Triggering content recommendation method and device

Abstract

The invention relates to the field of artificial intelligence, and particularly provides a beacon content recommendation method and device, and the method comprises the following steps: S1, obtaining a coding vector of beacon related data through employing a text embedding model based on the basic information of a beacon, and constructing a beacon vector library; s2, based on historical browsing data of the user, determining an effective browsing beacon, and calculating a corresponding weight of the effective browsing beacon; s3, determining a first feature tag set and a second feature tag set by using a primary and secondary factor analysis method based on the effective browsing tags of the user and the corresponding weights thereof; s4, constructing a user feature vector set by using a text embedding model based on the user first feature tag set and the first feature tag associated data set; and S5, according to the user feature tag set and the user feature vector set, designing a beacon recommendation strategy to recommend the potential interested beacon of the user. Compared with the prior art, the method has the advantages that the latent interested tags of the user can be accurately matched from the semantic level according to the browsing record of the user, and a more personalized tag recommendation service can be provided for the user.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and specifically provides a method and device for recommending standard information content. Background Art

[0002] How to effectively filter out the information that users are interested in from a large amount of information is a key challenge to improve user experience and enhance user stickiness. Traditional recommendation methods often rely on simple keyword matching or statistical analysis of user historical behavior. Although this method can reflect user preferences to a certain extent, it is limited by the shallow processing of content and is difficult to capture deep semantic information, resulting in a certain deviation between the recommended content and the actual interests and preferences of users.

[0003] In order to deeply analyze and mine the user's historical browsing information, determine the user's personalized preferences, and then achieve the purpose of recommending interesting information content to the user, the present invention proposes a semantic-based information content recommendation method. Summary of the invention

[0004] The present invention aims at solving the above-mentioned deficiencies of the prior art and provides a method for recommending standard information content with strong practicability.

[0005] A further technical task of the present invention is to provide a device for recommending standard information content that is reasonably designed, safe and applicable.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A method for recommending news content includes the following steps:

[0008] S1. Based on the basic information of the signal, the text embedding model is used to obtain the encoding vector of the signal-related data, and a signal vector library is constructed based on this;

[0009] S2. Determine valid browsing information based on the user's historical browsing data and calculate its corresponding weight;

[0010] S3, based on the user's effective browsing information and its corresponding weight, using the primary and secondary factor analysis method to determine the first feature tag set and the second feature tag set;

[0011] S4, constructing a user feature vector set based on the user's first feature label set using a text embedding model;

[0012] S5. According to the user feature tag set and the user feature vector set, a message recommendation strategy is designed based on vector retrieval technology to recommend messages that the user may be interested in.

[0013] Furthermore, in step S1, the process of initializing and building the signal vector library and the process of periodically updating the signal vector library are included;

[0014] The initialization construction process of the signal vector library is as follows:

[0015] For the existing bidding information data, based on the preset screening parameters and the release date of the bidding information, batch screen the bidding information data with a time interval less than the preset screening parameters from the relevant tables of the bidding information database, and obtain the valid data of the screened bidding information;

[0016] For the existing valid bidding information data obtained, the bidding information project name is vectorized using the text embedding model to obtain the corresponding encoding vector, and it is stored in the bidding information vector library together with its corresponding valid scalar data.

[0017] Furthermore, the periodic update process of the signal vector library is as follows:

[0018] Set up a scheduled task to successively execute the clearing operation of expired and invalid bid information data in the bid information vector library and the insertion operation of more effective bid information data. Specifically, for the bid information vector library, according to the bid information release date, delete the expired and invalid bid information data whose interval with the current date is greater than the preset screening parameter from the relevant data set;

[0019] For the bid information database, according to the bid information release date, the more valid bid information data with the release date of today is filtered out from the relevant data table, and the newly acquired coding vectors and corresponding valid data are inserted into the bid information vector library in batches according to the vector library initialization construction method.

[0020] Furthermore, in step S2, the unique identifier, project name, category label, region label, browsing start time and browsing end time of the user's historical browsing data are parsed and extracted, and based on the browsing start time of the browsing label, the data whose interval with the current time is less than the first time interval is filtered. If the number of filtered data exceeds the set selection number, the redundant browsing labels are filtered out again after arranging in descending order based on the browsing end time to determine the valid browsing labels.

[0021] Furthermore, for the effective browsing signal feature data, the time interval between the browsing start time and the current time is calculated, and based on the first time interval, it is converted into a signal first weight, which is used to measure the timeliness of the effective browsing signal;

[0022] Then, according to the browsing start time and the browsing end time, the browsing time of the corresponding valid browsing information is calculated, and it is converted into the second weight of the information based on the second time interval, which is used to measure the user's preference for the browsing information;

[0023] The product of the first weight and the second weight corresponding to the effective browsing information is used as its weight to measure the importance of the effective browsing information in the process of generating recommended content;

[0024] Specifically, the first weight and the second weight of the bid information are calculated as follows:

[0025]

[0026] Among them, w 1 and w 2 are the first weight and the second weight of the signal, respectively, t s and t e are the browsing start time and browsing end time respectively, c 1 and c 2 are the first time interval and the second time interval respectively, t now Indicates the current moment.

[0027] Furthermore, in step S3, based on the valid browsing information data, the sum of the weights of each category tag corresponding to all valid browsing information is calculated, and the ratio of the sum of the weights of each category tag to the sum of the weights of all valid browsing information is calculated as the weight of the category tag; the category tags are sorted in descending order according to the category tag weight values, and the category tags that meet the conditions are selected using the primary and secondary factor analysis method to construct the user's first feature tag set;

[0028] Accordingly, for the regional tags, the regional tags that meet the conditions are selected to construct the second feature tag set of the user.

[0029] Furthermore, the process of constructing the first feature label set and the second feature label set based on the primary and secondary factor analysis method is as follows: based on actual needs, set a cumulative weight threshold; based on the results of descending order of label weights, select the weights corresponding to the top-ranked labels in turn, add them up to calculate the cumulative weights, until the cumulative weight is greater than the cumulative weight threshold, and use the selected labels as user feature labels;

[0030] For the first feature tag set, determine the valid browsing information set corresponding to each category tag therein, and determine the recommendation ratio value corresponding to the category tag therein; use the valid browsing information set and the recommendation ratio value corresponding to the category tag as its associated data, and then obtain the first feature tag associated data set;

[0031] Among them, the calculation formula for the recommendation ratio corresponding to the category label is:

[0032]

[0033] r is the recommendation ratio corresponding to the category label, w is the weight of the category label, and w sum is the sum of the weights of the feature labels in the first feature label set.

[0034] Further, in step S4, for the user's first feature label associated data set, after preprocessing the information item names in the valid browsing information set corresponding to the category label, the preprocessed information item names are encoded using the text embedding model to obtain the encoding vector set corresponding to the category label, and according to the relative weights of the valid browsing information in the corresponding valid browsing information set, the vectors in the obtained encoding vector set are summed using a weighted method to determine the feature vector corresponding to the category label; based on this, the user feature vector set can be obtained;

[0035] Specifically, the calculation formula for the relative weight of the category label corresponding to the valid browsing information set is:

[0036]

[0037] where w c is the relative weight of the effective browsing signal, w is the weight of the effective browsing signal, and w sum It is the sum of the weights of all the bids in the valid browsing bid set corresponding to the category label.

[0038] Furthermore, in step S5, a recommendation strategy of information potentially of interest to the user and a recommendation strategy of information randomly recommended are included;

[0039] The recommendation strategy for the information that the user is potentially interested in is:

[0040] For the category labels and their corresponding feature vectors in the user's first feature label set, first, combine the recommendation ratio corresponding to the category label and the total preset recommendation quantity, and take the product of the two as the recommendation quantity corresponding to the category label;

[0041] Then, based on the category label and the second feature label set, a corresponding Boolean expression is generated as a vector search filter condition; the feature vector corresponding to the category label is used as the search vector, and the number of recommended messages corresponding to the category label is used as the search result limit. A conditional search is performed from the message vector library based on the vector search filter condition to obtain message information with high semantic similarity to the user's browsed messages in the effective browsed message set corresponding to the category label, as the user's potential interested message corresponding to the category label;

[0042] The random recommendation standard recommendation strategy is: based on the user feature vector set, determine the average value of the feature vector, and generate a random vector consistent with the dimension of the user feature vector based on the random value interval, and add them to obtain the current random feature vector;

[0043] A value between 0 and 0.1 is randomly generated, and the product of the value and the total preset recommendation number is used as the search result limit number; the current random feature vector is used as the search vector, and the search is performed from the message vector library according to the search result limit number, and the message content that has semantic relevance to the valid browsing message but has a small semantic relevance and changes randomly is obtained as the random recommendation message, thereby increasing the diversity of the recommendation results;

[0044] Combining the above two recommendation strategies, the summary results are randomly shuffled and used as user-personalized recommendation information data.

[0045] A device for recommending content of a newsletter, comprising: at least one memory and at least one processor;

[0046] The at least one memory is used to store a machine-readable program;

[0047] The at least one processor is used to call the machine-readable program to execute a method for recommending advertising content.

[0048] Compared with the prior art, the method and device for recommending information content of the present invention have the following outstanding beneficial effects:

[0049] Based on the text embedding model and vector database, the present invention constructs a standard message vector library with comprehensive category coverage and strong content timeliness. It can provide rich standard message candidate resources and accurate matching basis on the basis of achieving a deep understanding of the standard message content.

[0050] By carefully analyzing and deeply mining the user's historical browsing records, we can accurately extract user preference features; combining the text embedding model and the semantic-based news content recommendation strategy of text similarity search can accurately match the user's potential news of interest at the semantic level based on the user's browsing records, and provide users with more personalized news recommendation services. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Attached Figure 1 It is a flowchart of constructing a bid information vector library in a bid information content recommendation method;

[0053] Attached Figure 2 The invention discloses a flow chart of generating user recommended information in an information content recommendation method. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than 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.

[0055] A best embodiment is given below:

[0056] like Figure 1-2 As shown, a method for recommending news content in this embodiment has the following steps:

[0057] S1. Based on the basic information of the signal, the text embedding model is used to obtain the encoding vector of the signal-related data, and a signal vector library is constructed based on this;

[0058] Including the initialization construction process of the signal vector library and the periodic update process of the signal vector library;

[0059] The initialization construction process of the signal vector library is as follows:

[0060] For existing bidding information data, based on preset filtering parameters and according to the release date of the bidding information, batch filter the bidding information data with an interval with the current date less than the preset filtering parameters from the relevant tables of the bidding information database, and obtain multiple valid data such as the unique identification, project name, category label, regional label, release date, etc. of the filtered bidding information; for the existing valid bidding information data obtained, use the text embedding model to vectorize the bidding information project name to obtain the corresponding encoding vector, and store it together with its corresponding valid scalar data in the bidding information vector library.

[0061] The periodic update process of the signal vector library is as follows:

[0062] Set up a scheduled task to successively execute the clearing operation of expired and invalid signal data in the signal vector library and the insertion operation of more effective signal data.

[0063] Specifically, for the bid information vector library, based on the bid information release date, expired and invalid bid information data with a time interval greater than the preset filtering parameters from the current date are deleted from the relevant data set; for the bid information database, based on the bid information release date, the more valid bid information data with a release date of today is filtered out from the relevant data table, and the newly acquired coding vectors and corresponding valid data are batch inserted into the bid information vector library according to the vector library initialization construction method.

[0064] S2. Determine valid browsing information based on the user's historical browsing data and calculate its corresponding weight;

[0065] For the user's historical browsing data, the unique identifier of the browsing signal, project name, category label, region label, browsing start time and browsing end time and other valid features are analyzed and extracted, and based on the browsing start time of the browsing signal, the data with an interval with the current time less than the first time interval is filtered out. If the number of filtered data exceeds the set selection number, the redundant browsing signals are filtered out again after sorting in descending order based on the browsing end time to determine the valid browsing signals.

[0066] For the effective browsing information feature data, the time interval between the browsing start time and the current time is calculated, and based on the first time interval, it is converted into the first information weight, which is used to measure the timeliness of the effective browsing information; then, according to the browsing start time and the browsing end time, the browsing time corresponding to the effective browsing information is calculated, and based on the second time interval, it is converted into the second information weight, which is used to measure the user's preference for the browsing information; the product of the first weight corresponding to the effective browsing information and the second weight is used as its weight, which is used to measure the importance of the effective browsing information in the process of generating recommended content;

[0067] Specifically, the first weight and the second weight of the bid information are calculated as follows:

[0068]

[0069] Among them, w 1 and w 2 are the first weight and the second weight of the signal, respectively, s and t e are the browsing start time and browsing end time respectively, c 1 and c 2 are the first time interval and the second time interval respectively, t now Indicates the current moment.

[0070] S3, based on the user's effective browsing information and its corresponding weight, using the primary and secondary factor analysis method to determine the first feature tag set and the second feature tag set;

[0071] Based on the valid browsing signal data, the sum of the weights of each category label corresponding to all valid browsing signals is calculated, and its ratio to the sum of the weights of all valid browsing signals is calculated as the weight of the category label; the category labels are sorted in descending order according to the category label weight values, and the primary and secondary factor analysis method is used to select the category labels that meet the conditions to construct the user's first feature label set.

[0072] Correspondingly, for the regional tags, the regional tags that meet the conditions are selected according to the above method to construct the second feature tag set of the user.

[0073] Specifically, the process of constructing a user feature label set (a first feature label set and a second feature label set) based on the primary and secondary factor analysis method is as follows: based on actual needs, set a cumulative weight threshold (between 0 and 1); based on the results of descending sorting of the label weights, select the weights corresponding to the top-ranked labels in turn, add them up to calculate the cumulative weight, until the cumulative weight is greater than the cumulative weight threshold, and use the selected label as the user feature label.

[0074] For the user's first feature tag set, determine the valid browsing information set corresponding to each category tag, and determine the recommendation ratio corresponding to the category tag; use the valid browsing information set and recommendation ratio corresponding to the category tag as its associated data, and then the first feature tag associated data set can be obtained. The calculation formula for the recommendation ratio corresponding to the category tag is:

[0075]

[0076] Among them, r is the recommendation ratio corresponding to the category label, w is the weight of the category label, and w sum is the sum of the weights of the feature labels in the first feature label set.

[0077] S4, constructing a user feature vector set based on the user's first feature label set using a text embedding model;

[0078] For the user's first feature label associated data set, after preprocessing the information item names in the valid browsing information set corresponding to the category label, the preprocessed information item names are encoded using the text embedding model to obtain the encoding vector set corresponding to the category label, and according to the relative weights of the valid browsing information in the corresponding valid browsing information set, the weighted method is used to sum the vectors in the obtained encoding vector set to determine the feature vector corresponding to the category label; based on this, the user feature vector set can be obtained.

[0079] Specifically, the calculation formula for the relative weight of the category label corresponding to the valid browsing information set is:

[0080]

[0081] Among them, w c is the relative weight of the effective browsing signal, w is the weight of the effective browsing signal, and w sum It is the sum of the weights of all the bids in the valid browsing bid set corresponding to the category label.

[0082] S5. According to the user feature tag set and the user feature vector set, a message recommendation strategy is designed based on vector retrieval technology to recommend messages that the user may be interested in;

[0083] Including recommendation strategies for information that users are potentially interested in and random recommendation strategies;

[0084] The recommendation strategy for information that users may be interested in is:

[0085] For the category labels and their corresponding feature vectors in the user's first feature label set, first, combine the recommendation ratio value corresponding to the category label and the total preset recommendation quantity, and take the product value of the two as the number of recommended labels corresponding to the category label; then, based on the category label and the second feature label set, generate the corresponding Boolean expression as the vector retrieval screening condition; take the feature vector corresponding to the category label as the retrieval vector, and the number of recommended labels corresponding to the category label as the limit number of retrieval results, perform conditional retrieval from the label vector library based on the vector retrieval screening condition, and obtain the label information with high semantic similarity to the user's browsed labels in the effective browsed label set corresponding to the category label, as the user's potential interested labels corresponding to the category label.

[0086] The random recommendation strategy is:

[0087] Based on the user feature vector set, determine the average value of the feature vector, and generate a random vector consistent with the dimension of the user feature vector based on a random value interval, and add them up to obtain the current random feature vector; randomly generate a value from 0 to 0.1, and use the product of it and the total preset recommendation number as the search result limit; use the current random feature vector as the search vector, and search from the message vector library according to the search result limit to obtain message content that has semantic relevance to the effective browsing message but has a small semantic relevance and randomly changes, as the random recommendation message, thereby increasing the diversity of the recommendation results.

[0088] Combining the above two recommendation results, the summary results are randomly shuffled and used as user-personalized recommendation information data.

[0089] Based on the above method, a device for recommending content of a standard message in this embodiment includes: at least one memory and at least one processor;

[0090] The at least one memory is used to store a machine-readable program;

[0091] The at least one processor is used to call the machine-readable program to execute a method for recommending advertising content.

[0092] The above-mentioned specific implementations are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementations. Any technical solutions that conform to the above-mentioned specific implementations of the present invention and any appropriate changes or substitutions made by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0093] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for recommending information content, characterized in that: The steps are as follows: S1. Based on the basic information of the signal, the text embedding model is used to obtain the encoding vector of the signal-related data, and a signal vector library is constructed based on this; S2. Determine valid browsing information based on the user's historical browsing data and calculate its corresponding weight; S3, based on the user's effective browsing information and its corresponding weight, using the primary and secondary factor analysis method to determine the first feature tag set and the second feature tag set; S4, constructing a user feature vector set using a text embedding model based on the user's first feature tag set and the first feature tag associated data set; S5. According to the user feature tag set and the user feature vector set, a message recommendation strategy is designed based on vector retrieval technology to recommend messages that the user may be interested in.

2. A method for recommending information content according to claim 1, characterized in that: In step S1, it includes the initialization construction process of the signal vector library and the periodic update process of the signal vector library; The initialization construction process of the signal vector library is as follows: For the existing bidding information data, based on the preset screening parameters and the release date of the bidding information, batch screen the bidding information data with a time interval less than the preset screening parameters from the relevant tables of the bidding information database, and obtain the valid data of the screened bidding information; For the existing valid bidding information data obtained, the bidding information project name is vectorized using the text embedding model to obtain the corresponding encoding vector, and it is stored in the bidding information vector library together with its corresponding valid scalar data.

3. A method for recommending information content according to claim 2, characterized in that: The periodic update process of the signal vector library is as follows: Set up a scheduled task to successively execute the clearing operation of expired and invalid bid information data in the bid information vector library and the insertion operation of more effective bid information data. Specifically, for the bid information vector library, according to the bid information release date, delete the expired and invalid bid information data whose interval with the current date is greater than the preset screening parameter from the relevant data set; For the bid information database, according to the bid information release date, the more valid bid information data with the release date of today is filtered out from the relevant data table, and the newly acquired coding vectors and corresponding valid data are inserted into the bid information vector library in batches according to the vector library initialization construction method.

4. A method for recommending information content according to claim 3, characterized in that: In step S2, the user's historical browsing data is parsed and extracted, including the unique identifier, project name, category tag, region tag, browsing start time and browsing end time of the browsing signal, and based on the browsing start time of the browsing signal, the data whose interval with the current time is less than the first time interval is filtered out. If the number of filtered data exceeds the set selection number, the redundant browsing signals are filtered out again after being sorted in descending order based on the browsing end time to determine the valid browsing signals.

5. A method for recommending information content according to claim 4, characterized in that: For the effective browsing signal feature data, the time interval between the browsing start time and the current time is calculated, and based on the first time interval, it is converted into the first signal weight, which is used to measure the timeliness of the effective browsing signal; Then, according to the browsing start time and the browsing end time, the browsing time of the corresponding valid browsing information is calculated, and it is converted into the second weight of the information based on the second time interval, which is used to measure the user's preference for the browsing information; The product of the first weight and the second weight corresponding to the effective browsing information is used as its weight to measure the importance of the effective browsing information in the process of generating recommended content; Specifically, the first weight and the second weight of the bid information are calculated as follows: Among them, w1 and w2 are the first weight and the second weight of the signal, respectively. s and t e are the browsing start time and the browsing end time respectively, c1 and c2 are the first time interval and the second time interval respectively, t now Indicates the current moment.

6. A method for recommending information content according to claim 5, characterized in that: In step S3, based on the valid browsing information data, the sum of the weights of each category tag corresponding to all valid browsing information is calculated, and the ratio of the sum of the weights of each category tag to the sum of the weights of all valid browsing information is calculated as the weight of the category tag; the category tags are sorted in descending order according to the category tag weight values, and the category tags that meet the conditions are selected using the primary and secondary factor analysis method to construct the user's first feature tag set; Accordingly, for the regional tags, the regional tags that meet the conditions are selected to construct the second feature tag set of the user.

7. A method for recommending information content according to claim 6, characterized in that: The process of constructing the first feature label set and the second feature label set based on the primary and secondary factor analysis method is as follows: based on actual needs, set the cumulative weight threshold; based on the results of descending order of label weights, select the weights corresponding to the top-ranked labels in turn, add them up to calculate the cumulative weights, until the cumulative weight is greater than the cumulative weight threshold, and use the selected labels as user feature labels; For the first feature tag set, determine the valid browsing information set corresponding to each category tag therein, and determine the recommendation ratio value corresponding to the category tag therein; use the valid browsing information set and the recommendation ratio value corresponding to the category tag as its associated data, and then obtain the first feature tag associated data set; Among them, the calculation formula for the recommendation ratio corresponding to the category label is: r is the recommendation ratio corresponding to the category label, w is the weight of the category label, and w sum is the sum of the weights of the feature labels in the first feature label set.

8. A method for recommending information content according to claim 7, characterized in that: In step S4, for the user's first feature label associated data set, after preprocessing the information item names in the valid browsing information set corresponding to the category label, the preprocessed information item names are encoded using a text embedding model to obtain a coding vector set corresponding to the category label, and according to the relative weights of the valid browsing information in the corresponding valid browsing information set, the vectors in the obtained coding vector set are summed using a weighted method to determine the feature vector corresponding to the category label; based on this, the user feature vector set can be obtained; Specifically, the calculation formula for the relative weight of the category label corresponding to the valid browsing information set is: where w c is the relative weight of the effective browsing signal, w is the weight of the effective browsing signal, and w sum It is the sum of the weights of all the bids in the valid browsing bid set corresponding to the category label.

9. A method for recommending information content according to claim 7, characterized in that: In step S5, a recommendation strategy for information potentially of interest to the user and a recommendation strategy for information randomly recommended are included; The recommendation strategy for the information that the user is potentially interested in is: For the category labels and their corresponding feature vectors in the user's first feature label set, first, combine the recommendation ratio corresponding to the category label and the total preset recommendation quantity, and take the product of the two as the recommendation quantity corresponding to the category label; Then, based on the category label and the second feature label set, a corresponding Boolean expression is generated as a vector search filter condition; the feature vector corresponding to the category label is used as the search vector, and the number of recommended messages corresponding to the category label is used as the search result limit. A conditional search is performed from the message vector library based on the vector search filter condition to obtain message information with high semantic similarity to the user's browsed messages in the effective browsed message set corresponding to the category label, as the user's potential interested message corresponding to the category label; The random recommendation standard recommendation strategy is: based on the user feature vector set, determine the average value of the feature vector, and generate a random vector consistent with the dimension of the user feature vector based on the random value interval, and add them to obtain the current random feature vector; A value between 0 and 0.1 is randomly generated, and the product of the value and the total preset recommendation number is used as the search result limit number; the current random feature vector is used as the search vector, and the search is performed from the message vector library according to the search result limit number, and the message content that has semantic relevance to the valid browsing message but has a small semantic relevance and changes randomly is obtained as the random recommendation message, thereby increasing the diversity of the recommendation results; Combining the above two recommendation strategies, the summary results are randomly shuffled and used as user-personalized recommendation information data.

10. A device for recommending content of a standard message, characterized in that: include: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 9.

Images