Recommended content determination method

Abstract

The invention relates to the technical field of recommended content, and discloses a recommended content determination method, which comprises the following steps of: constructing a dual representation model, and respectively extracting feature vectors from two dimensions of content features and user behaviors; on the basis of a dual representation model, feature representation of new contents and new users is realized through knowledge migration and representation alignment, and a cold start adaptive network is constructed; personalized recommendation of new contents and new users is realized based on a cold start adaptive network and multi-objective optimization, and a recommended content determination system is constructed, so that recommendation of the new contents and the new users can be quickly and accurately realized under the condition that user behavior data is not completely reliable, the method can be seamlessly integrated into an existing recommendation system, and the user experience is improved. The system uses a unified frame when processing cold start and non-cold start scenes, the complexity of the system is reduced, the maintainability and expandability of the system are improved, the processing is free of offset, the processing consistency is remarkably improved, and the method is suitable for batch production of high-precision power transmission tower steel structure profiles.

Description

Technical Field

[0001] The present invention relates to the technical field of recommended content, and more specifically, it relates to a method for determining recommended content. Background Art

[0002] Traditional recommendation system methods often rely on user historical behavior data to construct a user interest model, regarding the recommendation system as a basic structure composed of user behavior data and simple recommendation algorithms; most of these methods are based on existing user behavior data recording technologies and are difficult to handle the higher requirements for efficient and accurate recommendations in modern digital marketing.

[0003] Therefore, how to quickly and accurately recommend new content and new users when user behavior data is not completely reliable has become an urgent problem to be solved. Summary of the Invention

[0004] The present invention provides a method for determining recommended content, which solves the technical problem of how to quickly and accurately recommend new content and new users when user behavior data is not completely reliable in the prior art.

[0005] The present invention provides a method for determining recommended content, including: Construct a dual representation model to extract feature vectors from two dimensions of content features and user behavior respectively; among them, use a content encoder to encode content information to obtain a content feature vector, construct a user behavior encoder to map the user historical interaction behavior sequence to a user interest vector, and construct a feature fusion module to map the content feature vector and the user interest vector to the same feature space; Based on the dual representation model, realize the feature representation of new content and new users through knowledge transfer and representation alignment, and construct a cold start adaptation network; among them, construct a content domain adaptation module to extract features of new content through transfer learning, construct a user domain adaptation module to extract features of new users through transfer learning, and construct a feature alignment module to ensure that the feature representations of new content and new users are in the same feature space as the feature representations of existing content and users; Based on the cold start adaptation network and multi-objective optimization, realize personalized recommendation for new content and new users, and construct a recommended content determination system; among them, construct a recommendation relevance calculation model to calculate the relevance score between users and content, construct a multi-objective optimization model to consider recommendation relevance, recommendation diversity and recommendation novelty at the same time, and determine a recommended content list for each user based on the optimization function.

[0006] Further, the content encoder includes at least one encoding network selected from the following group: A pre-trained language model network for text content encoding; Convolutional neural network for image content encoding; Vision Transformer model network for image content encoding; 3D convolutional neural network for video content encoding; 1D convolutional long short-term memory network for audio content encoding.

[0007] Furthermore, the working process of the user behavior encoder includes: Taking the historical interaction behavior sequence H of the user j = h j1 , h j2 , …, h jk , where H j represents the historical interaction behavior sequence, and h j1 , h j2 , …, h jk respectively represent the 1st, 2nd, …, kth interaction behaviors in the historical interaction behavior sequence, k represents the total number of interaction behaviors, and inputting it into the user behavior encoder f user to generate the user interest vector where, represents the interest vector representation of user u j ; The user behavior encoder includes a recurrent neural network or a long short-term memory network for capturing the temporal features of the user behavior sequence, and an attention mechanism network for learning the importance of interaction behaviors.

[0008] Furthermore, the working process of the feature fusion module includes: Inputting the content feature vector and the user interest vector into the feature fusion function g respectively to obtain the fused content feature vector and the user interest vector where, represents the feature vector representation of content c i , and g represents the feature fusion function; The feature fusion function includes a fully connected neural network and a normalization operation unit.

[0009] Furthermore, the working process of the content domain adaptation module includes: For the new content item c new , using the content encoder f content to obtain its initial feature vector Then, utilize the content domain adaptation function h content to map the initial feature vector to the adapted feature representation where Θ c represents the domain knowledge parameters learned from the content with rich interaction data, and represents the new content feature vector after domain adaptation; The content domain adaptation function includes a basic feature extraction layer and a domain adaptation layer, and the basic feature extraction layer shares parameters with the content encoder.

[0010] Furthermore, the working process of the user domain adaptation module includes: For the new user u new , utilize the limited data to generate the initial feature representation of the new user through the user domain adaptation function h user where Θ u represents the domain knowledge parameters learned from the users with rich interaction data, and represents the new user feature vector after domain adaptation; The user domain adaptation function includes an attention weighting mechanism for sparse user behavior and a multi-source data fusion mechanism.

[0011] Furthermore, the working process of the feature alignment module includes: Adjust the feature vectors of the new content and the new user to construct the feature alignment function align, such that where and respectively represent the sets of fused feature vectors of all existing content and users, and respectively represent the aligned new content and new user feature vectors; The feature alignment function adopts an adversarial learning method to achieve feature alignment by minimizing the feature space distribution difference.

[0012] Furthermore, the working process of the recommendation relevance calculation model includes: For the user u j and the content c i , calculate their relevance score score(u i , c j ): where sim represents the similarity calculation function. For the new user u newand new content c new , calculate the relevance using the alignment feature vector:

[0013] Furthermore, the working process of the multi-objective optimization model includes: Construct a multi-objective optimization function \(L = \alpha\cdot L\) relevance +\(\beta\cdot L\) diversity +\(\gamma\cdot L\) novelty ; where \(L\) relevance , \(L\) diversity and \(L\) novelty represent relevance loss, diversity loss and novelty loss respectively, \(\alpha\), \(\beta\) and \(\gamma\) are weight parameters, and their value ranges are \([0, 1]\), and satisfy \(\alpha+\beta+\gamma = 1\). The multi-objective optimization model adopts an optimization framework combining gradient descent and Lagrange multipliers, and adaptively adjusts the weights of each objective.

[0014] A computer-readable storage medium is used to store computer-readable instructions, which can run a recommended content determination method when read by a computer.

[0015] The beneficial effects of the present invention are as follows: Through the dual representation model and the cold start adaptation network, the present invention can generate effective feature representations for new content and new users, effectively reduce the exposure delay of new content, and improve the initial distribution efficiency of content; Secondly, it can provide personalized content recommendations for users when they first join the platform, improving the initial experience satisfaction of users and the platform retention rate; Thirdly, through the recommended content determination method of multi-objective optimization, the relevance, diversity and novelty of the recommendation are considered at the same time, avoiding the information cocoon effect of the recommendation system and improving the overall satisfaction of users with the recommended content; Finally, the present invention can be seamlessly integrated into the existing recommendation system, enabling the system to use a unified framework when dealing with cold start and non-cold start scenarios, reducing the complexity of the system, improving the maintainability and scalability of the system, and thus being able to quickly and accurately recommend new content and new users under the condition that user behavior data is not completely reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic flowchart of a recommended content determination method provided in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Reference will now be made to exemplary embodiments to discuss the subject matter described herein. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and that changes can be made to the functions and arrangements of the elements discussed without departing from the scope of protection of the content of this specification. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples may be combined in other examples.

[0018] As Figure 1 shown, in at least one embodiment of the present invention, a method for determining recommended content is disclosed, including: In the field of digital marketing, recommendation systems usually build user interest models based on user historical behavior data and recommend content to users based on this model. However, when new content or new users appear in the system, due to the lack of sufficient user behavior data, traditional deep learning recommendation algorithms cannot accurately model them, resulting in a decline in recommendation accuracy. This is the "cold start problem" in recommendation systems. This embodiment aims to solve the cold start problem of new content and new users in the field of digital marketing and achieve accurate personalized recommendations for new content and new users lacking user behavior data.

[0019] Step 1: Build a dual representation model to extract feature vectors from two dimensions of content features and user behavior respectively; among them, use a content encoder to encode content information to obtain a content feature vector, build a user behavior encoder to map the user historical interaction behavior sequence into a user interest vector, and build a feature fusion module to map the content feature vector and the user interest vector into the same feature space; In this step, to solve the cold start problem, build a dual representation model that can process both content features and user behavior data simultaneously. This model contains two parallel feature extraction paths: a content feature extraction path and a user behavior extraction path.

[0020] The specific implementation of the dual representation model is as follows: The model consists of three sub-networks: a content encoding network, a user behavior encoding network, and a feature fusion network. The content encoding network adopts a multi-layer neural network structure, including an input layer, multiple hidden layers, and an output layer. The hidden layers can be fully connected layers, convolutional layers, or attention layers, depending on the type of content. For example, for text content, a Transformer encoder with a multi-head self-attention mechanism can be used; for image content, a convolutional neural network with multiple convolutional operations can be used. The user behavior encoding network also adopts a multi-layer neural network structure, but particularly emphasizes sequence processing capabilities. It can use recurrent neural networks, LSTMs, or attention-based sequence models to capture the temporal features and interest change features in the user behavior sequence. The feature fusion network adopts a fully connected neural network structure to map feature vectors from different sources to the same feature space and ensures the comparability of feature vectors through normalization operations.

[0021] In a specific application scenario of digital content recommendation, such as the short video recommendation scenario, the working process of the dual representation model is as follows: For newly uploaded short video content, the content encoding network will extract the title text features of the video (through a text encoder), the video cover image features (through an image encoder), and the video content features (through a video encoder, which can include actions, scenes, and audio information in the video), and synthesize these multi-modal features into the content feature vector of the video. For users, the user behavior encoding network will analyze the video sequence that the user has watched historically, the content sequence of likes and comments, and the viewing duration distribution behavior data to generate a user interest vector that reflects the user's interests and preferences. Then, the feature fusion network maps the content feature vector and the user interest vector to the same feature space, enabling the system to calculate the relevance between the user and the content to achieve personalized recommendation.

[0022] Sub-step 1.1: Content feature extraction. For the content feature extraction path, a content encoder is used to encode the text, image, or other structured information of the content to obtain the feature representation vector of the content. For text content, a pre-trained language model such as BERT can be used for encoding; for image content, a convolutional neural network or a vision Transformer model can be used for encoding. Let the content set be C = c1, c2,..., c i , where C represents the content set, c1, c2,..., c i represent the 1st, 2nd,..., i-th content items respectively, i represents the total number of content items, and the content encoder is denoted as f content , then the content feature vector can be expressed as: where, represents content ci Characteristic vector representation.

[0023] Sub-step 1.2: User behavior extraction.

[0024] For the user behavior extraction path, construct a user behavior encoder to encode the historical interaction behavior sequence H of the user j = h j1 h j2 h jk where H j represents the historical interaction behavior sequence, and h j1 h j2 h jk respectively represent the 1st, 2nd,..., kth interaction behaviors in the historical interaction behavior sequence, k represents the total number of interaction behaviors, and the user behavior encoder is denoted as f user Then the user interest vector can be represented as: where represents the interest vector representation of user u j .

[0025] Sub-step 1.3: Feature fusion.

[0026] Construct a feature fusion module to map the content feature vector and the user interest vector to the same feature space so that they can be used for similarity calculation. Let the feature fusion function be g, then the fused content feature vector and user interest vector are respectively: where and are the fused content feature vector and user interest vector respectively.

[0027] Step 2: Based on the dual representation model, realize the feature representation of new content and new users through knowledge transfer and representation alignment, and construct a cold start adaptation network; Among them, construct a content domain adaptation module to extract the features of new content through transfer learning, construct a user domain adaptation module to extract the features of new users through transfer learning, and construct a feature alignment module to ensure that the feature representations of new content and new users are in the same feature space as the feature representations of existing content and users; For the cold start scenario, construct a dedicated cold start adaptation network, which can generate effective feature representations for new content and new users based on limited information.

[0028] The specific implementation method of the cold start adaptation network is as follows: The network consists of three components: a content domain adaptation sub-network, a user domain adaptation sub-network, and a feature alignment sub-network. The content domain adaptation sub-network adopts a transfer learning architecture, including a basic feature extraction layer (sharing parameters with the content encoding network in step 1) and a domain adaptation layer (composed of multiple fully connected layers and a normalization layer). The characteristic of this sub-network is to construct the feature extraction ability for new content by learning knowledge from existing content. The user domain adaptation sub-network also adopts a transfer learning architecture, including a basic feature extraction layer and a domain adaptation layer, but specifically designs a processing mechanism for sparse user behavior, such as attention weighting and multi-source data fusion. The feature alignment sub-network adopts an adversarial learning framework, including a feature transformation network and a discriminator network, and realizes the alignment of new content, new user features with existing content and user features by minimizing the difference in feature space distribution.

[0029] In the specific application scenario of an e-commerce platform, when a new product is just put on the shelf or a new user just registers, the working process of the cold start adaptation network is as follows: For a newly listed product, the content domain adaptation sub-network will first extract its basic attribute features (such as category, brand, price range), and then based on the interaction patterns of existing similar products, map the new product into the same feature space as the existing products to generate a product feature representation that can match user interests. For a newly registered user, the user domain adaptation sub-network will use the limited information provided by the user during the registration process (such as age, gender, initial interest category), combined with the behavior patterns of similar user groups, to generate the initial interest representation of the new user. The feature alignment sub-network ensures that these newly generated feature representations are consistent with the existing feature representations in the system in terms of distribution, so that the subsequent recommendation algorithm can seamlessly handle cold start and non-cold start scenarios.

[0030] Sub-step 2.1: Content domain adaptation. Construct a content domain adaptation module to extract features of new content through transfer learning. For a new content item c new , first use the content encoder in step 1 to obtain its initial feature vector Then, construct a content domain adaptation function h content , and map the initial feature vector of the new content to an adapted feature representation: where Θ c represents the domain knowledge parameters learned from content with rich interaction data, represents the new content feature vector after domain adaptation.

[0031] Sub-step 2.2: User domain adaptation.

[0032] Construct a user domain adaptation module to extract features of new users through transfer learning. For a new user u new , there may be only a very small amount of behavioral data or only user attribute data. Let these limited data be Construct a user domain adaptation function h user to obtain the initial feature representation of the new user: where Θ u represents the domain knowledge parameters learned from users with rich interaction data, and represents the feature vector of the new user after domain adaptation.

[0033] Sub-step 2.3: Feature alignment Construct a feature alignment module to ensure that the feature representations of new content and new users are in the same feature space as those of existing content and users. Construct a feature alignment function align to further adjust the feature vectors obtained in Step 2-1 and 2-2: where and represent the sets of fused feature vectors of all existing content and users in Step 1-3 respectively, and represent the aligned feature vectors of new content and new users respectively.

[0034] Step 3: Implement personalized recommendations for new content and new users based on the cold-start adaptation network and multi-objective optimization, and construct a recommended content determination system; Among them, construct a recommendation relevance calculation model to calculate the relevance score between users and content, construct a multi-objective optimization model to consider recommendation relevance, recommendation diversity, and recommendation novelty at the same time, and determine a list of recommended content for each user based on the optimization function. In this step, based on the feature representations obtained in Step 1 and Step 2, construct a recommended content determination system that can handle both cold-start and non-cold-start scenarios to ensure the overall performance of the recommendation system.

[0035] The specific implementation of the recommended content determination system is as follows: The system consists of three parts: a relevance calculation subsystem, a multi-objective optimization subsystem, and a recommended list generation subsystem. The relevance calculation subsystem adopts a multi-layer neural network structure, including a feature interaction layer and a similarity calculation layer. Among them, the feature interaction layer realizes the in-depth interaction between user features and content features through the attention mechanism or multi-layer perceptron, and the similarity calculation layer calculates the final relevance score based on the interaction result. The multi-objective optimization subsystem adopts an optimization framework that combines gradient descent and the Lagrange multiplier method, constructs a composite loss function that includes relevance, diversity, and novelty objectives, and adaptively adjusts the weights of each objective. The recommended list generation subsystem uses beam search or integer programming algorithms to generate an optimal recommended content list for each user on the premise of meeting the multi-objective optimization conditions.

[0036] In the specific application scenario of social media content recommendation, the working process of the recommended content determination system is as follows: After the user logs in to the social media platform, the system first calculates the initial relevance score based on the feature vectors of the user and the candidate content. Then, while considering relevance, the multi-objective optimization subsystem adds constraints on content diversity and novelty (such as ensuring that the recommended results include content on different topics and appropriately introducing new content types that the user has not encountered before) to generate the overall optimal objective function value. Finally, based on this optimization objective, the recommended list generation subsystem selects the most suitable content combination for the user from the candidate content pool and presents it to the user in a specific order. For newly registered users, the system uses the initial user feature representation generated in step two for relevance calculation and may increase the weight of content exploration to help new users quickly discover content categories they are interested in. For newly published content, the system uses the content feature representation generated in step two to recommend it to the user group that may be interested in it, accelerating the exposure and dissemination of new content.

[0037] Sub-step 3.1: Recommendation relevance calculation.

[0038] Build a recommendation relevance calculation model to calculate the relevance score between the user and the content. For user u j and content c i , calculate its relevance score score(uk, c i ): Among them, sim represents the similarity calculation function, which can be cosine similarity, dot product, or other similarity calculation methods. For new user u new and new content c new , calculate the relevance using the aligned feature vectors in step two: Sub-step 3.2: Multi-objective optimization.

[0039] Build a multi-objective optimization model, considering recommendation relevance, recommendation diversity, and recommendation novelty simultaneously. Build a multi-objective optimization function L: L = α·L relevance + β·L diversity + γ·L novelty Wherein, L relevance , L diversity and L novelty represent relevance loss, diversity loss, and novelty loss respectively. α, β, and γ are weight parameters, with a value range of [0, 1], and satisfy α + β + γ = 1, which are used to balance the importance of different objectives.

[0040] Sub-step 3.3: Determine the recommended content.

[0041] Based on the optimization function L, determine the recommended content list R u for each user u. First, calculate the relevance score between user u and all candidate contents c ∈ C′, where C′ is the set of candidate contents. Then, use the greedy algorithm or integer programming method to select a group of contents from the candidates to make the optimization function L reach the optimal value, and generate the final recommended content list R u .

[0042] For cold-start user u new , calculate the relevance score using the feature representation generated in step two, and determine the recommended content list in the same way Similarly, for the new content c to be recommended new , also calculate the relevance score using the feature representation generated in step two, and determine which users should be recommended this content.

[0043] The recommended content determination method implemented through the above steps can effectively solve the cold-start problem of the recommendation system in digital marketing. The specific technical effects are as follows: For the cold-start problem of new content, this method can, based on the characteristics of the content itself, through domain adaptation and feature alignment, generate a representation vector in the same feature space as the existing content, enabling the system to accurately recommend new content to users who may be interested in it in the absence of user behavior data, effectively reducing the exposure delay of new content and improving the initial distribution efficiency of the content.

[0044] For the cold-start problem of new users, this method can, based on the limited information of users, through domain adaptation and feature alignment, generate an effective user interest representation, enabling the system to provide personalized content recommendations for users when they first join the platform, improving the initial experience satisfaction of users and the retention rate of the platform.

[0045] A method for determining recommended content through multi-objective optimization. In this embodiment, not only the relevance of the recommendation is concerned, but also the diversity and novelty of the recommendation are considered simultaneously. It can provide more comprehensive and personalized recommendation results for users, avoid the information cocoon effect of the recommendation system, and improve the overall satisfaction of users with the recommended content.

[0046] The dual representation model and cold start adaptation network of this embodiment can be seamlessly integrated into the existing recommendation system, enabling the system to use a unified framework when processing cold start and non-cold start scenarios, reducing the complexity of the system, and improving the maintainability and scalability of the system.

[0047] In this section, taking an online short video platform as an example, the specific application of this embodiment in solving the cold start problem of content recommendation is demonstrated. There are a large number of new videos uploaded and new users registered on this platform every day, facing typical cold start challenges: Cold start of new videos: The platform adds more than 1 million short videos every day. These videos initially lack user viewing, liking, and commenting interaction data. How to recommend these new videos to users who may be interested in them in a timely manner is the primary problem faced by the platform.

[0048] Cold start of new users: The platform adds about 50,000 new registered users every day. These users lack historical behavior data, and it is difficult for the system to understand their interest preferences. How to provide personalized content recommendations for them and improve user retention rate is another key problem.

[0049] Before implementing this method, the platform adopted a traditional recommendation algorithm based on collaborative filtering. For newly uploaded videos, it mainly relied on the push of the creator's fans and random exposure to accumulate initial data; for new users, it provided recommended popular videos, lacking personalization. This led to a long exposure cycle for new videos, a poor experience for new users, and limited overall user growth and content ecosystem development of the platform.

[0050] In this short video platform, according to Step 1 of Embodiment 1, a dual representation model suitable for short video content was constructed. The model includes a content encoding network, a user behavior encoding network, and a feature fusion network. The structure of the content encoding network is shown in Table 1.

[0051] Table 1: Structure of the content encoding network of the short video platform Level Input Network structure Output dimension Video feature layer Original video frame sequence 3D-CNN+Transformer 256 Audio feature layer Original audio signal 1D-CNN+LSTM 128 Text feature layer Title, description, tags BERT pre-trained model 384 Feature fusion layer Multi-modal feature vector Attention fusion network 512 The structure of the user behavior encoding network is shown in Table 2: Table 2: Structure of the user behavior encoding network of the short video platform Level Input Network structure Output dimension Short-term interest layer Recently watched video sequence GRU+Self-attention mechanism 256 Long-term interest layer Historical viewing category statistics Multi-layer perceptron 128 Interaction behavior layer Like, comment, share records Convolutional neural network 128 Interest fusion layer Multi-dimensional interest vector Multi-head attention network 512 The feature fusion network adopts a two-layer fully connected network to map the content features and user interest features into the same 512-dimensional feature space. During the training process, 100 million user-video interaction records existing on the platform are used, and the contrastive learning method is adopted to minimize the distance of positive sample pairs (videos watched by users) and maximize the distance of negative sample pairs (videos not watched by users).

[0052] According to step 2 of Embodiment 1, a cold start adaptation network is constructed for the characteristics of the short video platform. The parameters of the content domain adaptation sub-network and the user domain adaptation sub-network of this network are shown in Table 3.

[0053] Table 3: Parameter Configuration of Cold Start Adaptation Network During the training process, new content and new user data within the most recent month on the platform are adopted, and the model effect is evaluated by simulating the cold start scenario (shielding part of the interaction data). The feature alignment sub-network adopts the adversarial training method to make the feature distributions of new content and new users as close as possible to the feature distributions of existing content and users.

[0054] According to step 3 of Embodiment 1, a recommended content determination system is constructed and deployed on the short video platform. The system configuration is shown in Table 4. Table 4: System Configuration of Recommended Content Determination During the system deployment process, the A / B test method is adopted. The platform users are randomly divided into a test group and a control group. The test group uses the recommended method of this embodiment, and the control group uses the original recommended method of the platform. The effect of this method is evaluated by comparing the performances of the two groups of users on various indicators.

[0055] Technical Effect Verification To verify the effect of this embodiment in solving the cold start problem, verification is carried out from two key technical effects: the exposure efficiency of new content and the retention rate of new users.

[0056] This embodiment can effectively solve the cold start problem of new content through the cold start adaptation network, significantly shorten the first exposure time of new content, and improve the content distribution efficiency. Table 5 shows the comparison of the new content exposure indicators between this method and the original method of the platform.

[0057] Table 5: Comparison of New Content Exposure Efficiency Indicator Original method This implementation Improvement ratio Time (hours) of the first effective exposure of the new video 8.4 2.1 75.0％ Exposure volume (average number of people) within 24 hours after upload 126 317 151.6％ Content first-day click-through rate (%) 2.8 5.3 89.3％ Content first-day interaction rate (%) 1.5 3.2 113.3％ Creator satisfaction score (on a 5-point scale) 3.2 4.5 40.6％ As can be seen from Table 5, compared with the original method, this embodiment can shorten the first effective exposure time of the new video from 8.4 hours to 2.1 hours, an increase of 75.0%; the average exposure volume within 24 hours of uploading increases from 126 person-times to 317 person-times, a growth of 151.6%. At the same time, since this method can recommend new content to users who are truly likely to be interested in it, the first-day click-through rate and interaction rate of the new content have also increased significantly, by 89.3% and 113.3% respectively. This has greatly improved the creator experience, and the creator satisfaction score has increased from 3.2 points to 4.5 points (on a 5-point scale), an increase of 40.6%.

[0058] Through the cold start adaptation network, this embodiment can generate effective interest representations based on the limited information of new users, thereby providing personalized content recommendations and significantly improving the experience satisfaction and retention rate of new users. Table 6 shows the comparison between this method and the original method of the platform in terms of new user retention metrics.

[0059] Table 6: Comparison of new user retention rates Indicator Original method This implementation Improvement ratio Average viewing duration (minutes) of new users on the first day 18.6 34.2 83.9％ Average number of videos watched by new users on the first day 12.4 23.8 92.0％ Next-day retention rate (%) 32.5 45.8 41.0％ 7-day retention rate (%) 18.3 29.6 61.8％ 30-day retention rate (%) 9.5 16.2 70.5％ As can be seen from Table 6, after using this embodiment, the average viewing duration of new users on the first day increases from 18.6 minutes to 34.2 minutes, an increase of 83.9%; the average number of videos viewed on the first day increases from 12.4 to 23.8, an increase of 92.0%. More importantly, the next-day retention rate of new users increases from 32.5% to 45.8%, an increase of 41.0%; the 7-day retention rate increases from 18.3% to 29.6%, an increase of 61.8%; the 30-day retention rate increases from 9.5% to 16.2%, an increase of 70.5%. This shows that this embodiment can effectively solve the cold start problem of new users, and through more accurate personalized recommendations, significantly improve the activity and retention rate of new users.

[0060] Based on the verification results of the above two key technical effects, the application of this embodiment in the short video platform has successfully solved the cold start problems of new content and new users, significantly improved the content distribution efficiency and user retention rate of the platform, and provided strong support for the content ecosystem and user growth of the platform.

[0061] The above describes the embodiments of the present invention, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more forms of the same embodiments, all of which fall within the protection scope of this embodiment.

Claims

1. A method for determining recommended content, characterized in that, Including the following steps: Construct a dual representation model to extract feature vectors from two dimensions of content features and user behavior respectively; among them, use a content encoder to encode content information to obtain a content feature vector, construct a user behavior encoder to map the user's historical interaction behavior sequence into a user interest vector, and construct a feature fusion module to map the content feature vector and the user interest vector into the same feature space; Based on the dual representation model, realize the feature representation of new content and new users through knowledge transfer and representation alignment, and construct a cold start adaptation network; among them, construct a content domain adaptation module to extract the features of new content through transfer learning, construct a user domain adaptation module to extract the features of new users through transfer learning, and construct a feature alignment module to ensure that the feature representations of new content and new users are in the same feature space as the feature representations of existing content and users; Based on the cold start adaptation network and multi-objective optimization, realize personalized recommendation for new content and new users, and construct a recommended content determination system; among them, construct a recommendation relevance calculation model to calculate the relevance score between the user and the content, construct a multi-objective optimization model to consider recommendation relevance, recommendation diversity and recommendation novelty at the same time, and determine a recommended content list for each user based on the optimization function.

2. The method for determining recommended content according to claim 1, characterized in that, The content encoder includes at least one encoding network selected from the following groups: A pre-trained language model network for text content encoding; A convolutional neural network for image content encoding; A vision Transformer model network for image content encoding; A three-dimensional convolutional neural network for video content encoding; A one-dimensional convolutional long short-term memory network for audio content encoding.

3. The method for determining recommended content according to claim 1, wherein The working process of the user behavior encoder includes: The historical interaction behavior sequence H of the user j = h j1 , h j2 , …, h jk , where H j represents the historical interaction behavior sequence, h j1 , h j2 , …, h jk respectively represent the 1st, 2nd, …, k-th interaction behaviors in the historical interaction behavior sequence, k represents the total number of interaction behaviors, and the input user behavior encoder f user is used to generate the user interest vector Among them, represents the interest vector representation of user u j . The user behavior encoder includes a recurrent neural network or a long short-term memory network for capturing the temporal features of the user behavior sequence, and an attention mechanism network for learning the importance of interaction behaviors.

4. The method for determining recommended content according to claim 1, characterized in that The working process of the feature fusion module includes: Input the content feature vector and the user interest vector into the feature fusion function g respectively, and obtain the fused content feature vector and the user interest vector Among them, represents the content c i in the form of a feature vector, and g represents a feature fusion function; The feature fusion function includes a fully connected neural network and a normalization operation unit.

5. The method for determining recommended content according to claim 1, wherein The working process of the content domain adaptation module includes: For the new content item c new , use the content encoder f content to obtain its initial feature vector Reuse the content domain adaptation function h content Map the initial feature vector to the adapted feature representation Among them, Θ c represents the domain knowledge parameters learned from the content with rich interaction data, and represents the new content feature vector after domain adaptation; The content domain adaptation function includes a basic feature extraction layer and a domain adaptation layer, and the basic feature extraction layer shares parameters with the content encoder.

6. The method for determining recommended content according to claim 1, characterized in that The working process of the user domain adaptation module includes: For new user u new , using limited data through the user domain adaptation function h user generate the initial feature representation of the new user Among them, Θ u represents the domain knowledge parameters learned from users with rich interaction data, and represents the new user feature vector after domain adaptation; The user domain adaptation function includes an attention weighting mechanism for sparse user behaviors and a multi-source data fusion mechanism.

7. A method for determining recommended content according to claim 1, characterized in that The working process of the feature alignment module includes: Adjust the feature vectors of new content and new users, and construct a feature alignment function align so that Among them, and respectively represent the set of integrated feature vectors of all existing content and users, and respectively represent the new content and new user feature vectors after alignment; The feature alignment function uses an adversarial learning method to achieve feature alignment by minimizing the difference in feature space distribution.

8. A method for determining recommended content according to claim 1, characterized in that, The working process of the recommendation relevance calculation model includes: For user u j and content c i , calculate their relevance score score(u j , c i ): where sim represents a similarity calculation function, for a new user u new and new content c new , the relevance is calculated using the aligned feature vectors:

9. The method for determining recommended content according to claim 1, wherein The working process of the multi-objective optimization model includes: Construct a multi-objective optimization function \(L = \alpha\cdot L\) relevance +\(\beta\cdot L\) diversity +\(\gamma\cdot L\) novelty ; Among them, L relevance , L diversity and L novelty respectively represent relevance loss, diversity loss, and novelty loss. α, β, and γ are weight parameters with a value range of [0, 1], and satisfy α + β + γ = 1. The multi-objective optimization model adopts an optimization framework combining gradient descent and the Lagrange multiplier method, and adaptively adjusts the weights of each objective.

10. A computer-readable storage medium, characterized in that, For storing computer-readable instructions, when the computer-readable instructions are read by a computer, they can run a method for determining recommended content as described in any one of claims 1-9.

Images