Exhibition content recommendation method, system and device based on big data analysis
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG MATRIX SPACE BUILDING TECHNOLOGY CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-21
Smart Images
Figure CN122432399A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of big data, and in particular to methods, systems and devices for recommending exhibition content based on big data analysis. Background Technology
[0002] Current exhibition content recommendation systems based on big data mostly adopt general recommendation architectures, primarily relying on static matching of user historical interactions and exhibit metadata. However, these methods generally suffer from a core flaw: they fail to effectively model the dynamic evolution of user interests and the deep correlation between the multimodal semantics of exhibition content. Specifically, the systems typically simplify user interests into fixed vectors, ignoring the differences in information intensity and time decay of different interactive behaviors (such as clicks, dwell times, and shares), and lack fine-grained control over multimodal feature fusion and similarity measurement during the matching phase. This results in recommendations that are insufficient in terms of accuracy, timeliness, and scenario adaptability.
[0003] Therefore, there is an urgent need for a technical solution that can dynamically integrate multi-source behavioral signals and multimodal content features to achieve accurate and interpretable personalized recommendations. Summary of the Invention
[0004] This application provides a method, system, and apparatus for recommending exhibition content based on big data analysis, which solves the technical problems of poor recommendation accuracy and timeliness caused by the static modeling of user interests and insufficient fusion of multimodal content features in existing technologies.
[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides a method for recommending exhibition content based on big data analysis, including: Collect multimodal raw data of the exhibition content to be recommended, wherein the multimodal raw data includes at least text description, image data and audio narration; The multimodal raw data are encoded by a pre-trained deep neural network model to obtain the corresponding text feature vector, image feature vector and audio feature vector; Based on the text feature vector, image feature vector, and audio feature vector, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network; wherein, the multimodal adaptive fusion network uses a nonlinear formula containing an exponential function and a normalization term to dynamically allocate the fusion weights of each modality; The number of valid historical interaction events of the target user within a preset time window is obtained, and the number of valid historical interaction events is compared with a preset interaction threshold to determine the user type of the target user. The user type includes cold start users and non-cold start users. Cold start users represent users whose historical interaction data is insufficient to support reliable interest modeling, and non-cold start users represent users whose historical interaction data is sufficient to support reliable interest modeling. The number of valid historical interaction events refers to the total number of interaction behaviors of the target user with the exhibition content within the preset time window that meet the validity determination criteria. Based on the user type of the target user, an interest vector of the target user is generated; wherein, the interest vector of a cold-start user is generated based on prior information, and the interest vector of a non-cold-start user is generated based on its historical interaction content. Calculate the matching score between the integrated content embedding vector and the interest vector; the matching score represents the correlation between each exhibition content and the target user's interest vector; All exhibition content to be recommended is sorted according to the matching score, and a predetermined number of the top-ranked exhibition content is pushed to the target user as the recommendation result.
[0006] Based on the above technical solutions, the exhibition content recommendation method based on big data analysis provided in this application addresses the highly dynamic and sparse nature of user interests in exhibition recommendation scenarios. Exhibition content itself contains multimodal information such as text, images, and audio. Traditional recommendation methods, employing static interest modeling and simple feature concatenation, struggle to accurately capture the deep semantic relationships between users and complex content, especially under cold-start or sparse interaction conditions. This application constructs a multimodal adaptive fusion network, dynamically allocating modal weights using a nonlinear mechanism with exponential functions and normalization terms, effectively improving the semantic completeness and discriminative power of content representation. Simultaneously, it identifies user types based on the number of effective interaction events within a preset time window, generating differentiated interest vectors for cold-start and non-cold-start users using prior information or historical behavior, balancing personalized and robust recommendations. This solution not only solves the problems of inaccurate recommendations and weak timeliness caused by rigid interest modeling and coarse multimodal fusion in existing technologies, but also significantly improves the system's usability and the relevance and diversity of recommendation results in data-sparse scenarios, thereby enhancing the user's exhibition experience and improving the distribution efficiency of exhibition resources.
[0007] In conjunction with the first aspect above, in one possible implementation, the encoding of the multimodal raw data includes: A pre-trained language model based on the Transformer architecture is used to semantically encode the text description, generating a dimension of... Text feature vectors ; Visual features are extracted from the image data using a convolutional neural network or a visual Transformer model, generating a dimension of... Image feature vectors ; A pre-trained audio model combining time-frequency analysis and deep sequence modeling is used to perform acoustic-semantic joint encoding on the audio narration, generating a dimension of [missing information]. audio feature vector ; The text feature vector, image feature vector, and audio feature vector are all mapped to the same semantic embedding space.
[0008] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the integrated content embedding vector includes: The text feature vector Image feature vectors and audio feature vectors The input is fed into a multimodal adaptive fusion network; The multimodal adaptive fusion network is based on the user's current context state vector. Using a nonlinear mapping function Calculate the context-dependent activation values for each modality. ;in, Characterizing multimodal data types, , This indicates vector concatenation; The integrated content embedding vector is obtained by weighted summation of the feature vectors of each modality. ;in, For fusion weighting coefficients; Wherein, the nonlinear mapping function It is implemented by a multilayer perceptron containing at least two fully connected layers and a nonlinear activation function, and the context state vector It includes at least one of the following: user access time, geographical location, and device interaction mode.
[0009] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the fusion weight coefficients includes: The fusion weight coefficients for each modality are calculated using the Softmax function. ;in, For temperature coefficient, ; Summation symbol The dummy variable in the code represents the traversal of the set of all modal types. The elements in.
[0010] In conjunction with the first aspect above, in one possible implementation, the method for determining the user type of the target user includes: Collect valid historical interaction events of the target user within a preset time window and count the number of valid historical interaction events; wherein, the valid historical interaction events include one or more of the user's actions of clicking, browsing, collecting, rating or sharing exhibition content, and exclude interaction records generated by automatic triggering or abnormal sessions; Compare the number of valid historical interaction events with a preset interaction threshold; If the number of valid historical interaction events is less than the preset interaction threshold, the target user is determined to be a cold start user; otherwise, the target user is determined to be a non-cold start user.
[0011] In conjunction with the first aspect above, in one possible implementation, the method for generating the interest vector of the cold-start user includes: Obtain the user's prior information, which includes at least one of the following: attribute information provided by the user during registration, demographic group labels, or default interest distribution preset by the system. The prior information is embedded and encoded to obtain the corresponding prior feature representation; Based on the prior feature representation, the initial interest vector of the target user is generated through the initialization mapping module; The method for generating the interest vectors of non-cold-start users includes: Obtain the user's historical interaction event sequence, which includes the user's interaction records with multiple exhibition contents within a preset time window. Each interaction record includes the identifier of the interacted exhibition content, the corresponding interaction type, and the interaction time. Based on the exhibition content identifier, multimodal semantic features of each exhibition content are obtained, wherein the multimodal semantic features include at least one of text description embedding vector, image visual feature vector, or structured label vector; Based on the interaction type and interaction time, dynamic weights are assigned to the multimodal semantic features of each exhibition content; wherein, different interaction types correspond to preset base weights, and the weights decrease with the interval between the interaction time and the current time in a decay function. The weighted multimodal semantic features are aggregated to generate the interest vector of the target user.
[0012] In conjunction with the first aspect above, in one possible implementation, the method for calculating the matching score includes: Based on the comprehensive content embedding vector and the interest vector, a similarity metric function is used in the embedding space to calculate the matching score between them. The similarity metric function is any one of normalized cosine similarity, scaled dot product with temperature coefficient, or weighted inner product of trainable weight matrices.
[0013] In conjunction with the first aspect above, in one possible implementation, the method for sorting all the exhibition content to be recommended includes: Obtain the matching score for each of the multiple exhibition contents to be recommended; The multiple pieces of content to be recommended are sorted in descending order according to their corresponding matching scores to generate a recommendation sequence; If two or more exhibition contents have the same matching score, they will be distinguished according to a preset secondary sorting rule, which includes at least one of the exhibition contents’ freshness, popularity, or geographical distance.
[0014] Secondly, this application provides an exhibition content recommendation system based on big data analysis, comprising: a collection module, an analysis module, a matching module, and a recommendation module; wherein, the collection module is used to collect multimodal raw data of the exhibition content to be recommended; the analysis module is used to encode the multimodal raw data respectively through a pre-trained deep neural network model to obtain corresponding text feature vectors, image feature vectors, and audio feature vectors; based on the text feature vectors, image feature vectors, and audio feature vectors, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network; the number of effective historical interaction events of the target user within a preset time window is obtained, and the number of effective historical interaction events is compared with a preset interaction threshold to determine the user type of the target user; based on the user type of the target user, an interest vector of the target user is generated; the matching module is used to calculate the matching degree score between the comprehensive content embedding vector and the interest vector; the recommendation module is used to sort all the exhibition content to be recommended according to the matching degree score, and push a predetermined number of the top-ranked exhibition content as recommendation results to the target user.
[0015] Thirdly, a device for recommending exhibition content based on big data analysis is provided, comprising: a communication unit and a processing unit; the communication unit is used to collect multimodal raw data of the exhibition content to be recommended; encode the multimodal raw data respectively through a pre-trained deep neural network model to obtain corresponding text feature vectors, image feature vectors, and audio feature vectors; calculate a comprehensive content embedding vector based on the text feature vectors, image feature vectors, and audio feature vectors through a multimodal adaptive fusion network; obtain the number of valid historical interaction events of the target user within a preset time window, and compare the number of valid historical interaction events with a preset interaction threshold to determine the user type of the target user; the processing unit generates an interest vector of the target user based on the user type of the target user; calculates the matching degree score between the comprehensive content embedding vector and the interest vector; sorts all the exhibition content to be recommended according to the matching degree score, and pushes a predetermined number of the top-ranked exhibition content as the recommendation result to the target user.
[0016] The exhibition content recommendation method, system, and device based on big data analysis provided in this application can construct a dynamic and accurate personalized recommendation engine by integrating multi-source heterogeneous data (such as user historical behavior, real-time interaction, exhibit metadata, spatiotemporal context, etc.), effectively solving the core pain points of traditional exhibition recommendation such as "information overload," "inaccurate interest matching," "difficulty in covering cold-start users," and "homogeneous recommendation results." This technology application utilizes big data processing technology to clean, model, and extract features from massive user-exhibit interaction logs, and then combines deep learning or collaborative filtering algorithms to generate high-dimensional user interest vectors and exhibit content embeddings; furthermore, it introduces a context-aware mechanism (such as time, location, device type) to achieve adaptive fusion, and adopts a learnable similarity measurement function to improve matching accuracy; finally, it is supplemented by secondary ranking strategies (such as freshness, popularity, geographical distance) to enhance recommendation diversity and scenario adaptability. Its beneficial effects are reflected in the following aspects: First, it significantly improves recommendation relevance and user satisfaction, increases exhibit exposure efficiency and visit duration; second, it effectively alleviates the cold start problem for new users and expands the service coverage; third, it optimizes the allocation of exhibition resources, helps curators understand audience preferences, and feeds back into content planning and spatial layout; and fourth, it builds a data-driven intelligent exhibition ecosystem, promoting the evolution of cultural exhibitions from "one-size-fits-all" to "personalized," combining technological advancement with commercial practicality.
[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description
[0018] Figure 1 A system architecture diagram of an exhibition content recommendation system based on big data analysis provided for embodiments of this application; Figure 2 A flowchart illustrating the exhibition content recommendation system method based on big data analysis provided in this application embodiment; Figure 3 This is a flowchart illustrating a comprehensive content embedding vector analysis method provided in an embodiment of this application. Detailed Implementation
[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] The exhibition content recommendation method based on big data analysis provided in this application can be applied to exhibition content recommendation systems based on big data analysis, such as... Figure 1 As shown, the communication system includes: a data acquisition module, an analysis module, a matching module, and a recommendation module; The data acquisition module is used to collect multimodal raw data of the content to be recommended for exhibition. The analysis module is used to encode the multimodal raw data using a pre-trained deep neural network model to obtain the corresponding text feature vectors, image feature vectors, and audio feature vectors. Based on text feature vectors, image feature vectors, and audio feature vectors, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network. Obtain the number of valid historical interaction events of the target user within a preset time window, and compare the number of valid historical interaction events with a preset interaction threshold to determine the user type of the target user; Generate the target user's interest vector based on the target user's user type; The matching module is used to calculate the matching score between the comprehensive content embedding vector and the interest vector; The recommendation module sorts all exhibition content to be recommended based on matching scores, and pushes a predetermined number of the top-ranked exhibition content as recommendation results to the target users.
[0021] To address the technical problems of poor recommendation accuracy and timeliness caused by static user interest modeling and insufficient multimodal content feature fusion in existing technologies, this application provides an exhibition content recommendation method based on big data analysis. This method includes: collecting multimodal raw data of the exhibition content to be recommended; encoding the multimodal raw data using a pre-trained deep neural network model to obtain corresponding text feature vectors, image feature vectors, and audio feature vectors; calculating a comprehensive content embedding vector based on the text feature vector, image feature vector, and audio feature vector through a multimodal adaptive fusion network; obtaining the number of valid historical interaction events of the target user within a preset time window and comparing the number of valid historical interaction events with a preset interaction threshold to determine the user type of the target user; generating an interest vector of the target user based on the user type of the target user; calculating the matching score between the comprehensive content embedding vector and the interest vector; ranking all exhibition content to be recommended according to the matching score, and pushing a predetermined number of the top-ranked exhibition content as recommendation results to the target user. Based on this, this application significantly improves the accuracy and generalization ability of exhibition recommendations through multimodal data fusion and adaptive user type modeling. This solution utilizes pre-trained deep networks to extract features from text, images, and audio respectively, and generates unified content embeddings through a multimodal adaptive fusion network, fully preserving the semantic integrity of heterogeneous information. Simultaneously, based on a comparison of the number of effective interaction events within a preset time window and a threshold, it accurately distinguishes between cold-start and non-cold-start users, dynamically generating interest vectors adapted to their behavioral patterns. This mechanism effectively mitigates recommendation bias caused by data sparsity for new users, avoiding "information cocoons." The matching degree calculation combines deep alignment of content and interests, resulting in ranking results that more closely reflect real preferences. The overall solution balances personalization, diversity, and system robustness, improving user experience and platform operational efficiency.
[0022] like Figure 2 As shown in the embodiments of this application, the exhibition content recommendation method based on big data analysis includes: S201. Collect multimodal raw data of the exhibition content to be recommended.
[0023] The multimodal raw data includes at least text descriptions, image data, and audio narration. Text descriptions summarize the core theme, exhibit information, and cultural background of the exhibition; image data visually presents the exhibition's visual elements, such as exhibits, exhibition hall layout, or promotional posters; and audio narration provides guided audio tours recorded by professional guides to convey deeper narratives and emotions.
[0024] In some implementations, text descriptions can be structurally extracted from museum websites, digital archives, or curatorial documents; image data can be obtained by systematically photographing exhibits and exhibition halls with high-resolution photography equipment, or by calling digital image resources provided by partner institutions; audio narration can be recorded in a recording studio by professional broadcasters based on standard narration scripts, and then processed with noise reduction and equalization to ensure clear sound quality.
[0025] It should be noted that, to ensure the effectiveness of subsequent multimodal fusion, the collected text, image, and audio data should be precisely aligned for the same exhibition unit (such as a single exhibit or a specific exhibition area), and a unified data identifier should be established to achieve cross-modal association. Simultaneously, all raw data must undergo standardized preprocessing after collection, including but not limited to text segmentation and cleaning, image size normalization and color correction, and audio sampling rate unification and silence clipping, thereby providing high-quality input for subsequent feature encoding and embedding generation.
[0026] S202. Encode the multimodal raw data using a pre-trained deep neural network model to obtain the corresponding text feature vector, image feature vector, and audio feature vector.
[0027] The text feature vector, image feature vector, and audio feature vector are obtained by encoding the multimodal raw data into the corresponding pre-trained deep neural network model. Each modal encoder runs independently but has the same output dimension to support subsequent alignment and fusion in a unified semantic space.
[0028] In some implementations, text feature vectors are generated by encoding exhibition text descriptions using pre-trained language models based on Transformer architectures (such as BERT, RoBERTa, or DeBERTa), and the [CLS] tag or the context representation after average pooling is taken as the final text embedding; image feature vectors are obtained by extracting global features from the input image after forward propagation using a visual Transformer or convolutional neural network (such as ResNet-50); audio feature vectors are generated by encoding audio narration waveforms or Mel spectrograms using pre-trained audio models (such as Wav2Vec 2.0, HuBERT, or Audio Spectrogram Transformer, AST), and outputting a fixed-length acoustic-semantic joint representation using a temporal pooling strategy.
[0029] It should be noted that although each modality uses a heterogeneous network structure for feature extraction, the output feature vectors are all constrained to the same dimension and mapped to a shared semantic embedding space through subsequent linear projection layers. In addition, to improve cross-modal semantic consistency, these pre-trained models can be fine-tuned on relevant corpora in the exhibition domain or jointly optimized under a multi-task learning framework, thereby enhancing the ability to understand cultural, artistic and historical contexts.
[0030] S203. Based on text feature vectors, image feature vectors, and audio feature vectors, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network.
[0031] Among them, the multimodal adaptive fusion network uses a nonlinear formula containing an exponential function and a normalization term to dynamically allocate the fusion weights of each mode.
[0032] It should be noted that the fusion process is not a simple splicing or multi-head attention mechanism, but rather achieves dynamic balance between modalities through a soft selection strategy. Its core innovation lies in modeling the weight generation as a differentiable normalized exponential function, which not only preserves gradient transferability to support end-to-end training, but also avoids the limitations of manually setting fixed fusion rules. In addition, this fusion network can be seamlessly integrated into the subsequent matching degree calculation module, ensuring that the comprehensive embedding vector has optimal discriminability and generalization ability in a unified semantic space.
[0033] S204. Obtain the number of valid historical interaction events of the target user within a preset time window, and compare the number of valid historical interaction events with a preset interaction threshold to determine the user type of the target user.
[0034] The number of valid historical interaction events refers to the total number of interaction behaviors that the target user has with the exhibition content within a preset time window (such as the last 30 days) and that meet the validity judgment criteria. These include, but are not limited to, explicit or implicit feedback such as clicking, browsing time exceeding a threshold, collecting, sharing, or completing audio guide listening. This number is used to compare with a preset interaction threshold. If it is less than the threshold, the target user is determined to be a cold start user; otherwise, it is determined to be a non-cold start user.
[0035] User types include cold-start users and non-cold-start users. Cold-start users represent users whose historical interaction data is insufficient to support reliable interest modeling, while non-cold-start users represent users whose historical interaction data is sufficient to support reliable interest modeling.
[0036] In some implementations, the preset time window can be dynamically adjusted according to the platform's activity cycle (e.g., extended to 45 days before and after holidays), while the preset interaction threshold can be set based on the quantile of the full user interaction distribution (e.g., taking the number of interactions corresponding to the 20th percentile as the threshold), or the critical point that causes a significant drop in recommendation performance can be determined through offline experiments. In addition, different interaction types can be assigned different weights (e.g., "completing the tour" is counted as 3 valid events, and "short click" is counted as 1 event) to more precisely reflect the intensity of user interest expression.
[0037] It is important to note that accurate user type identification is a crucial prerequisite for subsequent recommendation strategy bifurcation: for cold-start users, the system will prioritize recommendations based on content-side multimodal features and group preferences to avoid performance degradation caused by underfitting of personalized models; while for non-cold-start users, collaborative filtering or multi-task sequence modeling mechanisms based on deep interest modeling can be enabled. This determination logic not only improves the system's robustness but also provides a quantifiable and engineerable solution to the cold-start problem.
[0038] S205. Generate the target user's interest vector based on the target user's user type.
[0039] Specifically, the interest vectors of target users are generated using a differentiated strategy based on their user type: for cold-start users, due to the lack of sufficient historical interaction data, their interest vectors are constructed based on prior information (such as preference tags filled in during registration, demographic characteristics, device / regional context, or the average embedding of globally popular exhibitions); while for non-cold-start users, their interest vectors are dynamically generated by aggregating the comprehensive content embedding vectors corresponding to the exhibition content they have interacted with within a preset time window (such as weighted average, attention mechanism, or sequence modeling) to reflect their personalized preferences.
[0040] In some implementations, the interest vectors of cold-start users can be generated by a lightweight prior embedding mapping module. For example, the cultural preference distribution of the user's city, the centers of typical interest clusters corresponding to the age group, or the average embeddings of the platform's global Top-K popular exhibitions can be used as the initial interest representation. Meanwhile, the interest vectors of non-cold-start users can be weighted and fused from content embeddings in historical interaction sequences through a time-aware attention network, with recent interactions given higher weights, thereby capturing interest evolution trends. Furthermore, both types of interest vectors are projected into the same semantic space as the exhibition content embeddings to support subsequent similarity calculations.
[0041] It should be noted that this dual-path interest modeling mechanism effectively alleviates the "no data dilemma" in the cold start phase of the recommendation system, while ensuring that non-cold start users receive high-precision personalized services. More importantly, the generation process of the two types of interest vectors maintains semantic alignment, making it possible to uniformly use dot product or cosine similarity for candidate ranking in the future, thus avoiding matching deviations caused by inconsistent representation spaces.
[0042] S206. Calculate the matching score between the integrated content embedding vector and the interest vector; the matching score represents the correlation between each exhibition content and the target user's interest vector.
[0043] The matching score is obtained by calculating the similarity measure between the integrated content embedding vector and the target user interest vector, which is used to quantify the semantic relevance between specific exhibition content and user interests. This score is usually performed in a unified embedding space to ensure that the exhibition representation after cross-modal fusion is comparable to the user interest representation.
[0044] It should be noted that the matching score is not only the core input of the ranking module, but also the basic signal for subsequent re-ranking, diversity control and interpretability generation. More importantly, since the interest vectors of both cold-start users and non-cold-start users are mapped to the same semantic space, this matching mechanism can seamlessly support the unified recommendation process for all users, avoid introducing system architecture complexity due to different user types, and thus maintain the simplicity and scalability of engineering deployment while ensuring personalized effects.
[0045] S207. Sort all the exhibition content to be recommended according to the matching score, and push the top-ranked exhibition content to the target user as the recommendation result.
[0046] Based on the above technical solutions, the exhibition content recommendation method based on big data analysis provided in this application generally suffers from static interest modeling and coarse multimodal information fusion, making it difficult to accurately reflect users' dynamic preferences and the rich semantics of exhibition content. This is especially true when user interaction data is sparse or during the cold start phase, where recommendation performance significantly declines. This application introduces a multimodal adaptive fusion network, employing a nonlinear formula including exponential functions and normalization terms to dynamically adjust the fusion weights of text, image, and audio modalities, thereby generating a more discriminative and semantically consistent comprehensive content embedding vector. Simultaneously, it distinguishes between cold-start and non-cold-start users based on the number of effective interaction events within a preset time window, and constructs differentiated interest vectors based on prior knowledge or historical behavior accordingly, achieving more accurate and robust interest representation. This technical solution effectively overcomes the shortcomings of traditional methods in terms of recommendation accuracy, timeliness, and generalization ability. It not only improves the system's understanding of multimodal content but also enhances recommendation performance in low-interaction scenarios, ultimately achieving a more personalized, more relevant, and better-experienced exhibition content recommendation service.
[0047] In one possible implementation of this application embodiment, the above-mentioned S202 can be specifically described as follows: A pre-trained language model based on the Transformer architecture is used to semantically encode the text description, generating a dimension of... Text feature vectors ; Visual features are extracted from image data using convolutional neural networks or visual Transformer models, generating a dimension of... Image feature vectors ; A pre-trained audio model combining time-frequency analysis and deep sequence modeling is used to perform acoustic-semantic joint encoding on the audio narration, generating a dimension of [missing information]. audio feature vector ; In this context, text, image, and audio feature vectors are all mapped to the same semantic embedding space (i.e., dimension 1). Unified vector space Its core lies in achieving comparability and compatibility of cross-modal representations through a shared semantic alignment mechanism. This alignment is not a simple splicing or posterior projection, but rather relies on the joint optimization objective of pre-trained multimodal models (such as contrastive learning loss or cross-modal reconstruction loss) to make different modalities converge to a consistent abstract representation at the semantic level.
[0048] In some implementations, text encoding uses [CLS] vectors or sequence pooling outputs from a Transformer-based pre-trained language model (such as BERT or DeBERTa); image encoding can use a ViT (Vision Transformer) or ResNet+Transformer hybrid architecture, adding an adapter module in the last layer to align to the target dimension. Audio encoding employs time-frequency analysis (such as Mel spectrograms) as input to a deep sequence modeling network. For example, intermediate layer features from pre-trained Wav2Vec2.0 or Hubert are averaged over time and then mapped through a linear projection layer to... The three modules share a lightweight semantic alignment header (such as a two-layer MLP) and are jointly fine-tuned under a multi-task contrastive learning framework to ensure semantic consistency among modalities.
[0049] It is important to note that "mapping to the same semantic embedding space" does not merely refer to identical quantitative dimensions, but more importantly, semantic equivalence: that is, when two inputs from different modalities describe the same exhibition content (such as the text description and corresponding image of "Flying Apsaras in Dunhuang Murals"), their generated feature vectors should have high similarity in the embedding space. This characteristic is explicitly constrained by cross-modal contrastive loss (such as InfoNCE), significantly improving the reliability of subsequent fusion and matching stages. Furthermore, the unified space design greatly simplifies the interface complexity of downstream modules (such as adaptive fusion networks and interest matching modules), avoiding engineering redundancy caused by modal heterogeneity.
[0050] Based on the above technical solutions, existing semantic modeling of exhibition audio narration generally adopts traditional acoustic models or simple post-transcription processing, which makes it difficult to simultaneously capture the temporal dynamics and deep semantic structure of speech. Especially in the presence of background noise, speech rate variations, or technical terms, this can easily lead to sound-meaning mismatch and shallow feature representation, thus affecting the contribution of audio modalities in multimodal recommendation. This application specifically employs a pre-trained audio model combining "time-frequency analysis + deep sequence modeling" for joint acoustic-semantic encoding: time-frequency analysis (such as STFT or Mel spectrum) preserves the local time-frequency structure of speech, providing high-fidelity acoustic priors for the model; deep sequence models (such as Transformer or LSTM) model long-range dependent semantic evolution, achieving end-to-end mapping from acoustic signals to semantic representations. This approach avoids the two-stage error accumulation of traditional "transcription followed by encoding," significantly improving the semantic richness and context-awareness of audio features; simultaneously, it unifies the output dimension with text and image modalities. dThe feature vectors facilitate subsequent multimodal adaptive fusion, ensuring that each modality is semantically aligned and its weights are interpretable. This allows the system to more accurately understand the core content of the audio narration (such as exhibit background and historical context), enhance cross-modal reasoning capabilities, and ultimately improve the relevance, professionalism, and user immersion of the recommendation results, making it particularly suitable for knowledge-intensive exhibition scenarios.
[0051] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S203 can be specifically implemented through the following S301-S303, which are explained in detail below: S301, Transfer text feature vectors Image feature vectors and audio feature vectors The input is fed into a multimodal adaptive fusion network.
[0052] Among them, the multimodal adaptive fusion network is used to dynamically integrate text feature vectors. Image feature vectors and audio feature vectors To generate a unified and semantically rich integrated content embedding vector. The core of this network lies in its ability to adaptively allocate fusion weights based on the information confidence or contextual relevance of each modality, rather than using fixed weighting or simple splicing strategies, thereby effectively addressing the issue of modal integrity and quality differences among different exhibition items.
[0053] It is important to note that the key to multimodal adaptive fusion lies not only in structural design, but also in ensuring that the features of each modality are aligned in the same semantic embedding space during the aforementioned steps. If semantic consistency is not achieved, attention or weighting mechanisms may fail due to distribution shifts between modalities, or even introduce noise. Furthermore, to prevent the model from over-relying on a dominant modality (such as text), modality dropout or contrastive regularization terms can be introduced during training to force the network to learn complementary information, thereby ensuring the overall embedding vector after fusion. It can comprehensively and evenly reflect the multidimensional semantics of the exhibition content.
[0054] S302, Multimodal Adaptive Fusion Network is based on the user's current context state vector Using a nonlinear mapping function Calculate the context-dependent activation values for each modality. .
[0055] in, Characterizing multimodal data types, , Vector concatenation; nonlinear mapping function It is implemented by a multilayer perceptron containing at least two fully connected layers and a non-linear activation function, and the context state vector It includes at least one of the following: user access time, geographical location, and device interaction mode.
[0056] Wherein, the context state vector This is used to model the dynamic behavioral context of users in the current session (such as access time, geolocation, device interaction mode, etc.), and through a non-linear mapping function. Combine it with the feature vectors of each modality ( This process involves fusing the data to generate modality-specific context-aware activation values. This activation value can be regarded as a "context modulation signal" of the original modal features, which is used for the generation of subsequent adaptive fusion weights or feature recalibration.
[0057] In some implementations, nonlinear mapping functions A multilayer perceptron (MLP) consisting of two fully connected layers (containing ReLU or Swish activation) is used as the input, which is a concatenated vector. The output is a scalar activation value. .
[0058] It should be noted that the context state vector It doesn't just contain single-dimensional information, but integrates at least one of the following: user access time (e.g., hourly segment encoding, holiday markers), geographic location (e.g., city / venue-level embedding), and device interaction mode (e.g., touchscreen swipe frequency, dwell time distribution). Essentially, it transforms discrete / continuous behavioral signals into dense vector representations, thereby supporting end-to-end joint optimization. More importantly, this mechanism achieves modal-context decoupling modeling: that is, the modal preference of the same user for the same exhibition content may change significantly in different scenarios (e.g., browsing on a mobile device late at night vs. AR guided tour of an exhibition hall during the day) (e.g., relying more on audio narration than text and images). Its nonlinear mapping capability can precisely capture such dynamic interaction patterns, avoiding the biases caused by static fusion strategies.
[0059] S303. Perform a weighted summation of the feature vectors of each modality to obtain the comprehensive content embedding vector. .
[0060] in, The specific process for obtaining the weighting coefficients is as follows: The fusion weight coefficients for each modality are calculated using the Softmax function. ;in, For temperature coefficient, ; Summation symbol The dummy variable in the code represents the traversal of the set of all modal types. The elements in.
[0061] It should be pointed out that, This is a temperature coefficient used to control the "sharpness" of the weight distribution: when At that time, all Approaching a uniform distribution (equal-weight fusion); when At that time, the weights are concentrated at the maximum. The corresponding modality (hard selection). This design allows the model to smoothly adjust between "conservative fusion" and "focus on dominant modality" to adapt to the information reliability requirements of different scenarios.
[0062] Based on the above technical solutions, existing multimodal fusion methods often employ simple concatenation, average weighting, or fixed-weight fusion, which struggles to adapt to the dynamic preference differences of different users in different contexts (such as access time, geographical location, and device interaction mode), resulting in a "one-size-fits-all" fusion result and weakening the accuracy of personalized recommendations. This application proposes a method based on user context state vectors... Adaptive fusion mechanism: through nonlinear mapping function Dynamically generate activation values for each modality. Then, a weighted sum is obtained to obtain the comprehensive embedding. The weight Depend on Implicit learning, in essence, models the user context as conditional input for fusion decision-making. The key to this design lies in: Employing a perceptron structure with at least two fully connected layers and nonlinear activation, this approach possesses strong nonlinear fitting capabilities, enabling it to capture the complex mapping relationship between context and modality importance (e.g., nighttime access relies more on visual image information, while mobile users focus more on audio summaries). Compared to static fusion, this scheme achieves dynamic modality scheduling "based on person, time, and location," significantly improving the semantic adaptability and personalized expressive power of the fused representation; simultaneously, context vectors... Explicitly integrating key interaction signals such as time, location, and device enables the system to respond to behavioral heterogeneity in real-world scenarios, avoiding modal redundancy or information masking. Ultimately, while ensuring computational efficiency, it significantly enhances the relevance, contextual consistency, and user experience satisfaction of recommendation results.
[0063] In one possible implementation of this application embodiment, the above-mentioned S204 can be specifically implemented by the following S401-S402, which will be described in detail below: S401. Collect the valid historical interaction events of the target user within a preset time window and count the number of valid historical interaction events.
[0064] S402. Compare the number of valid historical interaction events with the preset interaction threshold; If the number of valid historical interaction events is less than the preset interaction threshold, the target user is determined to be a cold start user; otherwise, the target user is determined to be a non-cold start user.
[0065] Based on the above technical solutions, in exhibition recommendation systems, adopting a uniform interest modeling strategy for all users can lead to cold-start users (with sparse historical interactions) being incorrectly assigned high-confidence interest vectors, causing recommendation bias or even "information cocoons." Over-reliance on interaction data, on the other hand, can prevent new or inactive users from receiving effective recommendations for extended periods, reducing platform retention. This application addresses this by setting a preset interaction threshold and combining it with the number of valid historical interaction events within a preset time window (specifically limited to active behaviors such as clicks, browsing, favorites, ratings, and sharing, excluding automatically triggered or abnormal sessions) to objectively quantify user activity and accurately distinguish between cold-start and non-cold-start users. The core advantages of this solution are: firstly, rigorous definition, with "valid events" filtering out noisy data to avoid false signals interfering with judgment; secondly, configurable thresholds, dynamically adjustable according to business scenarios (e.g., lowering the threshold during large exhibitions to quickly activate new users), improving system adaptability; and thirdly, interpretable decision-making, based on a binary classification logic with clear quantitative comparisons, facilitating monitoring and attribution analysis. Thus, the system can enable prior knowledge-driven interest generation for cold-start users (such as based on exhibit categories, regional preferences, and time-of-day patterns), and personalized behavior modeling for existing users, achieving a refined "divide and conquer" recommendation strategy. This not only significantly improves the accuracy and satisfaction of first-time recommendations for new users, but also ensures the depth and diversity of recommendations for existing users, enhancing the overall robustness and business sustainability of the system.
[0066] In one possible implementation of this application embodiment, the above-mentioned S205 can be specifically described as follows: If the target user is a cold-start user, the method for generating the interest vector specifically includes the following process: Obtain the user's prior information, which includes at least one of the following: attribute information provided by the user during registration, demographic group labels, or default interest distributions preset by the system. Prior information is embedded and encoded to obtain the corresponding prior feature representation; Based on prior feature representation, the initial interest vector of the target user is generated through the initialization mapping module.
[0067] Among them, the user interest representation vector is constructed by aggregating the embedded exhibition content that users have interacted with in the past (such as clicked, favorited, or exhibits that have been viewed for a long time) and combining it with its contextual information (such as visit time and location). It usually uses attention mechanisms or temporal modeling networks (such as LSTM and Transformer) to encode the behavioral sequence in order to capture the user's long-term preferences and short-term interest dynamics. This vector serves as the core basis for personalized matching and directly affects the relevance and novelty of the recommendation results.
[0068] It should be noted that the quality of user interest representation is highly dependent on the sparsity and noise level of historical behavioral data. In cold start or low-frequency user scenarios, the interest representation vector may lack sufficient signals to support effective matching. In this case, it is necessary to introduce auxiliary information (such as demographic features and social relationships) or adopt strategies such as meta-learning and cross-domain transfer to enhance generalization ability. In addition, in order to avoid the model overfitting explicit feedback and ignoring potential interests, in practice, implicit feedback (such as scrolling speed and audio playback completion rate) is often combined with a multi-task learning framework to jointly optimize click-through rate, dwell time and diversity indicators, thereby constructing a more robust and comprehensive user interest profile.
[0069] If the target user is not a cold-start user, the method for generating the interest vector specifically includes the following process: Obtain the user's historical interaction event sequence. The historical interaction events include the user's interaction records with multiple exhibition contents within a preset time window. Each interaction record includes the identifier of the interacted exhibition content, the corresponding interaction type, and the interaction time. Based on the exhibition content identifier, obtain the multimodal semantic features of each exhibition content. The multimodal semantic features include at least one of the following: text description embedding vector, image visual feature vector, or structured label vector. Based on the interaction type and interaction time, dynamic weights are assigned to the multimodal semantic features of each exhibition content; different interaction types correspond to preset base weights, and the weights decrease with the interval between the interaction time and the current time in a decay function. The weighted multimodal semantic features are aggregated to generate the target user's interest vector.
[0070] The dynamic weight allocation mechanism is based on modeling the decay relationship between interaction type (such as click, favorite, share, long dwell time) and interaction time from the current moment: for each type of interaction... Preset base weights (Reflecting the information intensity of the behavior, such as "share" > "click"), and then introducing a time decay function (such as exponential decay). or power-law decay Dynamic adjustments are made to obtain the final weights. ; in, The interval between the time the interaction occurred and the current time (in hours or days). The decay rate hyperparameter is used to weight and aggregate the multimodal semantic features (text, image, structured tags) of the corresponding exhibition content, thereby generating a personalized signal of user interest in the content.
[0071] It should be noted that the semantic discriminative power of interaction types needs to be strongly coupled with the business scenario; for example, in a museum tour guide system, "audio narration playback completion rate" may be more indicative of interest than "image zoom count," therefore the basic weight should be... Static settings are not advisable; instead, dynamic calibration should be achieved through offline A / B testing or online reinforcement learning. Furthermore, if users engage in sudden high-value interactions (such as browsing five exhibits on the same topic late at night), relying solely on time decay may lead to the underestimation of recent low-frequency but high-intent behaviors. To address this, some implementations introduce local session-aware normalization: re-normalizing the weights of similar interactions within a sliding time window, or superimposing short-term interest gating signals (such as attention scores based on the three most recent interactions) to mitigate the "interest forgetting" bias caused by long-term decay, ensuring that the interest vector reflects historical accumulation while also responding swiftly to the latest changes in intent.
[0072] In one possible implementation of this application embodiment, the above-mentioned S206 can be specifically described as follows: Based on the integrated content embedding vector and interest vector, a similarity metric function is used in the embedding space to calculate the matching score between the two.
[0073] The similarity metric function can be any one of normalized cosine similarity, scaled dot product with temperature coefficient, or weighted inner product of trainable weight matrices.
[0074] It should be noted that the integrated content embedding vector and the interest vector have the same dimension.
[0075] The similarity metric function is used in the embedding space to calculate the composite content vector. User interest vector The matching score between them can be expressed as normalized cosine similarity (i.e., the inner product divided by the product of the moduli): This form naturally satisfies the range. This facilitates subsequent sorting and threshold determination; when a temperature coefficient is introduced... At this point, the similarity can be further scaled to control the sharpness of the match: This is equivalent to applying smoothing / focusing control to the Softmax weight distribution.
[0076] In some implementations, to enhance the model's ability to model nonlinear relationships, the similarity function can be replaced with a learnable bilinear interaction: ;in, The weight matrix is trainable, allowing heterogeneous interactions between features of different dimensions; furthermore, a multilayer perceptron (MLP) can be used to concatenate the vectors. It maps to scalar scores to capture higher-order cross features; in addition, in cold start scenarios, similarity scores can be fused with rule-based prior scores (such as content popularity and tag coverage) to form a hybrid matching strategy and improve the stability of initial recommendations.
[0077] It should be noted that dimensional consistency and distribution alignment of the embedding vectors are prerequisites for effective similarity calculation. and If two vectors are trained independently without sharing an optimization objective (such as separate training of the content encoder and the user encoder), their vector spaces may have a systematic shift, causing cosine similarity to fail. Therefore, in practice, a joint contrastive learning framework (such as InfoNCE loss) is usually adopted to force positive sample pairs (user-interaction content) to be close together in the embedding space and negative sample pairs to be far apart, thereby implicitly aligning their semantic spaces. At the same time, it is necessary to monitor the vector norm distribution (such as mean and variance) regularly to avoid similarity values being distorted due to gradient explosion or lack of normalization, which would affect the reliability of the ranking.
[0078] Based on the above technical solutions, in multimodal recommendation, if the matching degree calculation of the integrated content embedding vector and the user interest vector uses simple cosine similarity or dot product, it is easily affected by differences in vector scale, modal bias (such as image feature amplitude being much larger than text), and inconsistent semantic granularity, leading to matching results deviating from true relevance. For example, high-dimensional sparse text vectors may be dominated by low-dimensional dense image vectors, or short-term user interests may be masked by long-term preferences. This application proposes to use a learnable similarity measurement function (such as normalized cosine similarity, scaled dot product with temperature coefficient, or weighted inner product of training weight matrices) in the embedding space. Its core value lies in the fact that this function can adaptively calibrate the distribution characteristics and semantic importance of different modal embeddings, so that the matching process not only reflects geometric proximity but also reflects task-oriented relevance discrimination ability. For example, the temperature coefficient can adjust the sharpness of the similarity distribution and improve the discriminative power; the weighted matrix allows the model to learn "which dimensions are more critical for matching," achieving fine-grained semantic alignment. More importantly, this solution works in conjunction with the aforementioned dynamic fusion and cold start recognition modules: for cold start users, the temperature can be lowered to enhance generalization; for highly active users, the temperature can be increased to strengthen personalized discrimination. Ultimately, the matching score more accurately reflects "whether the content matches the current user's intent," significantly improving recommendation ranking quality, reducing false positives, and enhancing the system's generalization stability and interpretability across modalities and user scenarios.
[0079] In one possible implementation of this application embodiment, the above-mentioned S207 can be specifically implemented by the following S501-S502, which will be described in detail below. S501. Obtain the matching score for each of the multiple exhibition contents to be recommended; S502. Arrange multiple exhibition contents to be recommended in descending order according to their corresponding matching scores to generate a recommendation sequence; If two or more exhibition contents have the same matching score, they will be distinguished according to preset secondary sorting rules. The secondary sorting rules include at least one of the following: freshness, popularity, or geographical distance of the exhibition contents. For example, priority may be given to improving the freshness of the content (sorted in reverse order of the content's online time), or enhancing the balance of popularity (combined with the attenuation weighting of recent exposure / interaction frequency), or optimizing spatial affinity (sorted in ascending order of geographical distance between the user's current location and the exhibition hall), thereby injecting diversity, timeliness, and scene adaptability while maintaining core relevance.
[0080] In some implementations, secondary ranking can employ a multi-objective weighted fusion strategy: defining secondary feature vectors. And introduce learnable weights (Calculate secondary scores by dynamically adjusting offline optimization or online bandit strategies) ;in, For Min-Max or Z-score normalization operations, more advanced solutions model secondary ranking as a constrained optimization problem: maximizing the weighted sum of secondary indicators while ensuring that the primary ranking score is not lower than a threshold, or using a Pareto front screening mechanism to ensure that the recommendation results achieve an approximately optimal balance across multiple dimensions.
[0081] Based on the above technical solutions, ranking exhibitions solely based on a single matching score can easily lead to homogenized results: when multiple exhibits highly match user interests (e.g., all belonging to the same theme, era, or artist), the system cannot further distinguish their priorities, resulting in a lack of recommendation diversity and a monotonous user experience, especially affecting the exposure of long-tail content and the depth of knowledge exploration. This application introduces secondary ranking rules (at least one of freshness, popularity, and geographical distance), performing refined hierarchical ranking when matching scores are the same. Essentially, it decouples and models "relevance" and "contextual value": freshness can improve the visibility of newly added exhibits and promote the perception of content updates; popularity reflects group consensus and enhances social credibility; geographical distance is suitable for offline guided tour scenarios and optimizes the efficiency of the visitor flow. The advantages of this solution are: First, the rules are configurable and clearly business-oriented, allowing operations personnel to dynamically adjust secondary weights based on stage goals (such as attracting new products or activating regional users); second, it avoids hard threshold truncation, ensuring a smooth transition when primary matching scores are similar, thus guaranteeing the continuity and fairness of the ranking; third, it is compatible with both cold start and non-cold start strategies, emphasizing geographic / novelty for new users and combining historical preferences with weighted popularity for existing users. Therefore, the recommended sequence is not only "relevant" but also "hierarchical, surprising, and scenario-adaptable," significantly improving user dwell time, interaction depth, and the health of the platform ecosystem, achieving an upgrade from "precise matching" to "intelligent guidance."
[0082] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.
Claims
1. A method for recommending exhibition content based on big data analysis, characterized in that, include: Collect multimodal raw data of the exhibition content to be recommended, wherein the multimodal raw data includes at least text description, image data and audio narration; The multimodal raw data are encoded by a pre-trained deep neural network model to obtain the corresponding text feature vector, image feature vector and audio feature vector; Based on the text feature vector, image feature vector, and audio feature vector, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network; wherein, the multimodal adaptive fusion network uses a nonlinear formula containing an exponential function and a normalization term to dynamically allocate the fusion weights of each modality; The number of valid historical interaction events of the target user within a preset time window is obtained, and the number of valid historical interaction events is compared with a preset interaction threshold to determine the user type of the target user. The user type includes cold-start users and non-cold-start users. Cold-start users represent users whose historical interaction data is insufficient to support reliable interest modeling, and non-cold-start users represent users whose historical interaction data is sufficient to support reliable interest modeling. The number of valid historical interaction events refers to the total number of interaction behaviors of the target user with the exhibition content within the preset time window that meet the validity determination criteria. Based on the user type of the target user, an interest vector of the target user is generated; wherein, the interest vector of a cold-start user is generated based on prior information, and the interest vector of a non-cold-start user is generated based on its historical interaction content. Calculate the matching score between the integrated content embedding vector and the interest vector; the matching score represents the correlation between each exhibition content and the target user's interest vector; All exhibition content to be recommended is sorted according to the matching score, and a predetermined number of the top-ranked exhibition content is pushed to the target user as the recommendation result.
2. The exhibition content recommendation method based on big data analysis according to claim 1, characterized in that, The encoding of the multimodal raw data includes: A pre-trained language model based on the Transformer architecture is used to semantically encode the text description, generating a dimension of... Text feature vectors ; Visual features are extracted from the image data using a convolutional neural network or a visual Transformer model, generating a dimension of... Image feature vectors ; A pre-trained audio model combining time-frequency analysis and deep sequence modeling is used to perform acoustic-semantic joint encoding on the audio narration, generating a dimension of [missing information]. audio feature vector ; The text feature vector, image feature vector, and audio feature vector are all mapped to the same semantic embedding space.
3. The exhibition content recommendation method based on big data analysis according to claim 2, characterized in that, The method for obtaining the integrated content embedding vector includes: The text feature vector Image feature vectors and audio feature vectors The input is fed into a multimodal adaptive fusion network; The multimodal adaptive fusion network is based on the user's current context state vector. Using a nonlinear mapping function Calculate the context-dependent activation values for each modality. ;in, Characterizing multimodal data types, , This indicates vector concatenation; The integrated content embedding vector is obtained by weighted summation of the feature vectors of each modality. ;in, For fusion weighting coefficients; Wherein, the nonlinear mapping function It is implemented by a multilayer perceptron containing at least two fully connected layers and a nonlinear activation function, and the context state vector It includes at least one of the following: user access time, geographical location, and device interaction mode.
4. The exhibition content recommendation method based on big data analysis according to claim 3, characterized in that, The method for obtaining the fusion weight coefficients includes: The fusion weight coefficients for each modality are calculated using the Softmax function. ;in, For temperature coefficient, ; Summation symbol The dummy variable in the code represents the traversal of the set of all modal types. The elements in.
5. The exhibition content recommendation method based on big data analysis according to claim 1, characterized in that, The method for determining the user type of the target user includes: Collect valid historical interaction events of the target user within a preset time window and count the number of valid historical interaction events; wherein, the valid historical interaction events include one or more of the user's actions of clicking, browsing, collecting, rating or sharing exhibition content, and exclude interaction records generated by automatic triggering or abnormal sessions; Compare the number of valid historical interaction events with a preset interaction threshold; If the number of valid historical interaction events is less than the preset interaction threshold, the target user is determined to be a cold start user; otherwise, the target user is determined to be a non-cold start user.
6. The exhibition content recommendation method based on big data analysis according to claim 1, characterized in that, The method for generating the interest vector of the cold-start user includes: Obtain the user's prior information, which includes at least one of the following: attribute information provided by the user during registration, demographic group labels, or default interest distribution preset by the system. The prior information is embedded and encoded to obtain the corresponding prior feature representation; Based on the prior feature representation, the initial interest vector of the target user is generated through the initialization mapping module; The method for generating the interest vectors of non-cold-start users includes: Obtain the user's historical interaction event sequence, which includes the user's interaction records with multiple exhibition contents within a preset time window. Each interaction record includes the identifier of the interacted exhibition content, the corresponding interaction type, and the interaction time. Based on the exhibition content identifier, multimodal semantic features of each exhibition content are obtained, wherein the multimodal semantic features include at least one of text description embedding vector, image visual feature vector, or structured label vector; Based on the interaction type and interaction time, dynamic weights are assigned to the multimodal semantic features of each exhibition content; wherein, different interaction types correspond to preset base weights, and the weights decrease with the interval between the interaction time and the current time in a decay function. The weighted multimodal semantic features are aggregated to generate the interest vector of the target user.
7. The exhibition content recommendation method based on big data analysis according to claim 1, characterized in that, The method for calculating the matching score includes: Based on the comprehensive content embedding vector and the interest vector, a similarity metric function is used in the embedding space to calculate the matching score between the two. The similarity metric function is any one of normalized cosine similarity, scaled dot product with temperature coefficient, or weighted inner product of trainable weight matrices.
8. The exhibition content recommendation method based on big data analysis according to claim 1, characterized in that, The method for sorting all the exhibition content to be recommended includes: Obtain the matching score for each of the multiple exhibition contents to be recommended; The multiple pieces of content to be recommended are sorted in descending order according to their corresponding matching scores to generate a recommendation sequence; If two or more exhibition contents have the same matching score, they will be distinguished according to a preset secondary sorting rule, which includes at least one of the exhibition contents’ freshness, popularity, or geographical distance.
9. An exhibition content recommendation system based on big data analysis, operating based on the exhibition content recommendation method based on big data analysis as described in any one of claims 1-8, characterized in that, It includes a data collection module, an analysis module, a matching module, and a recommendation module; The acquisition module is used to collect multimodal raw data of the content to be recommended for exhibition; The analysis module is used to encode the multimodal raw data using a pre-trained deep neural network model to obtain corresponding text feature vectors, image feature vectors, and audio feature vectors. Based on the text feature vector, image feature vector, and audio feature vector, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network; The number of valid historical interaction events of the target user within a preset time window is obtained, and the number of valid historical interaction events is compared with a preset interaction threshold to determine the user type of the target user. Based on the user type of the target user, generate the target user's interest vector; The matching module is used to calculate the matching score between the comprehensive content embedding vector and the interest vector; The recommendation module is used to sort all the exhibition content to be recommended according to the matching score, and push a predetermined number of the top-ranked exhibition content as the recommendation results to the target user.
10. An exhibition content recommendation device based on big data analysis, operating based on the exhibition content recommendation method based on big data analysis according to any one of claims 1-8, characterized in that, Includes a communication unit and a processing unit; The communication unit is used to collect multimodal raw data of the content to be recommended for exhibition; The multimodal raw data are encoded by a pre-trained deep neural network model to obtain the corresponding text feature vector, image feature vector and audio feature vector; Based on the text feature vector, image feature vector, and audio feature vector, a comprehensive content embedding vector is calculated through a multimodal adaptive fusion network; The number of valid historical interaction events of the target user within a preset time window is obtained, and the number of valid historical interaction events is compared with a preset interaction threshold to determine the user type of the target user. The processing unit generates an interest vector for the target user based on the user type of the target user; Calculate the matching score between the integrated content embedding vector and the interest vector; All exhibition content to be recommended is sorted according to the matching score, and a predetermined number of the top-ranked exhibition content is pushed to the target user as the recommendation result.