Novel cold start recommendation method and device, electronic equipment and storage medium

Abstract

The invention relates to a novel cold start recommendation method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining novel data to be recommended and micro-drama data of a user, wherein the novel data to be recommended and the micro-drama data comprise information of the same type; based on an NLP model, respectively performing deep semantic feature extraction on the micro-drama data and the to-be-recommended novel data, and generating a micro-drama text vector and a novel text vector; based on a pre-trained multi-label model, performing label prediction according to the novel text vector, the micro-episode text vector and the acquired micro-episode label to obtain a novel prediction label; and through a hybrid recommendation engine, obtaining a semantic similarity score of the novel prediction tag and the micro-drama tag and a novel historical behavior score of the user, and generating a novel recommendation list for recommendation according to the semantic similarity score and the novel historical behavior score. According to the invention, the accuracy of novel recommendation can be improved.

Description

Technical Field

[0001] This application relates to the field of data push technology, and in particular to a novel cold start recommendation method, device, electronic device and storage medium. Background Technology

[0002] Current mainstream novel recommendation systems are essentially built upon single-user behavioral data within the novel domain. Therefore, when users of short dramas on a platform enter the novel section for the first time or rarely, despite the high correlation between short dramas and novels in terms of theme, narrative structure, and target audience preferences, and the rich user tags and complete user profiles accumulated by the short drama business, these are not reused across domains by the novel recommendation system, lacking cross-domain feature fusion capabilities. Furthermore, new user behavioral data in the novel section is sparse. As new users or those who have entered the novel field infrequently, the system has little or no novel-related behavioral data, such as clicks and reading time. It can only generate recommendations based on limited general strategies within the novel domain, further exacerbating the feature deficiency problem during the cold start phase, leading to a significant drop in recommendation accuracy.

[0003] It is evident that existing novel recommendation systems cannot reuse novels across different platforms by combining the same user's known micro-drama data, resulting in low accuracy in novel recommendations. Summary of the Invention

[0004] This application provides a novel cold start recommendation method, apparatus, electronic device, and storage medium to solve the technical problem that existing novel recommendation systems cannot achieve cross-domain reuse of novels on the same platform by combining the micro-drama data known to the same user, resulting in low novel recommendation accuracy.

[0005] According to one aspect of the embodiments of this application, this application provides a novel cold start recommendation method, the method comprising: acquiring novel data to be recommended and user micro-drama data, wherein the novel data to be recommended and the micro-drama data include similar information; performing deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on an NLP model to generate micro-drama text vectors and novel text vectors; performing tag prediction based on a pre-trained multi-tag model according to the novel text vectors, the micro-drama text vectors and the acquired micro-drama tags to obtain novel predicted tags; and obtaining the semantic similarity score between the novel predicted tags and the micro-drama tags and the user's novel historical behavior score through a hybrid recommendation engine, and generating a novel recommendation list for recommendation based on the semantic similarity score and the novel historical behavior score.

[0006] Optionally, obtaining the novel data to be recommended and the user's micro-drama data includes: obtaining the user's micro-drama metadata from a Hive table through a data call interface, and obtaining the metadata of the work to be recommended from the novel library; selecting the micro-drama title and micro-drama synopsis information from the user's micro-drama metadata, and selecting the novel title and novel synopsis information from the metadata of the work to be recommended; concatenating the micro-drama title and micro-drama synopsis information into text based on a preset data merging format to obtain the user's micro-drama data, and concatenating the novel title and novel synopsis information into text based on the preset data merging format to obtain the novel data to be recommended.

[0007] Optionally, the step of performing deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on the NLP model to generate micro-drama text vectors and novel text vectors includes: performing text segmentation on the micro-drama data and the novel data to be recommended based on the NLP model to obtain multiple micro-drama word data and multiple novel word data; performing vector transformation on the multiple micro-drama word data and the multiple novel word data based on a pre-trained language model to obtain multiple micro-drama word vectors and multiple novel word vectors; pooling the multiple micro-drama word vectors to obtain micro-drama sentence embedding vectors, and pooling the multiple novel word vectors to obtain novel sentence embedding vectors; and performing feature standardization on the micro-drama sentence embedding vectors and the novel sentence embedding vectors to obtain the micro-drama text vectors and the novel text vectors, respectively.

[0008] Optionally, the step of performing tag prediction based on the pre-trained multi-label model, according to the novel text vector, the micro-drama text vector, and the obtained micro-drama tags, to obtain the novel predicted tag includes: obtaining the micro-drama tag from the micro-drama metadata and establishing a mapping table between the micro-drama tag and the micro-drama title; inputting the novel text vector, the micro-drama tag, the micro-drama text vector, and the mapping table constructed by the micro-drama tag and the micro-drama title into the pre-trained multi-label model for tag prediction, so as to match the corresponding novel predicted tag to the novel text vector according to the user's micro-drama tag.

[0009] Optionally, the step of obtaining the semantic similarity score between the novel prediction tag and the micro-drama tag, as well as the user's novel history behavior score, through a hybrid recommendation engine, and generating a novel recommendation list for recommendation based on the semantic similarity score and the novel history behavior score, includes: comparing the novel prediction tag and the user's micro-drama tag using the NLP model based on the hybrid recommendation engine, and calculating the semantic similarity score using a preset similarity calculation function; obtaining the user's novel reading history data, predicting the historical behavior score based on the novel reading history data according to a preset novel rating model, and obtaining the novel historical behavior score; and generating a novel recommendation list for recommendation based on the user's semantic similarity score and the novel historical behavior score.

[0010] Optionally, the step of generating a novel recommendation list based on the user's semantic similarity score and novel historical behavior score includes: matching a first weight to the semantic similarity score and a second weight to the novel historical behavior score, wherein the sum of the first weight and the second weight is equal to 1; obtaining the user's historical novel clicks, constructing a weight adjustment function based on an exponential function, and dynamically adjusting the decay coefficient of the exponential function based on the historical novel clicks; adjusting the first weight and the second weight through the weight adjustment function to obtain the first weight adjustment value and the second weight adjustment value; calculating a novel recommendation score based on the user's semantic similarity score, the first weight adjustment value, the novel historical behavior score, and the second weight adjustment value; and generating the novel recommendation list based on the novel recommendation score.

[0011] Optionally, before performing label prediction based on the novel text vector, the micro-drama text vector, and the acquired micro-drama tags using the pre-trained multi-label model to obtain the novel predicted label, the method further includes: acquiring micro-drama sample data and novel sample data; performing deep semantic feature extraction on the micro-drama sample data and the novel sample data respectively based on the NLP model to generate micro-drama text vector samples and novel text vector samples; and training a multi-label model based on the acquired micro-drama sample tags, the micro-drama text vector samples, and the novel text vector samples to obtain the pre-trained multi-label model that outputs the novel sample predicted label.

[0012] According to another aspect of the embodiments of this application, this application provides a novel cold start recommendation device, the device comprising: a data acquisition module, configured to acquire novel data to be recommended and user micro-drama data, wherein the novel data to be recommended and the micro-drama data include similar information; a feature extraction module, configured to perform deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on an NLP model, generating micro-drama text vectors and novel text vectors respectively; a tag prediction module, configured to perform tag prediction based on a pre-trained multi-tag model, according to the novel text vectors, the micro-drama text vectors, and the acquired micro-drama tags, to obtain novel predicted tags; and a data recommendation module, configured to acquire the semantic similarity score between the novel predicted tags and the micro-drama tags and the user's novel historical behavior score through a hybrid recommendation engine, and generate a novel recommendation list for recommendation based on the semantic similarity score and the novel historical behavior score.

[0013] According to another aspect of the embodiments of this application, this application provides an electronic device, including: a processor, a memory, and a network interface. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory through the network interface, and the processor executes the machine-readable instructions to perform the steps of the novel cold start recommendation method.

[0014] According to another aspect of the embodiments of this application, this application provides a storage medium having processor-executable non-volatile program code, the program code causing the processor to perform the steps of the novel cold start recommendation method.

[0015] Compared with related technologies, the technical solutions provided in this application have the following advantages: The novel cold-start recommendation method provided in this application extracts deep semantic features from user micro-drama data and novel data to be recommended based on an NLP model. This not only improves semantic representation capabilities but also provides high-quality data support for subsequent multi-label model input. Through cross-domain label alignment, the obtained micro-drama tags, micro-drama text vectors, and novel text vectors are used as input to the multi-label model to establish the association between micro-drama tags and novel content, outputting corresponding novel prediction tags. This achieves preference transfer in the cold-start stage and obtains more accurate novel prediction tags for subsequent recommendations. Through a hybrid recommendation engine, it effectively solves the cold-start problem for new novel users by reusing the tag preferences accumulated by users in the micro-drama field to quickly generate accurate and personalized novel recommendations, avoiding the homogenization problem of general popular content pushes. It also enriches the recommendation feature dimensions for old novel users, ultimately improving the overall accuracy of the novel recommendation list and user satisfaction. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0017] Figure 1 This is a schematic diagram of the hardware environment for an optional novel cold start recommendation method provided according to an embodiment of this application; Figure 2 This is a schematic diagram of an optional novel cold start recommendation method provided according to an embodiment of this application; Figure 3 This is a schematic diagram of another optional novel cold start recommendation method provided according to an embodiment of this application; Figure 4 This is a system diagram of an optional cross-domain cold start recommendation system provided according to an embodiment of this application; Figure 5 This is a structural diagram of an optional novel cold start recommendation device provided according to an embodiment of this application; Figure 6 This is a schematic diagram of an optional electronic device structure provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To address the problems mentioned in the background art, according to one aspect of the embodiments of this application, an embodiment of a novel cold start recommendation method is provided.

[0020] like Figure 1 As shown, the above-mentioned novel cold start recommendation method can be applied to, for example... Figure 1The hardware environment shown is described. The system architecture 100 of the hardware environment includes a terminal device 101 and a server 103. The server 103 is connected to the terminal 101 via a network and can provide services to the terminal device 101 or clients installed on the terminal device 101. A database 105 can be set up on or independently of the server 103 to provide data storage services for the server 103. The network can include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0021] Users can use terminal device 101 to interact with server 103 via a network to receive or send messages. Various communication client applications can be installed on terminal device 101, such as web browsers, search applications, and instant messaging tools. Terminal device 101 can be various electronic devices with a display screen that support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers. Server 103 can be a server providing various services, such as a backend server supporting the pages displayed on terminal device 101.

[0022] It should be noted that the novel cold start recommendation method provided in this application embodiment is generally executed by a server and / or terminal device, and correspondingly, the novel cold start recommendation system is generally set up in the server / terminal device. Furthermore, it should be understood that... Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0023] like Figure 2 As shown, Figure 2 This is a flowchart illustrating a novel cold start recommendation method provided in an embodiment of the present invention. Taking the novel cold start recommendation method being executed by a server as an example, the novel cold start recommendation method includes the following steps: Step S202: Obtain the novel data to be recommended and the user's micro-drama data, wherein the novel data to be recommended and the micro-drama data include information of the same type.

[0024] Specifically, the aforementioned "novel data to be recommended" can refer to the data corresponding to all novels to be recommended to the user. The aforementioned "micro-drama data" can refer to the data corresponding to the micro-dramas the user has historically watched. Both the novel data to be recommended and the user's micro-drama data are from the same video platform. The inclusion of similar information between the novel data to be recommended and the micro-drama data means that they contain the same types of information. For example, if the novel data to be recommended is the novel title, then the micro-drama data will also be the micro-drama title; if the novel data to be recommended is both the novel title and a synopsis, then the micro-drama data will also be the micro-drama title and a synopsis. By obtaining similar information between novels and micro-dramas, it is easier to train the model to accurately transfer micro-drama tags to novels based on the correlation between the micro-drama data and the novel data to be recommended, thus achieving accurate novel recommendations.

[0025] Step S204: Based on the NLP model, perform deep semantic feature extraction on the micro-drama data and the novel data to be recommended, respectively, to generate micro-drama text vectors and novel text vectors.

[0026] Specifically, the aforementioned NLP (Natural Language Processing) model refers to a natural language processing model. The micro-drama data and the novel data to be recommended are presented in text form. The deep semantic feature extraction process includes: constructing an NLP pipeline, inputting the micro-drama data and the novel data to be recommended as text into the document assembler of the NLP model, converting them into document form, then loading a word segmentation model to segment the documents, and finally loading the BERT (Bidirectional Encoder Representations from Transformers) embedding layer of the NLP model, using the document and the word segmentation results as model input, and outputting sentence embedding vectors. For micro-dramas, the final output is a sentence-level vector corresponding to the micro-drama data, i.e., a micro-drama text vector; for novels, the output is a sentence-level vector corresponding to the novel data to be recommended, i.e., a novel text vector. It's important to note that each piece of micro-drama data corresponds to one micro-drama and one micro-drama text vector; similarly, each piece of novel data to be recommended corresponds to one novel to be recommended and one novel text vector.

[0027] Among them, the BERT model is trained through the masked language model task, which can generate deep bidirectional language representations by simultaneously utilizing the text information before and after the current position. It has a stronger language understanding ability and only uses the encoder part of the Transformer architecture to convert the input text into fixed-length vector representations. These vector representations contain rich semantic information, realizing a deep understanding and representation of the text, namely the micro-drama text vectors and novel text vectors mentioned above.

[0028] Step S206: Based on the pre-trained multi-label model, perform label prediction according to the novel text vector, the micro-drama text vector, and the obtained micro-drama labels to obtain the novel predicted labels.

[0029] Specifically, the aforementioned pre-trained multi-label model refers to a pre-trained multi-label model. Using this model, tag transfer can be performed based on the tags of the same user's historical viewing history of short dramas, enabling the tagging of novels and providing accurate novel recommendations when a user who watched a short drama enters the novel reading field as a new user.

[0030] The micro-drama tags refer to the tags associated with a user's viewing of a micro-drama within a specific historical period. This historical period corresponds to the timeframe for acquiring the user's micro-drama data, such as the past year. Micro-drama tags include, but are not limited to, "rebirth and revenge," "family feuds," "suspense and mystery," "sweet romance," and "workplace struggles." Each micro-drama has its own corresponding tag. It's important to note that for long-time micro-drama users who are new novel readers, they will only have micro-drama tags, not novel tags.

[0031] Furthermore, since the text vectors of micro-dramas and novels share certain similarities in text content—for example, semantic consistency in themes, character settings, and plots—this embodiment employs a multi-label model. By acquiring micro-drama tags and using the micro-drama tags, micro-drama text vectors, and novel text vectors as input to the model, a cross-domain content mapping from micro-drama tags to novel prediction tags is established. This enables accurate cross-domain migration based on micro-drama tags to novel prediction tags. Each novel prediction tag corresponds to a novel to be recommended.

[0032] In some examples, if there are too many micro-drama tags, a tag selection frequency can be set to filter out high-frequency micro-drama tags from the many micro-drama tags, which is more representative and helps to improve the accuracy of tag migration.

[0033] Step S208: Using a hybrid recommendation engine, obtain the semantic similarity score between the novel prediction tag and the micro-drama tag, as well as the user's novel history behavior score, and generate a novel recommendation list based on the semantic similarity score and the novel history behavior score for recommendation.

[0034] The hybrid recommendation engine refers to a method that combines multiple recommendation strategies and recommends novels based on the fusion result. In this embodiment, the hybrid recommendation engine combines the semantic similarity of the tags of novels and micro-dramas with the user's novel reading history for scoring. During scoring, the micro-drama viewing behavior data of micro-drama users can also be converted into relevant weights for novel recommendations to more accurately realize the preference transfer during the user's cold start phase.

[0035] Specifically, the semantic similarity score between the novel's predicted tags and the micro-drama tags can be calculated using similarity calculation functions, including but not limited to Jaccard similarity, cosine similarity, and edit distance. Obtaining the user's novel reading history score can refer to scoring based on the user's historical reading behavior data, which includes but is not limited to clicks, reading time, and consumption data. Scoring based on this historical reading behavior data can be done using custom scoring rules.

[0036] Furthermore, after calculating the semantic similarity score and the novel's historical behavior score, weights can be assigned to both. Combining the semantic similarity score, the novel's historical behavior score, and their respective weights yields the final score. Each novel to be recommended corresponds to one score, and a novel recommendation list is then generated based on the score rankings. Recommendations are then made to users based on this list.

[0037] In some examples, when there are many novels in the novel recommendation list, the top N novels with the highest ratings can be selected for recommendation, where N≥1. For example, the top 15 novels with the highest ratings can be selected for recommendation.

[0038] In this embodiment of the invention, deep semantic feature extraction is performed on the acquired user micro-drama data and novel data to be recommended based on an NLP model. This not only improves semantic representation capabilities but also provides high-quality data support for subsequent multi-label model input. Through cross-domain label alignment, the acquired micro-drama tags, micro-drama text vectors, and novel text vectors are used as input to the multi-label model to establish the association between micro-drama tags and novel content, outputting corresponding novel prediction tags. This achieves preference transfer during the cold start phase and obtains more accurate novel prediction tags for subsequent recommendations. Through a hybrid recommendation engine, the cold start problem for new novel users can be effectively solved. By reusing the tag preferences accumulated by users in the micro-drama field, accurate and personalized novel recommendations can be quickly generated, avoiding the homogenization problem of general popular content push. It also enriches the recommendation feature dimensions for old novel users, ultimately improving the overall accuracy of the novel recommendation list and user satisfaction.

[0039] In some optional embodiments, step S202 above includes: S2021, retrieves the user's micro-drama metadata from the Hive table and the metadata of works to be recommended from the novel library through the data call interface; S2022, Select the title and synopsis of the micro-drama from the user's micro-drama metadata, and select the title and synopsis of the novel from the metadata of the works to be recommended; S2023, based on a preset data merging format, the title of the micro-drama and the synopsis of the micro-drama are concatenated to obtain the user's micro-drama data; and based on the preset data merging format, the title of the novel and the synopsis of the novel are concatenated to obtain the novel data to be recommended.

[0040] Combination Figure 3 As shown, structured data can be extracted from micro-dramas and novels to provide high-quality input for subsequent modeling. Specifically, during data extraction, micro-drama metadata is obtained from Hive tables via the Spark SQL API. This metadata includes a unique identifier (id), title, brief information, and a list of tag IDs. The id is a unique identifier for the micro-drama on the platform or in the database, used to distinguish different works; the title is the name / title of the micro-drama, such as "A Sweet Little Drama" or "The CEO's Counterattack"; the brief is a short summary of the plot, such as "A corporate slave unexpectedly transmigrates into a princess and begins his counterattack"; and the tag IDs are a list storing the IDs corresponding to tags, serving as unique identifiers for the micro-drama's genre / type. Metadata for works to be recommended can also be obtained from the novel library, including the novel's id, title, and brief.

[0041] Furthermore, the titles and synopses of each micro-drama are read from the acquired micro-drama metadata, and the titles and synopses of novels are selected from the metadata of the works to be recommended. Further, the titles and synopses of each micro-drama are concatenated according to a preset data merging format to obtain the micro-drama data for each micro-drama; similarly, the titles and synopses of each novel are concatenated to obtain the novel data for each novel to be recommended. For example, a unified data format is constructed: item_id, title, brief, tags, type (episode / novel); text concatenation: concat(title,"."", brief).

[0042] In this embodiment, the title and synopsis of a micro-drama better reflect its specific content and corresponding tag type. Therefore, by retrieving the micro-drama's metadata and selecting its title and synopsis as the data basis for subsequent model predictions, more accurate prediction of novel tags can be achieved. Based on a preset data merging format, text is concatenated to form a single field, which retains the core identifier information of the title while supplementing the detailed information of the synopsis. This avoids the problems of insufficient title information and redundant synopsis information with a low proportion of core keywords. At the same time, integrating it into a single field simplifies the data processing flow, reduces the complexity and computational cost of modeling multiple fields separately, and facilitates unified NLP operations such as word segmentation and embedding vector generation. This allows the model to more comprehensively capture the content features and semantic information of the micro-drama, thereby improving the accuracy and reliability of subsequent novel prediction tag and micro-drama tag semantic similarity calculation and cross-domain recommendation matching, ultimately pushing content that better suits the user's preferences.

[0043] In some optional embodiments, step S204 above includes: S2041, Based on the NLP model, perform text segmentation on the micro-drama data and the novel data to be recommended to obtain multiple micro-drama word data and multiple novel word data; S2042, Based on the pre-trained language model, the multiple micro-drama word data and the multiple novel word data are respectively vectorized to obtain multiple micro-drama word vectors and multiple novel word vectors; S2043, pool the multiple micro-drama word vectors to obtain micro-drama sentence embedding vectors, and pool the multiple novel word vectors to obtain novel sentence embedding vectors; S2044, Perform feature normalization on the micro-drama sentence embedding vector and the novel sentence embedding vector respectively to obtain the micro-drama text vector and the novel text vector.

[0044] The pre-trained language models mentioned above include, but are not limited to, DistilBERT, BERT, ALBERT, and RoBERTa models. In this embodiment, the pre-trained language model is the BERT model. The BERT model has bidirectional context modeling capabilities, which can overcome the limitations of traditional word vectors in capturing contextual semantics. It can deeply mine the potential semantic connections between texts such as the title of a micro-drama and its synopsis, such as the synonymous semantics of "ancient costume counterattack" and "princess rises to power," and the near-synonymous connections of "sweet romance" and "warm love." The generated sentence-level vectors can fully carry the overall semantic information of the text.

[0045] Specifically, in combination Figure 3As shown, in the corresponding modeling layer, deep semantic feature extraction and alignment are performed based on an NLP model, which can integrate heterogeneous content features for pre-training a multi-label model. Specifically, during deep semantic feature extraction, text preprocessing is performed first: the text of both the micro-drama data and the novel data to be recommended is segmented into words, resulting in multiple micro-drama word data and multiple novel word data. Further word vectorization is performed: a pre-trained language model (BERT model) is used to convert each word into a dense vector (word embedding), resulting in multiple micro-drama word vectors and multiple novel word vectors. Sentence / document vectorization continues: pooling is used to aggregate the word vectors of each micro-drama into sentence-level vectors, obtaining the sentence embedding vector for each micro-drama, and pooling the word vectors of each novel into sentence-level vectors, obtaining the sentence embedding vector for each novel. Finally, feature standardization is performed: L2 normalization is applied to the sentence embedding vectors of both micro-drama and novel, obtaining micro-drama text vectors and novel text vectors respectively, ensuring consistent feature scale for easier subsequent calculations.

[0046] In this embodiment, deep semantic features are extracted using an NLP model, then converted into sentence-level vectors using a BERT model and normalized. This achieves accurate quantification and standardized expression of text semantic information. Normalization eliminates interference caused by differences in length and amplitude among different text vectors, unifies the scale range of vectors, and improves the accuracy of similarity calculation between vectors.

[0047] In some optional embodiments, step S206 above includes: S2061, Obtain the micro-drama tag from the micro-drama metadata, and establish a mapping table between the micro-drama tag and the micro-drama title; S2062, input the novel text vector, the micro-drama tag, the mapping table constructed by the micro-drama text vector and the micro-drama tag and the micro-drama title into the pre-trained multi-label model for tag prediction, so as to match the corresponding novel predicted tag to the novel text vector according to the user's micro-drama tag.

[0048] Specifically, the micro-drama metadata also includes micro-drama tags. Each micro-drama has a corresponding tag, meaning a mapping relationship can be established between the micro-drama tag ID and the micro-drama's unique identifier ID. Furthermore, the micro-drama tag can be determined based on the micro-drama title and synopsis information. Micro-drama tags can be extracted from the user's micro-drama metadata, allowing the creation of a mapping table between tag IDs and micro-drama titles. This clarifies all micro-drama titles corresponding to each tag, representing a one-to-many relationship. For example, the tag "historical drama" corresponds to multiple micro-drama titles such as "The Time-Traveling Princess is So Cool" and "The Secret History of the Ancient Palace," ensuring an accurate and comprehensive mapping.

[0049] Furthermore, the novel text vectors, micro-drama tags, micro-drama text vectors, and the mapping table constructed by micro-drama tags and micro-drama titles are input into the pre-trained multi-label model for tag prediction, resulting in predicted tags for novels without corresponding novels to be recommended. Specifically, according to the input format required by the model, the above data is uniformly loaded into the pre-trained multi-label model to complete the model input initialization. Importing the novel text vectors, micro-drama tags, micro-drama text vectors, and the mapping table constructed by micro-drama tags and micro-drama titles allows the model to obtain the association information between tags and corresponding micro-drama content titles, providing support for subsequent feature fusion. Furthermore, within the multi-label model: multi-feature fusion and semantic association mining are completed, and the model automatically performs deep processing on the imported multi-class data.

[0050] In some examples, the internal processing of the multi-label model includes: feature alignment and fusion of various inputs, integrating the semantic information of the novel text vector, the cross-domain semantic information of the micro-drama text vector, the category information of the micro-drama tags, and the association information between tags and micro-drama titles in the mapping table to form a multi-dimensional comprehensive feature matrix. Based on the tag classification logic learned in the pre-training stage, the model mines the hidden semantic associations in the comprehensive features, such as the matching relationship between the semantics of "princess's counterattack" in the novel text vector and the micro-drama tags "ancient costume" and "counterattack". The mapping table is used to further verify whether the micro-drama title corresponding to the tag matches the novel content. For each novel to be recommended, the model combines the fused comprehensive features to initially select a set of candidate tags highly related to the novel's content. Tag prediction: Outputting the final predicted tags for the novels to be recommended. After completing the internal feature processing and candidate tag selection, the model generates predicted tags for each novel to be recommended according to the output logic of multi-label classification.

[0051] In some examples, the model generates predicted tags for each novel to be recommended according to the output logic of multi-label classification. Specifically, this may include: the model scores the candidate tag set with confidence, and the higher the score, the higher the match between the tag and the novel content; the model selects tags that meet the score and are not duplicated according to a preset confidence threshold to form the final predicted tag set for each novel to be recommended. A novel may correspond to multiple predicted tags, such as the final predicted tags of a novel being "historical drama", "sweet romance", and "comeback". The above prediction process is then performed on all novels to be recommended in sequence, and finally the association result between each novel and the corresponding predicted tag is output, thus completing the overall tag prediction task.

[0052] In this embodiment, novel text vectors and micro-drama text vectors are generated by concatenating the title and synopsis. Micro-drama tags are extracted, and a mapping table between micro-drama tags and micro-drama titles is established. Multiple core features are obtained for tag prediction. By leveraging the cross-domain features related to micro-dramas and the tag mapping relationship, the deep semantic commonalities and matching logic between micro-dramas and novels in terms of theme, narrative, and tag system are fully explored. Then, the feature fusion and tag prediction capabilities of a pre-trained multi-tag model are used for prediction, which greatly improves the accuracy, completeness, and relevance of novel prediction tags. It can output high-quality tags that fit the content attributes and are suitable for cross-domain recommendation scenarios for each novel to be recommended. At the same time, the predicted tags are anchored to the content features of the novel itself and are adapted to the tag cognition and preference system of micro-drama users. This provides high-quality tag feature support for subsequent cross-domain tag semantic similarity calculation, user preference matching, and accurate recommendation ranking, ultimately improving the accuracy of novel recommendations and enhancing subsequent user stickiness.

[0053] In some alternative embodiments, prior to step S206 above, the method further includes: S101, Obtain micro-drama sample data and novel sample data; S102, Based on the NLP model, perform deep semantic feature extraction on the micro-drama sample data and the novel sample data respectively to generate micro-drama text vector samples and novel text vector samples; S103, based on the obtained micro-drama sample labels, micro-drama text vector samples and novel text vector samples, a multi-label model is trained to obtain the pre-trained multi-label model that outputs the predicted labels of novel samples.

[0054] Specifically, before practical application, multi-label model training is required. First, sufficient sample data from micro-dramas and novels are acquired. Both types of samples contain complete textual core content and label annotation information. The micro-drama sample data consists of the complete text of the micro-drama title and synopsis, along with corresponding real micro-drama labels. The novel sample data consists of the complete text of the novel title and synopsis, along with corresponding real novel labels. After screening, cleaning, and standardizing both types of sample data, the integrity and validity of the sample data are ensured, forming a standardized sample dataset that can be directly used for model training.

[0055] Furthermore, in the deep semantic feature extraction stage, based on a mature NLP model, independent deep semantic feature extraction processing is performed on the two types of sample data after processing. Semantic encoding is performed on the spliced ​​text of each micro-drama sample to mine the hidden deep semantic information such as theme, narrative, and content attributes in the text, generating standardized micro-drama text vector samples with complete semantic representation capabilities. Using the same NLP feature extraction logic and processing standards, the spliced ​​text of each novel sample is subjected to the same semantic encoding and feature extraction, generating standardized novel text vector samples with uniform dimensions and consistent representation, ensuring that the two types of text vector samples have the same semantic feature expression logic and comparable feature dimensions.

[0056] Furthermore, in the training preparation phase: all core input data required for training are integrated, and the labeled real micro-drama sample tags, extracted micro-drama text vector samples, and novel text vector samples are associated and integrated. Based on the micro-drama sample tags, using the micro-drama text vector samples as cross-domain semantic feature references, and using the novel text vector samples as training objects, a complete training dataset adapted to the multi-label classification task is constructed to achieve accurate matching between feature data and label data. Then, the integrated training dataset is input into the multi-label model to be trained for model training, and the semantic features of the micro-drama tags and the semantic features of the novel text are deeply fused and associated and mapped. During the training process, the model parameters are continuously iterated and optimized, and the prediction deviation from the novel text vectors to the corresponding labels is constantly corrected until the prediction accuracy and fitting effect of the model reach the preset standard. Finally, the training converges, and a pre-trained multi-label model that can stably output the predicted labels of novel samples is obtained.

[0057] In this embodiment, by performing standardized deep semantic feature extraction on micro-drama and novel samples, the consistency of semantic representation and feature dimension matching of the text vectors of the two types of samples are ensured. Then, by combining the real tags of micro-drama as supervision for multi-label model training, the pre-trained multi-label model can fully learn the semantic features and classification logic of the micro-drama tag system, and deeply grasp the cross-domain association matching rules between novel text semantics and micro-drama tags, thereby improving the accuracy and generalization ability of the model in predicting novel sample tags. At the same time, the tags and semantic features of micro-drama samples are fully reused, laying a high-quality model and data foundation for subsequent core links such as cross-domain content matching, user preference fusion, and accurate recommendation, thereby improving the matching efficiency and overall recommendation accuracy of cross-domain recommendation from the bottom up.

[0058] In some optional embodiments, step S208 above includes: S2081, Based on the hybrid recommendation engine, the novel prediction tags and the user's micro-drama tags are compared using the NLP model, and the semantic similarity score is calculated using a preset similarity calculation function; S2082, Obtain the user's novel reading history data, and based on the novel reading history data, predict the historical behavior score according to the preset novel rating model to obtain the novel historical behavior score; S2083, Generate a novel recommendation list based on the user's semantic similarity score and novel history behavior score, and make recommendations.

[0059] Specifically, in combination Figure 3 As shown, the aforementioned hybrid recommendation engine corresponds to a fusion layer and an output layer. It dynamically fuses cross-domain features to form a cross-domain cold-start recommendation system, generates a recommendation list to be interpreted, and finally recommends it to the user. Specifically, it can combine various rating data to score novel recommendations, and ultimately combine these ratings to make novel recommendations. In the cross-domain cold-start recommendation system, In some examples, calculating the semantic similarity score involves: performing pairwise cosine similarity calculations on each predicted novel tag and the micro-drama tag. Specifically, for each predicted tag vector of the novel to be recommended, a cosine similarity calculation is performed with the target micro-drama tag vector. The semantic similarity is determined by the angle between the vectors; the smaller the angle, the higher the cosine similarity score, indicating a higher semantic match between the two types of tags, and vice versa. After completing the matching calculation for all tag pairs, the complete semantic similarity scores for each micro-drama tag corresponding to a single novel are obtained. All the similarity scores obtained from the pairwise matching are then normalized, invalid and low-relevance scores are removed, and valid matching results are retained. This results in an accurate semantic similarity score between the predicted novel tag and the micro-drama tag for the corresponding novel, which can be directly used as the core reference for subsequent cross-domain recommendation matching.

[0060] In some examples, for users without a novel reading history, the aforementioned novel reading history data is empty. For users with a novel reading history, the process first involves accurately acquiring all relevant data, including but not limited to the user's core behavioral patterns for each novel. This includes valid behavioral data such as user clicks, reading time, completion rate, whether the user favorited, liked, reread, or skipped readings. Invalid, irrelevant behavioral data, such as accidental clicks and short-term dwell times, are then removed, completing data cleaning and filtering. The filtered user novel reading history data undergoes standardized preprocessing to unify statistical dimensions and quantification rules. Finally, a pre-defined novel rating model is invoked. This model is specifically designed for rating novel user behavior and has pre-learned and solidified the correlation logic between various user novel behaviors and preferences.

[0061] The preprocessed reading history data for each novel is then input into the preset novel rating model. Based on the input user's individual novel reading history data and combined with the model's built-in rating logic, the preset novel rating model calculates the user's behavioral data and quantifies the user's liking for each novel read. After the model completes the calculation, it outputs a corresponding rating result for each novel read by the user. This score is the novel history behavior score for each novel. A comprehensive novel history behavior score can be calculated by combining the individual novel history behavior scores, such as by calculating the average value.

[0062] Furthermore, by combining users' semantic similarity scores and novel-related historical behavior scores, novels that can be recommended are selected to generate a novel recommendation list for recommendation.

[0063] In this embodiment, a hybrid recommendation engine is used. Based on an NLP model, the semantic fit between tags is determined by similarity metric, resulting in more objective and accurate scoring results. This effectively identifies tag associations that are semantically similar but literal, while also eliminating invalid matches that are semantically unrelated but literal. Scoring prediction is based on users' actual novel reading history data, and the scoring results accurately and quantitatively reflect users' novel reading preferences and levels of liking. Novel recommendations are made by combining the semantic similarity scores of predicted novel tags and micro-drama tags with users' novel reading history scores. This achieves a deep integration and complementarity between users' potential cross-domain preferences and their actual reading preferences within the novel domain. This effectively solves the recommendation problem for new novel users with no historical behavior data during cold starts, and by relying on novel reading history scores, it accurately quantifies users' actual novel reading habits and explicit preferences, ensuring that recommended content matches users' existing reading preferences.

[0064] In some optional embodiments, step S2083 above includes: The semantic similarity score is matched with a first weight, and the novel's historical behavior score is matched with a second weight, the sum of the first weight and the second weight equals 1; Obtain the user's historical novel click count, construct a weight adjustment function based on the exponential function, and dynamically adjust the decay coefficient of the exponential function based on the historical novel click count; The first weight and the second weight are adjusted using the weight adjustment function to obtain the first weight adjustment value and the second weight adjustment value; A novel recommendation score is calculated based on the user's semantic similarity score, the first weight adjustment value, the novel's historical behavior score, and the second weight adjustment value. The novel recommendation list is generated based on the novel recommendation rating and then recommended.

[0065] In this context, the first weight is P1, and the second weight is P2, both satisfying the condition: P1 + P2 = 1. These two weights are matched to facilitate more precise control over rating bias. Regarding the novel's historical click count, new users have a click power of 0, while returning users have corresponding click counts. Then, a weight adjustment function is constructed based on an exponential function, dynamically adjusting the decay coefficient of the exponential function based on the novel's historical click count. The specific weight adjustment function is as follows: P1=p*e -1m*k (1) Where p represents the base weight, e -1*k Let m represent the exponential function, m represent the decay coefficient, and k represent the historical click count of the novel.

[0066] P1 is calculated according to the above formula (1). A change in P1 will cause a change in P2. Finally, the first weight adjustment value and the second weight adjustment value are recorded. Combined with the following formula (2), the novel recommendation score for each novel can be calculated: S = P1*S1 + P2*S2 (2) Where S represents the final novel recommendation score, S1 represents the semantic similarity score, and S2 represents the novel's historical behavior score.

[0067] Based on the above formula (2), the novel recommendation score of all novels is finally calculated. Then, a novel recommendation list is generated in descending order, and the top n novels in the novel recommendation list are selected for recommendation. For example, the top 10 novels are selected for recommendation. The following example will be used to illustrate this more clearly. For example, user Xiao Zhang usually watches short dramas on iQiyi and prefers content such as "ancient costume, sweet romance, and princess's counterattack". However, when he enters the novel channel for the first time, there is no record of reading or clicking on any novels. Novel to be recommended: "The Transmigrated Princess is So Cool", the genre tags are "ancient costume, sweet romance, and princess's counterattack", which is one of the novels promoted by the platform. User clicks: 0. Pre-configuration: Baseline weight coefficient p=0.8, decay coefficient m=0.1. Calculate the novel historical behavior score: The system attempts to predict the score based on Xiao Zhang's novel historical behavior through the novel recommendation model. Since there is no record of reading or clicking on any novels, the final novel historical behavior score is 0. The semantic similarity score was calculated by comparing Xiao Zhang's micro-drama tags (ancient costume, sweet romance, princess's counterattack) and the tags (ancient costume, sweet romance, princess's counterattack) of the novel to be recommended, "The Transmigrated Princess is So Cool". The two were completely consistent, and the cosine similarity was extremely high. The final semantic similarity score was 0.95, where the score range was 0-1. The closer to 1, the higher the matching degree. The weight P1 was calculated according to formula (1) and combined with Xiao Zhang's click volume (0) to obtain P2=0.8×1=0.8. At this time, the P2 value was high, indicating that the influence of cross-domain score should be taken into account. Since Xiao Zhang did not read the novel, it was necessary to rely on his micro-drama preference to deduce it. Based on the above (2), the final novel recommendation score was calculated. The score was substituted into the value: 0.8×0.95+(1-0.8)×0=0.76. This means that the matching degree between this novel and Xiao Zhang is as high as 0.76, which is a high matching degree. If the recommendation ranking is in the top 10, the system will recommend this novel to Xiao Zhang first.

[0068] In this embodiment, the final recommendation score formed by the weighted fusion of two types of scores can accurately quantify and rank candidate novels. This ensures that the recommendation results not only anchor the user's core novel preferences and align with reading habits, but also accurately reach the user's potential needs, significantly improving the accuracy and matching degree of the recommended content. At the same time, it fully utilizes the cross-domain content of platform micro-dramas and novels to strengthen cross-domain traffic conversion and user stickiness, thereby improving the overall accuracy of novel recommendations and enhancing user satisfaction.

[0069] Combination Figure 4As shown, the final cross-domain cold start recommendation system includes a data preprocessing layer, a feature calculation layer, and a hybrid recommendation engine. The data preprocessing layer is used for collecting metadata from micro-dramas and novels, cleaning novel metadata, and fusing heterogeneous data. The feature calculation layer is used for BERT semantic encoding, label system mapping, and multi-label model construction and prediction. The hybrid recommendation engine is used for novel behavior recommendation, calculating historical behavior scores for novels, using an NLP semantic channel to calculate semantic similarity scores, and a dynamic decay module to adjust the decay coefficient and calculate weights.

[0070] According to another aspect of the embodiments of this application, such as Figure 5 As shown, corresponding to the novel cold start recommendation method in the above embodiments, this embodiment provides a novel cold start recommendation device, the device comprising: Data acquisition module 501 is used to acquire novel data to be recommended and micro-drama data of users, wherein the novel data to be recommended and the micro-drama data include information of the same type. Feature extraction module 503 is used to perform deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on NLP model, and generate micro-drama text vector and novel text vector respectively; The tag prediction module 505 is used to predict the novel's predicted tags based on the novel's text vector, the micro-drama's text vector, and the obtained micro-drama tags, using a pre-trained multi-tag model. The data recommendation module 507 uses a hybrid recommendation engine to obtain the semantic similarity score between the novel prediction tag and the micro-drama tag, as well as the user's novel history behavior score. Based on the semantic similarity score and the novel history behavior score, it generates a novel recommendation list for recommendation.

[0071] It should be noted that in this embodiment, the data acquisition module 501 can be used to execute step S202 in this application embodiment, the feature extraction module 503 in this embodiment can be used to execute step S204 in this application embodiment, the label prediction module 505 in this embodiment can be used to execute step S206 in this application embodiment, and the data recommendation module 507 in this embodiment can be used to execute step S208 in this application embodiment.

[0072] It should be noted that the examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should also be noted that the above modules, as part of a device, can operate in environments such as... Figure 1 The hardware environment shown can be implemented either through software or through hardware.

[0073] It should be noted that the suffixes such as module, submodule, and subunit used to represent elements in the above-described device are only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, they can be used in combination.

[0074] According to another aspect of the embodiments of this application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the novel cold start recommendation method in any of the above embodiments.

[0075] According to another aspect of the embodiments of this application, this application also provides an electronic device, such as... Figure 6 As shown, the device includes a memory 601, a processor 603, and a network interface 605. The memory 601 stores a computer program that can run on the processor 603. The memory 601 and the processor 603 communicate through the network interface 605 and a communication bus 607. When the electronic device is running, the processor 603 and the memory 601 communicate through the network interface 605. When the processor 603 executes the computer program, it implements the steps of the novel cold start recommendation method described above.

[0076] The memory and processor in the aforementioned electronic device communicate with each other via a communication bus and a communication interface. The communication bus can be a peripheral component interconnect standard (PCI) bus or an extended industry standard structure (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. The memory can include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor. The aforementioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0077] It is understood that the embodiments described herein can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits, digital signal processors, digital signal processing devices, microprocessors, and other electronic units or combinations thereof for performing the functions described herein. For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. Software code can be stored in memory and executed by a processor.

[0078] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0079] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device or unit, and may be electrical, mechanical or other forms.

[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0081] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0082] It should be noted that, in this document, relational terms such as first, second, etc., are used only to distinguish one entity or operation from another entity or operation. The terms include, encompass, or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0083] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A novel cold start recommendation method, characterized in that, The method includes: Obtain data on novels to be recommended and user micro-drama data, wherein the data on novels to be recommended and the micro-drama data include information of the same type; Based on the NLP model, deep semantic features are extracted from the micro-drama data and the novel data to be recommended, respectively, to generate micro-drama text vectors and novel text vectors. Based on a pre-trained multi-label model, label prediction is performed according to the novel text vector, the micro-drama text vector, and the obtained micro-drama labels to obtain the novel predicted labels; Using a hybrid recommendation engine, the semantic similarity score between the novel prediction tag and the micro-drama tag, as well as the user's novel history behavior score, are obtained. A novel recommendation list is generated based on the semantic similarity score and the novel history behavior score for recommendation.

2. The novel cold start recommendation method according to claim 1, characterized in that, The acquisition of data on novels to be recommended and user micro-drama data includes: The system retrieves the user's micro-drama metadata from the Hive table and the metadata of works to be recommended from the novel library through the data call interface. Select the title and synopsis of the micro-drama from the user's micro-drama metadata, and select the title and synopsis of the novel from the metadata of the works to be recommended; Based on a preset data merging format, the titles and synopses of the micro-drama are concatenated to obtain the user's micro-drama data. Similarly, based on the preset data merging format, the titles and synopses of the novel are concatenated to obtain the novel data to be recommended.

3. The novel cold start recommendation method according to claim 2, characterized in that, The method involves performing deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on an NLP model, generating micro-drama text vectors and novel text vectors, including: Based on the NLP model, the micro-drama data and the novel data to be recommended are processed by text segmentation to obtain multiple micro-drama word data and multiple novel word data. Based on a pre-trained language model, multiple micro-drama word data and multiple novel word data are respectively vectorized to obtain multiple micro-drama word vectors and multiple novel word vectors; Pooling is performed on multiple micro-drama word vectors to obtain micro-drama sentence embedding vectors, and pooling is performed on multiple novel word vectors to obtain novel sentence embedding vectors; The micro-drama sentence embedding vector and the novel sentence embedding vector are respectively normalized to obtain the micro-drama text vector and the novel text vector.

4. The novel cold start recommendation method according to claim 2, characterized in that, The pre-trained multi-label model predicts novel labels based on the novel text vector, the micro-drama text vector, and the acquired micro-drama labels, including: Obtain the micro-drama tags from the micro-drama metadata and establish a mapping table between the micro-drama tags and the micro-drama titles; The novel text vector, the micro-drama tag, the mapping table constructed by the micro-drama text vector and the micro-drama tag and the micro-drama title are input into the pre-trained multi-label model for tag prediction, so as to match the corresponding novel predicted tag to the novel text vector according to the user's micro-drama tag.

5. The novel cold start recommendation method according to claim 1, characterized in that, The process involves using a hybrid recommendation engine to obtain the semantic similarity score between the predicted novel tag and the micro-drama tag, as well as the user's historical novel behavior score. A novel recommendation list is then generated based on the semantic similarity score and the historical novel behavior score for recommendation. This includes: Based on the hybrid recommendation engine, the novel prediction tags and the user's micro-drama tags are compared using the NLP model, and the semantic similarity score is calculated using a preset similarity calculation function. Obtain the user's novel reading history data, and based on the novel reading history data, predict the historical behavior score according to the preset novel rating model to obtain the novel historical behavior score; A novel recommendation list is generated based on the user's semantic similarity score and the novel's historical behavior score for recommendation.

6. The novel cold start recommendation method according to claim 5, characterized in that, The process of generating a novel recommendation list based on the user's semantic similarity score and novel history behavior score includes: The semantic similarity score is matched with a first weight, and the novel's historical behavior score is matched with a second weight, the sum of the first weight and the second weight equals 1; Obtain the user's historical novel click count, construct a weight adjustment function based on the exponential function, and dynamically adjust the decay coefficient of the exponential function based on the historical novel click count; The first weight and the second weight are adjusted using the weight adjustment function to obtain the first weight adjustment value and the second weight adjustment value; A novel recommendation score is calculated based on the user's semantic similarity score, the first weight adjustment value, the novel's historical behavior score, and the second weight adjustment value. The novel recommendation list is generated based on the novel recommendation rating and then recommended.

7. The novel cold start recommendation method according to any one of claims 1 to 6, characterized in that, Before performing tag prediction based on the novel text vector, the micro-drama text vector, and the obtained micro-drama tags using the pre-trained multi-label model to obtain the novel's predicted tag, the method further includes: Obtain sample data for micro-dramas and novels; Based on the NLP model, deep semantic features are extracted from the micro-drama sample data and the novel sample data respectively to generate micro-drama text vector samples and novel text vector samples. The multi-label model is trained based on the obtained micro-drama sample labels, micro-drama text vector samples, and novel text vector samples to obtain the pre-trained multi-label model that outputs the predicted labels of novel samples.

8. A novel cold start recommendation device, characterized in that, The device includes: The data acquisition module is used to acquire data on novels to be recommended and user micro-drama data, wherein the data on novels to be recommended and the micro-drama data include information of the same type. The feature extraction module is used to perform deep semantic feature extraction on the micro-drama data and the novel data to be recommended based on the NLP model, and generate micro-drama text vectors and novel text vectors respectively. The tag prediction module is used to predict the novel's predicted tags based on the novel's text vector, the micro-drama's text vector, and the obtained micro-drama tags, using a pre-trained multi-tag model. The data recommendation module uses a hybrid recommendation engine to obtain the semantic similarity score between the novel prediction tag and the micro-drama tag, as well as the user's novel history behavior score. Based on the semantic similarity score and the novel history behavior score, it generates a novel recommendation list for recommendation.

9. An electronic device, comprising: A processor, a memory, and a network interface, wherein the memory stores machine-readable instructions executable by the processor, characterized in that: when the electronic device is running, the processor communicates with the memory via the network interface, and the processor executes the machine-readable instructions to perform the steps of the novel cold start recommendation method as described in any one of claims 1 to 7.

10. A storage medium having processor-executable non-volatile program code, characterized in that, The program code causes the processor to execute the steps of the novel cold start recommendation method according to any one of claims 1 to 7.

Images