Design of novel recommendation method applied to e-commerce platform

Abstract

The invention discloses a design of a novel recommendation method applied to an e-commerce platform, and the method comprises the steps: adding an impact factor in the calculation of a distance to represent the correlation of multiple latitudes between articles, so as to avoid recommendation deviation caused by not considering the individual features of a user; the basic preference information of the user is obtained in the new user registration stage, so that the cold start problem caused by the addition of the new user in the existing recommendation method is solved; an improved restricted Boltzmann machine is adopted to fill a coefficient matrix of an article score of a user to increase information amount, so that recommendation deviation caused by incomplete metadata is solved, and the accuracy of a recommendation result is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce, and in particular to the design of a new recommendation method applied to an e-commerce platform. Background Art

[0002] For e-commerce platforms, recommendation methods are their core competitiveness. In the prior art, recommendation methods are mainly content-based recommendation methods. Content-based recommendation methods extract the metadata of recommended items based on user preferences, discover the correlation and rules between items, and recommend similar items to users. This method can effectively model user preferences and then make accurate recommendations based on the model. However, this recommendation method has the following disadvantages: First, the recommended items are all modeled after the user's preferences, resulting in the quality of the recommended items being too dependent on the completeness of the model. If the metadata obtained is not comprehensive, the recommended items may be far from the user's preferences; second, the items obtained are only recommended based on the item's own label, without considering the user's individual characteristics, so there will be differences when making recommendations; third, the addition of new users will cause a cold start problem, that is, for new users who have just entered the system, the historical behavior is empty, so it is impossible to make accurate recommendations for new users. Summary of the invention

[0003] The purpose of the present invention is to provide a design of a new recommendation method applied to an e-commerce platform to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a design of a new recommendation method applied to an e-commerce platform, comprising the following steps: step 1, analyzing the reasons for the low accuracy of the basic recommendation method; step 2, conceiving an optimization plan for the basic recommendation method; step 3, implementing the optimization plan; step 4, adding the optimized recommendation plan to the e-commerce platform; step 5, software testing; step 6, software optimization;

[0005] In the above step 1, the reasons why the basic recommendation method based on similarity judgment has low accuracy are analyzed in terms of data and classification;

[0006] In the above step 2, for the data reasons mentioned in step 1, the optimization plan is as follows: obtain the basic preferences of new users during the registration phase of new users, and use the improved restricted Boltzmann machine to fill the coefficient matrix of user ratings for items in data modeling; for the classification reasons mentioned in step 1, the optimization plan is as follows: add influencing factors to the calculation of the Euclidean distance of items to take into account individual differences in multiple dimensions;

[0007] In the above step 3, the existing basic recommendation method is optimized according to the optimization scheme proposed in step 2;

[0008] In the above step 4, the recommendation plan after optimization in step 3 is added to the e-commerce platform;

[0009] In the above step 5, the recommendation function of the e-commerce platform is tested. The specific testing method is: recruiting volunteers to simulate new user registration, and statistically analyzing the volunteers' satisfaction with the recommendation results;

[0010] In the above step six, the e-commerce platform software is optimized according to the test results in step five, and then the test in step five is performed, and the testing and optimization process is repeated until the satisfaction reaches the preset value.

[0011] Preferably, in step one, the recommendation method based on similarity is specifically as follows: judging whether the item is the user's preference by judging the similarity between the merchant's item information and the user's preference information; if so, recommending the item to the user, and recommending the user to the merchant corresponding to the item; if not, proceeding to judging the next item until a termination condition is reached, and the termination condition is specifically as follows: reaching a preset number of recommendations or completing similarity judgment of all items.

[0012] Preferably, in step 1, the data reasons include the following two points: first, the metadata of the basic recommendation method is incomplete, resulting in the lack of accuracy of the recommendation results; second, the lack of new user information leads to the lack of accuracy of the recommendation results of the basic recommendation method.

[0013] Preferably, in step 1, the reason for classification is specifically that the basic recommendation method does not take into account the differences in individual characteristics of users.

[0014] Preferably, in step 2, the specific method for obtaining the basic preferences of the new user during the registration stage of the new user is: during the registration stage of the new user, first guide the new user to complete the filling of basic information, thereby generating basic preference information from this information, and then provide multiple-choice questions related to the preferences for the user to choose, and improve the basic preference information according to the user's selection results.

[0015] Preferably, in step 2, the influencing factor is added to the calculation of the Euclidean distance of the items as follows:

[0016]

[0017] In the formula, i is always not equal to j, r ij is the Pearson correlation coefficient of data samples i and j, max_r represents the maximum value of the Pearson correlation coefficient of all data objects, and min_r represents the minimum value of the Pearson correlation coefficient of all data objects;

[0018] Therefore, the distance after adding the impact factor is:

[0019] D ij =d ij ×(1-g ij +0.05)

[0020] d ij is the standard Euclidean distance. In order to prevent the distance from being 0 when the impact factor value is 1, 0.05 is added at the end. After obtaining the distance calculation formula, you can start to divide the entire data set into clusters. After completion, find out the isolated points and eliminate them.

[0021] Preferably, the cluster division is specifically as follows: the minimum distance between classes needs to be greater than or equal to the distance between the largest internal element point of a class and the center of the class, and the minimum distance between classes is:

[0022]

[0023] The distance between the largest internal element point of a class and the center of the class:

[0024]

[0025] Where D is the distance after adding the impact factor, c i and c j Represent the centers of the two clusters, x i It means it belongs to cluster c i A data point of

[0026] The average distance between classes:

[0027]

[0028] The minimum distance between classes needs to be less than or equal to the average value, and the maximum distance between the internal element point of a class and the center of the class must also be less than the average value. Because the average distance between classes is calculated by accumulation and then averaging, there must be coefficients λ1 and λ2, so that

[0029]

[0030] Set up a function:

[0031]

[0032] h1 in the function is the minimum distance between classes in the optimal state, and h2 is the distance between the maximum internal element point of a class and the center of the class in the optimal state, so the minimum value of the function y should be h1. h2 is 0; because in the process of clustering, as the amount of data continues to increase, the divided classes can be approximated as circles, and the two classes are getting closer and closer. The best case is that the radius of the class circles of the two clusters is equal and the sum of the tangents is equal to the distance between the centers of the circles. In the best case, the following equations should be satisfied:

[0033]

[0034] Substituting it into the function y, we can find that when λ1=4 / 5, λ2=2 / 5, y is the smallest, so we can conclude that when the distance between a data element and the cluster center is greater than 4 / 5 times the average value, it will be separated from this class; when the distance between a data element and the cluster center is less than 2 / 5 times the average value, it will be classified into this class. In this way, a complete data set can be divided into k categories with relatively obvious differences.

[0035] Preferably, the method for defining the isolated point is determined by density. Density is calculated by calculating the number of neighbors that the data element has within a certain radius. In the process of continuous clustering, the cluster center is taken as the neighbor of the point, so that there will be an average value of the distance from the data point to the cluster center:

[0036]

[0037] n is the total number of data sets used, represents the total set of 2 selected from the data set, k represents the number of cluster centers, and D is the distance after adding the impact factor;

[0038] If satisfied:

[0039]

[0040] This means that data point i is a neighbor of class j. After looping through all the points, if the neighbor points of a data element account for more than 25% of all data points within a certain radius, then the data element is an isolated point.

[0041] Preferably, in step 2, the coefficient matrix of user ratings for items is filled using the modified restricted Boltzmann machine: the nodes of the visible layer and the hidden layer of the modified restricted Boltzmann machine are binary variables, consisting of 0 and 1, and the energy function of the model is:

[0042]

[0043] Where, the parameter θ of the improved restricted Boltzmann machine consists of the weight W between the visible layer and the hidden layer, and the biases a and b corresponding to the visible layer and the hidden layer, respectively, that is, θ = (W, a, b), n and m are the number of nodes of the visible layer v and the hidden layer h, respectively;

[0044] The joint distribution probability of the two layers can be obtained:

[0045]

[0046] Where z(θ) is the normalization factor:

[0047]

[0048] From the above formula, it can be concluded that after the user's rating matrix is ​​input into the model, the activation probability of the unit in the hidden layer is:

[0049]

[0050] Where σ() is the activation function, set to:

[0051]

[0052] The value of the i-th unit in the visible layer can be obtained:

[0053]

[0054] Preferably, in step six, the optimization method is specifically: optimizing the content of the multiple-choice questions that need to be filled in during the new user registration phase, increasing the number of multiple-choice questions, so that more accurate user preference information can be obtained.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention adds an influencing factor to the distance calculation to represent the multi-dimensional correlation between items, thereby avoiding recommendation bias caused by not taking into account the individual characteristics of the user; the present invention obtains the user's basic preference information during the new user registration stage, thereby solving the cold start problem of the existing recommendation method caused by the addition of new users; the present invention increases the amount of information by using an improved restricted Boltzmann machine to fill in the coefficient matrix of the user's item ratings, thereby solving the recommendation bias caused by incomplete metadata, thereby effectively improving the accuracy of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Design a step diagram for the recommended method of the present invention;

[0057] Figure 2 is a flow chart of the recommended method of the present invention;

[0058] Figure 3 This is a similarity judgment flow chart of the recommendation method of the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Please refer to the attached Figure 1 -Attached Figure 3 , an embodiment provided by the present invention: a design of a new recommendation method applied to an e-commerce platform, comprising the following steps: step 1, analyzing the reasons for the low accuracy of the basic recommendation method; step 2, conceiving an optimization plan for the basic recommendation method; step 3, implementing the optimization plan; step 4, adding the optimized recommendation plan to the e-commerce platform; step 5, software testing; step 6, software optimization;

[0061] In the above step 1, the reasons for the low accuracy of the basic recommendation method based on similarity judgment are analyzed in terms of data and classification; the recommendation method based on similarity is specifically: judging whether the item belongs to the user's preference by judging the similarity between the merchant's item information and the user's preference information, if so, recommending the item to the user, and recommending the user to the merchant corresponding to the item; if not, it will enter the judgment of the next item until the termination condition is reached, and the termination condition is specifically: reaching the preset number of recommendations or completing the similarity judgment of all items; the reasons in terms of data include the following two points: first, the metadata of the basic recommendation method is incomplete, resulting in the lack of accuracy of the recommendation results, and second, the lack of new user information leads to the lack of accuracy of the recommendation results of the basic recommendation method; the reasons in terms of classification are specifically: the basic recommendation method does not take into account the differences in individual user characteristics;

[0062] In the above step 2, for the data reasons mentioned in step 1, the optimization plan is as follows: obtain the basic preferences of new users during the registration stage of new users, and use the improved restricted Boltzmann machine to fill the coefficient matrix of user ratings for items in data modeling; for the classification reasons mentioned in step 1, the optimization plan is as follows: add influencing factors to the calculation of the Euclidean distance of items to take into account individual differences in multiple dimensions; the specific method for obtaining the basic preferences of new users during the registration stage of new users is as follows: first guide the new users to complete the filling of basic information during the registration stage of new users, so that basic preference information is generated from this information, and then provide multiple-choice questions related to preferences for users to choose, and improve the basic preference information according to the user's selection results; add influencing factors to the calculation of the Euclidean distance of items as follows:

[0063]

[0064] In the formula, i is always not equal to j, r ij is the Pearson correlation coefficient of data samples i and j, max_r represents the maximum value of the Pearson correlation coefficient of all data objects, and min_r represents the minimum value of the Pearson correlation coefficient of all data objects;

[0065] Therefore, the distance after adding the impact factor is:

[0066] D ij =d ij ×(1-g ij +0.05)

[0067] d ij is the standard Euclidean distance. In order to prevent the distance from being 0 when the impact factor is 1, 0.05 is added at the end. After obtaining the distance calculation formula, the entire data set can be divided into clusters. After completion, isolated points can be found and eliminated. The specific division of clusters is as follows: the minimum distance between classes must be greater than or equal to the distance between the largest internal element point of a class and the center of the class, and the minimum distance between classes is:

[0068]

[0069] The distance between the largest internal element point of a class and the center of the class:

[0070]

[0071] Where D is the distance after adding the impact factor, c i and c j Represent the centers of the two clusters, x i It means it belongs to cluster c i A data point of

[0072] The average distance between classes:

[0073]

[0074] The minimum distance between classes needs to be less than or equal to the average value, and the maximum distance between the internal element point of a class and the center of the class must also be less than the average value. Because the average distance between classes is calculated by accumulation and then averaging, there must be coefficients λ1 and λ2, so that

[0075]

[0076] Set up a function:

[0077]

[0078] h1 in the function is the minimum distance between classes in the optimal state, and h2 is the distance between the maximum internal element point of a class and the center of the class in the optimal state, so the minimum value of the function y should be h1. h2 is 0; because in the process of clustering, as the amount of data continues to increase, the divided classes can be approximated as circles, and the two classes are getting closer and closer. The best case is that the radius of the class circles of the two clusters is equal and the sum of the tangents is equal to the distance between the centers of the circles. In the best case, the following equations should be satisfied:

[0079]

[0080] Substituting it into the function y, we can find that when λ1=4 / 5, λ2=2 / 5, y is the smallest, so it can be concluded that when the distance between a data element and the cluster center is greater than 4 / 5 times the average value, it is divided out of this class; when the distance between a data element and the cluster center is less than 2 / 5 times the average value, it is classified into this class, thus dividing a complete data set into k classes with obvious differences; the definition of isolated points is determined by density, which is calculated by calculating the number of neighbors the data element has within a certain radius. In the process of continuous clustering, the cluster center is taken as the neighbor of the point, so that there will be an average value of the distance from the data point to the cluster center:

[0081]

[0082] n is the total number of data sets used, represents the total set of 2 selected from the data set, k represents the number of cluster centers, and D is the distance after adding the impact factor;

[0083] If satisfied:

[0084]

[0085] It means that data point i is a neighbor of class j. Loop through all points. If the neighbor points of a data element account for more than 25% of all data points within a certain radius, then the data element is an isolated point. Use the improved restricted Boltzmann machine to fill in the coefficient matrix of user ratings for items (as shown in Table 1). Specifically, the nodes of the visible layer and hidden layer of the improved restricted Boltzmann machine are binary variables, consisting of 0 and 1. The energy function of the model is:

[0086]

[0087] Where, the parameter θ of the improved restricted Boltzmann machine consists of the weight W between the visible layer and the hidden layer, and the biases a and b corresponding to the visible layer and the hidden layer, respectively, that is, θ = (W, a, b), n and m are the number of nodes of the visible layer v and the hidden layer h, respectively;

[0088] The joint distribution probability of the two layers can be obtained:

[0089]

[0090] Where z(θ) is the normalization factor:

[0091]

[0092] From the above formula, it can be concluded that after the user's rating matrix is ​​input into the model, the activation probability of the unit in the hidden layer is:

[0093]

[0094] Where σ() is the activation function, set to:

[0095]

[0096] The value of the i-th unit in the visible layer can be obtained:

[0097]

[0098] In the above step 3, the existing basic recommendation method is optimized according to the optimization scheme proposed in step 2;

[0099] In the above step 4, the recommendation plan after optimization in step 3 is added to the e-commerce platform;

[0100] In the above step 5, the recommendation function of the e-commerce platform is tested. The specific testing method is: recruiting volunteers to simulate new user registration, and statistically analyzing the volunteers' satisfaction with the recommendation results;

[0101] In the above step six, the e-commerce platform software is optimized according to the test results in step five, and then the test of step five is performed, and the testing and optimization process is repeated until the satisfaction reaches the preset value; wherein, the optimization method is specifically as follows: the content of the multiple-choice questions that need to be filled in during the new user registration phase is optimized, and the number of multiple-choice questions is increased so that it can obtain more accurate user preference information.

[0102] Table 1 Coefficient matrix of user ratings for items

[0103]

[0104] Note: I represents items, u represents users, and r represents the ratings given to different items by different users.

[0105] Based on the above, the advantages of the present invention are that, when the present invention is used, first, an influencing factor is added to the distance calculation to represent the multi-dimensional correlation between items, so as to avoid recommendation bias caused by not taking into account the individual characteristics of the user; secondly, the problem of cold start is solved by obtaining the basic preference information of the user; furthermore, the coefficient matrix of the user's rating of the item is filled by the improved restricted Boltzmann machine to increase the amount of information, so as to solve the recommendation bias caused by incomplete metadata; the present invention can reduce the interference factors that may exist in the method by removing isolated points after distance classification, and the present invention combines the improved restricted Boltzmann machine, the clustering method and the improved Euclidean distance, first calculates the distance, then performs clustering, and finally fills the coefficient matrix of the classified categories in sequence using the improved restricted Boltzmann machine, so as to improve the accuracy of the recommendation, and divides a large amount of data into small blocks by clustering, which is not only convenient for calculation, but also conducive to processing isolated and uncertain influencing factors, and can solve the problem of a sharp increase in calculation amount when the amount of data is huge.

[0106] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A design of a new recommendation method applied to an e-commerce platform includes the following steps: Step 1: Analyze the reasons for the low accuracy of the basic recommendation method; Step 2: Conceive an optimization plan for the basic recommendation method; Step 3: Implement the optimization plan; Step 4: Add the optimized recommendation plan to the e-commerce platform; Step 5: Software testing; Step 6: Software optimization; It is characterized by: In the above step 1, the reasons why the basic recommendation method based on similarity judgment has low accuracy are analyzed in terms of data and classification; In the above step 2, for the data reasons mentioned in step 1, the optimization plan is as follows: obtain the basic preferences of new users during the registration phase of new users, and use the improved restricted Boltzmann machine to fill the coefficient matrix of user ratings for items in data modeling; for the classification reasons mentioned in step 1, the optimization plan is as follows: add influencing factors to the calculation of the Euclidean distance of items to take into account individual differences in multiple dimensions; In the above step 3, the existing basic recommendation method is optimized according to the optimization scheme proposed in step 2; In the above step 4, the recommendation plan after optimization in step 3 is added to the e-commerce platform; In the above step 5, the recommendation function of the e-commerce platform is tested. The specific testing method is: recruiting volunteers to simulate new user registration, and statistically analyzing the volunteers' satisfaction with the recommendation results; In the above step six, the e-commerce platform software is optimized according to the test results in step five, and then the test in step five is performed, and the testing and optimization process is repeated until the satisfaction reaches the preset value.

2. According to claim 1, the design of a new recommendation method applied to an e-commerce platform is characterized by: In the step 1, the recommendation method based on similarity is specifically as follows: whether the item is a user's preference is determined by judging the similarity between the merchant's item information and the user's preference information; if so, the item is recommended to the user, and the user is recommended to the merchant corresponding to the item; if not, the next item is judged until the termination condition is met, and the termination condition is specifically as follows: reaching a preset number of recommendations or completing the similarity judgment of all items.

3. According to the design of a new recommendation method applied to an e-commerce platform according to claim 1, it is characterized by: In the step 1, the data reasons include the following two points: first, the metadata of the basic recommendation method is incomplete, resulting in the lack of accuracy of the recommendation results; second, the lack of new user information leads to the lack of accuracy of the recommendation results of the basic recommendation method.

4. According to claim 1, the design of a new recommendation method applied to an e-commerce platform is characterized by: In the step 1, the classification reason is specifically that the basic recommendation method does not take into account the individual characteristics of users.

5. According to claim 1, the design of a new recommendation method applied to an e-commerce platform is characterized by: In the step 2, the specific method for obtaining the basic preferences of the new user during the registration stage of the new user is as follows: during the registration stage of the new user, first guide the new user to complete the filling of basic information, thereby generating basic preference information from this information, and then provide multiple-choice questions related to the preferences for the user to choose, and improve the basic preference information according to the user's selection results.

6. The design of a new recommendation method applied to an e-commerce platform according to claim 1 is characterized by: In step 2, the influencing factor is added to the calculation of the Euclidean distance of the items as follows: In the formula, i is always not equal to j, r ij is the Pearson correlation coefficient of data samples i and j, max_r represents the maximum value of the Pearson correlation coefficient of all data objects, and min_r represents the minimum value of the Pearson correlation coefficient of all data objects; Therefore, the distance after adding the impact factor is: D ij =d ij ×(1-g ij +0.05) d ij is the standard Euclidean distance. In order to prevent the distance from being 0 when the impact factor value is 1, 0.05 is added at the end. After obtaining the distance calculation formula, you can start to divide the entire data set into clusters. After completion, find out the isolated points and eliminate them.

7. The design of a new recommendation method applied to an e-commerce platform according to claim 6 is characterized by: The cluster division is specifically as follows: the minimum distance between classes needs to be greater than or equal to the distance between the largest internal element point of a class and the center of the class, and the minimum distance between classes is: The distance between the largest internal element point of a class and the center of the class: Where D is the distance after adding the impact factor, c i and c j Represent the centers of the two clusters, x i It means it belongs to cluster c i A data point of The average distance between classes: The minimum distance between classes needs to be less than or equal to the average value, and the maximum distance between the internal element point of a class and the center of the class must also be less than the average value. Because the average distance between classes is calculated by accumulation and then averaging, there must be coefficients λ1 and λ2, so that Set up a function: h1 in the function is the minimum distance between classes in the optimal state, and h2 is the distance between the maximum internal element point of a class and the center of the class in the optimal state, so the minimum value of the function y should be h1. h2 is 0; because in the process of clustering, as the amount of data continues to increase, the divided classes can be approximated as circles, and the two classes are getting closer and closer. The best case is that the radius of the class circles of the two clusters is equal and the sum of the tangents is equal to the distance between the centers of the circles. In the best case, the following equations should be satisfied: Substituting it into the function y, we can find that when λ1=4 / 5, λ2=2 / 5, y is the smallest, so we can conclude that when the distance between a data element and the cluster center is greater than 4 / 5 times the average value, it will be separated from this class; when the distance between a data element and the cluster center is less than 2 / 5 times the average value, it will be classified into this class. In this way, a complete data set can be divided into k categories with relatively obvious differences.

8. The design of a new recommendation method applied to an e-commerce platform according to claim 6 is characterized by: The method for defining isolated points is determined by density. Density is calculated by calculating the number of neighbors that the data element has within a certain radius. In the process of continuous clustering, the cluster center is taken as the neighbor of the point, so that there will be an average value of the distance from the data point to the cluster center: n is the total number of data sets used, represents the total set of 2 selected from the data set, k represents the number of cluster centers, and D is the distance after adding the impact factor; If satisfied: This means that data point i is a neighbor of class j. After looping through all the points, if the neighbor points of a data element account for more than 25% of all data points within a certain radius, then the data element is an isolated point.

9. The design of a new recommendation method applied to an e-commerce platform according to claim 1 is characterized by: In the step 2, the coefficient matrix of the user's rating of the item is filled using the improved restricted Boltzmann machine: the nodes of the visible layer and the hidden layer of the improved restricted Boltzmann machine are binary variables, consisting of 0 and 1, and the energy function of the model is: Where, the parameter θ of the improved restricted Boltzmann machine consists of the weight W between the visible layer and the hidden layer, and the biases a and b corresponding to the visible layer and the hidden layer, respectively, that is, θ = (W, a, b), n and m are the number of nodes of the visible layer v and the hidden layer h, respectively; The joint distribution probability of the two layers can be obtained: Where z(θ) is the normalization factor: From the above formula, it can be concluded that after the user's rating matrix is ​​input into the model, the activation probability of the unit in the hidden layer is: Where σ() is the activation function, set to: The value of the i-th unit in the visible layer can be obtained:

10. The design of a new recommendation method applied to an e-commerce platform according to claim 1 is characterized by: In step six, the optimization method is specifically as follows: optimizing the content of the multiple-choice questions that need to be filled in during the new user registration phase, increasing the number of multiple-choice questions, and making it possible to obtain more accurate user preference information.

Images