It is supplemented that the first predictive model M1 also performs predictive analysis on the target mobile phone number in a hierarchical predictive manner to generate the operating system predictive result, which is similar to that generated in the aforementioned sub-step S41 to sub-step S44. The method of regional forecast results will not be repeated here.

After the first prediction model M1 fills in the corresponding prediction result in each missing feature field, so that the initial data D becomes the data to be evaluated D'without any missing feature field, the process proceeds to step S5.

In step S5, the processing unit 12 inputs the to-be-evaluated data D'into the second prediction model M2. Then, the flow proceeds to step S6.

In step S6, the second prediction model M2 is based on the relationship between the house number features of the data to be evaluated D'(that is, the area feature f1, the gender feature f2, and the operating system feature f3 shown in FIG. 4). Predictive analysis is performed on the relevance, and a decision result is generated and output. Moreover, the decision result indicates whether the target mobile phone number is a target marketing object suitable for receiving marketing.

After the second prediction model M2 generates the decision result, the process proceeds to step S7.

In step S7, the processing unit 12 determines whether to send a marketing message to the mobile phone device corresponding to the target mobile phone number according to the decision result. In this embodiment, the marketing message can be implemented as a marketing message, for example, But it is not limited to this.

More specifically, if the decision result indicates that the target mobile phone number is the target marketing target, the processing unit 12 will send the marketing message to the mobile phone device corresponding to the target mobile phone number for the target The mobile phone device corresponding to the mobile phone number outputs the marketing message for reference by the user corresponding to the target mobile phone number. On the other hand, if the decision result indicates that the target mobile phone number is not the target marketing target, the processing unit 12 does not send the marketing message to the mobile phone device corresponding to the target mobile phone number, that is, it is not the target mobile phone number. The user corresponding to the target mobile phone number performs marketing.

The above-mentioned steps S1 to S7 are the marketing target decision-making method implemented by the marketing system 1 of this embodiment.

It is supplemented that, for ease of description, this embodiment only uses a single piece of initial data D in step S3 for description. However, it should be understood that in an actual implementation aspect, the processing unit 12 can also perform step S3. Is to input multiple pieces of initial data D into the first prediction model M1 simultaneously in batches, so that the first prediction model M1 can process the initial data D in batches in step S4 to output multiple pieces of initial data D. The pens respectively correspond to the data to be evaluated D′ of the initial data D, and the second prediction model M2 can also generate multiple decision results corresponding to the data to be evaluated D′ in step S6 in batches.

Moreover, the implementation of the marketing system 1 as a single server in this embodiment is only an exemplary implementation. It should be understood that, in an actual implementation, the marketing system 1 can also be implemented as multiple servers. For example, the marketing system 1 may also include a first server for training and running the first prediction model M1 to output the data D'to be evaluated, and a first server electrically connected to the first server and used for training and A second server that runs the second predictive model M2 to output the decision result, and a third server that is electrically connected to the second server and used to receive the decision result to determine whether to send the marketing message, and not It is limited to this embodiment.

In addition, in another implementation aspect similar to this embodiment, each training material may further include multiple other reference features in addition to the region reference feature, the gender reference feature, and the operating system reference feature, and The other reference features may include, for example, an age reference feature, a product of interest reference feature, a mobile phone type reference feature, a short message URL click time reference feature, and a web browsing behavior reference feature.

More specifically, the age reference feature indicates, for example, the age range of the user corresponding to the reference mobile phone number, such as "under 18 years old", "18-25 years old", "25-35 years old" and so on. The reference feature of the product of interest, for example, indicates the type or model of the product that the user corresponding to the reference mobile phone number has browsed or purchased. The mobile phone type reference feature, for example, indicates the brand and/or model of the mobile phone device corresponding to the corresponding reference mobile phone number. The short message URL click time reference feature indicates, for example, that the mobile device corresponding to the reference mobile phone number receives a short message containing a specific URL, and the time range during which the specific URL is clicked, more specifically, the The click time reference feature of the SMS URL can indicate one of "morning", "afternoon" and "evening", or one of "Monday" to "Sunday", for example. The web browsing behavior reference feature indicates, for example, whether the mobile phone device corresponding to the reference mobile phone number is connected to a specific website through the specific URL, and the duration of time that the mobile phone device is connected to the specific website.

Therefore, in another embodiment, the processing unit 12 further trains the first prediction model M1 and the second prediction model M2 according to the other reference features mentioned above in step S2 of the marketing target decision-making method. In addition, in step S3 of this embodiment, the initial data D, for example, further includes a plurality of other characteristic fields, and the other characteristic fields include an age characteristic field for filling in an age characteristic. Position, one is used to fill in the feature field of the product of interest, one is used to fill in the feature field of the type of mobile phone, and one is used to fill in the text URL of the short message URL click time feature Click the time feature field, and a web browsing behavior feature field for filling in a web browsing behavior feature. In this way, the first prediction model M1 can analyze and predict the other feature fields according to the target mobile phone number in step S4, and the second prediction model M2 can also perform further analysis and prediction in step S6. The decision result is generated based on the house number characteristics filled in the other characteristic fields.

Next, the following describes the difference between a second embodiment of the marketing system 1 of the present invention and the first embodiment.

The difference between the second embodiment and the first embodiment is mainly that: in the second embodiment, the first prediction model M1 and the second prediction model M2 are trained with two different training data.

More specifically, in step S1 of the second embodiment, the training data obtained by the processing unit 12 are respectively used as a plurality of first training data, and the reference mobile phone number of each first training data is used as A first reference mobile phone number. Moreover, in the second embodiment, each first training material does not need to include the reference marketing result.

In addition, in step S1 of the second embodiment, the processing unit 12 also obtains multiple pieces of second training data, but it does not necessarily have to be obtained at the same time as the first training data. Each second training material includes a second reference mobile phone number consisting of multiple numbers, a plurality of reference features corresponding to the second reference mobile phone number, and a reference marketing corresponding to the second reference mobile phone number result. Wherein, each reference feature of each second training data is related to a user or a mobile device corresponding to the second reference mobile phone number, and the reference marketing result is indicated as "Yes" or "No", for example Whether the second reference mobile phone number is a target marketing object suitable for receiving marketing.

Further, in step S2 of the second embodiment, the processing unit 12 trains the first prediction model M1 with the first training data, and trains the second prediction model M2 with the second training data. In addition, the machine learning method of the first prediction model M1 and the second prediction model M2 is the same as that of the first embodiment, so it will not be repeated here.

In summary, by implementing the marketing target decision-making method, the marketing system 1 can predict whether the target mobile phone number is a target marketing target suitable for marketing according to the digital composition of the target mobile phone number, and predict whether the target mobile phone number is suitable for marketing. The marketing message is sent to the mobile phone device corresponding to the target mobile phone number only when the result is yes. As a result, compared to the prior art method of transmitting marketing content to all mobile phone numbers, the marketing system of this embodiment 1 It can effectively lock target marketing objects for more efficient and accurate marketing, thereby reducing the traffic and time costs required for marketing, and increasing the conversion rate of marketing. It is worth mentioning that even if the initial data D’s house number features are not complete and are missing, the first prediction model M1 can also use predictive analysis to fill in the missing parts of the initial data D, even if the initial data D is missing. All the feature fields in the initial data D are all missing feature fields. The first prediction model M1 can still predict the house number feature corresponding to each missing feature field, so as to ensure that the second prediction model M2 can be based on the complete The data to be evaluated D'is subjected to the predictive analysis of "whether it is the target marketing target", that is to say, in extreme applications, even if the initial data D only contains the target door number and does not contain any door number characteristics, the The marketing system 1 can still predict all the house number characteristics corresponding to the target house number first, and then generate the decision result corresponding to the initial data D based on the predicted house number characteristics, so it can indeed achieve the purpose of the invention.

However, the above are only examples of the present invention. When the scope of implementation of the present invention cannot be limited by this, all simple equivalent changes and modifications made in accordance with the scope of the patent application of the present invention and the content of the patent specification still belong to This invention patent covers the scope.

1: Marketing system 11: storage unit 12: Processing unit M1: The first predictive model M2: second predictive model D: Initial information D’: Materials to be evaluated C1: Area feature field C2: Gender characteristics field C3: Operating system feature field Cm1: The first missing feature field Cm2: The second missing feature field f1: Regional characteristics f2: gender characteristics f3: Operating system characteristics S1~S7: steps S21~S24: Sub-step S41~S44: Sub-step

Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, in which: Figure 1 is a block diagram of a first embodiment of the marketing system of the present invention; Figure 2 is a flowchart illustrating how the first embodiment implements the decision-making method for marketing objects; Fig. 3 is another flowchart illustrating how a first prediction model of the first embodiment performs machine learning; Fig. 4 is a schematic diagram illustrating how the first prediction model makes an initial data with missing feature fields a complete data to be evaluated; and Fig. 5 is another flowchart illustrating how the first embodiment generates a prediction result according to the digital composition of a target mobile phone number.

S1~S7: steps

Claims (8)

A marketing target decision-making method implemented by a marketing system. The marketing target decision-making method includes: (A) Obtaining multiple training data, each training data contains a reference mobile phone number composed of multiple numbers, and multiple respective and The reference feature corresponding to the reference mobile phone number, and each reference feature is related to a user or a mobile phone device corresponding to the reference mobile phone number; (B) training a first prediction model and a second prediction model, the The first prediction model is trained by machine learning by analyzing the relationship between the number composition of each reference mobile phone number and each reference feature corresponding to the reference mobile phone number; (C) Data is input into the first predictive model. The initial data includes a target mobile phone number composed of multiple numbers, and a plurality of feature fields corresponding to the target mobile phone number. Each feature field corresponds to the target mobile phone number. It is related to a user or a mobile phone device corresponding to the number, and is used to fill in a door number feature corresponding to the target mobile phone number, and each feature field that has not yet been filled with the door number feature in the feature fields The bit is regarded as a missing feature field; (D) For each missing feature field, the first prediction model generates a prediction result corresponding to the missing feature field according to the digital composition of the target mobile phone number, and then The prediction result is filled in the missing feature field as the door number feature corresponding to the target mobile phone number, so that the initial data becomes a piece of data to be evaluated without missing feature fields; (E) the data to be evaluated is entered into the missing feature field. The second prediction model; and (F) The second prediction model generates and outputs a decision result based on the correlation between the door number characteristics of the data to be evaluated, the decision result indicating whether the target mobile phone number is a target suitable for marketing Marketing target. The marketing target decision-making method according to claim 1, wherein: in step (A), each training material also includes a pair of reference marketing results that should refer to the mobile phone number, and the reference marketing result indicates the reference mobile phone number Whether it is a target marketing object suitable for marketing; and in step (B), the second predictive model is performed by analyzing the correlation between the reference features of each training data and the reference marketing result Learned and trained. The marketing target decision method according to claim 1, wherein in step (A), one of the reference features of each training material is used as a target reference feature, and, in step (B), for For the target reference features of the training data, the machine learning method of the first prediction model includes: for a first part of the numbers of each reference mobile phone number, analyzing the reference mobile phone numbers The relationship between the number composition method of the first part and the target reference features to generate the first correlation data corresponding to the first part and the target reference features; these numbers for each reference mobile phone number A second part of the second part of the reference mobile phone number, the analysis of the relationship between the digital composition of the second part of the reference mobile phone number and the target reference characteristics, to generate a correspondence between the second part and the target reference The second relevance data of the feature; and Calculate the correlation between the digital composition method of the reference mobile phone numbers and the target reference features at least according to the first correlation data and the second correlation data, so as to complete the machine learning of the target reference features . The marketing target decision method according to claim 1, wherein, in step (D), the method for generating the prediction result by the first prediction model includes: a first part of the numbers of the target mobile phone number Analyze the digital composition to generate the first speculation data corresponding to the first part. The first speculation data includes a first speculation result and a first confidence value indicating the estimated accuracy of the first speculation result ; Analyze the digital composition of a second part of the numbers of the target mobile phone number to generate a second guess data corresponding to the second part, the second guess data includes a second guess result, And a second confidence value indicating the estimated accuracy of the second prediction result; at least according to the magnitude relationship between the first confidence value and the second confidence value, and at least from the first prediction result and the second confidence value 2. The prediction result is determined from the prediction result. The marketing target decision-making method as described in claim 1, further comprising a step (F) after step (F): (H) if the decision result indicates that the target mobile phone number is the target marketing target, sending a marketing message to the target The mobile phone device corresponding to the mobile phone number is used for the mobile phone device corresponding to the target mobile phone number to output the marketing message. If the decision result indicates that the target mobile phone number is not the target marketing target, it will not be sent The marketing message to the mobile phone device corresponding to the target mobile phone number. The marketing target decision-making method according to claim 1, wherein: in step (A), the reference features of each training data include at least a region reference feature, a gender reference feature, and an operating system reference feature, and the region The reference feature indicates the residential area of the user corresponding to the reference mobile phone number, the gender reference feature indicates the gender of the user corresponding to the reference mobile phone number, and the operating system indicates the reference feature Corresponding to the operating system type of the mobile phone device corresponding to the reference mobile phone number; and in step (C), the feature fields include at least a region feature field, a gender feature field, and an operating system feature field In addition, the door number feature used to be filled in the area feature field is an area feature indicating the residential area of the user corresponding to the target mobile phone number, and is used to be filled in the gender feature field The door number feature of the bit is a gender feature indicating the gender of the user corresponding to the target phone number, and the door number feature used to be filled in the operating system feature field is a gender feature indicating the target phone The operating system characteristics of the operating system type of the mobile device corresponding to the door number. The marketing target decision-making method according to claim 1, wherein: in step (A), the training data is used as a plurality of first training data, and the reference mobile phone number of each first training data is used as a first training data. A reference mobile phone number; the marketing target decision-making method includes before step (B): (G) obtaining a plurality of second training data, and each second training data includes a second reference mobile phone door composed of multiple numbers Number, multiple each and the second reference hand The reference feature corresponding to the phone number, and the reference marketing result corresponding to the second reference phone number. Each reference feature of each second training data corresponds to a user or a user or a corresponding to the second reference phone number. The mobile phone device is related, and the reference marketing result indicates whether the second reference mobile phone number is a target marketing object suitable for marketing; and in step (B), the second prediction model is performed by analyzing each 2. The correlation between the reference features of the training data and the reference marketing result is trained by machine learning. A marketing system includes a storage unit and a processing unit electrically connected to the storage unit, and the processing unit can implement the marketing target decision method according to any one of claim items 1 to 7.

Images