Product recommendation method and device based on deep learning, equipment and medium

Abstract

The product recommendation method and device based on deep learning, equipment and medium provided by the application are related to the technical field of artificial intelligence, and are suitable for the financial field and the medical field. The method comprises the following steps: acquiring multi-source product data; wherein the multi-source product data comprises at least two modes of text mode, image mode and audio mode; performing feature construction on the multi-source product data to obtain a time sequence feature matrix; projecting each feature in the time sequence feature matrix to obtain a deep product feature; performing time sequence processing on the deep product feature through a bidirectional long short-term memory model to obtain a high-order semantic feature; performing attention calculation on the high-order semantic feature to obtain a weighted context feature; and predicting the weighted context feature through a target product prediction model to obtain a target recommended product. The application can solve the problem of poor stability of product recommendation, and improve the accuracy and stability of product recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the financial and medical fields. In particular, it relates to a product recommendation method, device, or medium based on deep learning. Background Technology

[0002] In the field of artificial intelligence technology, suitable products can be recommended to users based on product data such as price and sales volume. For example, in the financial sector, insurance products (such as car insurance from Company A) can be recommended to users based on product data. Similarly, in the medical field, medical equipment (such as ultrasound machines from Medical Company B) can be recommended to users based on product data.

[0003] Currently, product recommendation technologies primarily rely on time-series models such as ARIMA and GARCH to fit price series data. These models assume that product data is a stationary linear sequence, but actual product data often exhibits non-linear characteristics, leading to extremely poor model prediction stability. For example, ARIMA can only reflect short-term linear trends and cannot cope with sudden market jumps. Summary of the Invention

[0004] The main objective of this application is to propose a product recommendation method, apparatus, device, and medium based on deep learning, which can solve the problem of poor stability in product recommendations and improve the accuracy and stability of product recommendations.

[0005] To achieve the above objectives, a first aspect of this application proposes a product recommendation method based on deep learning, the method comprising: Acquire multi-source product data; wherein the multi-source product data includes at least two of the following modalities: text modality, image modality, and audio modality; Feature construction is performed on the multi-source product data to obtain a time-series feature matrix; Projecting each feature in the time-series feature matrix yields in-depth product features; The deep product features are processed temporally using a bidirectional long short-term memory model to obtain higher-order semantic features; Attention is calculated on the higher-order semantic features to obtain weighted context features; The target recommended product is obtained by predicting the weighted contextual features using the target product prediction model.

[0006] Optionally, the step of predicting the weighted contextual features using a target product prediction model to obtain the target recommended product includes: The reference product prediction model parameters in the preset blockchain are updated based on the multi-source product data to obtain the target product prediction model parameters, and the target product prediction model is constructed based on the target product prediction model parameters; the target product prediction model parameters are constructed based on historical product data. The recommended score of the candidate product is obtained by predicting the weighted contextual features of the candidate product using the target product prediction model. The candidate products are filtered based on the recommendation scores to obtain the target recommended product.

[0007] Optionally, the step of predicting the weighted contextual features of the candidate products using the target product prediction model to obtain the recommendation score of the candidate products includes: The target product prediction model is used to predict the price of the weighted context features and the candidate products to obtain the future short-term price. The target product prediction model is used to predict the volatility of the weighted context features and the candidate products to obtain the future volatility. The target product prediction model is used to predict the future risk probability by analyzing the weighted context features and the candidate products. The recommendation score is obtained by fusing the future short-term price, the future volatility, and the future risk probability.

[0008] Optionally, projecting each feature in the time-series feature matrix to obtain deep product features includes: The category of each feature in the time-series feature matrix is ​​determined to obtain the feature category; When the feature category is a discrete feature, feature embedding is performed on the features in the time-series feature matrix to obtain product auxiliary features; When the feature category is a continuous feature, a 1D convolution is performed on the features in the time-series feature matrix to obtain the local time features of the product. All the product auxiliary features and all the product local time features are concatenated to obtain the product concatenated features, and the product concatenated features are linearly mapped to obtain the deep product features.

[0009] Optionally, the bidirectional long short-term memory model includes a forward long short-term memory model and a reverse long short-term memory model; The process of performing time-series processing on the deep product features using a bidirectional long short-term memory model to obtain high-order semantic features includes: The deep product features and the first higher-order semantic state at the first time step are extracted using the positive long short-term memory model to obtain the first higher-order semantic state at the target time step; wherein, the first time step step is before the target time step ... The deep product features and the second higher-order semantic state at the second time step are extracted using the reverse long short-term memory model to obtain the second higher-order semantic state at the target time step; wherein, the target time step is before the second time step step. The higher-order semantic features at the target time are obtained by fusing the first higher-order semantic state and the second higher-order semantic state.

[0010] Optionally, the higher-order semantic features are higher-order semantic features at the target time, and there are multiple target times. The step of performing attention calculation on the higher-order semantic features to obtain weighted context features includes: Attention scoring is applied to the higher-order semantic features to obtain an importance score; The higher-order semantic features at each target time are summed based on the importance score to obtain the weighted context features.

[0011] Optionally, attention scoring is performed on the higher-order semantic features to obtain an importance score, including: The weights of the long short-term memory model are multiplied by the higher-order semantic features to obtain the first mapping feature; The preset state projection matrix is ​​multiplied by the higher-order semantic features to obtain the second mapping feature; The first mapping feature and the second mapping feature are added together to obtain a summation feature, and the summation feature is activated to obtain an activation feature; The importance score is obtained by multiplying the preset attention vector with the activation feature.

[0012] To achieve the above objectives, a second aspect of this application proposes a product recommendation device based on deep learning, the device comprising: The data acquisition module is used to acquire multi-source product data; wherein the multi-source product data belongs to at least two of the following modalities: text modality, image modality, and audio modality; The feature construction module is used to construct features from the multi-source product data to obtain a time-series feature matrix; The feature projection module is used to project each feature in the time-series feature matrix to obtain deep product features; The feature extraction module is used to perform temporal processing on the deep product features through a bidirectional long short-term memory model to obtain high-order semantic features; The attention calculation module is used to perform attention calculation on the higher-order semantic features to obtain weighted context features; The product prediction module is used to make predictions based on the weighted context features to obtain the target recommended product.

[0013] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the deep learning-based product recommendation method described in the first aspect.

[0014] To achieve the above objectives, a fourth aspect of the present application provides a storage medium, which is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the deep learning-based product recommendation method described in the first aspect.

[0015] This application proposes a deep learning-based product recommendation method, apparatus, and medium. First, it acquires multi-source product data, including at least one modality from text, images, and audio, providing rich information about the products. Next, it constructs features from the multi-source product data to obtain a temporal feature matrix. The features corresponding to the multi-source product data in this temporal feature matrix are arranged in chronological order, facilitating the subsequent capture of temporal dependencies. Then, it projects each feature in the temporal feature matrix to obtain deep product features, which helps in learning long-short-term dependencies and more robust temporal semantics. Next, it performs temporal processing on the deep product features using a bidirectional long short-term memory model to obtain higher-order semantic features, and performs attention calculations on these higher-order semantic features to obtain weighted contextual features. This yields a sequence-level representation that is most informative for prediction. Finally, it uses a target product prediction model to predict the weighted contextual features to obtain the target recommended product. In summary, this application integrates multi-source feature encoding, temporal modeling, and attention mechanisms to eliminate the impact of non-linear product data on the stability of product recommendations, thereby solving the problem of poor product recommendation stability and improving the accuracy and stability of product recommendations.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deep learning-based product recommendation method provided in an embodiment of this application; Figure 2 yes Figure 1 The flowchart for step 103 in the document; Figure 3 yes Figure 1 The flowchart for step 104 in the document; Figure 4 yes Figure 1 The flowchart for step 105 in the document; Figure 5 yes Figure 4 The flowchart for step 401 in the document; Figure 6 yes Figure 1 The flowchart for step 106 in the document; Figure 7 This is a block diagram of the module structure of the product recommendation device based on deep learning provided in the embodiments of this application; Figure 8 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0022] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). It is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information and image processing, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0023] Existing recommender systems are primarily based on statistical modeling and manual feature engineering. For example, they use time-series models such as ARIMA and GARCH to fit price series. These models assume that product data (such as market data) is a stationary linear sequence, but the actual market is strongly non-linearly influenced by sales strategies, sentiment, and geopolitical events, leading to extremely poor model prediction stability. For instance, ARIMA can only reflect short-term linear trends and cannot cope with sudden market jumps. Subsequently, recommender systems attempted to introduce machine learning models (such as Support Vector Machines (SVM) and Random Forests (RF), using technical indicators (such as MA, RSI, and MACD) as input features. While these models capture some non-linear relationships, they still rely on manually constructed features and static training sets, making it difficult to handle high-dimensional heterogeneous data (such as news sentiment, macroeconomic data, and cross-market signals) and real-time update requirements. A bigger problem is that most existing recommender systems operate on an offline decision-making model. Predictions are lagging, they cannot automatically learn about changes in market structure, and their response time is too long (typically >1 second) in high-frequency scenarios, making them unsuitable for algorithmic trading or real-time risk control systems. Therefore, building a recommender system that can be updated in real-time, learn dynamically, and respond with low latency has become an urgent need.

[0024] Based on this, embodiments of this application propose a product recommendation method, a product recommendation device, an electronic device, and a computer-readable storage medium based on deep learning. It integrates processing methods such as multi-source feature encoding, temporal modeling, and attention mechanisms, eliminating the impact of non-linear product data on the stability of product recommendation, thereby solving the problem of poor product recommendation stability and improving the accuracy and stability of product recommendation.

[0025] The product recommendation method based on deep learning provided in this application relates to the field of deep learning technology in artificial intelligence. This product recommendation method can be applied to terminals and servers, or it can be software running on the server. The server can be configured as a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The software can be an application implementing the product recommendation method based on deep learning, but is not limited to the above forms.

[0026] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include server computers, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0027] This application provides a product recommendation method, a product recommendation device, an electronic device, and a computer-readable storage medium based on deep learning. The specific embodiments are described below. First, the product recommendation method based on deep learning in this application is described.

[0028] It should be noted that in each specific embodiment of this application, when it is necessary to process data related to the user's identity or characteristics, such as image data, text data, and audio data, the user's permission or consent will be obtained first. Moreover, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards.

[0029] Reference Figure 1 , Figure 1 This is an optional flowchart of a product recommendation method based on deep learning provided in the embodiments of this application, which may include, but is not limited to, steps 101 to 107.

[0030] Step 101: Obtain multi-source product data; wherein, the multi-source product data includes at least two of the following modalities: text modality, image modality, and audio modality; Step 102: Construct features from the multi-source product data to obtain a time-series feature matrix; Step 103: Project each feature in the time-series feature matrix to obtain the deep product features; Step 104: Perform temporal processing on the deep product features using a bidirectional long short-term memory model to obtain high-order semantic features; Step 105: Perform attention calculation on the higher-order semantic features to obtain weighted context features; Step 106: Predict the weighted contextual features using the target product prediction model to obtain the target recommended product.

[0031] Steps 101 to 106, as illustrated in this embodiment, firstly, multi-source product data is acquired. This multi-source product data includes at least one modality among text, images, and audio, providing rich information about the product. Next, features are constructed from the multi-source product data to obtain a temporal feature matrix. The features corresponding to the multi-source product data in this temporal feature matrix are arranged in chronological order, facilitating the subsequent capture of temporal dependencies from this matrix. Then, each feature in the temporal feature matrix is ​​projected to obtain deep product features, which helps in learning long-short-term dependencies and more robust temporal semantics from these deep product features. Next, the deep product features are temporally processed using a bidirectional long short-term memory model to obtain higher-order semantic features, and attention is calculated on these higher-order semantic features to obtain weighted contextual features. In this way, a sequence-level representation most informative for prediction is obtained. Finally, the target product prediction model is used to predict the weighted contextual features to obtain the target recommended product. In summary, this application integrates multi-source feature encoding, temporal modeling, and attention mechanisms to eliminate the impact of non-linear product data on the stability of product recommendations, thereby solving the problem of poor product recommendation stability and improving the accuracy and stability of product recommendations.

[0032] In the insurance sector within the financial field, when recommending insurance products, multiple insurance companies offer health insurance. Recommendations can be made based on each company's health insurance product data (such as price series over a period of time, transaction volume, highest price, and lowest price).

[0033] In the context of clinical support and health management in the medical field, when recommending medical products, many medical companies offer portable ultrasound, portable electrocardiogram, and in vitro diagnostic equipment. These products can be recommended to doctors or patients based on medical product data (such as price series, transaction volume, highest price, and lowest price over a period of time).

[0034] In step 101 of some embodiments, multi-source product data is acquired. Multi-source product data refers to product-related data and information from different sources that has been integrated or aggregated for the purpose of describing, analyzing, and making decisions. Multi-source product data includes at least two of the following modalities: text modality, image modality, and audio modality. Multi-source product data includes product description text belonging to the text modality, product images belonging to the image modality, and product audio belonging to the audio modality.

[0035] The product description text specifically includes: insurance name, description, coverage period, waiting period, renewal terms, scope of coverage, exclusions, insurance liability details, maximum cumulative sum insured, premium structure, premium calculation instructions, discount and preferential terms, claims process, list of claims materials, and claims time limit.

[0036] Product images include: visual information related to the product / risk, such as: illustrations of insurance terms, screenshots of key fields in policy samples; visual attributes (color tone, style, brand consistency) of policy-related promotional images and product posters. Behavioral and interface signals include: visual relevance signals between in-app ad images and user clicks / browsing; and perceptible emotional / aesthetic signals of UI elements (relevance of color, typography style, and user preferences).

[0037] Product audio includes: customer voice recordings and conversation recordings. For example, verbal descriptions of customer needs, risk preferences, and health conditions; and tone of voice, urgency, and trust signals in customer service conversations.

[0038] In step 102 of some embodiments, feature construction is performed on the multi-source product data to obtain a time-series feature matrix. The multi-source product data can be represented as: , This represents the product data (such as product description text) at time t, which belongs to the text modality. This represents the product data at time t that belongs to the image modality (such as an insurance policy diagram). The product data at time t, which belongs to the audio modality (such as customer voice and dialogue recordings), includes the following feature construction: (1) Time alignment and missing value handling: Interpolation or forward padding is used to ensure synchronization: , (2)Standardization (sliding window mean and variance standardization for each column): ,in, It is the preprocessed feature of the j-th data in the multi-source product data. , This represents the mean and standard deviation of the j-th data point in the multi-source product data over the window length ω. This step yields the standardized, aligned time-series feature matrix. The data will then be fed into the embedding / encoding module to obtain a unified vector representation, which will facilitate the processing of different types of data by subsequent deep networks.

[0039] In one example, a multimodal contrastive learning encoder architecture (MM-CLR) is used to map all product information from multiple sources into a unified shared embedding space.

[0040] In step 103 of some embodiments, each feature in the time-series feature matrix is ​​projected to obtain deep product features.

[0041] In one embodiment, reference is made to Figure 2 Step 103 may include: Step 201: Determine the category of each feature in the time-series feature matrix to obtain the feature category; Step 202: When the feature category is discrete feature, feature embedding is performed on the features in the time series feature matrix to obtain product auxiliary features; Step 203: When the feature category is continuous feature, perform 1D convolution on the features in the time series feature matrix to obtain the local time features of the product. Step 204: Concatenate all product auxiliary features and all product local time features to obtain product concatenated features, and perform linear mapping on the product concatenated features to obtain deep product features.

[0042] Heterogeneous features are projected onto the same vector space to extract local temporal patterns (such as microstructural shocks and short-term peaks) and generate sequence-level representations. If the raw, standardized data is directly input into the LSTM, the performance may be affected by noise and scale. Performing embedding and local pattern extraction first can improve the signal-to-noise ratio and provide more meaningful input features for the attention mechanism. The specific process is as follows: 1. Embedding discrete / categorical features: ,in, For the m-th product auxiliary feature, The feature category is a discrete feature (market status, product type, etc.).

[0043] 2. Perform 1D convolution on continuous features to extract local temporal features (multiple different convolution kernels can be used in parallel, with different kernels capturing patterns at different time scales): ,in, Let t be the local time feature of the product at time k.

[0044] 3. Concatenate all product auxiliary features and all product local time features, and perform a linear mapping to obtain deep product features: ,in, Let t be the depth product feature at time t.

[0045] The advantage of the embodiments of steps 201 to 204 described above is that high-dimensional vector sequences at the sequence level can be obtained. The representation of each time point contains information on both discrete and continuous variables, which helps subsequent steps learn long-term and short-term dependencies and more robust temporal semantics, thereby improving the accuracy of product recommendations.

[0046] In step 104 of some embodiments, the deep product features are processed temporally using a bidirectional long short-term memory model to obtain higher-order semantic features.

[0047] Bidirectional Long Short-Term Memory (BiLSTM) refers to adding a backpropagation branch to one or more LSTM layers, allowing information to be passed from the past to the future and from the future back to the past in the time dimension. LSTM is a variant of Recurrent Neural Network (RNN) specifically designed to handle and model long-range dependencies in sequential data. The goal of LSTM is to remember past information that influences the current output in the time dimension while suppressing irrelevant or outdated information, thereby improving the performance of tasks such as sequence prediction and labeling. LSTM uses a gating mechanism (gating units) to control the memorization, updating, and forgetting of information, avoiding the gradient vanishing and gradient exploding problems that occur in traditional RNNs on long sequences. The unit state (memory unit) at each time step is responsible for accumulating useful information over time so that it can be used by the hidden state and output at the current time step when needed.

[0048] In one embodiment, the bidirectional long short-term memory model includes a forward long short-term memory model and a reverse long short-term memory model; refer to Figure 3 Step 104 may include: Step 301: Extract features from deep product features and the first higher-order semantic state at the first time step using a positive long short-term memory model to obtain the first higher-order semantic state at the target time step; wherein, the first time step is before the target time step. Step 302: Extract features from deep product features and the second higher-order semantic state at the second time step using the reverse long short-term memory model to obtain the second higher-order semantic state at the target time step; wherein, the target time step is before the second time step. Step 303: Fuse the first and second higher-order semantic states to obtain the higher-order semantic features at the target time.

[0049] The specific process for calculating higher-order semantic features is as follows: , , .in, This represents the higher-order semantic state at time t-1 output by the forward LSTM model. This represents the higher-order semantic state at time t+1 output by the inverse LSTM model. This represents the higher-order semantic state at time t output by the forward LSTM model. This represents the higher-order semantic state at time t output by the inverse LSTM model.

[0050] The benefit of the embodiments of steps 301 to 303 above is that BiLSTM can capture positive and negative temporal dependencies from deep product features, thereby learning long-short-term dependencies and more robust temporal semantics, thus improving the accuracy of product recommendations.

[0051] In step 105 of some embodiments, attention is calculated on the higher-order semantic features to obtain weighted context features. The purpose of this step is to automatically measure the importance of each historical moment to the current prediction using an attention mechanism, forming a weighted context vector. Because market signals are periodic and abrupt, certain historical moments (such as major news events) have a much greater impact on the future than others. The attention mechanism can dynamically assign weights, thereby improving the response to sudden / non-stationary events.

[0052] In one embodiment, the higher-order semantic features are the higher-order semantic features of the target time, and there are multiple target times. For example, the target time is one of t=0 to t=T, where T is a positive integer. In this embodiment, referencing Figure 4 Step 105 may include: Step 401: Perform attention scoring on the higher-order semantic features to obtain an importance score; Step 402: Add the high-order semantic features of each target time step according to the importance score to obtain the weighted context features.

[0053] In one embodiment, reference is made to Figure 5 Step 401 may include: Step 501: Multiply the weight values ​​of the long short-term memory model with the higher-order semantic features to obtain the first mapping feature; Step 502: Multiply the preset state projection matrix with the higher-order semantic features to obtain the second mapping feature; Step 503: Add the first mapping feature and the second mapping feature to obtain the summed feature, and activate the summed feature to obtain the activated feature; Step 504: Multiply the preset attention vector with the activation feature to obtain the importance score.

[0054] The specific process of calculating importance scores can be expressed as follows: .in, Rate the importance of time t. For activation function, Let be the state projection matrix, and let... Mapped to the same attention space; An attention vector is a set of weight values ​​applied to the model. These are the weight values ​​for the LSTM model; To amplify the index and the differences, the importance of strength is highlighted, and major events in the relevant fields (such as finance or healthcare) should be emphasized.

[0055] The advantage of the embodiments of steps 501 to 504 described above is that they can improve the accuracy of importance scoring.

[0056] In one embodiment, after step 504, the importance score can be normalized, which can be expressed as: , Assess the importance of time t after normalization. For the k-th importance score, there are a total of T importance scores.

[0057] In step 402, the weighted contextual features can be represented as: Among them, the weighted context vector It is a sequence-level representation that provides the most information for future predictions (containing both short-term shocks and long-term trend information).

[0058] The advantage of the embodiments of steps 401 to 402 described above is that they can obtain weighted contextual features that are rich in information and have strong temporal characteristics, thereby helping to improve the accuracy of product recommendations.

[0059] In step 106 of some embodiments, a prediction is made based on weighted contextual features to obtain the target recommended product.

[0060] In one embodiment, reference is made to Figure 6 Step 106 may include: Step 601: Update the reference product prediction model parameters in the preset blockchain according to the multi-source product data to obtain the target product prediction model parameters, and construct the target product prediction model according to the target product prediction model parameters; Step 602: Predict the weighted contextual features of candidate products using the target product prediction model to obtain the recommendation score of the candidate products; Step 603: Filter candidate products based on recommendation scores to obtain target recommended products.

[0061] In step 601, the reference product prediction model parameters are constructed based on historical product data and stored in a preset blockchain. After each acquisition of new multi-source product data, the reference product prediction model parameters in the blockchain can be updated based on the multi-source product data, and then the target product prediction model is constructed based on the updated target product prediction model parameters.

[0062] In one example, the product recommendation system continuously and automatically fine-tunes its model parameters based on the latest product data during operation, ensuring that the product prediction model always keeps pace with market trends. For instance, the patterns in the financial market change rapidly; if the model is not updated, the prediction results will quickly become inaccurate; however, updating too frequently will cause model instability. Therefore, this application adopts a "balanced" adaptive learning method, allowing the model to adapt quickly while remaining robust. The parameter update process is as follows: , For the updated model parameters, The learning rate is adaptive (automatically adjusted based on prediction error). The predicted loss for the current batch, The gradient represents the parameters. The product recommendation system fine-tunes the model parameters based on the error each time new data arrives, allowing the model to gradually "learn" new market characteristics, thereby continuously maintaining prediction accuracy and stability.

[0063] It is understandable that after updating the reference product prediction model parameters based on multi-source product data to obtain the target product prediction model parameters, the target product prediction model can be used as the latest reference product prediction model parameters recorded on the preset blockchain.

[0064] In one embodiment, step 602 may include: predicting the price of the weighted context features and candidate products using the target product prediction model to obtain the future short-term price; predicting the volatility of the weighted context features and candidate products using the target product prediction model to obtain the future volatility; predicting the risk of the weighted context features and candidate products using the target product prediction model to obtain the future risk probability; and fusing the future short-term price, future volatility, and future risk probability to obtain a recommendation score.

[0065] In this embodiment, the parallel output of future short-term prices, future volatility, and future risk probabilities enables the linkage between trading decisions and risk control signals. Because a single objective (e.g., predicting only prices) cannot simultaneously satisfy both trading and risk control needs, the target product prediction model can be a multi-task learning model. This model can share underlying representations, improve sample efficiency, and consider both returns and risks during prediction. The calculation formula is as follows: , , . This is a weight matrix used to predict future short-term prices. This is a weight matrix used to predict future volatility. This is a weight matrix used to predict the probability of future risks. , , All are bias vectors. For the error range, For future short-term prices, For future volatility, This represents the probability of future risks (abnormalities).

[0066] In one embodiment, future short-term prices, future volatility, and future risk probabilities can be imported into an online learning module to update the target product prediction model parameters in a preset blockchain.

[0067] In one embodiment, the deep learning-based product recommendation method may further include: immediately triggering a risk warning and generating a traceable record of the model's state when abnormal market fluctuations or excessive deviations in prediction results are detected. This ensures safety and compliance, preventing the model from "going berserk" under extreme market conditions and allowing for post-event tracing of the source of the problem. The detection method is as follows: , For predicted values, For the true value, As for the probability of future risks, and This is a preset threshold. When the prediction error or risk probability exceeds the threshold, an alarm is immediately issued, and the current model parameters are generated into a hash fingerprint and saved to the blockchain for subsequent auditing and tracing.

[0068] In summary, the present application achieves at least the following beneficial effects: 1. From static to dynamic: Traditional models have fixed parameters, while the present application possesses an adaptive learning mechanism that can reconverge within minutes after a sudden change in market structure. 2. From single prediction to multi-dimensional prediction: Simultaneously predicting price trends, volatility, and risks improves the accuracy of product recommendations. 3. From black box to transparency: Each model training result and parameter is hashed and uploaded to the blockchain, ensuring auditability, verifiability, and immutability. 4. From latency to real-time: Through asynchronous stream processing and a lightweight inference engine (TensorRT or ONNX Runtime), prediction latency is reduced to milliseconds, supporting high-frequency trading and real-time monitoring.

[0069] Please see Figure 7 This application also provides a product recommendation device based on deep learning, which can implement the above-mentioned product recommendation method based on deep learning. Figure 7 The block diagram of the deep learning-based product recommendation device provided in this application embodiment is shown. The device includes: a data acquisition module 701, a feature construction module 702, a feature projection module 703, a feature extraction module 704, an attention calculation module 705, and a product prediction module 706. Specifically, the data acquisition module 701 is used to acquire multi-source product data; wherein, the multi-source product data includes at least two modalities: text modality, image modality, and audio modality; the feature construction module 702 is used to construct features from the multi-source product data to obtain a temporal feature matrix; the feature projection module 703 is used to project each feature in the temporal feature matrix to obtain deep product features; the feature extraction module 704 is used to perform temporal processing on the deep product features using a bidirectional long short-term memory model to obtain high-order semantic features; the attention calculation module 705 is used to perform attention calculation on the high-order semantic features to obtain weighted context features; and the product prediction module 706 is used to predict the weighted context features using a target product prediction model to obtain a target recommended product.

[0070] It should be noted that the specific implementation of this deep learning-based product recommendation device is basically the same as the specific implementation of the deep learning-based product recommendation method described above, and will not be repeated here.

[0071] This application also provides an electronic device, which includes: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for communication between the processor and the memory. When the program is executed by the processor, it implements the aforementioned deep learning-based product recommendation method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0072] Please see Figure 8 , Figure 8The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 801 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 802 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 802 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 802 and is called and executed by the processor 801 using the deep learning-based product recommendation method of the embodiments of this application. The 803 input / output interface is used to implement information input and output. The communication interface 804 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 805 transmits information between various components of the device (e.g., processor 801, memory 802, input / output interface 803, and communication interface 804); The processor 801, memory 802, input / output interface 803, and communication interface 804 are connected to each other within the device via bus 805.

[0073] This application embodiment also provides a storage medium, which is a computer-readable storage medium for computer-readable storage. The storage medium stores one or more programs, which can be executed by one or more processors to implement the above-described deep learning-based product recommendation method.

[0074] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0075] This application proposes a deep learning-based product recommendation method, apparatus, and medium. First, it acquires multi-source product data, including at least one modality from text, images, and audio, providing rich information about the products. Next, it constructs features from the multi-source product data to obtain a temporal feature matrix. The features corresponding to the multi-source product data in this temporal feature matrix are arranged in chronological order, facilitating the subsequent capture of temporal dependencies. Then, it projects each feature in the temporal feature matrix to obtain deep product features, which helps in learning long-short-term dependencies and more robust temporal semantics. Next, it performs temporal processing on the deep product features using a bidirectional long short-term memory model to obtain higher-order semantic features, and performs attention calculations on these higher-order semantic features to obtain weighted contextual features. This yields a sequence-level representation that is most informative for prediction. Finally, it uses a target product prediction model to predict the weighted contextual features to obtain the target recommended product. In summary, this application integrates multi-source feature encoding, temporal modeling, and attention mechanisms to eliminate the impact of non-linear product data on the stability of product recommendations, thereby solving the problem of poor product recommendation stability and improving the accuracy and stability of product recommendations.

[0076] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0077] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0078] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0080] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0081] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0083] 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.

[0084] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0085] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0086] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) 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 programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A product recommendation method based on deep learning, characterized in that, The method includes: Acquire multi-source product data; wherein the multi-source product data includes at least two of the following modalities: text modality, image modality, and audio modality; Feature construction is performed on the multi-source product data to obtain a time-series feature matrix; Projecting each feature in the time-series feature matrix yields in-depth product features; The deep product features are processed temporally using a bidirectional long short-term memory model to obtain higher-order semantic features; Attention is calculated on the higher-order semantic features to obtain weighted context features; The target recommended product is obtained by predicting the weighted contextual features using the target product prediction model.

2. The method according to claim 1, characterized in that, The step of predicting the weighted contextual features using a target product prediction model to obtain the target recommended product includes: The reference product prediction model parameters in the preset blockchain are updated based on the multi-source product data to obtain the target product prediction model parameters, and the target product prediction model is constructed based on the target product prediction model parameters; the reference product prediction model parameters are constructed based on historical product data. The recommended score of the candidate product is obtained by predicting the weighted contextual features of the candidate product using the target product prediction model. The candidate products are filtered based on the recommendation scores to obtain the target recommended product.

3. The method according to claim 2, characterized in that, The step of predicting the weighted contextual features of candidate products using the target product prediction model to obtain the recommendation score of the candidate products includes: The target product prediction model is used to predict the price of the weighted context features and the candidate products to obtain the future short-term price. The target product prediction model is used to predict the volatility of the weighted context features and the candidate products to obtain the future volatility. The target product prediction model is used to predict the future risk probability by analyzing the weighted context features and the candidate products. The recommendation score is obtained by fusing the future short-term price, the future volatility, and the future risk probability.

4. The method according to any one of claims 1 to 3, characterized in that, The projection of each feature in the time-series feature matrix to obtain deep product features includes: The category of each feature in the time-series feature matrix is ​​determined to obtain the feature category; When the feature category is a discrete feature, feature embedding is performed on the features in the time-series feature matrix to obtain product auxiliary features; When the feature category is a continuous feature, a 1D convolution is performed on the features in the time-series feature matrix to obtain the local time features of the product. All the product auxiliary features and all the product local time features are concatenated to obtain the product concatenated features, and the product concatenated features are linearly mapped to obtain the deep product features.

5. The method according to any one of claims 1 to 3, characterized in that, The bidirectional long short-term memory model includes a forward long short-term memory model and a reverse long short-term memory model; The process of performing time-series processing on the deep product features using a bidirectional long short-term memory model to obtain high-order semantic features includes: The deep product features and the first higher-order semantic state at the first time step are extracted using the positive long short-term memory model to obtain the first higher-order semantic state at the target time step; wherein, the first time step step is before the target time step ... The deep product features and the second higher-order semantic state at the second time step are extracted using the reverse long short-term memory model to obtain the second higher-order semantic state at the target time step; wherein, the target time step is before the second time step step. The higher-order semantic features at the target time are obtained by fusing the first higher-order semantic state and the second higher-order semantic state.

6. The method according to any one of claims 1 to 3, characterized in that, The higher-order semantic features are the higher-order semantic features at the target time, and there are multiple target times. The attention calculation on the higher-order semantic features to obtain weighted context features includes: Attention scoring is applied to the higher-order semantic features to obtain an importance score; The higher-order semantic features at each target time are summed based on the importance score to obtain the weighted context features.

7. The method according to claim 6, characterized in that, Attention scoring is applied to the higher-order semantic features to obtain an importance score, including: The weights of the long short-term memory model are multiplied by the higher-order semantic features to obtain the first mapping feature; The preset state projection matrix is ​​multiplied by the higher-order semantic features to obtain the second mapping feature; The first mapping feature and the second mapping feature are added together to obtain a summation feature, and the summation feature is activated to obtain an activation feature; The importance score is obtained by multiplying the preset attention vector with the activation feature.

8. A product recommendation device based on deep learning, characterized in that, The device includes: The data acquisition module is used to acquire multi-source product data; wherein the multi-source product data includes at least two of the following modalities: text modality, image modality, and audio modality; The feature construction module is used to construct features from the multi-source product data to obtain a time-series feature matrix; The feature projection module is used to project each feature in the time-series feature matrix to obtain deep product features; The feature extraction module is used to perform temporal processing on the deep product features through a bidirectional long short-term memory model to obtain high-order semantic features; The attention calculation module is used to perform attention calculation on the higher-order semantic features to obtain weighted context features; The product prediction module is used to predict the weighted context features using the target product prediction model to obtain the target recommended product.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images