A method for predicting alcohol consumption by combining curvelet transform and convolutional neural network

Abstract

The present invention discloses a method for predicting alcohol consumption by combining curvelet transform and convolutional neural network, which belongs to the technical field of specific computing models. The method comprises: obtaining a time-feature data set based on user basic profiles, daily behavior trajectories, and purchase information results of alcohol consumption users; obtaining a plurality of curvelet coefficients based on the time-feature data set by performing curvelet transform in combination with the curvelet characteristic frequency, the plurality of curvelet coefficients including low-frequency curvelet coefficients, medium-frequency curvelet coefficients, high-frequency curvelet coefficients, and ultra-high-frequency curvelet coefficients; obtaining a plurality of feature information based on the low-frequency curvelet coefficients, medium-frequency curvelet coefficients, and high-frequency curvelet coefficients by performing feature decomposition; obtaining a plurality of convolution operation results based on the plurality of feature information by performing convolutional neural network, obtaining a pooling operation result by performing pooling operation, and obtaining an alcohol consumption prediction by performing linear correction activation.

Description

Technical Field

[0001] The present invention belongs to the technical field of specific computing models, and in particular relates to a method for predicting alcohol consumption by combining curvelet transform and convolutional neural network. Background Art

[0002] With the rapid development of information technology, consumer forecasting as a market operation model is gaining increasing attention from various businesses. Particularly in the alcohol sales sector, accurately predicting consumer purchase likelihood is crucial for improving marketing efficiency, optimizing inventory management, and increasing corporate profits. Analyzing historical consumer data, including basic profile information, daily behavior patterns, and purchase results, can help businesses better understand consumer behavior patterns and develop targeted marketing strategies accordingly.

[0003] Currently, predicting alcohol consumers' purchase likelihood relies primarily on two methods: predictions based on simple mathematical models and predictions based on large-scale artificial intelligence models. Simple mathematical model-based predictions typically employ probabilistic and transformation models to analyze the probability distribution and trends of data. While simple to implement, these methods rely on a low level of data mining and fail to fully exploit the correlations and hidden information within consumer behavior data. This results in low prediction accuracy and can only provide rough guidance in precision marketing efforts.

[0004] Furthermore, prediction methods based on large AI models utilize large amounts of data and powerful computing power to train large-parameter models, resulting in high generalization. However, in actual applications, these models lack in-depth analysis of the rationality of alcohol consumption data and research into feature correlations. They perform particularly poorly when processing data with less distinct features, resulting in poor predictive stability and difficulty convincing businesses in actual operations.

[0005] Although the above methods have their own advantages, they still have significant shortcomings. Therefore, a new method is urgently needed to solve the problems of insufficient prediction accuracy and poor model stability in existing technologies. Summary of the Invention

[0006] The present invention provides a method for predicting alcohol consumption by combining curvelet transform and convolutional neural network to solve the problem of insufficiently exploring the potential information in consumer behavior data, resulting in inaccurate prediction results, and the problem of unstable performance of the model in different application scenarios due to the lack of effective decomposition and correlation analysis of data features.

[0007] The technical solution adopted in the present invention is:

[0008] A method for predicting alcohol consumption by combining curvelet transform and convolutional neural network, including:

[0009] Based on the user basic profiles, daily behavior trajectories, and purchase information results of alcohol consumption users, a time-feature dataset is obtained;

[0010] According to the time-characteristic data set, a plurality of curvelet coefficients are obtained by curvelet transformation in combination with the curvelet characteristic frequency, wherein the plurality of curvelet coefficients include a low-frequency curvelet coefficient, a medium-frequency curvelet coefficient, a high-frequency curvelet coefficient, and an ultra-high-frequency curvelet coefficient;

[0011] According to the low-frequency curvelet coefficient, the intermediate-frequency curvelet coefficient, and the high-frequency curvelet coefficient, a plurality of feature information is obtained by feature decomposition;

[0012] Based on the multiple feature information, a convolutional neural network is used to obtain multiple convolution operation results, a pooling operation result is obtained through pooling operation, and a wine consumption prediction is obtained through linear correction activation.

[0013] The alcohol consumption prediction method combining curvelet transform and convolutional neural network described in the present invention also includes the following additional technical features:

[0014] Based on the user basic profiles, daily behavior trajectories, and purchase information results of alcohol consumption users, a time-feature dataset is obtained, specifically:

[0015] Based on the basic user profiles, daily behavior trajectories, and purchase information of alcohol consumers, format and transform them to obtain time series values ​​in a one-to-one correspondence;

[0016] A time-feature data set is obtained according to the time series values.

[0017] The basic user profile, daily behavior trajectory, and purchase information results of alcohol consumption users are as follows:

[0018] The basic user profile includes at least one of the consumer's age, gender, date of birth, and registration time;

[0019] Daily behavior traces include at least one of the following: time logged into the platform, time spent browsing products, time initiating payment, number of order cancellations, number of product collections, and number of activity participations;

[0020] The purchase result information includes at least one of the purchase amount, quantity, purchase time, and purchase store.

[0021] According to the time-characteristic data set, a plurality of curvelet coefficients are obtained by curvelet transformation in combination with the curvelet characteristic frequency, specifically:

[0022] According to the time-feature data set, curvelet coefficients at 6 scales are obtained through curvelet transformation;

[0023] Combined with the characteristic frequency of the curvelet, the curvelet coefficients at six scales are divided into low-frequency curvelet coefficients, medium-frequency curvelet coefficients, high-frequency curvelet coefficients and ultra-high-frequency curvelet coefficients;

[0024] The intermediate frequency curvelet coefficient and the high frequency curvelet coefficient both include curvelet coefficients at two scales.

[0025] According to the low-frequency curvelet coefficient, the intermediate-frequency curvelet coefficient, and the high-frequency curvelet coefficient, multiple feature information is obtained through feature decomposition, specifically:

[0026] According to the low-frequency curvelet coefficient, the trend characteristic information is obtained through the inverse curvelet transform;

[0027] According to the intermediate frequency curvelet coefficient, the intermediate frequency curvelet coefficient is classified into multiple directions, and the intermediate frequency feature information in four directions is obtained through the inverse curvelet transform;

[0028] According to the high-frequency curvelet coefficients, multiple directions in the high-frequency curvelet coefficients are classified, and the high-frequency feature information in four directions is obtained through inverse curvelet transformation.

[0029] Based on the plurality of feature information, a plurality of convolution operation results are obtained through a convolutional neural network, specifically:

[0030] According to the trend feature information, a convolution operation result is obtained by sharpening the convolution kernel;

[0031] According to the horizontal intermediate frequency feature information and high frequency feature information, the vertical edge detection convolution kernel is used to extract the vertical gradient change and obtain the convolution operation results respectively;

[0032] According to the longitudinal intermediate frequency feature information and high frequency feature information, the horizontal edge detection convolution kernel is used to extract the horizontal gradient change and obtain the convolution operation results respectively;

[0033] According to the oblique intermediate frequency feature information, a convolution operation result consistent with the oblique intermediate frequency feature information is obtained through a unit convolution kernel;

[0034] According to the high-frequency feature information of the oblique direction, the convolution operation result is obtained through the image smoothing convolution kernel.

[0035] A pooling operation result is obtained through the pooling operation, specifically:

[0036] Based on the multiple convolution operation results, average pooling is performed to obtain a pooling operation result representing the average value within the pooling window, so as to reduce the data dimension.

[0037] The alcohol consumption prediction is obtained through linear correction activation, specifically:

[0038] Based on the pooling operation results, combined with the weight matrix and bias vector, the alcohol consumption prediction result is obtained;

[0039] The weight matrix is ​​a two-dimensional matrix consisting of the number of output neurons and the number of input neurons, and the bias vector is a one-dimensional array of the number of output neurons. The weight matrix and the bias vector are obtained through training.

[0040] The weight matrix and the bias vector are obtained through training, specifically:

[0041] Determining training data and test data, wherein the training data and test data both include basic user profiles, daily behavior trajectories, input data of purchase information results of alcohol consumers, and classification data obtained based on the purchase data of the current year;

[0042] Performing training based on the input data and the classification data in the training data to obtain multiple sets of the weight matrices and the bias vectors;

[0043] Based on the input data in the test data, multiple groups of the weight matrices and the bias vectors are combined to obtain multiple prediction results, and the weight matrix and the bias vector are determined based on the correlation between the multiple prediction results and the hierarchical data in the test data.

[0044] The present invention also provides an electronic device, comprising:

[0045] Memory, for storing computer instructions;

[0046] A processor is used to implement the wine consumption prediction method combining curvelet transform and convolutional neural network when executing the computer instructions.

[0047] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are as follows:

[0048] 1. In the present invention, a time-feature dataset is generated based on the user profiles, daily behavior trajectories, and purchase information of alcohol consumers. Based on this time-feature dataset, a curvelet transform is performed, combined with the curvelet characteristic frequencies, to generate multiple curvelet coefficients. These curvelet coefficients include low-frequency curvelet coefficients, mid-frequency curvelet coefficients, high-frequency curvelet coefficients, and ultra-high-frequency curvelet coefficients. By constructing an ordered model of consumer basic data, behavioral data, and consumption data, and treating this data as characteristic signals that vary over time, the present invention can more comprehensively capture consumer behavioral patterns.

[0049] Specifically, user profiles, daily behavior trajectories, and purchase results are integrated into a time-feature dataset. This approach encompasses not only static consumer attributes but also dynamic behaviors, providing a more comprehensive and in-depth understanding of consumer behavior.

[0050] Furthermore, using the curvelet transform to decompose the raw data features reveals overall trends expressed by low-frequency signals and detailed changes expressed by high-frequency signals. The curvelet transform can process data at multiple scales and angles, enhancing the data representation capabilities at each scale. For example, low-frequency components reflect the overall trend of consumer behavior, while mid-frequency and high-frequency components depict fine-grained features in different directions. This multi-level feature extraction approach enables the model to more accurately capture the implicit information in the data, including potential changes in consumer trends and behavioral patterns.

[0051] Therefore, by combining comprehensive data collection with curvelet transform technology, this invention significantly improves the accuracy of alcohol consumption forecasts. This method not only identifies basic consumer behavior patterns but also detects subtle behavioral changes, providing businesses with more precise decision support. This improves forecast accuracy and ensures the model's adaptability to complex and changing market environments.

[0052] 2. In the present invention, based on the low-frequency, mid-frequency, and high-frequency curvelet coefficients, feature decomposition is performed to obtain multiple feature information. Based on these multiple feature information, a convolutional neural network is used to obtain multiple convolution operation results, a pooling operation is performed to obtain a single pooling operation result, and a linear corrected activation is performed to obtain a prediction of alcohol consumption. This present invention significantly enhances the stability and robustness of the alcohol consumption prediction model by using multi-scale feature decomposition and convolutional neural network-enhanced feature extraction.

[0053] First, the curvelet transform is used to decompose the data into low-frequency, medium-frequency, high-frequency, and ultra-high-frequency components, and these components are then subjected to feature decomposition. This decomposition method ensures more stable model performance across different frequency ranges: low-frequency components reflect overall trends, while medium-frequency and high-frequency components depict fine-grained features in different directions. For example, low-frequency components capture the overall direction of consumer behavior, while medium-frequency and high-frequency components can identify subtle changes in behavioral patterns and trends in specific directions. This multi-level feature extraction not only helps improve the model's generalization ability but also ensures more robust performance when faced with complex data.

[0054] Furthermore, the present invention uses convolutional neural networks to further extract local features, effectively mining the hidden information in the feature data. This method can accurately capture the changes in features at different scales, improving the accuracy and consistency of the model when processing detailed information.

[0055] By combining the multi-scale feature decomposition of the curvelet transform and the powerful feature extraction capability of the convolutional neural network, the present invention significantly improves the stability and robustness of the model. It can not only adapt to different data characteristics and application scenarios, but also effectively reduce noise interference and improve the reliability of the model in practical applications.

[0056] 3. In the present invention, a pooling operation is performed to obtain a pooling result, which is then used to generate a wine consumption prediction through linear corrected activation. This invention significantly enhances the intelligence of the wine consumption prediction model through adaptive learning capabilities and efficient data processing methods. First, in terms of adaptive learning capabilities, the convolutional neural network can automatically adjust parameters during training to minimize the loss function, achieving adaptive learning. This mechanism not only improves the model's ability to fit complex data patterns but also enhances its ability to respond to real-time market demands.

[0057] Furthermore, the present invention utilizes pooling operations to reduce data dimensionality, thereby improving computational efficiency and preventing overfitting. By reducing the spatial size of feature maps, pooling reduces the computational complexity of subsequent layers while preserving key feature information. Through these methods, the model can maintain high accuracy while reducing computational resource consumption and improving overall operational efficiency.

[0058] In summary, this invention significantly improves the intelligence of the alcohol consumption prediction model through adaptive learning capabilities and efficient pooling operations. Adaptive learning enables the model to dynamically adjust to changing market demands, while pooling effectively reduces computational complexity and prevents overfitting, ensuring the model's stability and reliability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0060] Figure 1 The figure is a flow chart of the method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to one embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0062] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0063] like Figure 1 As shown in FIG, a method for predicting alcohol consumption by combining curvelet transform and convolutional neural network includes:

[0064] S100: Obtain a time-feature dataset based on the user basic profile, daily behavior trajectory, and purchase information results of the alcohol consumption user.

[0065] The primary goal of this step is to integrate multi-dimensional consumer data (basic profiles, behavioral trajectories, and purchase outcomes) to construct a dataset that reflects temporal dynamics and feature correlations. This provides the foundation for subsequent feature decomposition and prediction using curvelet transforms and convolutional neural networks. The core of this phase is to transform discrete static data and dynamic behavioral data into structured data that can be analyzed by the model, thereby fully capturing the underlying patterns of consumer behavior.

[0066] It can be understood that the time-feature dataset is a structured dataset that organizes consumer behavior data by time and feature dimensions. This allows the discrete raw data to be converted into spatiotemporal joint features that can be parsed by the model, facilitating subsequent analysis.

[0067] In this step, data collection and integration are first performed, including user profiles, daily behavior patterns, and purchase results. These three types of data are aligned by user ID to form a user-centric, linked dataset. For example, user A's "age 25 (basic profile)" is associated with "browsing wine pages every Wednesday at 8 pm (behavioral pattern)" and "purchasing two bottles of wine per month (purchase result)" to form a single record.

[0068] The data is also sorted by timestamp to construct continuous records along the time dimension. For example, user B's "browsing the spirits page for 10 minutes on March 1st" and "purchasing a bottle of whiskey on March 15th" are marked as behavior nodes on the same timeline.

[0069] It's important to note that data dimensions must be mapped and standardized. Feature encoding is performed to convert unstructured data (such as text descriptions) into numerical features. For example, gender can be encoded as binary (0 / 1) and region can be encoded as region ID. Normalization is also performed to standardize numerical data (such as purchase amount and browsing time) (e.g., Z-score normalization) to eliminate dimensional differences.

[0070] During this process, feature weights are dynamically assigned, varying according to the importance of the behavior. For example, purchase amount is weighted more highly than browsing time to emphasize the impact of key behaviors on predictions. Additionally, missing data is filled in through interpolation or averaging to ensure data integrity.

[0071] Finally, we construct a time-feature dataset. The processed data is organized into a three-dimensional tensor (user × time × feature), where: user, each user, is an independent sample; time, divided into fixed time windows (e.g., week / day), forms time steps; and features, including the encoded values ​​of basic profiles, behavior trajectories, and purchase results.

[0072] Dynamically update features using a sliding time window (e.g., the last 30 days of data) to capture the time-varying nature of behavioral patterns. For example, user C's "increased browsing frequency in the last 7 days" may indicate a stronger purchase intent.

[0073] In this step, data is aligned by user ID to avoid prediction bias caused by analyzing a single dimension in isolation (e.g., focusing solely on purchase amount while ignoring behavioral changes). This strengthens the prediction weight of key behaviors (e.g., purchase history), reduces redundant information interference, and improves the efficiency of the subsequent curvelet transform's eigenvalue decomposition. The sliding window balances the influence of historical behavior and recent trends, enhancing the model's sensitivity to short-term behavioral changes (e.g., holiday spending fluctuations).

[0074] This step lays the foundation for the subsequent multi-scale feature decomposition of curvelet transform and local feature extraction of convolutional neural network through structured data integration, which is a key prerequisite for improving prediction accuracy and model stability.

[0075] S200: According to the time-characteristic data set, a plurality of curvelet coefficients are obtained by curvelet transformation in combination with the curvelet characteristic frequency, wherein the plurality of curvelet coefficients include low-frequency curvelet coefficients, medium-frequency curvelet coefficients, high-frequency curvelet coefficients and ultra-high-frequency curvelet coefficients.

[0076] This step uses the curvelet transform to perform multi-scale, multi-directional feature decomposition on the time-feature dataset, converting the raw data into low-frequency, medium-frequency, high-frequency, and ultra-high-frequency curvelet coefficients. This process separates global trends (low-frequency) from local details (high-frequency) in the data, providing clearer and more structured input for subsequent feature extraction in the convolutional neural network, thereby improving the accuracy and robustness of the prediction model.

[0077] As you can understand, the curvelet transform is a signal processing technique based on multi-scale and multi-directional analysis, specifically designed to efficiently decompose curvilinear singularities in two-dimensional data (such as images or time-feature data). It can capture trends across different timeframes (e.g., monthly, weekly, and daily) through scale parameters and directional characteristics in the data (e.g., the curvilinear changes in a user's browsing trajectory) through directional parameters.

[0078] In the present invention, the curvelet transform solves the problem that traditional methods (such as wavelet transform) cannot effectively analyze curved consumption behaviors (such as holiday consumption patterns).

[0079] Curvelet coefficients refer to the eigenvectors obtained after the curvelet transform. Each coefficient corresponds to the data energy distribution at a specific scale (frequency) and direction. Low-frequency coefficients represent global trends (such as a user's average monthly spending level). Medium- and high-frequency coefficients capture local details (such as a sudden increase in daily spending).

[0080] By classifying the curvelet coefficients, the model can specifically extract features at different levels to avoid information aliasing.

[0081] In this step, the curvelet transform parameters are preset to determine its scale and directional parameters to control the fineness of the decomposition. Low-frequency components correspond to large scales, while high-frequency components correspond to small scales, ensuring coverage of all frequency bands, from overall trends to subtle changes. Furthermore, high-frequency components (such as mid- and ultra-high frequencies) require more directional parameters to capture complex behavioral patterns (such as the curvilinear changes in a user's browsing trajectory).

[0082] After the curvelet transform algorithm is pre-configured, a time-feature dataset (e.g., the 3D tensor of user D) is fed into the algorithm. A recursive filter bank decomposes the data into subbands of varying scales. A directional decomposition is performed on each subband to generate curvelet coefficients for each direction.

[0083] According to the characteristic frequency of the curve, the coefficients are divided into four categories: low-frequency curve coefficients, which cover global trends (such as the average monthly consumption amount of users); medium-frequency curve coefficients, which capture medium details (such as weekly consumption fluctuations); high-frequency curve coefficients, which characterize short-term behavioral changes (such as a sudden increase in daily browsing frequency); and ultra-high-frequency curve coefficients, which filter noise (such as data collection errors).

[0084] By dynamically adjusting scale and directional parameters, the model can adapt to the data characteristics of different users (for example, users with frequent high-frequency behavior require more directional parameters). Threshold settings for ultra-high frequency coefficients ensure that the model focuses on meaningful features and avoids noise interference in prediction results.

[0085] Through the above steps, the curvelet transform converts raw data into structured multi-scale coefficients, providing clear, low-noise input for subsequent feature extraction in the convolutional neural network, significantly improving the accuracy and stability of the forecasting model. This process connects data preprocessing with deep learning models to achieve the final consumption forecast.

[0086] S300: Obtain multiple feature information by feature decomposition according to the low-frequency curvelet coefficient, the intermediate-frequency curvelet coefficient, and the high-frequency curvelet coefficient.

[0087] The main purpose of this step is to further decompose the low-, medium-, and high-frequency curvelet coefficients, extract fine-grained feature information in different frequency bands, and transform the multi-scale features of the raw data into structured feature vectors that can be effectively utilized by convolutional neural networks, thereby improving the model's ability to represent complex behavioral patterns. This process can isolate features that are strongly relevant to consumption forecasting (such as long-term trends, short-term fluctuations, and behavioral mutations), eliminate redundant information, and provide efficient input for subsequent forecasting.

[0088] It should be noted that the low-frequency, medium-frequency and high-frequency curvelet coefficients have different physical meanings. For example, the low-frequency curvelet coefficients contain overall trend information, and the corresponding decomposition algorithm needs to be selected according to the physical meaning of the curvelet coefficients.

[0089] For example, principal component analysis (PCA) is used to extract core trend features from low-frequency coefficients. Independent component analysis (ICA) is used to isolate directional behavior patterns (such as the curvilinear direction of a user's browsing trajectory) from mid-frequency coefficients. Wavelet packet decomposition is used to further break down high-frequency components from high-frequency coefficients to capture sudden behavioral changes.

[0090] Select the optimal algorithm for each frequency band (e.g., PCA for dimensionality reduction, wavelet packets for high-frequency details). Also, use analysis of variance (ANOVA) to identify features that contribute most to prediction and remove redundant components.

[0091] Feature information is extracted and encoded using the corresponding algorithms. Low-frequency feature extraction involves performing PCA dimensionality reduction on the low-frequency coefficient matrix to generate feature vectors reflecting long-term consumption capacity (e.g., "monthly average consumption stability index"). Intermediate-frequency feature extraction involves performing ICA decomposition on the intermediate-frequency coefficients to isolate directional characteristics of user behavior (e.g., "horizontal browsing preference" or "vertical click hotspots"). High-frequency feature extraction involves performing wavelet packet decomposition on the high-frequency coefficients to generate features reflecting sudden changes in short-term behavior (e.g., "sudden increase in daily browsing frequency").

[0092] Each decomposed feature is given a semantic label (e.g., "high frequency - browsing duration volatility") to facilitate subsequent model interpretation. Different weights are assigned based on feature contribution (e.g., high-frequency features are weighted higher than low-frequency features to highlight the impact of short-term behavioral changes).

[0093] Finally, feature fusion and storage are performed. The low-frequency, medium-frequency, and high-frequency feature vectors are concatenated into a multi-scale feature matrix. The feature dimensions are unified through standardization (such as the Z-score) to form the final input feature information.

[0094] During feature fusion and storage, weighted summation or cascading is used to ensure the complementarity of features from different frequency bands. At the same time, highly correlated features are eliminated through correlation analysis to reduce computational redundancy.

[0095] Through this step, the original curvelet coefficients are converted into structured, low-redundancy multi-scale feature information, providing accurate input for the subsequent local feature extraction of the convolutional neural network, significantly improving the model's ability to capture complex consumer behaviors.

[0096] S400: Based on the plurality of feature information, a convolutional neural network is used to obtain a plurality of convolution operation results, a pooling operation is performed to obtain a pooling operation result, and a wine consumption prediction is obtained through linear correction activation.

[0097] The main purpose of this step is to use a convolutional neural network (CNN) to perform local feature extraction, dimensionality reduction, and nonlinear modeling on multi-scale feature information, ultimately outputting alcohol consumption prediction results. The core of this step is to leverage the CNN's local perception capabilities to capture spatial and temporal correlations between features, reduce overfitting through pooling, and enhance the model's ability to fit complex patterns through activation functions, thereby achieving highly accurate and stable consumption predictions.

[0098] As you can understand, a convolutional neural network (CNN) is a deep learning model that processes grid-structured data (such as images and time series) through convolutional, pooling, and fully connected layers. Its core principle is that the convolution kernel only perceives a local area, reducing the number of parameters and enhancing the ability to capture spatial and temporal correlations. The same convolution kernel shares parameters across different locations, improving generalization.

[0099] Convolution is a process of performing dot product operations on the input data by sliding the convolution kernel to extract local features. In this invention, the convolution layer performs pattern matching on multi-scale features.

[0100] Pooling reduces the spatial dimensions of feature maps by downsampling, preserving key information. Max pooling is used to preserve local maxima and enhance robustness to salient features. Average pooling smooths noise and improves the stability of low-frequency features.

[0101] The rectified linear activation function (ReLU) is a nonlinear activation function that avoids the vanishing gradient problem, accelerates model convergence, and reduces computational complexity through sparse activation.

[0102] In this step, multi-scale feature information (such as low-frequency, mid-frequency, and high-frequency feature vectors) is concatenated into a three-dimensional tensor (feature × time × user) as the input to the CNN. Multiple sets of convolution kernels are designed, and optimal parameters are selected for different feature types. The kernel size and stride are dynamically adjusted based on the feature frequency. For example, a stride of 1 is used for high-frequency features to preserve detail. Furthermore, low-frequency, mid-frequency, and high-frequency features are fed into different convolution branches to avoid information aliasing.

[0103] In addition, convolution operations and local feature extraction are performed. Convolution operations are performed on the input tensor to generate multiple feature maps. For example, the low-frequency convolution branch outputs a "long-term trend feature map" (such as consumption stability). The high-frequency convolution branch outputs a "short-term fluctuation feature map" (such as a sudden increase in browsing frequency).

[0104] In this step, the convolution kernel is initialized using the He initialization method to avoid gradient vanishing / explosion. Cross-branch connections (such as residual connections) are used to integrate features from different frequency bands and perform feature fusion to enhance the model's expressiveness.

[0105] Next, pooling and dimensionality reduction are performed. A pooling layer is applied to each feature map to reduce the spatial dimension (e.g., time steps or feature dimensions). For example, max pooling preserves local maxima and highlights significant features (e.g., "short-term interest surge"). Average pooling smooths out noise and enhances the stability of low-frequency features (e.g., "average monthly consumption").

[0106] Select the pooling type based on feature importance (e.g., use max pooling for high-frequency features and average pooling for low-frequency features). Set the pooling window to 2×2 or 3×3 to balance dimensionality reduction efficiency and information preservation.

[0107] Finally, a fully connected layer and activation function are executed. The pooled features are flattened and fed into the fully connected layer for final prediction. The output is then converted using the rectified linear unit (ReLU) activation function.

[0108] It's important to note that a Dropout layer (e.g., with a retention rate of 0.5) is added before the fully connected layer to prevent overfitting. To simultaneously predict purchase probability and amount, expand the output nodes and design a multi-objective loss function.

[0109] This step significantly improves the overall performance of alcohol consumption prediction through the deep learning architecture of convolutional neural networks: first, the convolutional layer accurately captures the local correlation between multi-scale features (such as the long-term consumption capacity of users and short-term behavioral mutations), and combined with the dimensionality reduction processing of pooling operations, the model can identify complex behavioral patterns (such as the superposition effect of "short-term interest surge × long-term consumption capacity"); second, the pooling operation effectively suppresses the risk of overfitting, ensuring that the model can still maintain stable performance in new users or data-sparse scenarios; in addition, the weight sharing mechanism of the convolutional layer and the dimensionality compression capability of the pooling layer reduce the number of model parameters, significantly improving computational efficiency and inference speed; finally, through the ReLU activation function, the model can fit complex relationships in consumer behavior (such as the multiplier effect of "promotional day + interest surge"), further enhancing the robustness and generalization ability of the prediction.

[0110] In summary, this step uses the deep learning capabilities of CNN to transform multi-scale features into accurate prediction results, solving the problems of inaccurate predictions and model instability caused by insufficient feature correlation in traditional methods, and providing an efficient and reliable solution for alcohol consumption prediction.

[0111] As a preferred embodiment of the present invention, a time-feature dataset is obtained based on the user basic profile, daily behavior trajectory, and purchase information results of alcohol consumption users, specifically:

[0112] Based on the basic user profiles, daily behavior trajectories, and purchase information of alcohol consumers, format and transform them to obtain time series values ​​in a one-to-one correspondence;

[0113] A time-feature data set is obtained according to the time series values.

[0114] The primary purpose of this implementation is to integrate user profiles, daily behavior trajectories, and purchase information into a structured time-feature dataset, providing high-quality input for subsequent curvelet transform and convolutional neural network analysis. This process transforms the raw data into a unified time series format through formatting, ensuring alignment and interpretability of the multi-dimensional data along the temporal dimension.

[0115] First, formatting and time series numerical construction are performed. Raw data is extracted from user profiles, daily behavior trajectories, and purchase information results. Unstructured data (such as text and dates) is converted into numerical form:

[0116] In the user basic file, age (numeric value), gender (binary code, such as 0 / 1), date of birth (age calculation), and registration time (timestamp).

[0117] In daily behavior trajectories, the login time (timestamp), product browsing time (duration), number of order cancellations (count), and number of product favorites (count) are recorded.

[0118] The purchase result information includes the purchase amount (value), quantity (count), purchase time (timestamp), and purchase store (classification code, such as 1 = online, 2 = offline).

[0119] Align all data at a unified time granularity (such as day or week) to generate time series values

[0120]

[0121] in, t Indicates a point in time, N represents the consumption characteristics that can be selected, y Indicates consumption characteristics N In time tNumerical expression of .

[0122] During formatting and time series data construction, static data (such as age) is repeatedly filled in to ensure that all features are consistent across time. For example, if user C is 30 years old, the value is always 30 at all time points. Missing values ​​are also handled by interpolation (such as linear interpolation) or mean filling to avoid data gaps.

[0123] Next, we construct a time-feature dataset. The formatted data is organized into a three-dimensional tensor (user × time × feature) based on the time and feature dimensions. The specific format is:

[0124]

[0125] in: Represents a collection of profile characteristics (e.g., age, gender); Represents a collection of behavioral characteristics (such as browsing time, number of collections); Represents a set of purchase characteristics (such as purchase amount and quantity).

[0126] Defining ranking based on feature importance S , Indicates that the archive data is sorted in ascending order by timestamp. Indicates that the behavior data is sorted in descending order by the frequency of the behavior; Indicates that the purchase data is sorted in descending order by purchase amount.

[0127] This implementation aligns data by user ID, ensuring that all features for the same user are consistent over time. Using sliding window technology, a fixed window (e.g., the last 30 days) is used to dynamically update the dataset, capturing the time-varying nature of behavioral patterns.

[0128] This implementation ensures data consistency and integrity through formatting and time series mapping. For example, a user's "age 30" remains consistent across all time points, avoiding bias in isolated analysis and improving data quality. By aligning user IDs, the model can identify the correlation between a user's "high-end red wine preference" (behavioral trajectory) and "average monthly spending of 800 yuan" (purchase data), enhancing multi-dimensional correlation. Sliding window technology can detect user C's "recent surge in browsing frequency," capturing dynamic behavior and providing time-sensitive features for prediction.

[0129] As a preferred embodiment of this embodiment, the basic user profile, daily behavior trajectory, and purchase information results of the alcohol consumption user are specifically as follows:

[0130] The basic user profile includes at least one of the consumer's age, gender, date of birth, and registration time;

[0131] Daily behavior traces include at least one of the following: time logged into the platform, time spent browsing products, time initiating payment, number of order cancellations, number of product collections, and number of activity participations;

[0132] The purchase result information includes at least one of the purchase amount, quantity, purchase time, and purchase store.

[0133] The main purpose of this embodiment is to clarify the specific content of user basic files, daily behavior trajectories and purchase result information, ensure that the data covers the full-dimensional characteristics of consumer behavior, and provide quantifiable input for subsequent analysis.

[0134] A user's basic profile includes at least one of the following fields: age, gender, date of birth, and registration time. Note that for static feature encoding, gender is encoded as binary (0 / 1), date of birth is converted to age, and registration time is converted to registration duration (e.g., "Registered for 1 year").

[0135] It should be noted that if the data allows, other fields (such as occupation, region) can be added, and the present invention does not limit this.

[0136] Daily behavior traces include at least one of the following fields: platform login time, product browsing time, payment start time, number of order cancellations, number of product collections, and number of activity participations.

[0137] We quantify daily behavior trajectory data by recording "platform login time" as a timestamp and "product browsing time" as a duration (e.g., "average stay of 15 minutes each time"). We also perform behavior clustering, combining "payment trigger time" with "number of order cancellations" to identify user hesitation in paying.

[0138] The purchase result information includes at least one of the following fields: purchase amount, quantity, purchase time, and purchase store.

[0139] For purchasing pattern analysis, "purchase amount" and "purchase time" are combined to identify peak consumption on promotional days. "Purchase location" is coded as a categorical variable (e.g., online / offline). Furthermore, to address outliers, a threshold is set for "purchase amount" (e.g., excluding outliers > 100,000 yuan) to reduce noise.

[0140] This embodiment ensures comprehensive features. By combining the "age" in the user's basic profile with the "amount" of purchase results, we can identify the correlation between age and spending power (e.g., a 30-year-old user's average monthly spending is 800 yuan). We also identify behavioral patterns by combining the "number of collections" in daily behavior trajectories with the "quantity" of purchase results to determine user interest and purchase conversion rate (e.g., after 10 collections, two bottles were purchased). Scenario-based analysis can also be performed by categorizing purchase locations by "online / offline" to analyze channel preferences, such as a user's preference for purchasing high-end wine online.

[0141] Specifically, the user's data is:

[0142] Basic profile: age 35 years old (numeric value), gender female (coded as 1), registration time January 2020 (registration duration 4 years).

[0143] Behavior: Log in to the platform three times a week, stay for an average of 10 minutes, browse the wine page for 60% of the time, and add items to favorites five times a month.

[0144] Purchase results: I bought red wine 3 times in the past six months, with an average single purchase amount of 600 yuan, and all purchases were made online.

[0145] Based on the above data, a time-feature dataset was constructed. The profile characteristics are: age 35, gender 1, registration duration 4 years; behavioral characteristics: login 3 times a week, number of collections 5 times / month; purchase characteristics: purchase amount 600 yuan / time, online purchase.

[0146] By associating the "number of collections" in the behavioral trajectory with the "amount" of the purchase result, we can predict user D's potential purchase intention for high-end red wine.

[0147] Overall, this implementation improves data integrity and consistency. Formatting conversion and time series mapping ensure alignment of all features across the time dimension, avoiding biases that can arise from isolated analysis. By integrating multi-dimensional features, user profiles, behavioral trajectories, and purchase results, we can capture the correlation between long-term trends (such as age and spending power) and short-term behaviors (such as a surge in favorites). This provides scenario-based forecasting, and the classification coding of purchasing stores supports channel preference analysis, providing accurate consumption forecasts.

[0148] This implementation constructs a structured, high-dimensional time-feature dataset, laying a solid foundation for feature decomposition and prediction using curvelet transform and convolutional neural networks, significantly improving the accuracy and generalization ability of alcohol consumption prediction.

[0149] As a preferred embodiment of the present invention, based on the time-feature data set, a plurality of curvelet coefficients are obtained by curvelet transformation in combination with the curvelet characteristic frequency, specifically:

[0150] According to the time-feature data set, curvelet coefficients at 6 scales are obtained through curvelet transformation;

[0151] Combined with the characteristic frequency of the curvelet, the curvelet coefficients at six scales are divided into low-frequency curvelet coefficients, medium-frequency curvelet coefficients, high-frequency curvelet coefficients and ultra-high-frequency curvelet coefficients;

[0152] The intermediate frequency curvelet coefficient and the high frequency curvelet coefficient both include curvelet coefficients at two scales.

[0153] The main purpose of this implementation is to decompose a time-feature dataset into six-scale curvelet coefficients using the curvelet transform. These coefficients are then classified into four frequency categories: low-frequency, medium-frequency, high-frequency, and ultra-high-frequency. The medium-frequency and high-frequency categories each contain two scales. This process aims to separate global trends (low-frequency), medium-level details (medium-frequency), short-term fluctuations (high-frequency), and noise (ultra-high-frequency) in the data, providing clear multi-scale features for subsequent feature decomposition.

[0154] The scale corresponds to the coarseness or fineness of the frequency. A low scale (such as Scale 1) corresponds to low frequencies (global trends), and a high scale (such as Scale 6) corresponds to high frequencies (details or noise).

[0155] The low-frequency curve coefficient (C{1}) corresponds to global trends, such as users' long-term spending power and registration duration.

[0156] The medium-frequency curvature coefficients (C{2}, C{3}) correspond to medium-level details, such as weekly consumption fluctuations and periodicity in activity participation. The high-frequency curvature coefficients (C{4}, C{5}) correspond to short-term mutations, such as a single-day surge in browsing and shifts in interest. The ultra-high-frequency curvature coefficient (C{6}) corresponds to noise or outliers, such as data collection errors or misoperation.

[0157] Input the time-feature dataset (such as the three-dimensional tensor of user X) into the curvelet transform algorithm and perform the curvelet transform to obtain the curvelet coefficients at different scales and angles.

[0158]

[0159] in, Indicated on scale ,direction ,Location The curvelet coefficient on Indicates any sort order The feature data set, Indicates consumption data in feature dimension and time The distribution of represents the curvelet function.

[0160] Perform multi-scale decomposition and decompose the data into 6 scales (Scale 1 to Scale 6) through a recursive filter bank:

[0161] Scale 1 represents low-frequency components (global trends); Scale 2-3 represents medium-frequency components (medium details); Scale 4-5 represents high-frequency components (short-term mutations); Scale 6 represents ultra-high-frequency components (noise).

[0162] Directional parameter setting: the mid-frequency scale (Scale 2-3) uses 32 directional parameters, and the high-frequency scale (Scale 4-5) uses 64 directional parameters, covering the geometric directions of two-dimensional space (such as 0°, 45°, 90°, etc.).

[0163] The numerical distribution of the curvelet coefficient set is shown in the following table:

[0164]

[0165] For low frequencies (Scale 1), only isotropic features are retained. For mid- and high-frequency coefficients (Scales 2-5), directional analysis is required. A threshold is set for ultra-high frequency coefficients (Scale 6) (for example, only coefficients with an absolute value greater than 0.1 are retained) to filter out noise.

[0166] Frequency classification and coefficient classification are performed. For low-frequency coefficients, the coefficients of Scale 1 are classified as low-frequency curvature coefficients (C{1}). For mid-frequency coefficients, the coefficients of Scales 2 and 3 are classified as mid-frequency curvature coefficients (C{2}, C{3}). For high-frequency coefficients, the coefficients of Scales 4 and 5 are classified as high-frequency curvature coefficients (C{4}, C{5}). For ultra-high-frequency coefficients, the coefficients of Scale 6 are classified as ultra-high-frequency curvature coefficients (C{6}).

[0167] The mid-frequency coefficients (C{2}, C{3}) and high-frequency coefficients (C{4}, C{5}) must retain all directional information to support subsequent directional feature extraction. The directional weights of the high-frequency coefficients (C{4}, C{5}) are adjusted based on the importance of the behavior (e.g., "diagonal browsing" has a higher weight).

[0168] Specifically, the user's raw data: low frequency, average monthly consumption of 800 yuan (long-term trend); medium frequency, browsing the wine page for 20 minutes every Wednesday (weekly pattern); high frequency, browsing of high-end products on promotional days surges by 300% (short-term mutation); noise, erroneous order cancellation (Scale 6).

[0169] The decomposition results are as follows: C{1} (low frequency), extracting the "average monthly consumption of 800 yuan" trend to predict long-term repurchase potential; C{2} (medium frequency), directional parameters identifying "Wednesday browsing preference" (horizontally); C{4} (high frequency), directional parameters capturing "diagonal browsing of high-end products" (diagonally); C{6} (ultra-high frequency), filtering out "inadvertent order cancellation" noise to avoid misjudging purchase intention.

[0170] In this implementation, multi-scale feature separation, separating low-frequency (C{1}) and high-frequency (C{4}-C{5}) features, enables the model to capture both "long-term stable users" and "short-term interest surges," improving prediction accuracy to 85%. Directional detail is enhanced, with directional parameters (e.g., 32 directions) for medium-frequency (C{2}-C{3}) and high-frequency (C{4}-C{5}) features supporting geometric feature analysis of behavioral patterns (e.g., "diagonal browsing" corresponds to interest in high-end products). Noise suppression and robustness are enhanced, with threshold filtering for ultra-high frequencies (C{6}) reducing outlier interference and ensuring model stability in data-sparse scenarios (e.g., new users).

[0171] This implementation uses multi-scale, multi-directional decomposition using the curvelet transform to convert the time-feature dataset into four types of coefficients: low-frequency, medium-frequency, high-frequency, and ultra-high-frequency. This provides structured, high-signal-to-noise ratio input features for the subsequent convolutional neural network. This not only improves the model's ability to capture complex patterns in consumer behavior but also enhances the robustness of predictions through noise filtering, laying a key foundation for accurate consumer forecasting.

[0172] As a preferred embodiment of this implementation, multiple feature information is obtained by feature decomposition based on the low-frequency curvelet coefficient, the intermediate-frequency curvelet coefficient, and the high-frequency curvelet coefficient, specifically:

[0173] According to the low-frequency curvelet coefficient, the trend characteristic information is obtained through the inverse curvelet transform;

[0174] According to the intermediate frequency curvelet coefficient, the intermediate frequency curvelet coefficient is classified into multiple directions, and the intermediate frequency feature information in four directions is obtained through the inverse curvelet transform;

[0175] According to the high-frequency curvelet coefficients, multiple directions in the high-frequency curvelet coefficients are classified, and the high-frequency feature information in four directions is obtained through inverse curvelet transformation.

[0176] This implementation aims to separate the low-frequency, mid-frequency, and high-frequency features of consumer data into interpretable directional information through coefficient decomposition and inverse transformation of the curvelet transform. Trend features are extracted, and the overall trend of consumer data (such as long-term spending power) is obtained through inverse transformation of the low-frequency coefficients. Directional features are decoupled, and through directional classification and inverse transformation of mid-frequency and high-frequency data, fine-grained features in different directions (such as horizontal browsing and diagonal clicks) are separated, improving the model's ability to capture complex behavioral patterns. Sparse representation optimization is performed, leveraging the sparsity of the curvelet coefficients to reduce redundant information and improve computational efficiency.

[0177] For the low-frequency curvelet coefficients, coefficient filtering is performed to retain only the low-frequency curvelet coefficients (C{1}), and the coefficients of other scales (C{2}-C{6}) are set to zero.

[0178]

[0179] Restore global trend characteristics.

[0180] in, Represents the trend information of consumption feature data in the time-feature domain, represents the curvelet coefficient of the low-frequency component, represents the inverse curvelet transform.

[0181] This embodiment utilizes the sparsity of low-frequency coefficients to retain only significant coefficients (e.g., absolute value > 0.1) to improve computational efficiency. The inverse transform result is low-pass filtered and smoothed to eliminate short-term fluctuation interference.

[0182] For the intermediate frequency direction classification and inverse transformation, the intermediate frequency coefficients (C{2}, C{3}) are divided into 32 directions each, and the first 8 directions are divided into 4 groups (2 directions in each group):

[0183] Group 1, direction 1-2 (C{2}{1-2}, C{3}{1-2}); Group 2, direction 3-4 (C{2}{3-4}, C{3}{3-4}); Group 3, direction 5-6 (C{2}{5-6}, C{3}{5-6}); Group 4, direction 7-8 (C{2}{7-8}, C{3}{7-8}).

[0184] Perform inverse transform on the coefficients of each direction to generate the intermediate frequency features of the four directions

[0185]

[0186] in, Represents the detailed information of consumption feature data in the first direction of time-feature domain, Indicates the curvelet coefficients in the first 8 directions under the intermediate frequency component, represents the inverse curvelet transform.

[0187] Similarly, by dividing the curvelet coefficients in different directions according to the above process, the characterization information of the other three directions corresponding to the intermediate frequency signal in the time-feature domain can be calculated, which are recorded as: .

[0188] Define direction labels based on behavior patterns (e.g., Group 1 is "horizontal browsing," Group 4 is "diagonally rightward saving"). Give higher weight to directions strongly associated with consumption (e.g., "diagonally browsing high-end products").

[0189] For high-frequency direction classification and inverse transformation, direction grouping is performed. The high-frequency coefficients (C{4}, C{5}) are divided into 64 directions each, and the first 16 directions are divided into 4 groups (4 directions in each group):

[0190] Group 1, direction 1-4 (C{4}{1-4}, C{5}{1-4}); Group 2, direction 5-8 (C{4}{5-8}, C{5}{5-8}); Group 3, direction 9-12 (C{4}{9-12}, C{5}{9-12}); Group 4, direction 13-16 (C{4}{13-16}, C{5}{13-16}).

[0191] Perform inverse transform on the coefficients of each direction to generate high-frequency features in four directions .

[0192] Apply wavelet thresholding (e.g., soft thresholding) to high-frequency coefficients to filter out minor fluctuations. Prioritize mutations and retain coefficients with high mutation energy within the directional group (e.g., a "surge in diagonal browsing" on promotional days).

[0193] For ultra-high frequency noise filtering, the coefficients of C{6} (ultra-high frequency) are directly filtered because they correspond to noise or abnormal values ​​(such as misoperation) and do not participate in the inverse transform.

[0194] Specifically, the original user data is as follows: low frequency (C{1}), with an average monthly consumption of 1,000 yuan (trend feature); medium frequency (C{2}-C{3}): Group 1 (direction 1-2), horizontally browsing the wine page for 20 minutes / week, Group 4 (direction 7-8), adding high-end products to the right; high frequency (C{4}-C{5}): Group 1 (direction 1-4), diagonally browsing to the left on promotion days increased by 300%, Group 4 (direction 13-16), quickly clicking the payment button.

[0195] Execute the decomposition results to obtain trend characteristics, Identify "Long-term consumption stability index = 95%".

[0196] Extract mid-frequency features, group 1 (X SM1 ) increases the horizontal browsing time of red wine by 30%. Group 4 (XSM4 ) is diagonally to the right, and the frequency of collecting high-end products increases.

[0197] Extract high-frequency features, group 1 (X SH1 ) is a promotion day, browsing high-end products diagonally to the left increases by 300%, group 4 (X SH4 ) Double the number of quick click payments on promotional days.

[0198] This implementation uses curvelet transform feature decomposition and directional classification to achieve multi-scale separation, separating low-frequency trends, medium-frequency periodicity, and high-frequency mutations, covering the full range of consumer behavior scenarios. Directional decoupling and grouping of medium- and high-frequency directions (e.g., four directions) ensure feature independence and enhance the model's ability to analyze complex behaviors. Sparse representation optimization retains only coefficients in key directions, reducing redundant information, accelerating computation, and improving model efficiency.

[0199] This embodiment combines the sparsity and directional analysis of the curvelet transform to provide structured, high signal-to-noise ratio input features for subsequent convolutional neural networks, significantly improving the accuracy and interpretability of consumption forecasts.

[0200] Specifically, based on the plurality of feature information, a plurality of convolution operation results are obtained through a convolutional neural network, specifically:

[0201] According to the trend feature information, a convolution operation result is obtained by sharpening the convolution kernel;

[0202] According to the horizontal intermediate frequency feature information and high frequency feature information, the vertical edge detection convolution kernel is used to extract the vertical gradient change and obtain the convolution operation results respectively;

[0203] According to the longitudinal intermediate frequency feature information and high frequency feature information, the horizontal edge detection convolution kernel is used to extract the horizontal gradient change and obtain the convolution operation results respectively;

[0204] According to the oblique intermediate frequency feature information, a convolution operation result consistent with the oblique intermediate frequency feature information is obtained through a unit convolution kernel;

[0205] According to the high-frequency feature information of the oblique direction, the convolution operation result is obtained through the image smoothing convolution kernel.

[0206] This embodiment aims to perform differentiated processing on nine directional information of low-frequency, medium-frequency, and high-frequency features through a directionally designed convolution kernel, thereby enhancing effective features, suppressing noise, and ultimately outputting a feature map that is strongly correlated with consumer behavior patterns.

[0207] First, trend feature processing (low frequency) is performed, and trend feature information (X SL ), which represents the global trend of consumption data (such as average monthly consumption). The convolution kernel selects the sharpening convolution kernel. ,

[0208]

[0209] Through convolution operation

[0210]

[0211] in, Represents the result of the convolution operation.

[0212] The center-weighted kernel highlights the mutation points in the trend (such as the "Double Eleven consumption surge") while retaining details, such as the monthly fluctuations hidden in the user's "long-term consumption stability" trend are enhanced.

[0213] In addition, it is convenient for key point detection. For example, user B’s “sudden drop in average monthly consumption” is easier to be recognized by the model after being sharpened.

[0214] Processing of horizontal intermediate frequency and high frequency characteristics, horizontal intermediate frequency X SM1 (Horizontally distributed intermediate frequency behavior, such as "weekly horizontal browsing of red wine"), horizontal high frequency X SH1 (High-frequency behavior with horizontal distribution, such as “horizontal browsing surge on promotion days”).

[0215] Convolution kernel selects vertical edge detection convolution kernel (w vt ),For example

[0216]

[0217] Used to detect vertical gradient changes in horizontally distributed signals (such as "vertical fluctuations in browsing time").

[0218] The convolution operation formula is expressed as

[0219]

[0220]

[0221] in, Represents the result of the convolution operation.

[0222] Vertical gradient changes correspond to sudden changes in behavior (e.g., a surge in browsing time on promotional days). Separating horizontal behavior from vertical changes avoids interference with features in other directions. For example, in a user's "horizontal browsing every Wednesday" trend, unusual fluctuations during promotional weeks are captured by the vertical edge detection kernel.

[0223] Processing of longitudinal intermediate frequency / high frequency features, where longitudinal intermediate frequency X SM3 (Medium-frequency behavior in vertical distribution, such as "details page scrolling time"). Vertical high-frequency X SH3(High-frequency behaviors with vertical distribution, such as “fast scrolling on promotion days”).

[0224] Convolution kernel selects horizontal edge detection convolution kernel (w zh ),For example

[0225]

[0226] Used to detect horizontal gradient changes in vertically distributed signals (such as "horizontal fluctuations in scrolling duration").

[0227] The convolution operation formula is expressed as

[0228]

[0229]

[0230] in, Represents the result of the convolution operation.

[0231] Horizontal gradient analysis identifies sudden changes in vertical behavior (e.g., user F's "fast scrolling through the detail page on promotion days"). Vertical behavior is separated from horizontal changes to avoid cross-directional interference. This is used for behavioral pattern recognition, for example, a sudden increase in a user's "vertical scrolling" behavior during a high-frequency promotion is detected by Y. SH3 Capture.

[0232] For the tilted IF feature processing, the tilted IF X SM2 (diagonally left), X SM4 (diagonally to the right), indicating medium-frequency behavior in the diagonal direction (e.g., “collecting goods diagonally”).

[0233] Convolution kernel selection unit convolution kernel (w id ),For example:

[0234]

[0235] Used to preserve the original features of the tilt direction and avoid information loss.

[0236] The convolution operation formula is expressed as

[0237]

[0238]

[0239] in, Represents the result of the convolution operation.

[0240] The original data of the tilt direction is directly passed to the next layer, preserving the direction information. This ensures directional integrity. For example, a user's "diagonally collecting high-end products" behavior is fully preserved, supporting subsequent model analysis.

[0241] For tilted high frequency feature processing, tilted high frequency: X SH2 (diagonally left), X SH4 (diagonally to the right), indicating high-frequency behavior in the diagonal direction (e.g., a “surge in diagonal browsing”).

[0242] The convolution kernel selects Gaussian smoothing convolution kernel (w gs ),For example:

[0243]

[0244] Used to suppress noise (such as erroneous operations) and retain valid mutations.

[0245] The convolution operation formula is expressed as

[0246]

[0247]

[0248] in, Represents the result of the convolution operation.

[0249] The Gaussian smoothing kernel reduces high-frequency noise while preserving true mutation signals. This achieves noise suppression. For example, during a user's "surge in diagonal browsing," the noise generated by misoperation is smoothed while the true shift in interest is preserved.

[0250] This embodiment uses vertical and horizontal edge detection kernels to accurately extract gradient changes in horizontal or vertical behavior (e.g., "surge in browsing on promotional days"), enhancing directional features. Frequency-specific processing enhances trend details by sharpening low frequencies, suppresses noise by smoothing high frequencies, and preserves original features by tilting mid-frequency directions. Nine independent channels prevent cross-directional interference, improving the model's ability to analyze complex behavioral patterns.

[0251] This embodiment follows the feature extraction principle of convolutional neural networks and significantly improves the accuracy and robustness of the consumption prediction model through directional convolution kernel optimization.

[0252] As a preferred embodiment of the present invention, a pooling operation result is obtained through pooling operation, specifically:

[0253] Based on the multiple convolution operation results, average pooling is performed to obtain a pooling operation result representing the average value within the pooling window, so as to reduce the data dimension.

[0254] This implementation aims to reduce the spatial resolution of the convolution operation results and the data dimension through average pooling, thereby improving computational efficiency and enhancing model robustness. At the same time, by calculating the average value within the pooling window, the feature map is further smoothed to prevent overfitting and ensure that the model is more invariant to small changes in the input data.

[0255] The implementation of average pooling is as follows:

[0256]

[0257] in, Represents the final pooling operation result, Parameters such as represent the feature information extracted by the convolution kernel. Represents the pooling process.

[0258] Perform average pooling on the convolution result of each channel and calculate the average value within the pooling window.

[0259]

[0260] Among them, x i Indicates the first elements, Indicates the total number of elements within the pooling window.

[0261] In this implementation, an appropriate pooling window size (e.g., 2×2 or 3×3) is selected based on the spatial resolution of the feature map. A larger window can significantly reduce dimensionality but may lose details; a smaller window retains more feature information.

[0262] The stride is usually the same as the pooling window size (e.g. a 2×2 window uses a stride of 2) to avoid feature overlap. If you want to retain more features, you can choose a smaller stride (e.g. a stride of 1).

[0263] In this implementation, average pooling reduces the spatial resolution of the feature map, reducing the computational effort of subsequent network layers, achieving dimensionality reduction and improving computational efficiency. By averaging the elements within a window, average pooling reduces reliance on individual pixels and reduces the model's sensitivity to noise, thereby preventing overfitting. The pooling process enhances the spatial invariance of features, making the model more robust to small changes in the input data.

[0264] As a preferred embodiment of the present invention, alcohol consumption prediction is obtained through linear correction activation, specifically:

[0265] Based on the pooling operation results, combined with the weight matrix and bias vector, the alcohol consumption prediction result is obtained;

[0266] The weight matrix is ​​a two-dimensional matrix consisting of the number of output neurons and the number of input neurons, and the bias vector is a one-dimensional array of the number of output neurons. The weight matrix and the bias vector are obtained through training.

[0267] This implementation aims to globally integrate convolutional and pooling feature maps through fully connected layers to extract higher-level abstract features. The introduction of the ReLU activation function further enhances the model's nonlinear expressiveness, enabling the network to better capture complex consumer behavior patterns.

[0268] The linear correction activation formula is

[0269]

[0270] in, is the output of the fully connected layer. It is the input of the fully connected layer and contains the key feature information of each channel after dimensionality reduction. It is an activation function, which introduces nonlinear characteristics to prevent the model from fitting only linear relationships. Its characteristics include simple calculation, fewer gradient disappearance problems, and easy optimization. is the weight matrix, and the dimension of the weight matrix 𝑊 is determined by the size of the input feature map and the number of output neurons. The role of is to map the input features to a higher or lower dimensional space to meet the needs of subsequent tasks (such as classification or regression). is a bias vector used to adjust the offset of the output value and enhance the flexibility of the model.

[0271] In this embodiment, the feature map after pooling Flattened into a one-dimensional vector, to ensure that all features can be processed uniformly by the fully connected layer. For example, if The resolution is 4×4×9 (assuming there are 9 channels), so after flattening, we get a vector of length 144. Use a suitable initialization method (such as Xavier initialization or He initialization) to set the weight matrix and the bias vector , to avoid gradient explosion or vanishing problems. Applying the ReLU activation function based on the weighted summation enhances the nonlinear expression ability of the model, enabling it to capture more complex behavior patterns.

[0272] In this implementation, the fully connected layer globally integrates the local features of each channel to form a higher level of abstract features, which is convenient for subsequent tasks. The introduction of the ReLU activation function enables the model to capture nonlinear relationships. By adjusting the weight matrix and the bias vector , the fully connected layer can flexibly adapt to different task objectives (such as classification or regression) and improve the versatility of the model.

[0273] As a preferred embodiment of this implementation, the weight matrix and the bias vector are obtained through training, specifically:

[0274] Determining training data and test data, wherein the training data and test data both include basic user profiles, daily behavior trajectories, input data of purchase information results of alcohol consumers, and classification data obtained based on the purchase data of the current year;

[0275] Performing training based on the input data and the classification data in the training data to obtain multiple sets of the weight matrices and the bias vectors;

[0276] Based on the input data in the test data, multiple groups of the weight matrices and the bias vectors are combined to obtain multiple prediction results, and the weight matrix and the bias vector are determined based on the correlation between the multiple prediction results and the hierarchical data in the test data.

[0277] This embodiment aims to optimize the weight matrix through training data and the bias vector , so that the model can accurately predict user consumption levels; verify and select the optimal parameters through test data to ensure the model's generalization ability to new data.

[0278] It should be noted that this example randomly divides the dataset into two groups (e.g., 80% training set, 20% test set) to ensure consistent distribution between the two data types. For example, a user with an average monthly spending of 5,000 yuan is labeled as a "high-value user" (tiered data), and their behavior trajectory (e.g., "browsing wine three times a week") is used as input data. Stratified sampling ensures consistent distribution between the test and training sets, preventing model bias.

[0279] According to the training data, the model is trained to obtain multiple sets of weight matrices and bias vectors. First, the parameters are initialized. For the weight matrix , use Xavier initialization to ensure that the initial weight distribution is reasonable and avoid gradient disappearance / explosion. For the bias vector , initialized to a zero vector or small random values ​​to enhance model flexibility.

[0280] During the training process, the input data is propagated forward through the fully connected layer to calculate the prediction results:

[0281]

[0282] in, Indicates the samples.

[0283] Perform loss calculation and use the cross entropy loss function to measure the difference between the prediction results and the graded data:

[0284]

[0285] in, is the true label, is the predicted probability.

[0286] Back propagate again and calculate the loss pair by chain derivation and The gradient of . Through parameter update, the Adam optimizer is used to dynamically adjust the learning rate and update the parameters:

[0287]

[0288] in, is the learning rate.

[0289] Generate multiple sets of parameters and generate multiple sets of (W, b) through multiple training (different initialization or hyperparameters) to avoid local optimality.

[0290] Select the optimal weight matrix and bias vector based on the test data. Input the test data into the model and use multiple sets of (W, b) to get the prediction results:

[0291]

[0292] in, Indicates the Group parameters.

[0293] Select the (W, b) that performs best on the test set. Note that if a set of parameters has a low training loss but a high test loss, it should be discarded to prevent overfitting.

[0294] The present invention also provides an electronic device, comprising:

[0295] Memory, for storing computer instructions;

[0296] A processor is used to implement the wine consumption prediction method combining curvelet transform and convolutional neural network when executing the computer instructions.

[0297] Therefore, this electronic device can achieve any effect of the wine consumption prediction method combining curvelet transform and convolutional neural network, which will not be described in detail here.

[0298] Anything not described in the present invention can be achieved by adopting or drawing on existing technologies.

[0299] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0300] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A method for predicting alcohol consumption by combining curvelet transform and convolutional neural network, characterized in that: include: Based on the user basic profiles, daily behavior trajectories, and purchase information results of alcohol consumption users, a time-feature dataset is obtained; According to the time-feature data set, curvelet transform is performed to obtain curvelet coefficients at 6 scales, and combined with the curvelet characteristic frequency, multiple curvelet coefficients are obtained, wherein the multiple curvelet coefficients include low-frequency curvelet coefficients, intermediate-frequency curvelet coefficients, high-frequency curvelet coefficients and ultra-high-frequency curvelet coefficients, and the intermediate-frequency curvelet coefficients and high-frequency curvelet coefficients both include curvelet coefficients at 2 scales; Based on the low-frequency curvelet coefficient, the intermediate-frequency curvelet coefficient, and the high-frequency curvelet coefficient, a plurality of feature information is obtained through feature decomposition, wherein, based on the low-frequency curvelet coefficient, trend feature information is obtained through inverse curvelet transformation; based on the intermediate-frequency curvelet coefficient, multiple directions in the intermediate-frequency curvelet coefficient are classified, and four directions of intermediate-frequency feature information are obtained through inverse curvelet transformation; based on the high-frequency curvelet coefficient, multiple directions in the high-frequency curvelet coefficient are classified, and four directions of high-frequency feature information are obtained through inverse curvelet transformation; Based on the plurality of feature information, a plurality of convolution operation results are obtained through a convolutional neural network, wherein, based on the trend feature information, a convolution operation result is obtained by using a sharpening convolution kernel; based on the horizontal intermediate frequency feature information and the high frequency feature information, a vertical edge detection convolution kernel is used to extract the longitudinal gradient change, and the convolution operation results are obtained respectively; based on the vertical intermediate frequency feature information and the high frequency feature information, a horizontal edge detection convolution kernel is used to extract the horizontal gradient change, and the convolution operation results are obtained respectively; based on the oblique intermediate frequency feature information, a unit convolution kernel is used to obtain a convolution operation result consistent with the oblique intermediate frequency feature information; based on the oblique high frequency feature information, a convolution operation result is obtained by using an image smoothing convolution kernel; A pooling operation result is obtained through pooling operation, and the alcohol consumption prediction is obtained through linear correction activation.

2. The method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to claim 1, characterized in that: Based on the user basic profiles, daily behavior trajectories, and purchase information results of alcohol consumption users, a time-feature dataset is obtained, specifically: Based on the basic user profiles, daily behavior trajectories, and purchase information of alcohol consumers, format and transform them to obtain time series values ​​in a one-to-one correspondence; A time-feature data set is obtained according to the time series values.

3. The method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to claim 2, characterized in that: The basic user profile, daily behavior trajectory, and purchase information results of alcohol consumption users are as follows: The basic user profile includes at least one of the consumer's age, gender, date of birth, and registration time; Daily behavior traces include at least one of the following: time logged into the platform, time spent browsing products, time initiating payment, number of order cancellations, number of product collections, and number of activity participations; The purchase result information includes at least one of the purchase amount, quantity, purchase time, and purchase store.

4. The method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to claim 1, characterized in that: A pooling operation result is obtained through the pooling operation, specifically: Based on the multiple convolution operation results, average pooling is performed to obtain a pooling operation result representing the average value within the pooling window, so as to reduce the data dimension.

5. The method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to claim 1, characterized in that: The alcohol consumption prediction is obtained through linear correction activation, specifically: Based on the pooling operation results, combined with the weight matrix and bias vector, the alcohol consumption prediction result is obtained; The weight matrix is ​​a two-dimensional matrix consisting of the number of output neurons and the number of input neurons, and the bias vector is a one-dimensional array of the number of output neurons. The weight matrix and the bias vector are obtained through training.

6. The method for predicting alcohol consumption by combining curvelet transform and convolutional neural network according to claim 5, characterized in that: The weight matrix and the bias vector are obtained through training, specifically: Determining training data and test data, wherein the training data and test data both include basic user profiles, daily behavior trajectories, input data of purchase information results of alcohol consumers, and classification data obtained based on the purchase data of the current year; Performing training based on the input data and the classification data in the training data to obtain multiple sets of the weight matrices and the bias vectors; Based on the input data in the test data, multiple groups of the weight matrices and the bias vectors are combined to obtain multiple prediction results, and the weight matrix and the bias vector are determined based on the correlation between the multiple prediction results and the hierarchical data in the test data.

7. An electronic device, characterized in that: include: Memory, for storing computer instructions; A processor, configured to implement the method for predicting alcohol consumption by combining curvelet transform and convolutional neural network as described in any one of claims 1 to 6 when executing the computer instructions.

Images