Precision marketing method and system based on large model

Abstract

The present invention discloses a precision marketing method and system based on a large model, which belongs to the fields of data processing technology and marketing technology. It uses user information, historical browsing records and historical search records as the basis, adopts a large model to analyze short videos, and simultaneously performs user similarity analysis, and combines the two analyses to generate final marketing content, thereby realizing multimodal data fusion marketing. It can effectively realize precision marketing even in the absence of historical behavior data and advertising interaction records, and can significantly improve the effect and efficiency of marketing. It has broad application prospects and market value.

Description

Technical Field

[0001] The present invention belongs to the fields of data processing technology and marketing technology, and specifically relates to a precision marketing method and system based on a large model. Background Art

[0002] Short videos are an emerging form of internet content, typically ranging from a few seconds to several minutes in length. Their speed, intuitiveness, and ease of dissemination have rapidly made them an integral part of daily life. Short videos offer a diverse range of content, encompassing entertainment, education, lifestyle, news, and other areas, satisfying users' viewing needs during their fragmented time. Through short video platforms, users can not only watch content but also participate in creation and interaction, forming a powerful social network. The popularity of short videos has also driven the development of related industries, becoming a new frontier for advertising and marketing. With technological advancements and evolving user needs, the short video industry continues to innovate and develop. At the same time, short video advertising not only serves as a primary source of revenue for short video platforms but also provides advertisers with a high-quality delivery channel. Therefore, how to deliver short video ads to interested users and achieve effective marketing results has become a critical issue. Short video platforms can effectively improve advertising performance through recommendation algorithms. However, existing technologies typically rely on personal data and historical behavioral data from users. This leads to a sharp decline in the performance of recommendation algorithms in the absence of historical behavioral data and ad interaction records, making it difficult to achieve targeted marketing. Summary of the Invention

[0003] The present invention provides a large-scale model-based precision marketing method and system to solve the problem that the existing technology cannot effectively implement marketing due to the lack of historical behavior data and advertising interaction records.

[0004] In one aspect, the present invention provides a precision marketing method based on a large model, comprising:

[0005] Obtain user information, historical browsing records, and historical search records; wherein the historical browsing records and historical search records are short videos;

[0006] Determine the similarity between any two users based on the user information and historical browsing records, and determine a target similar user corresponding to each user based on the similarity between the any two users;

[0007] Based on a first historical search record corresponding to a target similar user corresponding to the user, a large model is used to extract a first video feature in the first historical search record, and first marketing content is generated based on the first video feature;

[0008] According to the second historical search record corresponding to the user, extract the second video feature in the second historical search record using the large model, and generate the second marketing content according to the second video feature;

[0009] The first marketing content and the second marketing content are weighted and sorted to obtain final marketing content, and the final marketing content is pushed to the user.

[0010] Furthermore, the similarity between any two users is determined based on the user information and historical browsing records, including:

[0011] According to the user's corresponding historical browsing records, obtain the tags added when the historical browsing records are published, and obtain the user's corresponding browsing tag set;

[0012] The similarity between any two users is determined based on the user information corresponding to the users and the browsing tag set.

[0013] Furthermore, determining the similarity between any two users based on the user information corresponding to the users and the browsing tag set includes:

[0014] Determining, based on the user information corresponding to the user, the gender similarity, age similarity, height similarity, and weight similarity between the user and other users;

[0015] According to the browsing tag set corresponding to the user, the interest similarity between the user and other users is determined as:

[0016]

[0017] in, Indicates the interest similarity between user A and other user B, represents the TD-IDF value corresponding to the hth browsing tag of user A, H represents the total number of browsing tags corresponding to user A, represents the TD-IDF value of other user B regarding the hth browsing tag, and when there are no h browsing tags in the browsing tag set corresponding to other user B, Set to 0;

[0018] The similarity between any two users is determined based on the gender similarity, age similarity, height similarity, weight similarity, and interest similarity.

[0019] Furthermore, determining a target similar user corresponding to each user based on the similarity between any two users includes:

[0020] For any user, determine other users whose similarity with the user is greater than a preset user similarity threshold, and obtain target similar users corresponding to the user.

[0021] Furthermore, based on the first historical search record corresponding to the target similar user corresponding to the user, a large model is used to extract a first video feature in the first historical search record, and first marketing content is generated based on the first video feature, including:

[0022] For any first historical search record corresponding to a target similar user corresponding to the user, extract frames from the short video corresponding to the first historical search record to obtain multiple video frames corresponding to the first historical search record;

[0023] For any video frame corresponding to the first historical search record, the large model is used to describe the video frame to generate a first video frame content text;

[0024] Filtering the first video frame content text corresponding to the first historical search record using a mean filtering method to obtain the filtered first video frame content text;

[0025] For the filtered first video frame content text corresponding to the first historical search record, BERT is used to generate a text feature vector to obtain the first video feature corresponding to the first historical search record;

[0026] Obtaining a third video feature of the video to be marketed, and determining the similarity between the first video feature corresponding to the first historical search record and the third video feature of the video to be marketed, to obtain the similarity between the first historical search record and the video to be marketed;

[0027] The first marketing content is determined based on the similarity between the first historical search record and the video to be marketed.

[0028] Furthermore, the first video frame content text corresponding to the first historical search record is filtered using a mean filtering method to obtain the filtered first video frame content text, including:

[0029] For any first video frame content text corresponding to the first historical search record, obtaining a cosine similarity score between the first video frame content text and other first video frame content texts;

[0030] For any first video frame content text corresponding to the first historical search record, obtain an average of all cosine similarity scores corresponding to the first video frame content text to obtain a mean similarity score corresponding to the first video frame content text;

[0031] For the first video frame content text corresponding to the first historical search record, the first video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered first video frame content text is used as the filtered first video frame content text.

[0032] Furthermore, based on the second historical search record corresponding to the user, a large model is used to extract a second video feature in the second historical search record, and second marketing content is generated based on the second video feature, including:

[0033] According to any second historical search record corresponding to the user, frame extraction processing is performed on the short video corresponding to the second historical search record to obtain multiple video frames corresponding to the second historical search record;

[0034] For any video frame corresponding to the second historical search record, the large model is used to describe the video frame to generate a second video frame content text;

[0035] Filtering the second video frame content text corresponding to the second historical search record using a mean filtering method to obtain the filtered second video frame content text;

[0036] For the filtered second video frame content text corresponding to the second historical search record, BERT is used to generate a text feature vector to obtain the second video feature corresponding to the second historical search record;

[0037] Obtaining a third video feature of the video to be marketed, and determining the similarity between the second video feature corresponding to the second historical search record and the third video feature of the video to be marketed, to obtain the similarity between the second historical search record and the video to be marketed;

[0038] The second marketing content is determined based on the similarity between the second historical search record and the video to be marketed.

[0039] Furthermore, the second video frame content text corresponding to the second historical search record is filtered using a mean filtering method to obtain the filtered second video frame content text, including:

[0040] For any second video frame content text corresponding to the second historical search record, obtaining a cosine similarity score between the second video frame content text and other second video frame content texts;

[0041] For any second video frame content text corresponding to the second historical search record, obtain an average of all cosine similarity scores corresponding to the second video frame content text to obtain a mean similarity score corresponding to the second video frame content text;

[0042] For the second video frame content text corresponding to the second historical search record, the second video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered second video frame content text is used as the filtered second video frame content text.

[0043] Furthermore, weighted sorting is performed on the first marketing content and the second marketing content to obtain final marketing content, and the final marketing content is pushed to the user, including:

[0044] Assigning the same recommendation score to both the first marketing content and the second marketing content, thereby obtaining the recommendation score of the first marketing content and the recommendation score of the second marketing content;

[0045] When the tags between the first marketing content and the second marketing content are the same, the recommendation score of the first marketing content and the recommendation score of the second marketing content are both added with a preset recommendation score increment to obtain the recommendation score of the first marketing content after one update and the recommendation score of the second marketing content after one update;

[0046] For any second marketing content, obtain the similarity between the second marketing content and the first marketing content to obtain the marketing similarity between the second marketing content and the first marketing content;

[0047] For any second marketing content, when the marketing similarity between the second marketing content and the first marketing content is greater than a preset marketing similarity threshold, the corresponding first marketing content is used as the target marketing content, and the target marketing content corresponding to the second marketing content is obtained;

[0048] Transferring the recommendation score of the target marketing content to the corresponding second marketing content based on the recommendation score of the first marketing content after the first update and the recommendation score of the second marketing content after the first update, thereby obtaining the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update; wherein, after the recommendation score of the target marketing content is transferred, the recommendation score of the target marketing content is reset to zero;

[0049] According to the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update, the first marketing content and the second marketing content are arranged in descending order of the recommendation scores to obtain sorted marketing content;

[0050] Based on the marketing content after the sorting, the first K marketing contents are taken as the final marketing content, and the final marketing content is pushed to the user.

[0051] On the other hand, the present invention provides a precision marketing system based on a large model, comprising: an information acquisition module, a similar user acquisition module, a first marketing module, a second marketing module, and a comprehensive marketing module;

[0052] The information acquisition module is used to obtain user information, historical browsing records, and historical search records; wherein the historical browsing records are the N latest short videos whose viewing progress of the user exceeds a preset viewing threshold; the historical search records are the M latest short videos whose viewing progress of the user exceeds a preset viewing threshold;

[0053] The similar user acquisition module is used to determine the similarity between any two users based on the user information and historical browsing records, and determine the target similar user corresponding to each user based on the similarity between the any two users;

[0054] The first marketing module is configured to extract first video features from the first historical search records corresponding to target similar users corresponding to the user using a large model, and generate first marketing content based on the first video features;

[0055] The second marketing module is configured to extract second video features from the second historical search records corresponding to the user using the large model, and generate second marketing content based on the second video features;

[0056] The integrated marketing module is used to perform weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and push the final marketing content to the user.

[0057] The present invention provides a large-model-based precision marketing method and system, which uses a large model to analyze short videos based on user information, historical browsing records, and historical search records, and simultaneously performs user similarity analysis. The two analyses are combined to generate the final marketing content, thereby realizing multimodal data fusion marketing. It can effectively realize precision marketing even in the absence of historical behavioral data and advertising interaction records, and can significantly improve the effectiveness and efficiency of marketing. It has broad application prospects and market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0059] Figure 1 A flowchart of a large-model-based precision marketing method provided in an embodiment of the present invention.

[0060] Figure 2A schematic diagram of the structure of a large-scale model-based precision marketing system provided in an embodiment of the present invention.

[0061] In the accompanying drawings, 201 is an information acquisition module, 202 is a similar user acquisition module, 203 is a first marketing module, 204 is a second marketing module, and 205 is an integrated marketing module.

[0062] The above drawings have shown specific embodiments of the present invention, which will be described in more detail below. These drawings and the text description are not intended to limit the scope of the present invention in any way, but rather to illustrate the concept of the present invention to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0063] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0064] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0065] like Figure 1 As shown, an embodiment of the present invention provides a precision marketing method based on a large model, including:

[0066] S101. Obtain user information, historical browsing records, and historical search records; wherein the historical browsing records and historical search records are all short videos;

[0067] The historical browsing records are short videos whose viewing progress exceeds the preset viewing progress threshold, and the number of historical browsing records collected is set to N, and the collection is based on the current moment (i.e., the marketing moment); the historical search records are short videos that are searched and whose viewing progress exceeds the preset viewing progress threshold, and the number of historical search records collected is set to M, and the collection is based on the current moment (i.e., the marketing moment).

[0068] In recent years, short video ads have become a major source of commercial revenue for short video platforms and have also provided advertisers with a high-quality distribution channel. Therefore, how to push short video ads to interested users and achieve good marketing results has become a key issue. Therefore, the embodiments of the present invention primarily target short video platforms for marketing, and both historical browsing and search records should be short videos.

[0069] At the same time, historical browsing and search records should reflect the content that users are most interested in. Therefore, a viewing threshold should be set. Only when the user's viewing progress exceeds this viewing threshold will participation be calculated. The latest short videos refer to the short videos that users have watched at historical moments based on marketing moments.

[0070] S102: determining the similarity between any two users based on the user information and historical browsing records, and determining target similar users corresponding to each user based on the similarity between the any two users;

[0071] Through user information and historical browsing records, it is possible to determine whether the behaviors of two users are similar. If they are similar, the target similar users corresponding to the users can be determined, and marketing can be carried out based on the target similar users.

[0072] S103: extracting first video features from a first historical search record corresponding to a target similar user corresponding to the user using a large model, and generating first marketing content based on the first video features;

[0073] The first marketing content is generated based on user similarity, and user similarity is mainly obtained by relying on user information and historical browsing records. Therefore, compared with traditional similarity algorithms, the embodiment of the present invention can achieve precision marketing even when the user does not have historical behavior data.

[0074] S104: extracting second video features from the second historical search record using the large model based on the user's corresponding second historical search record, and generating second marketing content based on the second video features;

[0075] The second historical search record indicates the content that the user is more interested in at a historical moment. This part of the data is processed and identified through a large model to generate the second marketing content, so that even if the user does not purchase or search for products, it can further achieve precision marketing.

[0076] S105: Perform weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and push the final marketing content to the user.

[0077] The first marketing content and the second marketing content are both generated based on user data. By weighted sorting of the first marketing content and the second marketing content, more advertising videos that users are interested in can be marketed, effectively improving marketing accuracy.

[0078] The present invention provides a large-model-based precision marketing method and system, which uses a large model to analyze short videos based on user information, historical browsing records, and historical search records, and simultaneously performs user similarity analysis. The two analyses are combined to generate the final marketing content, thereby realizing multimodal data fusion marketing. It can effectively realize precision marketing even in the absence of historical behavioral data and advertising interaction records, and can significantly improve the effectiveness and efficiency of marketing. It has broad application prospects and market value.

[0079] In an embodiment of the present invention, determining the similarity between any two users based on the user information and historical browsing records includes:

[0080] According to the user's corresponding historical browsing records, obtain the tags added when the historical browsing records are published, and obtain the user's corresponding browsing tag set;

[0081] It is worth noting that the browsing tags corresponding to historical browsing records can be the same. For any browsing tag, the more historical browsing records corresponding to the browsing tag, the more it indicates that the user is more inclined to browse the short video style corresponding to the tag, so marketing can be carried out based on this.

[0082] The similarity between any two users is determined based on the user information corresponding to the users and the browsing tag set.

[0083] In an embodiment of the present invention, determining the similarity between any two users based on the user information corresponding to the users and the browsing tag set includes:

[0084] Determining, based on the user information corresponding to the user, the gender similarity, age similarity, height similarity, and weight similarity between the user and other users;

[0085] For example, the gender similarity can be: if the two users have the same gender, the gender similarity is considered to be 100%, otherwise it is 0%.

[0086] Age similarity can be calculated by setting an age difference threshold. When the age difference between two users is greater than or equal to this threshold, the age similarity between the two users is considered 0%. Similarly, when the age difference is zero, the age similarity between the two users is considered 100%. When the age difference varies within this threshold, the percentage of change is used as the age similarity. For example, if the age difference threshold is set to 10, the age similarity between the two users is 0% when the age difference is 10; if the age difference is 9, the age similarity between the two users has changed by 10%, so the age similarity between the two users is 10%.

[0087] Height similarity can be: A height difference threshold can be set in advance. When the height difference between two users is greater than or equal to this height difference threshold, the height similarity between the two users is considered 0%. Similarly, when the height difference is zero, the height similarity between the two users is considered 100%. When the height difference changes within this threshold, the percentage of change is used as the height similarity.

[0088] Weight similarity can be determined by setting a weight difference threshold. When the weight difference between two users is greater than or equal to this threshold, the weight similarity between the two users is considered 0%. Similarly, when the weight difference is zero, the weight similarity between the two users is considered 100%. When the weight difference changes within this threshold, the percentage of change is used as the weight similarity.

[0089] According to the browsing tag set corresponding to the user, the interest similarity between the user and other users is determined as:

[0090]

[0091] in, Indicates the interest similarity between user A and other user B, represents the TD-IDF value corresponding to the hth browsing tag of user A, H represents the total number of browsing tags corresponding to user A, represents the TD-IDF value of other user B regarding the hth browsing tag, and when there are no h browsing tags in the browsing tag set corresponding to other user B, Set to 0;

[0092] The TF-IDF value is obtained by multiplying the term frequency (TF) and the inverse document frequency (IDF) and is used to evaluate the importance of a word to a document set or a corpus.

[0093] The similarity between any two users is determined based on the gender similarity, age similarity, height similarity, weight similarity, and interest similarity.

[0094] Optionally, the gender similarity, age similarity, height similarity, weight similarity, and interest similarity can be directly added together to obtain the similarity between the two users. Alternatively, different weight coefficients can be set for the gender similarity, age similarity, height similarity, weight similarity, and interest similarity, respectively, and the gender similarity, age similarity, height similarity, weight similarity, and interest similarity are multiplied by their respective weight coefficients and summed to obtain the similarity between the two users. These two similarity acquisition methods can be selected based on actual needs. In the embodiment of the present invention, the second method is preferably used to obtain user similarity.

[0095] In the embodiment of the present invention, determining a target similar user corresponding to each user according to the similarity between any two users includes:

[0096] For any user, determine other users whose similarity with the user is greater than a preset user similarity threshold, and obtain target similar users corresponding to the user.

[0097] Optionally, for any user, L other users with the greatest similarity may be selected as target similar users.

[0098] In an embodiment of the present invention, based on a first historical search record corresponding to a target similar user corresponding to the user, a large model is used to extract a first video feature in the first historical search record, and first marketing content is generated based on the first video feature, including:

[0099] For any first historical search record corresponding to a target similar user corresponding to the user, extract frames from the short video corresponding to the first historical search record to obtain multiple video frames corresponding to the first historical search record;

[0100] For example, based on a preset data collection frequency, video frames of the short video corresponding to the first historical search record may be collected, thereby obtaining multiple video frames corresponding to the first historical search record.

[0101] For any video frame corresponding to the first historical search record, the large model is used to describe the video frame to generate a first video frame content text;

[0102] Optionally, the large model in the embodiment of the present invention can be set to LLaVA (Large Language and Vision Assistant) or LLM (Large Language Model), with LLaVA being preferred in the embodiment of the present invention. However, it is worth noting that the aforementioned large model is merely an example of the embodiment of the present invention. While ensuring equivalent or similar functionality, other large models can also be used to describe video frames, leveraging the large model's ability to analyze and understand image scene content, and enriching the understanding of image content through textual content, thereby improving the model's overall understanding of video content and ultimately achieving marketing.

[0103] Filtering the first video frame content text corresponding to the first historical search record using a mean filtering method to obtain the filtered first video frame content text;

[0104] In the process of identifying the content text of the first video frame, there may be a problem that the content text of the first video frame does not match the scene information of the video frame. Therefore, an embodiment of the present invention adopts a mean filtering method to filter out the content text of the first video frame that does not match the scene information of the video frame, thereby improving marketing accuracy.

[0105] For the filtered first video frame content text corresponding to the first historical search record, BERT (pre-trained language model) is used to generate a text feature vector to obtain the first video feature corresponding to the first historical search record;

[0106] Optionally, all text feature vectors of the first historical search record may be concatenated in chronological order to determine the first video feature corresponding to the first historical search record.

[0107] Obtaining a third video feature of the video to be marketed, and determining the similarity between the first video feature corresponding to the first historical search record and the third video feature of the video to be marketed, to obtain the similarity between the first historical search record and the video to be marketed;

[0108] The method for obtaining the third video feature is the same as the method for obtaining the first video feature and the second video feature, and will not be repeated here. The cosine similarity between the first video feature and the third video feature can be determined to obtain the similarity between the first historical search record and the video to be marketed.

[0109] The first marketing content is determined based on the similarity between the first historical search record and the video to be marketed.

[0110] For example, based on the similarity between the first historical search record and the video to be marketed, Q videos to be marketed with the greatest similarity can be determined to obtain Q first marketing contents.

[0111] In an embodiment of the present invention, the first video frame content text corresponding to the first historical search record is filtered using a mean filtering method to obtain the filtered first video frame content text, including:

[0112] For any first video frame content text corresponding to the first historical search record, obtaining a cosine similarity score between the first video frame content text and other first video frame content texts;

[0113] For any first video frame content text corresponding to the first historical search record, obtain an average of all cosine similarity scores corresponding to the first video frame content text to obtain a mean similarity score corresponding to the first video frame content text;

[0114] For the first video frame content text corresponding to the first historical search record, the first video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered first video frame content text is used as the filtered first video frame content text.

[0115] The embodiment of the present invention adopts a mean filtering method to filter out the first video frame content text that is inconsistent with the video frame scene information, thereby improving marketing accuracy.

[0116] In an embodiment of the present invention, based on a second historical search record corresponding to a user, extracting a second video feature in the second historical search record using a large model, and generating second marketing content based on the second video feature include:

[0117] According to any second historical search record corresponding to the user, frame extraction processing is performed on the short video corresponding to the second historical search record to obtain multiple video frames corresponding to the second historical search record;

[0118] For any video frame corresponding to the second historical search record, the large model is used to describe the video frame to generate a second video frame content text;

[0119] Filtering the second video frame content text corresponding to the second historical search record using a mean filtering method to obtain the filtered second video frame content text;

[0120] For the filtered second video frame content text corresponding to the second historical search record, BERT is used to generate a text feature vector to obtain the second video feature corresponding to the second historical search record;

[0121] Obtaining a third video feature of the video to be marketed, and determining the similarity between the second video feature corresponding to the second historical search record and the third video feature of the video to be marketed, to obtain the similarity between the second historical search record and the video to be marketed;

[0122] The second marketing content is determined based on the similarity between the second historical search record and the video to be marketed.

[0123] The method for obtaining the second marketing content is similar to the method for obtaining the first marketing content, and will not be further described in this embodiment of the present invention.

[0124] In an embodiment of the present invention, the second video frame content text corresponding to the second historical search record is filtered using a mean filtering method to obtain the filtered second video frame content text, including:

[0125] For any second video frame content text corresponding to the second historical search record, obtaining a cosine similarity score between the second video frame content text and other second video frame content texts;

[0126] For any second video frame content text corresponding to the second historical search record, obtain an average of all cosine similarity scores corresponding to the second video frame content text to obtain a mean similarity score corresponding to the second video frame content text;

[0127] For the second video frame content text corresponding to the second historical search record, the second video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered second video frame content text is used as the filtered second video frame content text.

[0128] In an embodiment of the present invention, weighted sorting is performed on the first marketing content and the second marketing content to obtain final marketing content, and the final marketing content is pushed to the user, including:

[0129] Assigning the same recommendation score to both the first marketing content and the second marketing content, thereby obtaining the recommendation score of the first marketing content and the recommendation score of the second marketing content;

[0130] When the tags between the first marketing content and the second marketing content are the same, the recommendation score of the first marketing content and the recommendation score of the second marketing content are both added with a preset recommendation score increment to obtain the recommendation score of the first marketing content after one update and the recommendation score of the second marketing content after one update;

[0131] The first marketing content is the short video content that similar target users are interested in, and the second marketing content is the short video content that the users themselves are interested in. Therefore, when the two contents are similar or close, they are more worthy of recommendation.

[0132] It is worth noting that when the same tag corresponds to multiple first marketing contents and second marketing contents, the recommendation score is added with a preset recommendation score increment.

[0133] For any second marketing content, obtain the similarity between the second marketing content and the first marketing content to obtain the marketing similarity between the second marketing content and the first marketing content;

[0134] Optionally, the marketing similarity between the second marketing content and the first marketing content can be determined by using a large model-based video feature extraction method to determine the video features, and then determining the marketing similarity based on the video features of the second marketing content and the video features of the first marketing content, that is, the cosine similarity between the video features.

[0135] For any second marketing content, when the marketing similarity between the second marketing content and the first marketing content is greater than a preset marketing similarity threshold, the corresponding first marketing content is used as the target marketing content, and the target marketing content corresponding to the second marketing content is obtained;

[0136] Transferring the recommendation score of the target marketing content to the corresponding second marketing content based on the recommendation score of the first marketing content after the first update and the recommendation score of the second marketing content after the first update, thereby obtaining the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update; wherein, after the recommendation score of the target marketing content is transferred, the recommendation score of the target marketing content is reset to zero;

[0137] Compared with the first marketing content, the second marketing content can better bring out content that users are interested in. Therefore, when the first marketing content is similar to the second marketing content, the recommendation score of the second marketing content can be improved, thereby improving the marketing accuracy.

[0138] According to the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update, the first marketing content and the second marketing content are arranged in descending order of the recommendation scores to obtain sorted marketing content;

[0139] Based on the marketing content after the sorting, the first K marketing contents are taken as the final marketing content, and the final marketing content is pushed to the user.

[0140] It is worth noting that, in the process of pushing the final marketing content to the user, the content can be pushed in sequence, and the pushing method is defined by the staff in actual application.

[0141] like Figure 2 As shown, the embodiment of the present invention provides a precision marketing system based on a large model, including: an information acquisition module 201, a similar user acquisition module 202, a first marketing module 203, a second marketing module 204 and a comprehensive marketing module 205;

[0142] The information acquisition module 201 is used to obtain user information, historical browsing records, and historical search records; wherein the historical browsing records are the N latest short videos whose viewing progress of the user exceeds a preset viewing threshold; the historical search records are the M latest short videos whose viewing progress of the user exceeds a preset viewing threshold;

[0143] The similar user acquisition module 202 is used to determine the similarity between any two users based on the user information and historical browsing records, and determine the target similar user corresponding to each user based on the similarity between the any two users;

[0144] The first marketing module 203 is configured to extract first video features from the first historical search records corresponding to target similar users corresponding to the user using a large model, and generate first marketing content based on the first video features;

[0145] The second marketing module 204 is configured to extract second video features from the second historical search records corresponding to the user using a large model, and generate second marketing content based on the second video features;

[0146] The integrated marketing module 205 is configured to perform weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and push the final marketing content to the user.

[0147] The embodiment of the present invention provides a large model-based precision marketing system that can execute the above-mentioned precision marketing method. Its principles and beneficial effects are similar and will not be repeated here.

[0148] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art not disclosed herein. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and variations can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A precision marketing method based on a large model, characterized by: include: Obtain user information, historical browsing records, and historical search records; wherein the historical browsing records and historical search records are short videos; Determine the similarity between any two users based on the user information and historical browsing records, and determine a target similar user corresponding to each user based on the similarity between the any two users; Based on a first historical search record corresponding to a target similar user corresponding to the user, a large model is used to extract a first video feature in the first historical search record, and first marketing content is generated based on the first video feature; According to the second historical search record corresponding to the user, extract the second video feature in the second historical search record using the large model, and generate the second marketing content according to the second video feature; performing weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and pushing the final marketing content to the user; Determine the similarity between any two users based on the user information and historical browsing records, including: According to the user's corresponding historical browsing records, obtain the tags added when the historical browsing records are published, and obtain the user's corresponding browsing tag set; Determining the similarity between any two users based on the user information corresponding to the users and the browsing tag set; Determining the similarity between any two users based on the user information corresponding to the users and the browsing tag set includes: Determining, based on the user information corresponding to the user, the gender similarity, age similarity, height similarity, and weight similarity between the user and other users; According to the browsing tag set corresponding to the user, the interest similarity between the user and other users is determined as: in, Indicates the interest similarity between user A and other user B, represents the TD-IDF value corresponding to the hth browsing tag of user A, H represents the total number of browsing tags corresponding to user A, represents the TD-IDF value of other user B regarding the h-th browsing tag, and when the h-th browsing tag does not exist in the browsing tag set corresponding to other user B, Set to 0; Determine the similarity between any two users based on the gender similarity, age similarity, height similarity, weight similarity, and interest similarity; Based on a first historical search record corresponding to a target similar user corresponding to the user, a large model is used to extract a first video feature in the first historical search record, and first marketing content is generated based on the first video feature, including: For any first historical search record corresponding to a target similar user corresponding to the user, extract frames from the short video corresponding to the first historical search record to obtain multiple video frames corresponding to the first historical search record; For any video frame corresponding to the first historical search record, the large model is used to describe the video frame to generate a first video frame content text; Filtering the first video frame content text corresponding to the first historical search record using a mean filtering method to obtain the filtered first video frame content text; For the filtered first video frame content text corresponding to the first historical search record, BERT is used to generate a text feature vector to obtain the first video feature corresponding to the first historical search record; Obtaining a third video feature of the video to be marketed, and determining the similarity between the first video feature corresponding to the first historical search record and the third video feature of the video to be marketed, to obtain the similarity between the first historical search record and the video to be marketed; The first marketing content is determined based on the similarity between the first historical search record and the video to be marketed.

2. The precision marketing method based on a large model according to claim 1 is characterized in that: Determining target similar users corresponding to each user based on the similarity between any two users includes: For any user, determine other users whose similarity with the user is greater than a preset user similarity threshold, and obtain target similar users corresponding to the user.

3. The precision marketing method based on a large model according to claim 1, characterized in that: The first video frame content text corresponding to the first historical search record is filtered using a mean filtering method to obtain the filtered first video frame content text, including: For any first video frame content text corresponding to the first historical search record, obtaining a cosine similarity score between the first video frame content text and other first video frame content texts; For any first video frame content text corresponding to the first historical search record, obtain an average of all cosine similarity scores corresponding to the first video frame content text to obtain a mean similarity score corresponding to the first video frame content text; For the first video frame content text corresponding to the first historical search record, the first video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered first video frame content text is used as the filtered first video frame content text.

4. The large model-based precision marketing method according to claim 3, characterized in that: According to the second historical search record corresponding to the user, a large model is used to extract a second video feature in the second historical search record, and second marketing content is generated according to the second video feature, including: According to any second historical search record corresponding to the user, frame extraction processing is performed on the short video corresponding to the second historical search record to obtain multiple video frames corresponding to the second historical search record; For any video frame corresponding to the second historical search record, the large model is used to describe the video frame to generate a second video frame content text; Filtering the second video frame content text corresponding to the second historical search record using a mean filtering method to obtain the filtered second video frame content text; For the filtered second video frame content text corresponding to the second historical search record, BERT is used to generate a text feature vector to obtain the second video feature corresponding to the second historical search record; Obtaining a third video feature of the video to be marketed, and determining the similarity between the second video feature corresponding to the second historical search record and the third video feature of the video to be marketed, to obtain the similarity between the second historical search record and the video to be marketed; The second marketing content is determined based on the similarity between the second historical search record and the video to be marketed.

5. The large-scale model-based precision marketing method according to claim 4, characterized in that: The second video frame content text corresponding to the second historical search record is filtered using a mean filtering method to obtain the filtered second video frame content text, including: For any second video frame content text corresponding to the second historical search record, obtaining a cosine similarity score between the second video frame content text and other second video frame content texts; For any second video frame content text corresponding to the second historical search record, obtain an average of all cosine similarity scores corresponding to the second video frame content text to obtain a mean similarity score corresponding to the second video frame content text; For the second video frame content text corresponding to the second historical search record, the second video frame content text whose similarity score average is less than the preset text similarity threshold is filtered out, and the unfiltered second video frame content text is used as the filtered second video frame content text.

6. The large model-based precision marketing method according to claim 4, characterized in that: Performing weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and pushing the final marketing content to the user, including: Assigning the same recommendation score to both the first marketing content and the second marketing content, thereby obtaining the recommendation score of the first marketing content and the recommendation score of the second marketing content; When the tags between the first marketing content and the second marketing content are the same, the recommendation score of the first marketing content and the recommendation score of the second marketing content are both added with a preset recommendation score increment to obtain the recommendation score of the first marketing content after one update and the recommendation score of the second marketing content after one update; For any second marketing content, obtain the similarity between the second marketing content and the first marketing content to obtain the marketing similarity between the second marketing content and the first marketing content; For any second marketing content, when the marketing similarity between the second marketing content and the first marketing content is greater than a preset marketing similarity threshold, the corresponding first marketing content is used as the target marketing content, and the target marketing content corresponding to the second marketing content is obtained; Transferring the recommendation score of the target marketing content to the corresponding second marketing content based on the recommendation score of the first marketing content after the first update and the recommendation score of the second marketing content after the first update, thereby obtaining the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update; wherein, after the recommendation score of the target marketing content is transferred, the recommendation score of the target marketing content is reset to zero; According to the recommendation score of the first marketing content after the second update and the recommendation score of the second marketing content after the second update, the first marketing content and the second marketing content are arranged in descending order of the recommendation scores to obtain sorted marketing content; Based on the marketing content after the sorting, the first K marketing contents are taken as the final marketing content, and the final marketing content is pushed to the user.

7. A precision marketing system based on a large model, which is capable of executing the precision marketing method based on a large model as described in any one of claims 1 to 6, characterized in that: include: Information acquisition module, similar user acquisition module, first marketing module, second marketing module and comprehensive marketing module; The information acquisition module is used to obtain user information, historical browsing records, and historical search records; wherein the historical browsing records are the N latest short videos whose viewing progress of the user exceeds a preset viewing threshold; the historical search records are the M latest short videos whose viewing progress of the user exceeds a preset viewing threshold; The similar user acquisition module is used to determine the similarity between any two users based on the user information and historical browsing records, and determine the target similar user corresponding to each user based on the similarity between the any two users; The first marketing module is configured to extract first video features from the first historical search records corresponding to target similar users corresponding to the user using a large model, and generate first marketing content based on the first video features; The second marketing module is configured to extract second video features from the second historical search records corresponding to the user using the large model, and generate second marketing content based on the second video features; The integrated marketing module is used to perform weighted sorting on the first marketing content and the second marketing content to obtain final marketing content, and push the final marketing content to the user.

Images