TWI783613B - Digital marketing decision-making system and digital marketing decision-making method - Google Patents

Abstract

A digital marketing decision-making system includes a database server and an analysis server. The analysis server uses multiple bag-of-words models to generate multiple reference text type scores for each reference advertisement data, and generates a reference text classification result for each reference advertisement data, and trains and generates a copy matching customer similarity model, and for The candidate advertising copy of each candidate advertising data uses the bag of words model to generate multiple candidate copy type scores, and generates a candidate copy classification result for each candidate advertising data, and uses the copy matching customer similarity model to generate a copy matching customer similarity analysis As a result, multiple accumulated clicks are generated, and one of them is selected from the candidate advertisement data as a target advertisement data according to the classification result of the candidate copy and the accumulated clicks.

Description

Digital marketing decision-making system and digital marketing decision-making method

The present invention relates to an artificial intelligence system, in particular to a digital marketing decision-making system. The invention also relates to a digital marketing decision-making method.

At present, e-commerce, web streaming and other businesses, when generating product purchase suggestions or video viewing suggestions for customers, use the recommendation system to model the customer's purchase history or viewing history, and then use the This system generates next purchase or viewing suggestions for customers' reference.

The technology usually used in general recommendation systems is FM (Factorization Machine)/FunkSVD (Funk Singular Value Decomposition) in the matrix decomposition algorithm.

The biggest disadvantage of the matrix decomposition algorithm is the problem of cold start (Cold Start), that is to say, when there are new videos or new customers, the recommendation system cannot immediately reflect the new videos and new customers, resulting in some distortion .

How to develop a new digital marketing decision-making system that can improve the above-mentioned shortcomings of the prior art is a subject to be further discussed in the present invention.

Therefore, the object of the present invention is to provide a digital marketing decision-making system.

Another object of the present invention is to provide a digital marketing decision-making method.

Therefore, the digital marketing decision-making system of the present invention includes a database server and an analysis server.

The database server stores multiple pieces of reference advertisement data, multiple pieces of reference customer background data related to multiple reference customers, multiple pieces of reference copy click records, multiple pieces of candidate advertisement data, and a piece of information related to a target customer Target customer background information, each reference advertising material is related to one of multiple reference products and includes a reference advertising copy, each reference copy click record is related to one of these reference customers and is related to the reference advertising materials Firstly, each candidate advertisement data is related to one of the plurality of candidate products and includes a candidate advertisement copy. The analysis server is electrically connected to the database server through a communication network.

The analysis server extracts a plurality of keywords from the reference advertisement text of the reference advertisement data.

The analysis server trains a plurality of original bag-of-words models to generate a plurality of bag-of-words models corresponding to a plurality of copywriting categories according to the keywords and a plurality of classification settings respectively corresponding to the keywords, and each classification setting indicates one of those types of copywriting.

For each reference advertisement data, the analysis server uses the bag-of-words models to generate a plurality of reference text type scores, and each reference text type score corresponds to the reference advertisement data and the used bag-of-words model.

For each reference advertisement data, according to the corresponding reference text type scores, the analysis server generates a reference text classification result corresponding to the reference advertisement data and indicating one of the text types.

The analysis server trains an original text matching customer similarity model according to the reference customer background data to generate a copy matching customer similarity model.

The analysis server uses the bag-of-words models for the candidate advertising copy of each candidate advertising data to generate a plurality of candidate copy type scores, each candidate copy type score corresponding to the candidate advertising data and the bag of words used Model.

For each candidate advertisement data, according to the corresponding scores of the candidate text types, the analysis server generates a candidate text classification result corresponding to the candidate advertisement data and indicating one of the text types.

The analysis server uses the copywriting matching customer similarity model based on the target customer background information to generate a copywriting matching customer similarity analysis result, and the copywriting matching customer similarity analysis result is related to the target customer background among the reference customer background information The one with the closest data.

The analysis server generates a plurality of corresponding to The cumulative number of clicks for this copy category.

The analysis server selects one of the candidate advertisement data as a target advertisement data according to the classification results of the candidate texts and the accumulated click times.

In some implementations, the analysis server uses latent semantic analysis technology to extract the keywords from the reference advertisement data.

In some implementations, the database server also stores multiple reference product click records, and multiple reference customer page browsing tracks respectively related to the reference customers, and each reference product click record is related to the reference One of the customers and related to one of the reference products and indicating whether to click or not, each reference customer page browsing track contains multiple dwell times respectively related to multiple pages and multiple Views.

The analysis server trains an original slot prediction model to generate a slot prediction model based on the background information of the reference customers, the click records of the reference products, and the page browsing tracks of the reference customers. The slot prediction model includes multiple Corresponding to the reference slot click rate data of the reference customers, each reference willingness rate data includes a plurality of reference slot click rates respectively corresponding to a plurality of slots.

The analysis server uses the slot prediction model based on the background information of the target customer to generate a plurality of customer similarities respectively related to the reference customers, and each customer similarity indicates that the target customer is related to the reference customer and according to the customer similarities, use the position prediction model to generate a plurality of predicted click willingness rates respectively corresponding to the positions.

The analysis server selects one of the positions as a target position according to the predicted click-through rate, and the target position is used to place the target advertisement data.

In some implementation aspects, the slot prediction model is a deep factorization machine model.

In some implementations, each reference product and each candidate product is related to one of a plurality of product categories, and the database server also stores a plurality of reference customer page browsing tracks respectively related to the reference customers, And multiple reference order records, each reference customer page browsing track includes multiple stay times and multiple page views respectively related to these pages, these pages are related to these product categories, Each reference order record is related to one of the reference customers and to one of the reference products and indicates whether to place an order and the order amount.

The analysis server trains an original product familiarity prediction model to generate a product familiarity prediction model according to the reference customer page browsing tracks and the reference order records.

The analysis server trains an original familiarity prediction customer similarity model to generate a familiarity prediction customer similarity model according to the reference customer background data.

The analysis server uses the familiarity prediction customer similarity model according to the background information of the target customer to generate a familiarity prediction customer similarity analysis result, and the familiarity prediction customer similarity analysis result is related to the background information of the reference customers and the The one with the closest background profile of the target customer.

The analysis server predicts the reference customer's page browsing track and the reference order record or the reference order records related to the reference customer's background information related to the reference customer's background information related to the customer similarity analysis result according to the familiarity, Using the product familiarity prediction model, a plurality of product familiarity prediction data respectively corresponding to the product categories are generated.

The analysis server selects one of the candidate advertisement data as the target advertisement data according to the product familiarity prediction data in addition to the classification results of the candidate texts and the accumulated click times.

In some implementation aspects, the product familiarity prediction model is a logistic regression model.

In some implementations, each reference product and each candidate product is related to one of a plurality of product categories, and the database server also stores a plurality of reference product click records, and a plurality of click records respectively related to the reference customers. One reference customer financial data, multiple reference customer page browsing tracks related to these reference customers, one target customer financial data related to the target customer, and one target customer page browsing track related to the target customer, each The reference product click record is related to one of the reference customers and is related to one of the reference products and indicates whether the click is on or not. Each reference customer page browsing track and the target customer page browsing track include multiple The dwell time of multiple pages and the number of views respectively related to those pages.

The analysis server trains an original click willing rate prediction model to generate a click willing rate prediction model based on the reference product click records, the reference customer financial data and the reference customer page browsing track, and the click willing rate prediction model includes A plurality of reference willingness rate data respectively corresponding to the reference customers, each reference willingness rate data includes a plurality of reference click willingness rates corresponding to the product categories respectively.

According to the target customer's financial information and the target customer's page browsing track, the analysis server uses the click willing rate prediction model to generate a plurality of customer similarities respectively related to the reference customers, and each customer similarity indicates the target The similarity between the customer and the related reference customer, and according to the customer similarity, using the click willing rate prediction model to generate a plurality of predicted click willing rates respectively corresponding to the product categories.

In addition to the classification results of the candidate texts and the accumulated click times, the analysis server also selects one of the candidate advertisement data as the target advertisement data according to the predicted click willingness rate.

In some implementation aspects, the click willing rate prediction model is a deep factorization machine model.

The digital marketing decision-making method of the present invention is implemented by a digital marketing decision-making system. The digital marketing decision-making system includes a database server and an analysis server. Multiple reference customer background materials, multiple reference copy click records, multiple candidate advertisement data, and one target customer background data related to one target customer. Each reference advertisement data is related to multiple reference products. One also includes a reference advertising copy, each reference copy click record is related to one of the reference customers and related to one of the reference advertising materials, and each candidate advertising material is related to one of multiple candidate products and include a candidate advertisement text, the analysis server is electrically connected to the database server through a communication network, and the method includes: the analysis server extracts a plurality of keywords from the reference advertisement text of the reference advertisement data; the The analysis server trains a plurality of original bag-of-words models to generate a plurality of bag-of-words models corresponding to a plurality of copywriting categories according to the keywords and a plurality of classification settings respectively corresponding to the keywords, and each classification setting indicates the one of such copy types; the analysis server uses the bag-of-words model for each reference advertisement data to generate a plurality of reference copy type scores, and each reference copy type score corresponds to the reference advertisement data and the used bag-of-words model; the analysis server generates a reference text classification result corresponding to the reference advertisement data and indicating one of the text types according to the corresponding reference text type scores for each reference advertisement data; the analysis The server trains an original copy matching customer similarity model based on the reference customer background data to generate a copy matching customer similarity model; the analysis server uses the bag of words model for the candidate advertising copy of each candidate advertising data, generating a plurality of candidate copy type scores, each candidate copy type score corresponding to the candidate advertisement data and the bag-of-words model used; for each candidate advertisement data, according to the corresponding candidate copy type scores, Generate a candidate copy classification result corresponding to the candidate advertisement data and indicating one of the copy types; the analysis server uses the copy matching customer similarity model according to the target customer background information to generate a copy matching customer similarity analysis As a result, the copy matching customer similarity analysis result is related to the reference customer background information that is the closest to the target customer background information; the analysis server matches the reference customer background information related to the customer similarity analysis result according to the copywriting The click records of the reference texts related to the reference client and the classification results of the reference texts generate a plurality of cumulative clicks corresponding to the text categories; and the analysis server classifies the candidate texts according to As a result and the accumulated click times, one of the candidate advertisement data is selected as a target advertisement data.

The effect of the present invention lies in: by training and using the bag-of-words models and the copy matching customer similarity model, it is possible to predict the type of copy that the target customer prefers while improving the shortcomings of the cold start of the previous technology. Furthermore, by training Generating and using the position prediction model can predict the position that the target customer is most willing to click on while improving the shortcomings of the previous technology cold start. Furthermore, the product familiarity prediction model and the familiarity degree can be generated and used through training. The predictive customer similarity model can predict the product category that the target customer is more familiar with while improving the shortcomings of the cold start of the previous technology. In addition, through training and using the click willing rate prediction model, it can be predicted that the target customer is more likely to click The desired product category also improves the shortcomings of the previous technology cold start, and then achieves a good personalized marketing effect.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same numerals.

Referring to FIG. 1 , an embodiment of the digital marketing decision-making system 100 of the present invention includes a database server 1 and an analysis server 2 . The analysis server 2 is electrically connected to the database server 1 through a communication network.

The database server 1 stores multiple pieces of reference advertisement data, multiple pieces of reference customer background data related to multiple reference customers, multiple pieces of reference copy click records, multiple pieces of candidate advertisement data, and a piece of information related to a target customer. Target customer background information, multiple reference product click records, multiple reference customer page browsing tracks respectively related to these reference customers, multiple reference order records, multiple reference customer financial materials respectively related to these reference customers, A target customer's financial information related to the target customer, and a target customer's page browsing track related to the target customer.

Each reference advertisement material is related to one of multiple reference products (such as financial products) and includes a reference advertisement copy (such as including subject and content). Each reference advertisement data includes text identification data, reference product identification data and reference advertisement text, for example.

The background information of each reference client includes, for example, ID card number, age, family, job, and rating.

Each reference copy click record is related to one of the reference customers and to one of the reference advertising materials. Each reference copy click record includes, for example, an ID card number, reference product identification information, reference advertisement copy (for example, including subject matter and content), and touch date.

Each candidate advertisement data is related to one of multiple candidate products (such as financial products) and includes a candidate advertisement copy (for example, including subject and content). Each reference product and each candidate product is associated with one of the plurality of product categories.

The background information of each target customer includes, for example, ID card number, age, family, job, and rating.

Each reference product click record is associated with one of the reference customers and is associated with one of the reference products and indicates click or not. Each reference product click record includes, for example, an ID card number, reference product identification information, click or not, and touch date.

Each page browsing track of the reference customer and the page browsing track of the target customer include a plurality of dwell times respectively related to a plurality of pages and a plurality of browsing times respectively related to the pages. These pages are related to these product categories. The browsing track of each reference customer page and the page browsing track of the target customer include, for example, the ID number, the stay time (such as the stay time of page A, the stay time of page B...), the number of visits (such as the number of page A visits, page B number of visits...) and month of visits.

Each reference order record is related to one of the reference customers and to one of the reference products and indicates whether to place an order and the order amount. Each reference order record includes, for example, an ID card number, reference product identification information, order status, order amount, and contact time.

Each reference customer financial data and the target customer financial data include, for example, an asset data, a loan data, multiple consumption records and multiple deposit and withdrawal records. The asset information includes, for example, ID number, total asset amount, investment asset amount, and deposit asset amount. The loan information includes, for example, an ID card number, a card loan amount, a fixed loan amount, and an unsecured loan amount. Each consumption record includes, for example, an ID card number, consumption items (such as transportation, catering, gas stations, department stores, etc.) and consumption date. Each deposit and withdrawal record includes, for example, the ID card number, transaction type (deposit or withdrawal), transaction amount and transaction date.

Referring to FIG. 1 and FIG. 2 (including FIG. 2A , FIG. 2B and FIG. 2C ), the steps of an advertisement recommendation program are described below. First, as shown in step S01, the analysis server 2 extracts a plurality of keywords from the reference advertisement copy of the reference advertisement data. In this embodiment, the analysis server 2 uses Latent Semantic Indexing (LSI) technology to extract the keywords from the reference advertisement data.

Then, as shown in step S02, the analysis server 2 trains the keywords according to the keywords and a plurality of classification settings (such as price type, service type, current affairs type, gift type and limited type) respectively corresponding to the keywords. A plurality of original bag of words (BOW) models generate a plurality of bag of words models respectively corresponding to a plurality of text categories, and each classification setting indicates one of the text categories.

Next, as shown in step S03, the analysis server 2 uses the bag-of-words models to generate a plurality of reference text type scores for each reference advertisement data, and each reference text type score corresponds to the reference advertisement data and the used The bag-of-words model of .

Next, as shown in step S04, for each reference advertisement data, the analysis server 2 generates a reference text corresponding to the reference advertisement data and indicating one of the text types according to the corresponding scores of the reference text types classification results. For example, the analysis server 2 selects the one with the highest score from the corresponding reference text type scores, and generates the reference text classification result according to the text category corresponding to the bag-of-words model corresponding to the reference text type score with the highest score.

Next, as shown in step S05 , the analysis server 2 trains an original copy matching customer similarity model according to the reference customer background data to generate a copy matching customer similarity model. In this embodiment, the copywriting matching customer similarity model is a Cosine Similarity model.

Next, as shown in step S06, the analysis server 2 uses the bag-of-words models for the candidate advertisement copy of each candidate advertisement data to generate a plurality of candidate copy type scores, each candidate copy type score corresponding to the candidate Advertising materials and the bag-of-words model used.

Next, as shown in step S07, for each candidate advertisement data, the analysis server 2 generates a candidate copy corresponding to the candidate advertisement data and indicating one of the copy types according to the corresponding scores of the candidate copy types classification results. For example, the analysis server 2 selects the candidate with the highest score from the corresponding scores of the candidate document types, and generates the candidate document classification result according to the document category corresponding to the bag-of-words model corresponding to the candidate document type with the highest score.

Next, as shown in step S08, the analysis server 2 uses the copywriting matching customer similarity model to generate a copywriting matching customer similarity analysis result based on the background information of the target customer, and the copywriting matching customer similarity analysis result is related to the reference Among the background information of the customer, the one that is closest to the background information of the target customer.

Next, as shown in step S09, the analysis server 2 clicks on the records of the reference documents related to the reference customers related to the reference customer background information related to the customer similarity analysis results related to the customer similarity analysis results, and the reference documents As a result of the classification, a plurality of accumulated click times respectively corresponding to the copywriting categories are generated.

Next, as shown in step S10, the analysis server 2 trains an original product familiarity prediction model to generate a product familiarity prediction model according to the reference customer page browsing tracks and the reference order records. In this embodiment, the product familiarity prediction model is a Logistic Regression model using XGBoost.

Next, as shown in step S11, the analysis server 2 trains an original familiarity prediction customer similarity model to generate a familiarity prediction customer similarity model according to the reference customer background data. In this embodiment, the familiarity prediction customer similarity model is a Cosine Similarity model.

Next, as shown in step S12, the analysis server 2 uses the familiarity prediction customer similarity model to generate a familiarity prediction customer similarity analysis result according to the target customer background information, and the familiarity prediction customer similarity analysis result is related to The one of the reference customer background information that is most similar to the target customer background information.

Next, as shown in step S13, the analysis server 2 predicts the reference customer's page browsing track and the reference order related to the reference customer's background information related to the reference customer's background information related to the customer similarity analysis result according to the familiarity record or the reference order records, use the product familiarity prediction model to generate a plurality of product familiarity prediction data respectively corresponding to the product categories.

Next, as shown in step S14, the analysis server 2 trains an original click willing rate prediction model to generate a click willing rate prediction based on the click records of the reference products, the financial data of the reference customers and the page browsing tracks of the reference customers A model, the click willing rate prediction model includes a plurality of reference willing rate data respectively corresponding to the reference customers, and each reference willing rate data includes a plurality of reference click willing rates respectively corresponding to the product categories. In this embodiment, the click willing rate prediction model is a Deep Factorization Machine (Deep Factorization Machine; DeepFM) model.

Next, as shown in step S15, the analysis server 2 uses the click willingness rate prediction model to generate a plurality of customer similarities respectively related to the reference customers according to the financial data of the target customer and the page browsing track of the target customer, Each customer similarity indicates the similarity between the target customer and the related reference customer, and according to the customer similarity, using the click willing rate prediction model to generate a plurality of predicted clicks respectively corresponding to the product categories willingness rate.

Next, as shown in step S16, the analysis server 2 selects one of the candidate advertisements from the candidate advertisements according to the classification results of the candidate advertisements, the cumulative number of clicks, the product familiarity prediction data and the predicted click willingness rate. One serves as a target advertising material.

Referring to FIG. 1 and FIG. 3 , the steps of a site recommendation procedure are described below. First, as shown in step S21, the analysis server 2 trains an original slot prediction model to generate a slot prediction model according to the background information of the reference customers, the click records of the reference products and the page browsing tracks of the reference customers, The placement prediction model includes a plurality of reference placement click-through rate data respectively corresponding to the reference customers, and each reference willing rate data includes a plurality of reference placement click-through rates respectively corresponding to a plurality of placements. In this embodiment, the slot prediction model is a Deep Factorization Machine (Deep Factorization Machine; DeepFM) model.

Next, as shown in step S22, the analysis server 2 generates a plurality of customer similarities respectively related to the reference customers according to the background information of the target customer using the position prediction model, and each customer similarity indicates the The similarity between the target customer and the related reference customer, and according to the similarity of the customers, using the position prediction model to generate a plurality of predicted click willingness rates respectively corresponding to the positions.

Next, as shown in step S23, the analysis server 2 selects one of the positions as a target position according to the predicted click willingness rates, and the target position is used to place the target advertisement data. In this embodiment, the analysis server 2 selects the position corresponding to the highest predicted click willingness rate as the target position.

To sum up, the digital marketing decision-making system 100 of the present invention generates and uses the bag-of-words model and the copy matching customer similarity model through training, so as to predict the type of copy that the target customer prefers and improve the shortcomings of the prior art cold start, Furthermore, by generating and using the position prediction model through training, it is possible to predict the position that the target customer is most willing to click on while improving the shortcomings of the cold start of the previous technology. Furthermore, through training and using the product familiarity The prediction model and the familiarity prediction customer similarity model can predict the product category that the target customer is more familiar with while improving the shortcomings of the cold start of the previous technology. In addition, through training and using the click willing rate prediction model, it can be predicted The target customers are more likely to click on the product category while improving the disadvantages of the cold start of the previous technology, thereby achieving a good personalized marketing effect, so the purpose of the present invention can indeed be achieved.

But the above-mentioned ones are only embodiments of the present invention, and should not limit the scope of the present invention. All simple equivalent changes and modifications made according to the patent scope of the present invention and the content of the patent specification are still within the scope of the present invention. Within the scope covered by the patent of the present invention.

100: Digital Marketing Decision System 1: Database server 2: Analysis server S01~S16: Steps S21~S23: Steps

Other features and effects of the present invention will be clearly presented in the implementation manner with reference to the drawings, wherein: Fig. 1 is a schematic diagram of a hardware connection relationship of an embodiment of the digital marketing decision-making system of the present invention; Fig. 2 (comprising Fig. 2A, Fig. 2B and Fig. 2C) is a flow chart of this embodiment, illustrating a click willingness rate prediction model, multiple bag-of-words models, a copywriting matching customer similarity model, and a product familiarity prediction The steps of the model and a familiarity prediction customer similarity model; and FIG. 3 is another flow chart of this embodiment, illustrating the steps involved in a slot prediction model.

100: Digital Marketing Decision System

1: Database server

2: Analysis server

Claims (10)

A digital marketing decision-making system, comprising: a database server, storing multiple reference advertisement data, multiple reference customer background data respectively related to multiple reference customers, multiple reference copy click records, multiple candidate advertisement data, And a target customer background information related to a target customer, each reference advertising material is related to one of multiple reference products and includes a reference advertising copy, and each reference copy click record is related to one of these reference customers And related to one of the reference advertising data, each candidate advertising data is related to one of a plurality of candidate products and includes a candidate advertising copy; and an analysis server, electrically connected to the database server through a communication network ; The analysis server extracts a plurality of keywords from the reference advertisement copy of the reference advertisement data; The analysis server trains a plurality of original The bag-of-words model generates a plurality of bag-of-words models respectively corresponding to a plurality of copywriting categories, and each classification setting indicates one of the copywriting categories; the analysis server uses the bag-of-words models for each reference advertisement data to generate A plurality of reference text type scores, each reference text type score corresponds to the reference advertisement data and the bag-of-words model used; for each reference advertisement data, the analysis server generates according to the corresponding reference text type scores a reference copy classification result corresponding to the reference advertisement material and indicating one of the copy types; The analysis server trains an original copy matching customer similarity model according to the reference customer background information to generate a copy matching customer similarity model; the analysis server uses the bag of words for the candidate advertising copy of each candidate advertising data A model for generating a plurality of candidate copy type scores, each candidate copy type score corresponding to the candidate advertisement data and the bag-of-words model used; the analysis server for each candidate advertisement data, according to the corresponding candidate copy type Scores to generate a candidate copy classification result corresponding to the candidate advertisement data and indicating one of the copy types; the analysis server uses the copy matching customer similarity model according to the target customer background information to generate a copy matching customer Similarity analysis results, the copy matching customer similarity analysis results are related to the reference customer background information that is most similar to the target customer background information; the analysis server matches the reference customer related to the customer similarity analysis results according to the copywriting The click records of the reference texts related to the reference client related to the background information, and the classification results of the reference texts, generate a plurality of cumulative clicks corresponding to the types of the texts; the analysis server According to the classification results and the cumulative number of clicks, one of the candidate advertisement data is selected as a target advertisement data; the database server also stores multiple reference product click records, and multiple records related to these reference customers. A reference customer page browsing track, each reference product click record is related to one of these reference customers And related to one of these reference products and indicating whether to click or not, each reference customer page browsing track includes a plurality of dwell times respectively related to multiple pages and a plurality of browsing times respectively related to these pages; The analysis server trains an original slot prediction model to generate a slot prediction model based on the background information of the reference customers, the click records of the reference products and the page browsing tracks of the reference customers. The slot prediction model includes a plurality of corresponding In the reference position click rate data of the reference customers, each reference willingness rate data includes a plurality of reference position click rates corresponding to multiple positions; the analysis server uses the target customer background information to use the A bit prediction model, generating a plurality of customer similarities respectively related to the reference customers, each customer similarity indicates the similarity between the target customer and the related reference customer, and according to the customer similarities, using the The placement prediction model generates a plurality of predicted click willingness rates corresponding to the placements; the analysis server selects one of the placements as a target placement according to the predicted click willingness ratios, and the target The slot is used to place the target advertisement data. The digital marketing decision-making system as described in claim 1, wherein the analysis server uses latent semantic analysis technology to extract the keywords from the reference advertisement data. The digital marketing decision-making system as claimed in Claim 1, wherein the position prediction model is a deep factorization machine model. The digital marketing decision-making system as claimed in claim 1, wherein each reference product and each candidate product is related to one of a plurality of product categories, the data The library server also stores multiple reference customer page browsing tracks and multiple reference order records respectively related to these reference customers. Each reference customer page browsing track includes multiple stay times and A number of page views respectively related to the pages related to the product categories, each reference order record related to one of the reference customers and related to one of the reference products and instructed Order or not and order amount; the analysis server trains an original product familiarity prediction model to generate a product familiarity prediction model according to the reference customer page browsing track and the reference order records; the analysis server generates a product familiarity prediction model according to the With reference to the customer background data, train an original familiarity prediction customer similarity model to generate a familiarity prediction customer similarity model; the analysis server uses the familiarity prediction customer similarity model according to the target customer background data to generate a familiarity According to the familiarity prediction customer similarity analysis result, the familiarity prediction customer similarity analysis result is related to the one of the reference customer background data that is the closest to the target customer background data; the analysis server uses the familiarity prediction customer similarity analysis result Using the product familiarity prediction model to generate multiple corresponding to The product familiarity prediction data of these product categories; in addition to the classification results of the candidate texts and the accumulated clicks, the analysis server also based on the product familiarity prediction data, from the Waiting for the candidate advertisement data to select one of them as the target advertisement data. The digital marketing decision-making system according to claim 4, wherein the product familiarity prediction model is a logistic regression model. The digital marketing decision-making system as described in Claim 1, wherein each reference product and each candidate product is related to one of a plurality of product categories, and the database server also stores a plurality of reference product click records, respectively related Multiple reference customer financial data on these reference customers, multiple reference customer page browsing tracks related to these reference customers, a target customer financial data related to the target customer, and a target customer related to the target customer Customer page browsing track, each reference product click record is related to one of the reference customers and related to one of the reference products and indicates whether the click is made, each reference customer page browsing track and the target customer page browsing Tracks include multiple dwell times and multiple page views respectively related to multiple pages; the analysis server clicks records of the reference products, financial information of the reference customers and pages of the reference customers Browsing tracks, training an original click willingness rate prediction model to generate a click willingness rate prediction model, the click willingness rate prediction model includes a plurality of reference willingness rate data respectively corresponding to the reference customers, and each reference willingness rate data includes multiple The reference click willingness rates corresponding to the product categories; the analysis server uses the click willing rate prediction model based on the financial data of the target customer and the page browsing track of the target customer to generate multiple Customer similarity, each customer similarity indicates the similarity between the target customer and the related reference customer, and based on According to the customer similarity, use the click willingness rate prediction model to generate a plurality of predicted click willingness rates corresponding to the product categories; in addition to the classification results of the candidate copy and the accumulated click times, the analysis server, Also, according to the predicted click-through rate, one of the candidate advertisement materials is selected as the target advertisement data. The digital marketing decision-making system as described in Claim 6, wherein the click willing rate prediction model is a deep factor decomposition machine model. A digital marketing decision-making method, implemented by a digital marketing decision-making system, the digital marketing decision-making system includes a database server and an analysis server, the database server stores a plurality of pieces of reference advertisement data, respectively related to a plurality of Multiple reference customer background information, multiple reference copy click records, multiple candidate advertising materials, and a target customer background information related to a target customer. Each reference advertising material is related to one of the multiple reference products. and includes a reference advertising copy, each reference copy click record is related to one of the reference customers and related to one of the reference advertising materials, each candidate advertising material is related to one of a plurality of candidate products and Including a candidate advertisement copy, the analysis server is electrically connected to the database server through the communication network, and the database server also stores a plurality of reference product click records, and a plurality of reference customers respectively related to the reference customers Page browsing track, each reference product click record is related to one of the reference customers and is related to one of the reference products and indicates whether to click or not, each reference customer page browsing track contains multiple A dwell time of a page and a plurality of page views respectively related to the pages, the method includes: The analysis server extracts a plurality of keywords from the reference advertisement copy of the reference advertisement data; the analysis server trains a plurality of original words according to the keywords and a plurality of classification settings respectively corresponding to the keywords The bag model generates a plurality of bag-of-words models respectively corresponding to a plurality of copywriting categories, and each classification setting indicates one of the copywriting categories; the analysis server uses the bag-of-words models for each reference advertisement data to generate a plurality of bag-of-words models A reference copy type score, each reference copy type score corresponds to the reference advertisement data and the bag-of-words model used; the analysis server generates a reference copy type score for each reference advertisement data according to the corresponding reference copy type scores A reference copy classification result corresponding to the reference advertisement data and indicating one of the copy types; the analysis server trains an original copy matching customer similarity model according to the reference customer background data to generate a copy matching customer similarity model; The analysis server uses the bag-of-words models for the candidate advertising copy of each candidate advertising data to generate a plurality of candidate copy type scores, each candidate copy type score corresponding to the candidate advertising data and the bag of words used model; the analysis server, for each candidate advertisement data, generates a candidate text classification result corresponding to the candidate advertisement data and indicating one of the text types according to the corresponding scores of the candidate text types; the analysis server According to the background information of the target customer, use the copywriting matching customer similarity model to generate a copywriting matching customer similarity analysis result As a result, the copy matching customer similarity analysis result is related to the reference customer background information that is the closest to the target customer background information; the analysis server matches the reference customer background information related to the customer similarity analysis result according to the copywriting The click records of the reference texts related to the reference client and the classification results of the reference texts generate multiple cumulative clicks corresponding to the text categories; the analysis server classifies the candidate texts according to the classification results and the cumulative number of clicks, select one of the candidate advertisement data as a target advertisement data; based on the background information of the reference customers, the click records of the reference products and the page browsing track of the reference customers, the analysis server, Training an original position prediction model to generate a position prediction model, the position prediction model includes a plurality of reference position click-through rate data corresponding to the reference customers, and each reference willingness rate data includes a plurality of information corresponding to multiple The click-through rate of reference slots of each slot; the analysis server uses the slot prediction model based on the background information of the target customer to generate a plurality of customer similarities respectively related to the reference customers, and each customer similarity indicates The similarity between the target customer and the relevant reference customer, and according to the customer similarity, use the slot prediction model to generate a plurality of predicted click-through rates corresponding to the slots; and the analysis server According to the predicted click-through rates, one of the positions is selected as a target position, and the target position is used to place the target advertisement data. The digital marketing decision-making method as described in claim item 8, wherein each reference product and each candidate product is related to one of a plurality of product categories, and the database server also stores a plurality of information respectively related to the reference customers One reference customer page browsing track, and multiple reference ordering records, each reference customer page browsing track contains multiple stay times related to multiple pages and multiple page views respectively related to these pages, each reference The order record is related to one of the reference customers and related to one of the reference products and indicates whether to place an order and the order amount. The method also includes: the analysis server according to the page browsing track of the reference customers and the reference order records, train an original product familiarity prediction model to generate a product familiarity prediction model; the analysis server trains an original familiarity prediction customer similarity model to generate a familiarity based on the reference customer background information Predictive customer similarity model; the analysis server uses the familiarity predictive customer similarity model based on the background information of the target customer to generate a familiarity predictive customer similarity analysis result, and the familiarity predictive customer similarity analysis result is related to the reference The one of the customer background data that is closest to the target customer background data; the analysis server predicts the reference customer page related to the reference customer background data related to the reference customer related to the customer similarity analysis result based on the familiarity The browsing track and the reference order record or the reference order records use the product familiarity prediction model to generate a plurality of product familiarity prediction data corresponding to the product categories; and the analysis server, in addition to the Candidate copy classification results and the According to the accumulated number of clicks, one of the candidate advertising materials is selected as the target advertising data according to the product familiarity prediction data. The digital marketing decision-making method as described in Claim 8, wherein each reference product and each candidate product is related to one of a plurality of product categories, and the database server also stores a plurality of reference product click records, respectively related Multiple reference customer financial data on these reference customers, multiple reference customer page browsing tracks related to these reference customers, a target customer financial data related to the target customer, and a target customer related to the target customer Customer page browsing track, each reference product click record is related to one of the reference customers and related to one of the reference products and indicates whether the click is made, each reference customer page browsing track and the target customer page browsing The track includes a plurality of dwell times and a plurality of page views respectively related to the pages, and the method also includes: the analysis server clicks on the records of the reference products, the financial data of the reference customers and These reference customer page browsing tracks train an original click willing rate prediction model to generate a click willing rate prediction model. The click willing rate prediction model includes multiple reference willing rate data corresponding to the reference customers respectively. The rate data includes a plurality of reference click willingness rates corresponding to the product categories; the analysis server uses the click willingness rate prediction model based on the financial data of the target customer and the page browsing track of the target customer to generate a number of click-through rate prediction models respectively corresponding to The customer similarity of these reference customers, each customer similarity indicates the similarity between the target customer and the related reference customer, and according to the customer similarity, using the click willingness rate prediction model, how many a predicted click-through rate corresponding to the product category; and the analysis server selects from the candidate advertisement data based on the predicted click-through rate in addition to the classification results of the candidate copy and the accumulated click times One of them is used as the target advertising material.

Images