TWI714213B - User type prediction system and method thereof - Google Patents

Abstract

A user type prediction system includes a data transmission module, a data classification module, a user type tag module, and a user type prediction module. The data transmission module is configured to receive first browsing domain data of complete data users and corresponding webpage content, and receive second browsing domain data of a partial data user. The data classification module is configured to classify the first browsing domain data and the corresponding webpage content, and generate classification results. The user type tag module is configured to generate a type tag according to the first browsing domain data and the classification results. The user type prediction module is configured to compare the second browsing domain data with the first browsing domain data to generate a first comparison result, and label the partial data user according to the first comparison result to predict the user type of the partial data user.

Description

User profile prediction system and method

The present invention relates to a system and method for predicting user patterns, and in particular to a system and method for predicting the lifestyle patterns of users with only partial user data.

In this era of technology that uses big data for analysis to conduct behavioral analysis for consumers, telecommunications companies have also begun to conduct big data analysis for the content of users’ web browsing in order to understand the user profile of each user. Users do further services, such as: advertising. Since telecom operators use the text and keyword terms that users contact when they visit the page, they are used as the basis for judging consumers' lifestyle preferences. Therefore, telecom operators need sufficient web page data to analyze the strength of the corresponding keywords based on the frequency of the characters contacted by users. If only a limited amount of data is used for analysis, there will be a risk of partiality. Therefore, it is necessary to use a sufficient amount of web content as the object of analysis, for example: at least about 30 web pages browsed every month.

However, the amount of analytical data currently available to telecom operators is really limited. Because some restricted web content cannot be obtained, such as pages that require personal login information, URL content obtained without relevant permissions (such as Facebook, Gmail...), crawling content cannot be obtained due to the token security agreement , High-risk websites (such as pornography, violence), do not comply with telecommunications information security regulations, jump pages/relay pages without substantial information. If we analyze its network behavior characteristics based on the data of 100,000 users, the number of users with sufficient amount of web page text content per month accounts for about 10% of the total number of users, and only the domain Browsing records but the monthly webpage text content is insufficient accounted for about 75% of the total number of users, and the remaining number of users has no online related records. With the amount of data mentioned above, it is really difficult for the telecommunications industry to make an accurate analysis, so there are indeed issues that must be improved.

In view of this, one scope of the present invention is to provide a user profile prediction system to solve the problems of the prior art. The user profile prediction system of the present invention is used to predict the user profile of some data users who use the network service, and some data users use the network service to generate a plurality of second browsing domain data. The user pattern prediction system includes a data transmission module, a data classification module, a user pattern labeling module, and a user pattern prediction module. The data transmission module is used for receiving plural first browsing domain data of a plurality of complete data users and plural web content corresponding to the first browsing domain data, and receiving second browsing domain data of some data users. The data classification module is connected to the data transmission module. The data classification module is used to classify the first browsed domain data and the corresponding webpage content, and generate a plurality of classification results respectively corresponding to the first domain data. The user type label module is connected to the data classification module. The user type tag module is used to generate type tags corresponding to the complete data user based on the first browsed domain data of the complete data user and the corresponding classification result. The user profile prediction module is connected to the user profile tag module. The user pattern prediction module is used to compare the second browsing domain data of the partial data user with the first browsing domain data of the complete data user to generate the first comparison result, and to compare parts according to the first comparison result Data users paste at least one of a plurality of type tags to predict the user types of some data users.

Among them, the user type prediction module can be used to compare the first browsing domain data with the second browsing domain data. When the composition of the second browsing domain data falls first among users with complete data When the first browsing domain data of the complete data user is covered, the user type tag module will paste the type label corresponding to the first complete data user to some data users, and then predict the user type of some data users .

Among them, the user type prediction module may further include a prediction model to compare the second browsing domain data with the first browsing domain data, and according to the first comparison result, a plurality of type tags are attached to some data users At least one of them. The prediction model can use the nearest neighbor algorithm (KNN) for comparison calculation.

Among them, the K of the nearest neighbor algorithm (KNN) can be 1.

Among them, the prediction model can be updated by machine learning based on the first browsed domain data of the user with complete data and the webpage content corresponding to the first browsed domain data.

Among them, the user type prediction system of the present invention may further include a user database connected to a user type prediction module. The user database can be used to store multiple user data of complete data users and partial data users. The user type prediction module compares the user data of the part of the data user with the user data of the complete data user to generate a second comparison result, and revises each type tag according to the second comparison result to correct the part of the data The user type of the user, thereby improving the prediction accuracy of the user type of the part of the data user.

Another category of the present invention is to provide a user profile prediction method to solve the problems of the prior art. The user type prediction method is used to predict the user type of some data users who use the network service, and some data users use the network service to generate plural second browsing domain data. The user type prediction method includes the following steps: receiving plural first browsing domain data and corresponding plural web content of a plurality of complete data users, and second browsing domain data of some data users; classifying the first browsing domain Data and corresponding web content, and Generate a plurality of classification results corresponding to the first domain data; according to the first browsed domain data of the complete data user and the corresponding classification results, generate the type label corresponding to the complete data user; compare the second of some data users Browsing the domain data and the first browsing domain data of the complete data user to generate the first comparison result; according to the first comparison result, paste at least one of a plurality of type tags on part of the data users to predict the part The user type of the data user.

Among them, the step of predicting the user types of some data users further includes the following sub-steps: when the first comparison result is that the composition of the second browsing domain data falls on the first complete data user among the complete data users 1. When browsing the coverage of the domain data, paste the type label corresponding to the first complete data user to some data users to predict the user type of some data users.

Among them, after the step of predicting the user profile of some data users, it further includes the following steps: comparing the plural user data of the complete data user with the user data of some data users to generate a second comparison result; according to the second comparison Revise the type label to the result to modify the user type of some data users, thereby improving the prediction accuracy of the user type of some data users.

Wherein, in the step of comparing the first browsed domain data with the second browsed domain data, the nearest neighbor algorithm (KNN) is used for comparison calculation.

Compared with the prior art, the user profile prediction system and method of the present invention have the following advantages: 1. By establishing a prediction model with a small number of users, the computer computing load is reduced. 2. Breaking through the threshold limitation that the existing crawling text can only obtain part of the data, the user profile prediction system of the present invention can analyze the user profile of users with complete data that account for 10% of the total users based on the existing crawling text, and To establish a predictive model to predict the user profile of some data users who account for 75% of the total users, thereby effectively expanding applications (such as advertising).

1‧‧‧User profile prediction system

11‧‧‧Data Transmission Module

12‧‧‧Data Classification Module

13‧‧‧User Type Label Module

14‧‧‧User profile prediction module

141‧‧‧Predictive model

15, 21‧‧‧User database

2‧‧‧Telecom Database

S1-S7‧‧‧Step

S51‧‧‧Substep

Fig. 1 is a functional block diagram of a user profile prediction system according to a specific embodiment of the present invention.

Fig. 2 is a functional block diagram of a user profile prediction system according to another specific embodiment of the present invention.

Figure 3 is a flow chart of the steps of a user profile prediction method according to a specific embodiment of the present invention.

FIG. 4 is a flowchart of the steps of a user profile prediction method according to another specific embodiment of the present invention.

Fig. 5 is a flowchart of the steps of a user profile prediction method according to another specific embodiment of the present invention.

In order to make the advantages, spirit and features of the present invention easier and clearer to understand, the following embodiments will be used for detailed and discussion with reference to the accompanying drawings. It should be noted that these examples are only representative examples of the present invention. However, it can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

The terms used in the various embodiments disclosed in the present invention are only used for the purpose of describing specific embodiments, and are not used to limit the various embodiments disclosed in the present invention. The singular form as used herein also includes the plural form, unless the context clearly indicates otherwise. Unless otherwise defined, all terms (including technical and scientific terms) used in this specification have the same meanings as commonly understood by those of ordinary skill in the art to which various embodiments disclosed in the present invention belong. The above-mentioned terms (such as those defined in commonly used dictionaries) will be interpreted as having the same meaning as the contextual meaning in the same technical field, and will not be interpreted as having idealized or overly formal meanings, Unless clearly defined in the various embodiments disclosed in the present invention.

Compared with the prior art, the user profile prediction system 1 and method of the present invention are transparent The text and keyword words that the user actually browses on the page are used as the basis for judging the preference of the user type label. The user type prediction system 1 and method of the present invention first classifies and affixes type tags to 10% of the total users who have complete data users with sufficient browsing content to browse web pages to determine the user type of complete data users . Then, the analysis results of complete data users are used as seed points to compare with some data users who account for 75% of total users with insufficient data, and use this to predict the user patterns of some data users who account for 75% of total users.

Please refer to FIG. 1. FIG. 1 is a functional block diagram of a user profile prediction system 1 according to a specific embodiment of the present invention. In this specific embodiment, the user profile prediction system 1 includes a data transmission module 11, a data classification module 12, a user profile labeling module 13, and a user profile prediction module 14. The data transmission module 11 is connected to the telecommunication database 2 of the telecommunication company, and is used to receive the plural first browsing domain data of the plural complete data users stored in the telecommunication database 2 and the corresponding first browsing domain data Multiple web content, and multiple second browsing domain data from users receiving partial data. Among them, the first browsing domain data, web content, and second browsing domain data are generated by full data users and some data users using network services, respectively. The data classification module 12 is connected to the data transmission module 11. The data classification module 12 is used to classify the first browsed domain data and the corresponding webpage content, and generate a plurality of classification results respectively corresponding to the first domain data. The user type label module 13 is connected to the data classification module 12. The user type tag module 13 is used to generate type tags corresponding to the complete data user based on the first browsed domain data of the complete data user and the corresponding classification result. The user profile prediction module 14 is connected to the user profile tag module 13. The user pattern prediction module 14 is used to compare the second browsing domain data of the partial data user with the first browsing domain data of the complete data user to generate the first comparison result, and compare it according to the first comparison result Some data users affix at least one of a plurality of type tags, and then the forecasting department The user type of the data user. In practical applications, the data transmission module 11, the data classification module 12, the user profile labeling module 13, and the user profile prediction module 14 can be built on the central processing unit or system processing chip of a computer or server.

Furthermore, the data classification module 12 may include a classification dictionary for crawling, segmenting and using the first browsed domain data and corresponding web content generated by the complete data user through the network service provided by the telecommunications company The content of each webpage is analyzed by the word frequency and compared with the classification dictionary to distinguish its categories. For example, tennis is classified as sports, Michael Jordan is classified as sports stars, and lipsticks are classified as beauty. , Presidents are classified as political and so on. Next, the user profile labeling module 13 uses the composition of the classification result to label the user with complete data. For example, the composition of the classification result includes makeup, feminine care, and feminine products, which are labeled as female. Labels and so on include lifestyle labels related to age, gender, thoughts and preferences. The user profile prediction module 14 may further include a prediction model 141. The prediction model 141 is established by using the first browsing domain data of the user with complete data and the web content corresponding to the first browsing domain data, and is updated by machine learning. The prediction model 141 uses the complete data user as a seed point, and compares the second browsing domain data of some data users with the first browsing domain data of the complete data user to generate the first comparison result.

In actual application, the prediction model 141 can, based on the first comparison result, put the composition of the plural second browsing domain data within the coverage of the first browsing domain data of the first complete data user among the complete data users , That is, when the composition of the second browsing domain data of this part of the data user is similar to the composition of the first browsing domain data of the first complete data user, the user type tag module 13 is the type corresponding to the first complete data user Status tags are attached to some data users to predict the user types of some data users.

In addition to the aforementioned prediction methods, the prediction model 141 can classify each second domain data that falls within the coverage of each first browse domain data into the same classification result as the corresponding first browse domain data. Then, the user profile labeling module 13 uses the classification result to attach profile labels to some data users. Finally, the prediction model 141 predicts the user profile of some data users based on the distribution of their profile tags. In the above two prediction processes, one is to compare the composition of multiple browsing domain data to predict the user types of some data users, and the other is to compare a single browsing domain data to predict the users of some data users Type. Those with ordinary knowledge in the field can choose the appropriate method according to their needs.

The predictive model 141 of the present invention attempts to calculate different machine learning methods: K-Nearest Neighbor (KNN), Decision Tree (Decision Tree), and Logistic regression. The overall browsing data collected by the industry is divided into training set and test set for comparison and evaluation. The evaluation index formulas respectively include accuracy (Precision) = number of users with real conditions in the data/number of users with identification conditions; Recall = number of users with real conditions in the data/number of users with identification conditions; Accuracy = (data The number of real users + the number of real unconditional users in the data)/the total number of users; F1-SCORE=2/1/recall rate+1/accuracy rate OR (2*accuracy rate*recall rate)/accuracy rate+recall rate ). Among them, the accuracy rate (Precision) is an evaluation index of accuracy, that is, the percentage of the data identified as the positive class actually being the positive class. Recall is the percentage of correctly identified positive data in the actual positive data. The disadvantage of Accuracy is that it cannot represent the potential distribution of any relevant test data, and the imbalance of classification cannot be considered. F1-SCORE is a weighted evaluation index of the harmonic mean of accuracy and recall.

Perform parameter test with K-Nearest Neighbor (KNN), K Take 1, 3, 5, 7, 9 respectively, and divide the data into training set and test set 7:3, and repeat 10 times (CV=10) for comparison and evaluation to obtain the following test data.

最後如上表所示，針對F1-score進行綜合評估，可以發現K=1時，雖然針對男性與壯年之預測結果都比K=3、5、7、9差，但差別不大，且針對女性、 年輕以及老年之預測結果都有比較好的表現，因此，以K=1的距離所得到的指標表現最為理想。

Finally, as shown in the above table, a comprehensive evaluation of F1-score can be found when K=1, although the prediction results for men and adults are worse than K=3, 5, 7, and 9, but the difference is not large, and for women , The prediction results of young and old have better performance. Therefore, the performance of the index obtained with the distance of K=1 is the most ideal.

Similarly, divide the data into training set and test set at 7:3, and repeat the test 10 times (CV=10) with Logistic regression analysis.

以F1-score評估指標來看，羅吉斯迴歸分析雖然於男性及壯年之預測結果比近鄰演算法好，但近鄰演算法與羅吉斯迴歸分析的差異不超過16%。然而，羅吉斯迴歸分析卻於女性、年輕及老年之預測結果表現與近鄰演算法差異500%以上。因此，近鄰演算法相較於羅吉斯迴歸分析具有較佳表現。綜上所述，於最佳實施例中，本發明之預測模型141選用近鄰演算法(KNN)，且K=1進行比對計算。

According to the F1-score evaluation index, although the logistic regression analysis predicts better than the nearest neighbor algorithm for men and middle-aged people, the difference between the nearest neighbor algorithm and the logistic regression analysis does not exceed 16%. However, Logis regression analysis differs by more than 500% in the prediction results of women, young and old, and the nearest neighbor algorithm. Therefore, the nearest neighbor algorithm has better performance than Logis regression analysis. In summary, in the preferred embodiment, the prediction model 141 of the present invention uses the nearest neighbor algorithm (KNN), and K=1 for comparison calculation.

In order to improve the accuracy of the results predicted by the user profile prediction system 1 of the present invention, the user profile prediction system 1 of the present invention may further include a user database 15, which uses the complete data stored in the user database 15 and For the user data of some data users, the user types of some data users predicted by the prediction model 141 are further revised. In the embodiment in FIG. 1, the user database 15 is connected to the user profile prediction module 14. 14 user profile prediction modules Compare the user data of some data users according to the user data of the complete data user to generate a second comparison result, and modify each type label according to the second comparison result, and then modify the user type of some data users. In addition, please refer to FIG. 2, which is a functional block diagram of a user profile prediction system according to a specific embodiment of the present invention. The difference from the embodiment of FIG. 1 is that the user database 21 of the embodiment of FIG. 2 is set in the telecommunication database 2 to save the storage space of the user profile prediction system. The device in the embodiment of FIG. 2 that is the same as that of FIG. 1 will not be described again.

In detail, the user information includes the user's gender, age, occupation, education, etc. The prediction model 141 can further use the user information of the user with complete data, the first browsed domain data and the content of the webpage for machine learning, and then find the user information The relationship with the first browse domain data and web content. The user profile prediction system 1 of the present invention uses the prediction model 141 to further predict the user profile of some data users, calculate the relationship between the user data and the second browsing domain data, and use the prediction model 141 to calculate the relationship between the user data and the second browsing domain data, and use the prediction model 141 to calculate The shape is revised.

Please refer to FIG. 3, which is a flow chart of the steps of a user profile prediction method according to a specific embodiment of the present invention. As shown in Figure 2, the user profile prediction method includes the following steps: Step S1: Receive a plurality of first browsing domain data and corresponding plural web content of a plurality of complete data users, and a second browsing network of some data users Domain data; Step S2: Classify the first browsed domain data and the corresponding web content, and generate a plurality of classification results corresponding to the first domain data; Step S3: According to the complete data user’s first browsed domain data and Corresponding classification results to generate type tags corresponding to users with complete data; Step S4: Compare the second browsing domain data of partial data users with the first browsing domain data of complete data users to generate a first comparison result Step S5: According to the first comparison result, paste at least one of a plurality of type tags on part of the data users, and then predict the user type of some data users. Among them, user profile prediction The method can be achieved with the aforementioned user profile prediction system, so the same content as the aforementioned will not be repeated here.

In detail, please refer to FIG. 4, which is a flowchart of the steps of a user profile prediction method according to another specific embodiment of the present invention. As shown in FIG. 4, in step S5, it further includes a sub-step S51: when the first comparison result is that the composition of the second browsing domain data falls in the first browsing domain of the first complete data user among the complete data users When the data is within the scope of coverage, the type label corresponding to the first complete data user is attached to some data users to predict the user type of some data users. The embodiment shown in FIG. 3 uses the aforementioned first prediction method, and uses the composition of a plurality of second browsing domain data as a comparison basis to compare the composition of a plurality of first browsing domain data. When it is compared that the composition of the second browsing domain data falls within the coverage of the first complete data user among the complete data users, it can be considered that the type label of this part of the data user should be the same as that of the first complete data user. The type labels are the same. Among them, the preferred embodiment of the present invention uses K-Nearest Neighbor (KNN), and K=1 to define the coverage of the first browsed domain data. Of course, those who are generally knowledgeable in the field can also perform the comparison in the second prediction method mentioned above, and it is not limited to this.

In order to improve the accuracy of the user profile prediction method of the present invention, the user profile prediction method of the present invention further includes a correction step. Please refer to FIG. 5, which is a flowchart of the steps of the user profile prediction method according to another specific embodiment of the present invention. As shown in FIG. 5, after step S5, it further includes step S6: comparing the user data of the complete data user with the user data of some data users to generate a second comparison result; step S7: according to the second comparison As a result, the type label is modified to modify the user type of some data users, thereby improving the accuracy of the prediction of the user type of some data users. In detail, the user profile prediction method of the present invention first compares the browsing domain data of the user to predict the user profile of some data users. Then, do it with user information Further corrections to ensure the accuracy of the forecast.

Compared with the prior art, the user profile prediction system 1 and method of the present invention establishes the prediction model 141 with a small number of users, thereby reducing the computational load of the computer. Furthermore, the user profile prediction system 1 and method of the present invention breaks through the threshold limitation that the existing crawling text can only obtain part of the data. The user profile prediction system 1 of the present invention can analyze and account for the existing crawling text. The user profile of 10% of the total users with complete data, and the establishment of a prediction model 141 to predict the user profile of some data users accounting for 75% of the total users, which can effectively expand the application (such as advertising). The user profile prediction system 1 of the present invention will provide big data for prediction and filling of missing data, so as to improve the data volume and accuracy that can be calculated by big data. In addition, the user profile prediction system 1 and method of the present invention further includes a correction function to further modify the user profile using user data, thereby ensuring the accuracy of the prediction.

Through the detailed description of the above specific embodiments, it is hoped that the characteristics and spirit of the present invention can be described more clearly, and the scope of the present invention is not limited by the specific embodiments disclosed above. On the contrary, its purpose is to cover various changes and equivalent arrangements within the scope of the patent application for the present invention.

1‧‧‧User profile prediction system

11‧‧‧Data Transmission Module

12‧‧‧Data Classification Module

13‧‧‧User Type Label Module

14‧‧‧User profile prediction module

141‧‧‧Predictive model

15‧‧‧User Database

Claims (10)

A user type prediction system for predicting a user type of a part of data users who use a network service. The part of the data user uses the network service to generate a plurality of second browsing domain data. The user type The state prediction system includes: a data transmission module for receiving a plurality of first browsing domain data of a plurality of complete data users and a plurality of web content corresponding to the first browsing domain data, and receiving the partial data users The second browsing domain data; a data classification module connected to the data transmission module, the data classification module is used to classify the first browsing domain data and the corresponding web content, and generate separate Corresponding to a plurality of classification results of the first domain data; a user type tag module connected to the data classification module, and the user type tag module is used for the first browsing of each user with the complete data The domain data and the corresponding classification results generate a type label corresponding to each complete data user; and a user type prediction module connected to the user type tag module, and the user type prediction module is used for Compare the second browsing domain data of the partial data users with the first browsing domain data of the complete data users to generate a first comparison result, and the first comparison result is based on the first comparison result Some data users attach at least one of the type tags to predict the user type of each part of the data users. For example, the user type prediction system described in item 1 of the scope of patent application, wherein the user type prediction module is used to compare the first browsing domain data with the second browsing domain data, when the When the composition of the second browsing domain data falls within the coverage of the first browsing domain data of one of the complete data users, the user type tag module will The type label of the first complete data user should be affixed to this part of the data user to predict the user type of the part of the data user. For example, the user type prediction system described in item 1 of the scope of patent application, wherein the user type prediction module further includes a prediction model to compare the second browsing domain data with the first browsing domain data. The first comparison result is generated, and at least one of the type labels is attached to the part of the data users according to the first comparison result, and the prediction model uses the nearest neighbor algorithm (KNN) for comparison calculation. The user profile prediction system described in item 3 of the scope of patent application, wherein the K of the nearest neighbor algorithm (KNN) is 1. For example, the user type prediction system described in item 3 of the scope of patent application, wherein the prediction model is based on the first browsing domain data of the complete data users and the web pages corresponding to the first browsing domain data The content is updated with machine learning. For example, the user type prediction system described in item 1 of the scope of the patent application further includes a user database connected to the user type prediction module, and the user database is used to store the complete data users and the partial data A plurality of user data of a user, the user type prediction module compares the user data of the partial data users with the user data of the complete data users to generate a second comparison result, and according to the second comparison result The comparison result modifies each of the type tags to modify the user type of the part of the data user, thereby improving the prediction accuracy of the user type of the part of the data user. A user type prediction method for predicting a user type of a part of data users who use a network service. The part of the data user uses the network service to generate a plurality of second browsing domain data. The user type The state prediction method includes the following steps: Receive plural first browsed domain data and corresponding plural webpage contents of plural complete data users, and the second browsed domain data of this part of the data users; classify the first browsed domain data and corresponding And generate a plurality of classification results corresponding to the first domain data; according to the first browse domain data of each complete data user and the corresponding classification results, the corresponding classification results are generated A type tag of the complete data user; compare the second browsing domain data of the partial data users with the first browsing domain data of the complete data users to generate a first comparison result; And, according to the first comparison result, at least one of the type tags is attached to the part of the data users to predict the user type of each part of the data users. For example, in the user type prediction method described in item 7 of the scope of patent application, the step of predicting the user type of each part of the data user further includes the following sub-steps: when the first comparison result is the first comparison result 2. When the composition of the browsing domain data falls within the coverage of the first browsing domain data of one of the first complete data users, it will correspond to the type tag of the first complete data user Post to this part of the data user to predict the user type of the part of the data user. For example, the user type prediction method described in item 7 of the scope of patent application, after the step of predicting the user type of each part of the data user, further includes the following step: comparing the plural user data of the complete data user Generate a second comparison result with the user data of the part of the data user; and revise each type tag according to the second comparison result to modify the user type of the part of the data user, thereby improving the part The data user’s prediction of the user type is accurate degree. For example, the user type prediction method described in item 8 of the scope of patent application, wherein in the step of comparing the first browsing domain data with the second browsing domain data, the nearest neighbor algorithm (KNN) is used to perform Comparison calculation.

Images