CN-122027871-A - User portrait-based new media content personalized generation method and system

Abstract

The embodiment of the invention provides a user portrait-based new media content personalized generation method and system, which belong to the technical field of new media content generation, wherein a shadow anchor portrait is constructed to convert cross-platform user portraits from single-point determination estimation to multi-anchor confidence set estimation, and confidence constraint generation is introduced in a generation stage, so that personalized contents which are consistent in cross-platform theme and respectively meet different platform context constraints can be output under the scenes of inconsistent cross-platform behaviors and higher portrait uncertainty; meanwhile, the collaborative process of generation, verification, reflow update and re-planning/regeneration is realized through the cross-platform consistency verification and trigger type correction mechanism, so that the cross-platform content operation efficiency and the user experience are improved. Therefore, the method solves the problem that when identity drift and uncertainty exist in cross-platform user portraits, the traditional method is difficult to stably generate personalized content which is consistent with each platform and accords with the context of each platform.

Inventors

• HUANG LILING
• Ding Yangxue

Assignees

• 成都职业技术学院

Dates

Publication Date

20260512

Application Date

20260416

Claims (10)

1. 1. A method for personalized generation of new media content based on user portraits, the method comprising: Acquiring a cross-platform behavior event sequence of the same target user on a plurality of new media platforms, and executing time window aggregation and vectorization fusion to form a cross-platform behavior fingerprint sequence; Constructing a shadow anchor portrait based on the cross-platform behavior fingerprint sequence, and deriving a unified portrait representation and a portrait uncertainty representation from the shadow anchor portrait; generating a platform preference representation of each platform and calculating a platform adaptation strength based on the uniform representation, the representation uncertainty representation and a platform context descriptor of each platform to obtain an adapted representation for each platform content generation; Generating a content planning result of each platform based on the content atomic map, the adaptive representation and the platform constraint set; And executing confidence constraint generation based on the content planning result, the shadow anchor portrait and the platform constraint set to generate personalized new media content of each platform, and executing cross-platform consistency check and trigger type correction on the personalized new media content of each platform and outputting.
2. 2. The method for personalized generation of new media content based on user portraits of claim 1, wherein obtaining a cross-platform behavior event sequence of the same target user on a plurality of new media platforms, performing time window aggregation and vectorization fusion to form a cross-platform behavior fingerprint sequence, comprises: acquiring a cross-platform behavior event sequence of the same target user on a plurality of new media platforms, representing the cross-platform behavior event sequence as a time-ordered event set, and recording behavior event association information for each event; The behavior event association information comprises a platform identifier, a behavior action type, a content atom reference identifier, an event time stamp and an event intensity index; Dividing the event set according to preset sliding time windows, and summarizing local behavior statistical vectors corresponding to the platforms in each time window respectively; and performing gate-controlled weighted aggregation on the local behavior statistical vectors of each platform to obtain window-level behavior fingerprint vectors of corresponding time windows, and sequentially forming the window-level behavior fingerprint vectors of each time window into the cross-platform behavior fingerprint sequence.
3. 3. The user portrayal-based new media content personalization generation method of claim 1, wherein constructing a shadow anchor portrayal based on the cross-platform behavioral fingerprint sequence and deriving a unified portrayal representation and a portrayal uncertainty representation from the shadow anchor portrayal comprises: Generating a plurality of candidate portrait anchors based on the cross-platform behavior fingerprint sequence, and establishing corresponding portrait center characterization, portrait uncertainty characterization and confidence weights for each candidate portrait anchor to form the shadow anchor portrait; Calculating the matching likelihood of each time window behavior fingerprint in the cross-platform behavior fingerprint sequence relative to each candidate portrait anchor point, and updating the confidence weight according to the matching likelihood; And based on the updated confidence weights, performing weighted fusion on the representation center characterization and the representation uncertainty characterization of each candidate representation anchor point to output uniform representation characterization and representation uncertainty characterization.
4. 4. The method for personalized creation of new media content based on user portraits of claim 3, further comprising the steps of constructing a cross-platform soft alignment coupling matrix and performing a consistency weighted update of the confidence weights based on the cross-platform soft alignment coupling matrix, the steps comprising: in each time window, extracting contribution characterization of each platform to window level behavior fingerprint vectors respectively, and constructing a matching cost matrix based on the distance between different platform contribution characterization; Performing entropy regular optimal transmission solution on the matching cost matrix to obtain the cross-platform soft alignment coupling matrix; and calculating a coupling consistency factor based on the cross-platform soft alignment coupling matrix, and weighting the matching likelihood by using the coupling consistency factor to update the confidence weight.
5. 5. The method for personalized generation of new media content based on user portraits of claim 1, Wherein generating a platform preference representation for each platform and computing a platform adaptation strength based on the uniform representation, the representation uncertainty representation, and a platform context descriptor for each platform to obtain an adapted representation for each platform content generation comprises: Building a platform context descriptor for each platform and mapping the unifying portrait representations to platform preference representations of the platform based on the platform context descriptors; and calculating the platform adaptation strength of each platform based on the portrait uncertainty characterization, and fusing the platform preference characterization and the unified portrait characterization by using the platform adaptation strength to obtain the adaptation portrait characterization of the platform.
6. 6. The user portrayal-based new media content personalization generation method of claim 1, wherein generating a content planning result for each platform based on a content atomic map, the adapted portrayal representation, and a set of platform constraints comprises: acquiring a content atomic map, configuring semantic characterization for each content atom in the content atomic map, and configuring transfer feasibility characterization for the joinable relationship among the content atoms; Generating a platform constraint set based on the platform context descriptor, and constraining the platform constraint set to the structural length, the presentation style and the interactive structure of the content planning; Under the constraint of the platform constraint set, performing path search on the content atomic map based on the adaptive representation to obtain a content atomic sequence of each platform as a content planning result.
7. 7. The user portrayal-based new media content personalization generating method of claim 1, wherein performing a confidence constraint generation to generate each platform personalized new media content based on the content planning result, the shadow anchor portrayal, and the platform constraint set comprises: Generating a structured generation prompt based on a platform content planning result, and taking the structured generation prompt and a platform constraint set together as a generation input; Respectively calculating the matching scores of the candidate content fragments obtained in the generation process and each candidate portrait anchor point in the shadow anchor portrait, and aggregating the matching scores based on the confidence weights to obtain the shadow anchor point consistency scores of the candidate content fragments; and in the process of generating and decoding, re-weighting the candidate output probabilities based on the shadow anchor consistency scores, and simultaneously, performing constraint penalty on the candidate outputs based on the platform constraint set to output personalized new media contents of each platform.
8. 8. The user portrait based new media content personalization generation method of claim 1 wherein performing cross-platform consistency check and triggered correction on each platform personalized new media content comprises: Respectively calculating content semantic characterizations of the personalized new media content of each platform, and calculating cross-platform theme consistency loss based on the content semantic characterizations of each platform; Calculating a shadow anchor point consistency statistic value of a generating process for each platform personalized new media content respectively, and calculating cross-platform anchor point consistency loss based on the shadow anchor point consistency statistic value; And carrying out weighted fusion on the cross-platform theme consistency loss and the cross-platform anchor point consistency loss to obtain the cross-platform consistency loss, and executing reflow updating on the confidence weight of the shadow anchor point portrait based on the personalized new media content of each platform when the cross-platform consistency loss exceeds a threshold value, and synchronously adjusting portrait uncertainty characterization.
9. 9. The personalized new media content generation method according to claim 1, wherein the triggered modification further comprises the steps of triggered re-planning and re-generation, and outputting the personalized new media content of each platform when the termination condition is satisfied: When the cross-platform consistency loss exceeds a preset threshold, recalculating the unified portrait characterization and the portrait uncertainty characterization based on the shadow anchor portrait after reflow updating, and recalculating the adaptive portrait characterization of each platform based on the updated unified portrait characterization and portrait uncertainty characterization; Regenerating a content planning result based on the updated adaptive representation, the content atomic map and the platform constraint set, and executing confidence constraint generation based on the updated shadow anchor representation to obtain regenerated personalized new media content of each platform; And executing cross-platform consistency check again on the regenerated personalized new media content of each platform, and outputting the final personalized new media content of each platform when the cross-platform consistency loss does not exceed a preset threshold or the regeneration times reach a preset upper limit.
10. 10. A user portrayal based new media content personalization generating system for performing the user portrayal based new media content personalization generating method according to any one of claims 1-9, the system comprising: the acquisition unit is used for acquiring a cross-platform behavior event sequence of the same target user on a plurality of new media platforms, and performing time window aggregation and vectorization fusion to form a cross-platform behavior fingerprint sequence; the construction unit is used for constructing a shadow anchor portrait based on the cross-platform behavior fingerprint sequence and deriving a unified portrait representation and a portrait uncertainty representation from the shadow anchor portrait; a computing unit for generating a platform preference representation for each platform and computing a platform adaptation strength based on the uniform representation, the representation uncertainty representation, and a platform context descriptor for each platform to obtain an adapted representation for each platform content generation; the generation unit is used for generating a content planning result of each platform based on the content atomic map, the adaptive portrait representation and the platform constraint set; And the output unit is used for executing confidence constraint generation to generate personalized new media contents of each platform based on the content planning result, the shadow anchor point portrait and the platform constraint set, executing cross-platform consistency check and triggering correction on the personalized new media contents of each platform, and outputting.

Description

User portrait-based new media content personalized generation method and system Technical Field The invention relates to the technical field of new media content generation, in particular to a user portrait-based new media content personalized generation method and system. Background With the continuous evolution of new media platform forms such as short videos, graphics and texts, live broadcast and the like, the same user often presents differentiated content consumption and interaction behaviors on different platforms. For the content operation and intelligent delivery scenes, if user portraits consistent with each other across platforms cannot be formed, the problems of platform splitting, inaccurate touch, low content multiplexing efficiency and the like are easy to occur. Particularly in practical application, the content propagation rules, recommendation distribution mechanisms, interaction inlets and expression structures of different platforms are obviously different, and the presentation effect and user acceptance of the same content on different platforms can be obviously offset, so that personalized generation strategies based on single-platform portraits or simple combined portraits are difficult to stably work. In the prior art, one type of scheme typically collects user behavior and trains portraits in a single platform, and then generates content based on the portraits. The scheme has the defect that when the behavior mode of a user on another platform is obviously inconsistent with that of the current platform or when the situation of sparse behavior, cold start, short-term interest mutation and the like exists on a certain platform, the single-platform portraits are difficult to provide reliable basis for cross-platform content generation. Another class of schemes attempts to splice or regularize multi-platform data, such as by account mapping, label merging, or simple weighting, to form a unifying portrait. However, in a cross-platform scenario, the user behavior often has identity drift and context drift, namely, on one hand, the behavior density, interaction strength and consumption link of the same user on different platforms are large, so that the combined portraits are easily led by a high-noise platform or a high-frequency platform, and on the other hand, the constraints of different platforms on the content structure, the length and the interaction mode are different, so that even if the portraits are consistent, the content generation may not conform to the platform context, and the effect is poor. Furthermore, some existing methods rely on a general generation model only to carry out shallow constraints such as length or sensitive words on output in the generation stage, and lack a generation constraint mechanism taking cross-platform portrait stability as a core. When uncertainty exists in the cross-platform image, the generated model is easy to output contents with semantic conflict or inconsistent interest directions on different platforms, so that the problems of inconsistent cross-platform operation caliber, user perception cracking, even negative feedback accumulation and the like are caused. Therefore, how to stably generate personalized content which can be executed by a platform and is consistent across platforms in a special scene of unstable cross-platform portrait and remarkable platform context difference becomes a technical problem to be solved. Disclosure of Invention The invention aims to provide a new media content personalized generation method based on user portraits, which at least solves the problem that the traditional method is difficult to stably generate personalized content which is consistent across platforms and accords with the context of each platform when identity drift and uncertainty exist in the cross-platform user portraits. In order to achieve the aim, the first aspect of the invention provides a user portrait-based new media content personalized generation method which is characterized by comprising the steps of obtaining a cross-platform behavior event sequence of the same target user on a plurality of new media platforms, executing time window aggregation and vectorization fusion to form a cross-platform behavior fingerprint sequence, constructing a shadow anchor portrait based on the cross-platform behavior fingerprint sequence, deriving a unified portrait representation and a portrait uncertainty representation from the shadow anchor portrait, generating platform preference representation of each platform and calculating platform adaptation strength based on the unified portrait representation, the portrait uncertainty representation and platform context descriptors of each platform to obtain adaptation portrait representation for each platform content generation, generating content planning results of each platform based on a content atomic map, the adaptation portrait representation and a platform constraint set, executing confidence con