A recommendation method based on deep learning DCP

Abstract

The present invention discloses a recommendation method based on deep learning DCP, which belongs to the technical field of commodity recommendation. The method comprises the following steps: obtaining an online platform user interaction behavior data set, preprocessing the online platform user interaction behavior data set to obtain a preprocessed data set; constructing an initial recommendation model, training the initial recommendation model based on the preprocessed data set to obtain a recommendation model; inputting user data into the recommendation model to obtain recommended commodity information, and pushing the recommended commodity information to the user. The recommendation model constructed by the present invention can well identify and process the nonlinear relationship between input features, improve the relevance of the recommendation, and make the recommendation more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of commodity recommendation, and in particular to a recommendation method based on deep learning DCP. Background Art

[0002] The rapid development of Internet technology has brought great convenience to people's lives, but it has also led to a sharp increase in the amount of data, which has made users face difficulties in information selection. To meet this challenge, recommendation algorithm models have emerged. Their main function is to use user portraits to filter out content that meets the user's interests from a large amount of data and generate a personalized recommendation list. This not only effectively alleviates the problem of information overload, but also improves the relevance and attractiveness of the content. In recent years, the rapid development of deep learning technology has promoted the rapid development of recommendation systems, especially in the field of product recommendation. By introducing deep learning models to analyze user behavior sequences, such algorithms can more accurately characterize user portraits and generate more accurate recommendations based on them. However, the accuracy of product recommendations in existing recommendation methods is not very high, and the expected recommendation effect cannot be achieved. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a recommendation method based on deep learning DCP to solve the problems existing in the above-mentioned prior art.

[0004] To achieve the above object, the present invention provides a recommendation method based on deep learning DCP, comprising the following steps:

[0005] Acquire an online platform user interaction behavior dataset, and preprocess the online platform user interaction behavior dataset to obtain a preprocessed dataset;

[0006] Constructing an initial recommendation model, and training the initial recommendation model based on the preprocessed data set to obtain a recommendation model;

[0007] The user data is input into the recommendation model to obtain recommended product information, and the recommended product information is pushed to the user.

[0008] Preferably, the method for pretreatment comprises:

[0009] LabelEncoder is used to assign a unique integer value to different category labels of behavior type data in the online platform user interaction behavior dataset;

[0010] The embedding vectorization method is used to map the high-dimensional category features in the online platform user interaction behavior dataset into a dense vector of fixed length.

[0011] Preferably, the initial recommendation model includes an embedding and feature stacking layer, a maximum-mean attention mechanism layer, a bitwise attention mechanism layer, a deep feedforward network layer and a combined output layer;

[0012] The embedding and feature stacking layer is used to map the preprocessed data set to an embedding space to obtain an embedding vector;

[0013] The maximum-mean attention mechanism layer is used to purify and refine the information of the embedded vector by calculating the weight distribution to obtain the feature vector;

[0014] The bit-by-bit attention mechanism layer is used to perform a refined weight assignment on each feature vector to obtain the contribution of each feature vector to the final prediction in each dimension;

[0015] The deep feedforward network layer includes multiple fully connected layers and introduces nonlinearity through activation functions to capture high-order interactions in the embedding vector;

[0016] The combined output layer is used to integrate the output results of the maximum-mean attention mechanism layer, the bit-by-bit attention mechanism layer and the deep feedforward network layer to obtain the predicted output.

[0017] Preferably, the expression for obtaining the comprehensive features by integrating the maximum-mean attention mechanism layer and the bit-by-bit attention mechanism layer is:

[0018] F MMB =σ(W MMB ·F concat +b MMB )

[0019] In the formula, σ represents the Sigmoid activation function, F concat represents the combination of the embedding vectors processed by the max-mean attention layer and the bitwise attention layer, W MMB represents the weights learned by the model, b MMB Represents the bias term.

[0020] Preferably, the expression for obtaining the feature vector is:

[0021] F MM =α·max(E)+(1-α)·mean(E)

[0022] Where α represents a hyperparameter, E represents an embedding vector set, max(E) represents the maximum value operation on the embedding vector set, and mean(E) represents the mean operation on the embedding vector set.

[0023] Preferably, the expression for performing refined weight allocation is:

[0024] F B =WB ⊙E

[0025] Where E represents the embedding vector set, ⊙ represents element-wise multiplication, and W B represents the learned dimension-specific attention weights.

[0026] Preferably, the evaluation indicators for training the initial recommendation model include the area under the receiver operating characteristic curve and the logistic loss, and the expression of the logistic loss is:

[0027]

[0028] In the formula, N represents the total number of samples, y i represents the actual label of the i-th sample. If the actual label is positive, it is 1, otherwise it is 0. i is the probability that the model predicts that the i-th sample is a positive class.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] The present invention discloses a recommendation method based on deep learning DCP, which belongs to the technical field of commodity recommendation. The method comprises the following steps: obtaining an online platform user interaction behavior dataset, preprocessing the online platform user interaction behavior dataset to obtain a preprocessed dataset; constructing an initial recommendation model, training the initial recommendation model based on the preprocessed dataset to obtain a recommendation model; inputting user data into the recommendation model to obtain recommended commodity information, and pushing the recommended commodity information to the user. The DCP model of the present invention can better identify and process the nonlinear relationship between input features by integrating the maximum average pooling attention mechanism, thereby improving the relevance and efficiency of the recommendation. It also breaks through the limitation that simple linear models cannot effectively capture complex user behaviors, and provides a more in-depth method for understanding and predicting user click behaviors, so that the recommended commodities are more in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0032] Figure 1 A schematic diagram of a recommendation model according to an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of a maximum-mean attention mechanism layer and a bitwise attention mechanism layer in an embodiment of the present invention;

[0034] Figure 3 This is a comparison chart of loss values ​​of various models in an embodiment of the present invention;

[0035] Figure 4 AUC value comparison diagram of different structures in the embodiment of the present invention;

[0036] Figure 5 The figure shows the comparison of LogLoss values ​​of different structures in the embodiments of the present invention. DETAILED DESCRIPTION

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Embodiment 1

[0040] This embodiment provides a recommendation method based on deep learning DCP, including the following steps:

[0041] Acquire an online platform user interaction behavior dataset, and preprocess the online platform user interaction behavior dataset to obtain a preprocessed dataset;

[0042] Constructing an initial recommendation model, and training the initial recommendation model based on the preprocessed data set to obtain a recommendation model;

[0043] The user data is input into the recommendation model to obtain recommended product information, and the recommended product information is pushed to the user.

[0044] In this embodiment, two widely recognized large-scale online platform user interaction behavior datasets, Criteo and Avazu, are selected as the basis for simulating user behavior, mainly considering their convenience for comparison with other models, and the similarity between the interaction data of these two datasets and the reader station data. In view of the fact that the online platform of the reader station has not yet been officially launched and lacks a large amount of user data, it has become a feasible strategy to use these datasets to simulate user behavior and optimize the performance of the recommendation system accordingly. The Criteo dataset provides rich user click behavior records, which record the interaction between users and advertisements in detail, and contains approximately 45 million user behavior data. The data is divided into a training set and a test set, where the training set is used for model construction and adjustment, and the test set is used to evaluate the performance of the model. Usually, the training set contains the data of the previous 7 days, and the test set contains the data of the following day, which is done to ensure the time continuity and effectiveness of the experiment. The data records include not only the behavior type (such as click or no click) and timestamp, but also 13 continuous features and 26 categorical features, which include extensive information about users and ads, such as click frequency, continuous features of historical interaction times, and categorical features such as user ID, ad ID, and product category, as shown in Table 1.

[0045] Table 1

[0046]

[0047] The Avazu dataset focuses on the field of mobile advertising, tracking the click behavior of users on mobile devices, and provides about 40 million records. Features include device type, application context, ad display time, etc. There are a small number of continuous features and up to 24 category features. The details are shown in Table 2. In terms of the use of the dataset, Avazu's training set covers 11 days of data, while the test set is the data of the day after the training set time. This division helps to simulate the scenario of real-world advertising and test the generalization ability of the model on unknown data. By conducting experiments on these datasets, we can not only comprehensively evaluate the performance of the CTR model, but also deeply understand the key factors that affect user click behavior, and provide data support for the optimization of product recommendation strategies.

[0048] Table 2

[0049]

[0050] The data preprocessing method in this embodiment is as follows:

[0051] When processing categorical features in a data set, that is, data with no size relationship and no partial order relationship between feature values, LabelEncoder is often used for processing. The principle of LabelEncoder is to assign a unique integer value to each different category label. LabelEncoder implements encoding by assigning a unique number to each category state. This encoding method is very effective when processing non-numerical categorical data, especially when there are many possible values ​​for categorical features. LabelEncoder effectively reduces the sparsity of the data by mapping each category to an integer, and greatly reduces the dimension of the feature space compared to one-hot encoding. In summary, LabelEncoder is an effective tool for processing categorical features, especially when there are many feature categories that cause one-hot encoding to produce excessive sparsity problems. It effectively reduces the dimension of the feature space by assigning a unique integer to each category, while maintaining the information richness of the data set. At the same time, for the high-dimensional feature of the advertising category in the Avazu dataset, considering that the direct use of LabelEncoder may not effectively capture the potential relationship between categories, the embedding vectorization method is used for this type of feature. Each advertising category is mapped to a fixed-length dense vector, which not only reduces the dimension of the feature but also retains the semantic relationship between categories to a certain extent. This hybrid strategy - using LabelEncoder for behavior types and embedding vectorization for high-dimensional category features - provides an effective way to process and analyze complex user behavior data, providing a richer and more effective feature representation for model training and prediction.

[0052] The model architecture in this embodiment is as follows:

[0053] like Figure 1As shown in the figure, it consists of five layers in total: Embedding and Feature Stacking Layer (EFSL), Max-Mean Attention Mechanism Layer (MMAML), Bit-wise Attention Mechanism Layer (BAML), Deep Feedforward Network Layer (DFNL), and Top-level Combined Output Layer (TCOL). At the base layer of the architecture, the Embedding and Feature Stacking Layer (EFSL), the model takes dense features and sparse features and maps them to the embedding space to generate dense embedding vectors. This transformation provides a unified low-dimensional representation for the high-dimensional feature learning of the upper layer, laying the foundation for the automation of feature engineering. The secondary layer, the Max-Mean Attention Mechanism Layer (MMAML), purifies and refines the information of the embedding vector by calculating the weight distribution. This mechanism focuses on identifying and amplifying specific parts of the feature vector that have the greatest influence on the output prediction, and uses maximization or averaging operations to achieve this goal. The following bitwise attention mechanism layer (BAML) further refines the weight assignment for each feature vector to evaluate its contribution to the final prediction in each dimension. This bitwise attention evaluation allows the model to be optimized at the feature dimension level, enhancing the selectivity and sensitivity of the model to features. The deep feedforward network layer (DFNL), composed of multiple fully connected layers, introduces nonlinearity through activation functions, allowing the model to capture high-order interactions in the input data. Each layer increases the expressiveness of the model through nonlinear transformations, so that more complex data patterns can be learned. Finally, the top combined output layer (TCOL) combines the processing results of all previous layers and forms the final prediction output through a fully connected layer and subsequent Sigmoid activation function. This layer not only integrates multi-dimensional feature information, but also unifies feature learning at different levels into a prediction target to output prediction results in the form of probability.

[0054] Max-Mean and Bit-wise Attention Network In the improvement of the Deep & Cross Network (DCN) model, the Cross Network part was replaced by the Max-Mean and Bit-wise Attention Network to enhance the model's recognition and learning of feature importance. Max-Mean Attention provides a balanced feature focus strategy by combining the maximization operation (capturing the strongest feature signal) with the mean operation (preserving the overall effect of the feature). Compared with the original CrossNetwork, this attention-driven approach brings many advantages. It provides a fine-grained feature importance assessment, allowing the model to capture more subtle data patterns. In addition, the hybrid strategy of Max-Mean Attention not only captures the most significant signals, but also balances the comprehensive contribution of all features, thereby providing the model with a comprehensive feature understanding. The flexibility of this mechanism also provides adaptability for different tasks and data characteristics. Compared with the static feature crossover of Cross Network, the Max-Mean and Bit-wise Attention mechanism can show higher performance and adaptability when dealing with complex data relationships, which may bring significant improvements in prediction accuracy and model generalization. Figure 2 shown.

[0055] The Max-Mean Attention part solves the problem of emphasizing key features while maintaining the overall feature context by integrating maximization and mean operations, as shown in Equation 1.

[0056] F MM =α·max(E)+(1-α)·meam(E) (1)

[0057] Where E represents the set of embedding vectors, max(E) and mean(E) represent the maximum and mean operations respectively, and the hyperparameter α balances the influence of the maximum and mean values. This method provides a technique to balance the most prominent feature signal with the overall feature contribution, which is suitable for diverse feature sets.

[0058] The Bit-wise Attention part specifically addresses the challenge of refining the feature importance evaluation on each feature dimension, as shown in Equation 2.

[0059] F B =W B ⊙E (2)

[0060] Where W Bis the learned dimension-specific attention weight, ⊙ represents element-wise multiplication, which allows the model to adjust the weights of each dimension in an element-wise manner, thus providing fine-grained control over complex feature interactions.

[0061] Finally, the combination layer generates a comprehensive feature representation by combining the advantages of the above two attention mechanisms, as shown in Equation 3.

[0062] F MMB =σ(W MMB ·F concat +b MMB ) (3)

[0063] In the formula, F concat It is a combination of the embedding vector after Max-Mean and Bit-wise processing, W MMB is the weight learned by the model, b MMB is the bias term, and σ is the Sigmoid activation function, which is used to map the comprehensive features to the predicted probability. This multi-level attention structure improves the accuracy and adaptability of the model in dealing with feature importance and interaction, thereby improving the overall prediction performance.

[0064] In this embodiment, the evaluation indicators for training the initial recommendation model include (Area Under the Curve) and Log Loss (Logarithmic Loss);

[0065] AUC refers to the area under the receiver operating characteristic curve (ROC Curve). ROC curve is a

[0066] A tool for evaluating the performance of classification models. It represents the performance of the model through the True Positive Rate (TPR) and False Positive Rate (FPR) at different thresholds.

[0067] The true positive rate (TPR) refers to the proportion of positive classes correctly predicted by the model among all actual positive classes. The calculation formula is Among them, TP is the number of true positive cases and FN is the number of false negative cases.

[0068] The false positive rate (FPR) refers to the proportion of all actual negative classes that the model incorrectly predicts as positive classes. The calculation formula is Among them, FP is the number of false positives and TN is the number of true negatives.

[0069] The ROC curve is plotted by plotting TPR and FPR at different classification thresholds, and AUC is the area under this curve. The value of AUC ranges from 0 to 1, where a higher AUC value indicates better model performance because it indicates that the model achieves a higher true positive rate while maintaining a low false positive rate.

[0070] LogLoss, also known as logical loss or logarithmic loss, is a method for evaluating the performance of classification models, especially for binary classification. Log Loss measures the difference between the probability distribution predicted by the model and the actual label.

[0071] For binary classification, the calculation formula of Log Loss is shown in Formula 4:

[0072]

[0073] In the formula, N represents the total number of samples, y i represents the actual label of the i-th sample. If the actual label is positive, it is 1, otherwise it is 0. i is the probability that the model predicts that the i-th sample is a positive class. The smaller the value of Log Loss, the better the performance of the model. This is because Log Loss penalizes the wrong prediction of the actual category, especially when the model is very confident in predicting the wrong category (that is, the predicted probability is close to 0 or 1), the penalty will be more severe.

[0074] This embodiment also conducts a series of experiments on hybrid training models based on deep learning on Criteo and Avazu datasets, which covers models including LR (Logistic Regression), FM (Factorization Machine), AFM (Attentional Factorization Machine), IFM (Interactive Factorization Machine), DNN (Deep Neural Network), Wide & Deep, DeepFM, DCN (Deep & Cross Network) and DCP (Deep Click Predictor) constructed by the present invention. During the training process, the learning rate, dropout rate, embedding dimension, number of neural units, activation function and other hyperparameters were finely adjusted, and the Adam optimizer was used to perform first-order gradient optimization on the loss function. The Adam optimizer performs well in learning rate adjustment due to its adaptability, especially in more effective use of gradient information when processing sparse data.

[0075] The model evaluation uses two indicators, AUC value and Log Loss value, to evaluate the model's ability to distinguish between positive and negative samples and the accuracy of the predicted probability.

[0076] In order to verify the reliability of the model proposed in this embodiment, this embodiment compares the proposed model with the following models in terms of AUC and Log Loss.

[0077] (1) LR (Linear Regression): In the field of machine learning, especially in classification tasks, logistic regression is a special linear regression model used for binary classification problems.

[0078] (2) FM (Factorization Machines): is a general prediction model that is particularly effective on sparse datasets. FM can simulate the factorization of any real-valued matrix and is therefore well suited for processing data with a large number of sparse features.

[0079] (3) AFM (Attentional Factorization Machine): Combining the characteristics of the factorization machine (FM) and the attention mechanism, AFM enhances the ability of the factorization machine in feature interaction learning by introducing the attention mechanism, enabling it to distinguish the importance of different feature interactions.

[0080] (4) IFM (Interactive Factorization Machine): IFM is a sample-aware FM model that improves FM from a new perspective and optimizes the representation of features in different scenarios.

[0081] (5) DNN (Deep Neural Networks): Widely used in various machine learning tasks, DNN can learn and model complex data patterns by simulating the working mode of neurons in the human brain.

[0082] (6) Wide & Deep: Aims to simultaneously utilize the memory capacity of linear models and the generalization ability of deep neural networks, and is particularly suitable for recommendation systems.

[0083] (7) DeepFM: is a machine learning model that combines factorization machine (FM) and deep neural network (DNN). DeepFM is mainly used in recommendation systems and click-through rate prediction tasks, especially when processing large-scale sparse data with rich feature interactions.

[0084] (8) Deep & Cross Network: It aims to provide a model that is both efficient and highly interpretable by combining a deep neural network (for capturing high-order feature combinations) with a specific cross network structure (for explicitly modeling feature interactions).

[0085] The DCP (Deep Click Predictor) model proposed in the present invention has the following key parameter settings: the deep neural network part adopts four fully connected layers, and the network structure is configured as 200-200-200-200, which ensures that the model can effectively handle complex feature representations. The model uses the Adam optimizer. The batch size of the model is set to 256, allowing a large number of samples to be effectively processed in each iteration. The size of the embedding layer is set to 12 dimensions, aiming to efficiently encode and process a large number of discrete features. The learning rate is set to 0.001, which balances the convergence speed and learning efficiency of the model. The entire training process is set to span 50 Epochs to ensure that the model has sufficient time to learn and refine features.

[0086] Table 3 in this example compares the performance of the proposed DCP model with the other eight models on the Criteo and Avazu datasets. Through this comparison, the following conclusions can be drawn:

[0087] Table 3

[0088]

[0089] According to the data in Table 3, the DCP model achieved a significant improvement of about 6.59% in AUC value compared with the traditional LR model on the Criteo dataset, and also showed a performance increase of about 1.39% in comparison with the DeepFM model. At the same time, compared with the DCN model, the DCP model also achieved an improvement of about 1.45%. On the Avazu dataset, the CBN model further confirmed its superiority, achieving an improvement of about 2.2% compared with the FM model, and also an improvement of about 2.1% in comparison with the AFM model. This significant performance improvement demonstrates the strong adaptability and prediction accuracy of the CBN model when processing different types of datasets. In terms of the LogLoss evaluation index, the DCP model showed an improvement of 0.52% over the Wide$Deep model on the Criteo dataset, and a decrease of 0.3% relative to the IFM model on the Avazu dataset. Overall, the CBN model performed well on multiple evaluation indicators and improved prediction accuracy.

[0090] Table 4

[0091]

[0092] Table 4 shows a comparative analysis of the final converged Loss values ​​of different models on the Criteo and Avazu datasets. These Loss values ​​are the average values ​​taken from 50 convergence iterations. A careful study of the table shows that on the Criteo dataset, the DCP model shows the lowest convergence Loss, which indicates that DCP has better performance in minimizing the Loss function compared to other models in the evaluation. The performance ranks among the top 8 models. Taking the Criteo dataset as the research object, the changes in the loss function of each model during the training process are shown in the figure below. Figure 3 As shown in the figure, the linear regression (LR) model has limited ability in feature learning, so its loss function value is relatively high. The DCP model shows a faster convergence speed and maintains a relatively low loss value throughout the training process. The final convergence value is about 0.4556, which is lower than the other seven models.

[0093] In this embodiment, the specific impact of the deep neural network structure on the model performance is also studied in depth. The study focuses on four models based on deep neural networks (DNN): DeepFM, Wide & Deep, Deep & Cross Network (DCN) and DCP. By fixing the total number of neurons to 800 and keeping the architecture of the four hidden layers unchanged, a detailed comparative analysis of different network shapes was carried out. When evaluating the performance of the model, Log Loss and AUC indicators were selected as the measurement criteria, and the final convergence value of each model was used as the evaluation basis. Such a research design aims to accurately reveal how the number of single-layer neurons affects the performance of the deep learning model based on DNN, and provides a scientific basis for the structural design of deep neural networks. The compared deep network shapes are as follows: (1) Structure 1 (increasing structure): four-layer structure, the number of neurons increases layer by layer, respectively 100, 150, 200, 350. (2) Structure 2 (uniform structure): four-layer homogeneous structure, the number of neurons in each layer is 200. (3) Structure 3 (decreasing structure): four-layer structure, the number of neurons decreases layer by layer, namely 350, 200, 150, and 100. (4) Structure 4 (increasing first and then decreasing structure): four-layer structure, the number of neurons increases first and then decreases, namely 150, 300, 250, and 100.

[0094] from Figure 4It can be seen from the AUC performance data that different configurations of deep network structures have a significant impact on model performance. In the DCP model, structure one (increasing structure) achieved an AUC value of 0.7920, while structure three (decreasing structure) slightly decreased to 0.7910. This shows that the two extreme configurations of the model, which concentrate more neurons in the shallowest layer and more neurons in the deepest layer, may limit its performance in deep and shallow feature learning. On the other hand, structure two (uniform structure) provides a more balanced distribution of neurons for the DCP model, resulting in the highest AUC value of 0.7971, which reveals the importance of achieving balance between the layers of the model for optimizing performance. For structure four (increasing first and then decreasing structure), the AUC value of the DCP model dropped slightly to 0.7890, which may reflect certain limitations of the model in feature learning at the shallowest and deepest levels.

[0095] exist Figure 5 In the displayed Log Loss indicator, we can observe how the difference in the distribution of neuron structures affects the performance of the model. The Log Loss of the DCP model under structure one (increasing structure) is 0.4693, while it is slightly improved to 0.4686 in structure three (decreasing structure). Although the differences are not large, they indicate that extreme neuron distribution may lead to some compromises in performance. At the same time, structure two (uniform structure) provides the best performance of 0.4635 for the DCP model in terms of Log Loss, confirming the effectiveness of the uniformly distributed network structure in improving performance. The Log Loss of structure four (increasing and then decreasing structure) is 0.4701, showing that concentrating more neurons in the middle layer may have a certain impact on the overall learning efficiency of the model. When the two performance indicators of AUC and Log Loss are combined for comprehensive consideration, the uniformly distributed neuron structure structure two (uniform structure) provides the most balanced and efficient performance for the DCP model. This network configuration not only supports effective feature learning and information transfer between different layers of the model, but also enhances the model's ability to interpret complex relationships in the data, which is crucial to the performance of deep learning models. It also ensures that the model not only achieves strong performance on existing datasets, but also has the ability to generalize to new and unknown data challenges.

[0096] In summary, the DCP model significantly enhances the performance of feature extraction by introducing the maximum average pooling and bitwise attention mechanism. This mechanism enables the model to focus more on the most critical elements in the data, thereby significantly improving the sensitivity to important information and prediction accuracy. In addition, this method also optimizes the generalization ability of the model and ensures effective processing of novel data. At the same time, this combination improves the flexibility and adaptability of the model, enabling it to better cope with various data changes.

[0097] In this embodiment, in order to accurately evaluate the contribution of each component in the proposed DCP model, a series of ablation experiments are also conducted. The purpose of the ablation experiment is to observe the changes in model performance by removing key components in the model one by one, so as to reveal the impact of each component on the overall performance of the model. The experimental settings are as follows: First, the complete DCP model is used as the baseline. Then, several variant models are constructed respectively, each of which removes a key component of DCP, such as the Max–Mean Pooling Attention Mechanism or the Bit-wise Attention Mechanism. By comparing the performance of these variant models with that of the complete model, the role of each component can be evaluated. In the ablation experiment, the following model variants are constructed and evaluated: (1) DCP without Max-Mean Attention: This version omits the Max–Mean Pooling Attention Mechanism, but keeps all other components unchanged. (2) DCP without Bit-wise Attention: This version removes the Bit-wise Attention Mechanism and retains all other components. (3) DCP without Attention Mechanism: This version removes both the Max–Mean Pooling and Bit-wise Attention Mechanisms, with the purpose of evaluating the impact of the attention mechanism on the overall performance of the model. (4) DCP complete model: The complete DCP model including all designed components and mechanisms.

[0098] Each variant model was evaluated on the same training and test datasets, using the same evaluation metrics AUC and Log Loss as the full model. The experimental results are shown in Table 5 below:

[0099] Table 5

[0100]

[0101] As can be seen from Table 5, when the DCP model removes the maximum average pooling attention mechanism, its AUC value drops by about 0.5%, and Log Loss also rises to about 0.4% on the Criteo dataset. This shows that the maximum average pooling attention mechanism plays an important role in improving model performance. Similarly, when the bitwise attention mechanism is removed, the model's AUC on the Criteo dataset drops by about 0.46%, while Log Loss rises slightly by about 0.3%. Although this change is not as significant as the maximum average pooling mechanism, it also proves the contribution of the bitwise attention mechanism to performance. When both attention mechanisms are removed, the model's AUC on the Criteo dataset drops more significantly, by about 1.5%, and Log Loss increases by about 1.78%, showing the largest drop in model performance. This further emphasizes the key role of the attention mechanism in improving performance in the DCP model. The same trend also appears on the Avazu dataset. After removing the maximum average pooling attention mechanism, the AUC drops by about 0.51% and the Log Loss rises by about 0.61%. When the bitwise attention mechanism is removed, the AUC slightly decreases by about 0.48% and the Log Loss slightly increases by about 0.33%. When both attention mechanisms are removed, the performance decreases most significantly, with the AUC decreasing by about 2.13% and the Log Loss increasing by about 0.24%. These data changes clearly show that the attention mechanism is crucial for the DCP model to improve its ability to parse data and its performance.

[0102] In general, the ablation experiment results clearly demonstrate the effectiveness of the maximum average pooling attention mechanism and the bitwise attention mechanism in the DCP model. The introduction of these mechanisms significantly enhances the model's extraction and learning of key features, thereby improving the model's performance in complex click-through rate prediction tasks. Through these experiments, the role of each component in the model is effectively verified.

[0103] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A recommendation method based on deep learning DCP, characterized in that: The following steps are involved: Acquire an online platform user interaction behavior dataset, and preprocess the online platform user interaction behavior dataset to obtain a preprocessed dataset; Constructing an initial recommendation model, and training the initial recommendation model based on the preprocessed data set to obtain a recommendation model; The initial recommendation model includes an embedding and feature stacking layer, a maximum-mean attention mechanism layer, a bitwise attention mechanism layer, a deep feedforward network layer, and a combined output layer; The embedding and feature stacking layer is used to map the preprocessed data set to an embedding space to obtain an embedding vector; The maximum-mean attention mechanism layer is used to purify and refine the information of the embedded vector by calculating the weight distribution to obtain the feature vector; The bit-by-bit attention mechanism layer is used to perform a refined weight assignment on each feature vector to obtain the contribution of each feature vector to the final prediction in each dimension; The deep feedforward network layer includes multiple fully connected layers and introduces nonlinearity through activation functions to capture high-order interactions in the embedding vector; The combined output layer is used to synthesize the output results of the maximum-mean attention mechanism layer, the bit-by-bit attention mechanism layer and the deep feedforward network layer to obtain a predicted output; The expression of the comprehensive features obtained by combining the maximum-mean attention mechanism layer and the bit-by-bit attention mechanism layer is: F MMB =σ(W MMB ·F concat +b MMB ) In the formula, σ represents the Sigmoid activation function, F concat represents the combination of the embedding vectors processed by the max-mean attention layer and the bitwise attention layer, W MMB represents the weights learned by the model, b MMB represents the bias term; The expression for obtaining the characteristic vector is: F MM =α·max(E)+(1-α)·mean(E) In the formula, α represents a hyperparameter, E represents an embedding vector set, max(E) represents the maximum value operation on the embedding vector set, and mean(E) represents the mean operation on the embedding vector set; The expression for the refined weight allocation is: F B =W B ⊙E Where E represents the embedding vector set, ⊙ represents element-wise multiplication, and W B represents the learned dimension-specific attention weights; The user data is input into the recommendation model to obtain recommended product information, and the recommended product information is pushed to the user.

2. The recommendation method based on deep learning DCP according to claim 1, characterized in that: The method for pre-processing comprises: LabelEncoder is used to assign a unique integer value to different category labels of behavior type data in the online platform user interaction behavior dataset; The embedding vectorization method is used to map the high-dimensional category features in the online platform user interaction behavior dataset into a dense vector of fixed length.

3. The recommendation method based on deep learning DCP according to claim 1, characterized in that: The evaluation indicators for training the initial recommendation model include the area under the receiver operating characteristic curve and the logistic loss. The expression of the logistic loss is: In the formula, N represents the total number of samples, y i represents the actual label of the i-th sample. If the actual label is positive, it is 1, otherwise it is 0. i is the probability that the model predicts that the i-th sample is a positive class.

Images