Recommendation methods, training methods, devices, media, and products based on Wide&Deep models

Abstract

This application relates to a recommendation method, training method, device, medium, and product based on the Wide&Deep model. The method includes: integrating product features of the target recommended product, profile features of the target user, skill and expertise features of customer service agents, and contextual features of the recommendation scenario to construct a multi-dimensional feature vector. This vector is then input into the Wide&Deep neural network for click-through rate (CTR) prediction. The Wide side of the model performs linear modeling of sparse features, while the Deep side performs non-linear fusion of category and continuous features. A weighted CTR score is output for the target recommended product. Finally, personalized recommendations are achieved based on the CTR scores of multiple target recommended products. By fusing multi-dimensional features, the accuracy of CTR prediction is significantly improved. Combined with the memory and generalization capabilities of the Wide&Deep model, the adaptability of the recommendation system to complex scenarios is enhanced, thereby achieving dynamic and accurate personalized recommendations and improving conversion rates and user experience.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a recommendation method, training method, device, medium and product based on the Wide&Deep model. Background Technology

[0002] Existing product recommendation methods are primarily based on user profiles and rely on dynamic modeling using historical user behavior data. However, this method has the following problems: 1. The feature dimension is too limited, relying solely on user profiles, which cannot fully reflect the user's real-time needs; 2. The lack of fusion between agent features and contextual features results in recommendation results that lack context adaptability; 3. The model is outdated and cannot respond promptly to changes in user behavior; 4. Limited accuracy of recommendations impacts customer satisfaction and business conversion rates.

[0003] Therefore, there is an urgent need for a product recommendation method that can integrate multi-dimensional features, update in real time, and accurately predict user intent. Summary of the Invention

[0004] To address the aforementioned problems, embodiments of this application provide a recommendation method based on the Wide&Deep model, the recommendation method comprising: A multi-dimensional feature vector of the target recommended product is obtained. The multi-dimensional feature vector includes the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product. The multidimensional feature vector is input into a target Wide&Deep neural network model to obtain the click-through rate prediction score of the target user for the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform non-linear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score. The multidimensional feature vectors corresponding to different target recommended products are input into the target Wide&Deep neural network model to obtain the click pass rate prediction scores of multiple target recommended products; Based on the click-through rate prediction scores of the multiple target recommended products, products are recommended to the target user.

[0005] In one embodiment, the user profile feature vector includes at least one of the following: user profile data features, communication behavior data features, package usage features, system service data features, user interaction features, and contract status features; The agent feature vector includes at least one of the following: agent skill features, historical conversion rate features, average call duration features, and customer satisfaction rating features. The context feature vector includes at least one of time features and session context features; The product feature vector includes at least one of the following: product tag, product discount, product category, and product lifecycle stage.

[0006] In one implementation, inputting the multidimensional feature vector into a target Wide&Deep neural network model to obtain a predicted click-through rate score for the target user on the target recommended product includes: Configure a first weight vector for the sparse binary feature vector corresponding to the sparse binary feature, and perform a weighted summation on the sparse binary feature vector based on the first weight vector to obtain the first score; Configure a second weight vector for the combined feature vector corresponding to the category feature and the continuous feature, and perform a weighted summation on the combined feature vector based on the second weight vector to obtain the second score; The sum of the first score and the second score is input into an activation function to obtain the click-through rate prediction score of the target recommended product.

[0007] In one implementation, recommending products to the target user based on the click-through rate prediction scores of the plurality of target recommended products includes: The click pass rate prediction scores of the multiple target recommended products are sorted to obtain the recommended product ranking results; Based on the ranking results of the recommended products, a list of recommended products and corresponding recommendation messages for each product are generated.

[0008] In one implementation, recommending products to the target user based on the click-through rate prediction scores of the plurality of target recommended products further includes: The recommended sales scripts for each product are sent to the customer service agent so that the agent can recommend the products to the target user.

[0009] According to a second aspect of the embodiments of this application, a training method for a Wide&Deep model is provided, the training method comprising: Multiple multidimensional feature vectors of the target recommended product are obtained. The multidimensional feature vectors include the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product. Each multidimensional feature vector is input into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform non-linear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score. The target Wide&Deep neural network model is trained using multiple multidimensional feature vectors as samples and the click pass rate prediction score as labels.

[0010] In one implementation, training the target Wide&Deep neural network model using multiple multidimensional feature vectors as samples and the click pass rate prediction score as a label includes: The multidimensional feature vector is input into the target Wide&Deep neural network model, and forward propagation calculation is performed to obtain the click pass rate prediction score; Based on the predicted score of the click pass rate and the actual score of the sample, the loss value of the target Wide&Deep neural network model is calculated using the cross-entropy loss function; Based on the loss value, the gradients of each parameter of the target Wide&Deep neural network model are calculated using the backpropagation algorithm; An adaptive optimization algorithm is used to update the weights and bias parameters of the target Wide&Deep neural network model according to the gradient until the target Wide&Deep neural network model converges, thus obtaining the trained target Wide&Deep neural network model.

[0011] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method of any one of the first aspects or the steps of the method of any one of the second aspects.

[0012] According to a fourth aspect of the embodiments of this application, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in any one of the first aspects or the steps of the method described in any one of the second aspects.

[0013] According to a fifth aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of any of the methods in the first aspect or the steps of any of the methods in the second aspect.

[0014] The embodiments of this application achieve the following beneficial effects: These embodiments integrate the product features of the target recommended product, the profile features of the target user, the skill and expertise features of the customer service agent, and the contextual features of the recommendation scenario to construct a multi-dimensional feature vector. This vector is input into the Wide&Deep neural network for click-through rate (CTR) prediction. The Wide side of the model performs linear modeling of sparse features, while the Deep side performs non-linear fusion of category and continuous features. The weighted output is the CTR score of the target recommended product. Finally, personalized product recommendations are achieved based on the ranking of the CTR scores of multiple target recommended products. By fusing multi-dimensional features, the accuracy of CTR prediction is significantly improved. Combined with the memory and generalization capabilities of the Wide&Deep model, the adaptability of the recommendation system to complex scenarios is enhanced, thereby achieving dynamic and accurate personalized recommendations and improving conversion rates and user experience.

[0015] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a recommendation method based on the WIDE&DEEP model provided in an embodiment of this application.

[0017] Figure 2 This is a flowchart of a recommendation method based on the WIDE&DEEP model provided in an embodiment of this application.

[0018] Figure 3 This is a flowchart of a recommendation method based on the WIDE&DEEP model provided in an embodiment of this application.

[0019] Figure 4 This is a flowchart of a recommendation method based on the WIDE&DEEP model provided in an embodiment of this application.

[0020] Figure 5 This is a flowchart of a training method for a WIDE&DEEP model provided in one embodiment of this application.

[0021] Figure 6 This is a flowchart of a training method for a WIDE&DEEP model provided in one embodiment of this application.

[0022] Figure 7 This is a block diagram of an electronic device provided in one embodiment of this application. Detailed Implementation

[0023] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this application.

[0024] It should be understood that the term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description.

[0025] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependencies. The modifications of "one" and "multiple" mentioned in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless explicitly stated in the context, they should be understood as "one or more". In the description of this application, unless otherwise stated, "multiple" means two or more, and other quantifiers are similar; "at least one item", "one item or multiple items", or similar expressions refer to any combination of these items, including any combination of single or multiple items.

[0026] Although operations or steps are described in a specific order in the accompanying drawings in the embodiments of this application, this should not be construed as requiring these operations or steps to be performed in the specific order or serial order shown, or requiring all of the shown operations or steps to be performed to obtain the desired result. In the embodiments of this application, these operations or steps may be performed serially; they may be performed in parallel; or a portion of these operations or steps may be performed.

[0027] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information. It is understood that before using the technical solutions disclosed in the embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0028] First, the application scenario of this application is explained. With the continuous opening of the communications market and increasingly fierce competition, dynamic click-through rate (CTR) modeling recommendation systems based on user profiles are widely adopted to improve user experience and customer retention. However, this recommendation method, which relies solely on user profiles, has several limitations. User profiles are typically built based on users' historical behavior and preferences, but this data may be incomplete and unable to cover all user needs and potential interests. For example, users may exhibit different needs in certain specific situations than their historical behavior, and user profiles may fail to capture these changes. Updates to user profiles may lag behind changes in user behavior, resulting in recommendation results that cannot reflect users' latest preferences in a timely manner. CTR modeling based on user profiles mainly focuses on users' click behavior, but click-through rate may not accurately reflect users' actual needs and satisfaction. Although dynamic CTR modeling can adjust recommendation results based on real-time data, its adjustment mechanism may rely too heavily on short-term data, neglecting users' long-term preferences and potential needs. CTR modeling typically only focuses on users' historical behavior and preferences, ignoring the impact of contextual factors (such as time, location, device, etc.) on user decisions. To overcome these limitations, it is necessary to comprehensively consider multi-dimensional data, introduce a context-aware mechanism, and dynamically adjust weights based on agent and user characteristics to improve the real-time performance and user engagement of the recommendation system, thereby providing more accurate and personalized recommendation services.

[0029] Traditional product recommendation methods rely on a limited number of feature dimensions in their model processing. Recommendation systems that depend solely on user features may fail to accurately recommend suitable products. As market environments and user needs constantly evolve, recommendation methods based solely on user features may struggle to adapt to these changes in a timely manner, failing to effectively improve user loyalty. Users may become dissatisfied due to inaccurate or untimely recommendations, thereby reducing loyalty and retention rates. To improve recommendation accuracy, this solution expands product recommendation from single-dimensional features to multi-dimensional features.

[0030] This application's solution predicts and analyzes users' service intentions in advance and provides dynamic intervention strategies. Its core is to collect multi-dimensional customer data, integrating basic customer information (such as city of origin, service plan, average monthly spending), consumption behavior (such as plan usage, data usage, call duration), and service records (such as historical inquiries and service processing records) to form a comprehensive customer profile. This profile is dynamically updated based on real-time customer behavior and feedback to ensure accuracy and timeliness. Dynamic adjustments are made by analyzing data such as agent click-through rate, agent characteristics, historical conversion rate, problem resolution rate, and customer satisfaction. Furthermore, contextual feature data (billing date, beginning / end of month / holidays, best-selling products, product information) is used to analyze the products needed by the customer. Through these strategies, precise matching of customer needs can be achieved, improving agent conversion rates, customer satisfaction, and overall service efficiency. The following describes this application with specific embodiments.

[0031] Figure 1 This is a flowchart illustrating a recommendation method based on the WIDE&DEEP model provided in one embodiment of this application. Figure 1 As shown in the embodiments of this application, a recommendation method based on the Wide&Deep model is provided, which may include the following steps: In step S10, a multi-dimensional feature vector of the target recommended product is obtained. The multi-dimensional feature vector includes the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product.

[0032] For example, user profile feature data of target users for the recommended products can be collected first. This includes personal information (city of origin, package, average monthly consumption), behavioral data (such as purchase history, browsing history), preferences (such as product preferences, service preferences), and feedback information (such as customer feedback, reviews, complaints), generating a user profile feature vector. Different weights can be assigned to different types of behavior (such as purchase, browsing), generally with higher weights for behaviors that require greater user investment. Recent user behavior is also more important than older behavior. Next, customer service agent feature data for recommending the target products to target users can be collected. This includes skills and expertise, historical performance, etc., generating an agent feature vector. For example, agent features can include categorical and numerical features. Categorical features (skill tags, team affiliation, etc.) are processed using feature hashing. The feature name and feature value are combined into a string, and then the hash value h is calculated using the MurmurHash function. The hash value h is modulo the fixed length d of the vector to obtain an integer index between 0 and d-1. This index indicates the position of the feature in the final vector v. Index = h % d. If a feature may appear multiple times in a dataset, the value at that index is incremented by 1 each time the same index is calculated. Numerical features (historical conversion rate, average call duration, customer satisfaction rating, etc.) are standardized using Z-score. Combining all processed categorical and numerical features generates a fixed-length, low-dimensional, dense feature vector (agent feature vector) representing the agent. Then, contextual feature data is collected when the agent recommends a product to the target user: billing date, beginning / end of month / holiday, etc., to generate a contextual feature vector. These features are often converted to Boolean features or sine / cosine encoded for periodic times (such as weekends). Additionally, categorical features such as conversation context (such as customer call channels) are encoded and incorporated into the vector. Next, product feature data for the target product is collected: product tags, product discounts, product category (data traffic, voice, broadband, etc.), product lifecycle stage, etc., to generate a product feature vector. These features are often converted to Boolean features. Finally, the above four vectors are integrated into a multi-dimensional feature vector for the target product.

[0033] In step S20, the multidimensional feature vector is input into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target user for the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform nonlinear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score.

[0034] For example, a first weight vector can be configured for the sparse binary feature vector corresponding to the sparse binary feature vector of the multidimensional feature vector, and the sparse binary feature vector can be weighted and summed based on the first weight vector to obtain the first score. Then, a second weight vector can be configured for the combined feature vector corresponding to the categorical feature and the continuous feature, and the combined feature vector can be weighted and summed based on the second weight vector to obtain the second score. Finally, the sum of the first score and the second score is input into the activation function to obtain the click pass rate prediction score of the target recommended product.

[0035] In step S30, the multi-dimensional feature vectors corresponding to different target recommended products are input into the target Wide&Deep neural network model to obtain the click pass rate prediction scores of multiple target recommended products.

[0036] In step S40, products are recommended to the target user based on the click-through rate prediction scores of the multiple target recommended products.

[0037] For example, the click-through rate prediction scores of multiple target recommended products can be sorted first to obtain the recommended product ranking results, and then a list of recommended products and the corresponding recommendation text for each product can be generated based on the recommended product ranking results.

[0038] This application integrates product features of the target recommended product, profile features of the target user, skill and expertise features of customer service agents, and contextual features of the recommendation scenario to construct a multi-dimensional feature vector. This vector is input into a Wide&Deep neural network for click-through rate (CTR) prediction. The Wide side of the model performs linear modeling of sparse features, while the Deep side performs non-linear fusion of category and continuous features. The weighted output is the CTR score of the target recommended product. Finally, personalized recommendations are achieved based on the CTR scores of multiple target recommended products. By fusing multi-dimensional features, the accuracy of CTR prediction is significantly improved. Combined with the memory and generalization capabilities of the Wide&Deep model, the adaptability of the recommendation system to complex scenarios is enhanced, thereby achieving dynamic and accurate personalized recommendations and improving conversion rates and user experience.

[0039] In one possible implementation, the user profile feature vector may include at least one of the following: user profile data features, communication behavior data features, package usage features, system service data features, user interaction features, and contract status features. The agent feature vector may include at least one of the following: agent skill features, historical conversion rate features, average call duration features, and customer satisfaction rating features. The context feature vector may include at least one of the following: time features and session context features. The product feature vector utilizes at least one of the following: product tags, product discounts, product categories, and product lifecycle stages.

[0040] This application embodiment, through refined feature acquisition and processing methods, enables multi-dimensional feature vectors to more comprehensively and accurately characterize user, agent, and scenario attributes, providing high-quality data support for subsequent model calculations and further improving the adaptability of recommendation results.

[0041] Figure 2 This is a flowchart illustrating a recommendation method based on the WIDE&DEEP model provided in one embodiment of this application. Figure 2 As shown, inputting the multidimensional feature vector into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target user for the target recommended product may include the following steps: In step S201, a first weight vector is configured for the sparse binary feature vector corresponding to the sparse binary feature, and the sparse binary feature vector is weighted and summed based on the first weight vector to obtain the first score.

[0042] For example, the input to the Wide sub-network layer mainly consists of the cross product of sparse binary features. For instance, features such as user package type, contract status, agent skill tags, and call channels are combined in pairs or groups using the FTRL optimizer. In this step, a first weight vector W_wide is configured for the sparse binary feature vector X_wide corresponding to the sparse binary features, and a weighted sum is performed on the sparse binary feature vector X_wide based on the first weight vector W_wide to obtain the first score y_wide. For example, the first score y_wide can be obtained by the following formula: y_wide = W_wide X_wide + b1 Formula 1 Where b1 is the bias term of the Wide sub-network layer.

[0043] In step S202, a second weight vector is configured for the combined feature vector corresponding to the category feature and the continuous feature, and the combined feature vector is weighted and summed based on the second weight vector to obtain the second score.

[0044] For example, the Deep sub-network layer directly inputs numerical feature vectors, while categorical feature vectors are transformed into low-dimensional dense vectors through their respective embedding layers. All low-dimensional dense vectors and normalized continuous features are concatenated into a long feature vector, which is then input into the Deep sub-network layer. In this step, a second weight vector W_deep is configured for the combined feature vector X_deep corresponding to the categorical and continuous features, and the combined feature vector X_deep is weighted and summed based on the second weight vector W_deep to obtain the second score y_deep. For example, the second score y_deep can be obtained by the following formula: y_deep = (W_ deep Formula 2 (X_deep + b2) Where b2 is the bias term of the Deep sub-network layer. This is the activation function.

[0045] In step S203, the sum of the first score and the second score is input into an activation function to obtain the click-through rate prediction score of the target recommended product.

[0046] In this step, the sum of the first score y_wide and the second score y_deep is input into the Sigmoid activation function to obtain the click-through rate prediction score p_i for the target recommended product i. For example, the click-through rate prediction score p_i for the target recommended product i can be obtained by the following formula: p_i=Sigmoid(y_wide+y_deep) formula 3 The mathematical expression for the Sigmoid activation function can be represented by the following formula: Formula 4 in, Let be the sigmoid activation function, and let e be an exponential function with the natural number e as the base.

[0047] The calculation process in this application takes into account both linear and nonlinear features, making the score prediction more consistent with the actual situation, providing a reliable quantitative basis for subsequent product recommendations, and ensuring the scientific nature of the recommendation decision.

[0048] Figure 3 This is a flowchart illustrating a recommendation method based on the WIDE&DEEP model provided in one embodiment of this application. Figure 3 As shown, recommending products to the target user based on the click-through rate prediction scores of the multiple target recommended products may include the following steps: In step S401, the click pass rate prediction scores of the multiple target recommended products are sorted to obtain the recommended product ranking results.

[0049] For example, the click pass rate prediction scores p_i of multiple target recommended products can be sorted in descending order from high to low to obtain the recommended product ranking results.

[0050] In step S402, a list of recommended products and corresponding recommendation scripts for each product are generated based on the ranking results of the recommended products.

[0051] This application collects the click-through rate prediction scores of all target recommended products, sorts them according to their scores, and forms a clear ranking of recommended products. Based on the ranking results, the top-ranked high-quality products are selected to generate a structured list of recommended products, and a recommendation script that matches the characteristics and user needs of each product is matched.

[0052] This score-based recommendation method ensures that products most likely to interest users are prioritized, while accompanying wording makes the recommendations more targeted, increases user acceptance of the recommended products, and improves recommendation efficiency and effectiveness.

[0053] Figure 4 This is a flowchart illustrating a recommendation method based on the WIDE&DEEP model provided in one embodiment of this application. Figure 4 As shown, recommending products to the target user based on the click-through rate prediction scores of the multiple target recommended products may further include the following steps: In step S403, the recommended scripts corresponding to each product are sent to the customer service agent so that the customer service agent can recommend products to the target user.

[0054] For example, after generating a list of recommended products and corresponding recommendation messages, the system automatically pushes the recommendation messages for each product to the customer service agents responsible for recommending products to target users. After receiving the messages, the agents can combine the key information contained in the messages, such as product advantages and suitable scenarios, to conduct precise and efficient recommendation communication with users. The messages provide agents with professional communication guidance, reduce communication costs, make recommendations more persuasive, help users quickly understand product value, and thus improve the success rate of recommendation conversion.

[0055] Figure 5 This is a flowchart illustrating a training method for a WIDE&DEEP model provided in one embodiment of this application. Figure 5 As shown in the embodiments of this application, a training method for a Wide&Deep model is provided, which may include the following steps: In step S500, multiple multidimensional feature vectors of the target recommended product are obtained. The multidimensional feature vectors include the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product.

[0056] In step S510, each multidimensional feature vector is input into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform nonlinear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score.

[0057] In step S520, the target Wide&Deep neural network model is trained using multiple multidimensional feature vectors as samples and the click pass rate prediction score as labels.

[0058] For example, the multidimensional feature vector can be input into the target Wide&Deep neural network model first, and forward propagation can be performed to obtain the click pass rate prediction score. Then, based on the click pass rate prediction score and the sample's true score, the cross-entropy loss function can be used to calculate the loss value of the target Wide&Deep neural network model. Based on the loss value, the gradient of each parameter of the target Wide&Deep neural network model can be calculated using the backpropagation algorithm. Finally, an adaptive optimization algorithm can be used to update the weights and bias parameters of the target Wide&Deep neural network model according to the gradient until the target Wide&Deep neural network model converges, thus obtaining the trained target Wide&Deep neural network model.

[0059] This application collects multi-dimensional feature vectors from multiple target users as training samples. These samples comprehensively cover feature combinations from different users, agents, and scenarios. Each sample is input into a target Wide&Deep neural network model. The features are processed through Wide and Deep sub-network layers respectively, and the output layer is fused to obtain the click-through rate prediction score for each sample. Using these multi-dimensional feature vectors as training samples and the corresponding click-through rate prediction scores as labels, the model is continuously trained iteratively. The multi-sample training process allows the model to fully learn the correlation patterns between different features, continuously improving the model's ability to predict click-through rates and providing solid model support for subsequent accurate recommendations.

[0060] Figure 6 This is a flowchart illustrating a training method for a WIDE&DEEP model provided in one embodiment of this application. Figure 6 As shown, training the target Wide&Deep neural network model using multiple multidimensional feature vectors as samples and the click pass rate prediction score as a label may include the following steps: In step S5201, the multidimensional feature vector is input into the target Wide&Deep neural network model, and forward propagation calculation is performed to obtain the click pass rate prediction score.

[0061] For example, the input multidimensional feature vector can be processed through the connection weights between layers in the network ( ) and bias ( The process involves linear combination layer by layer, followed by an activation function. A nonlinear transformation is performed to obtain a predicted value. The formula is as follows: Formula 5 in, It can be any layer in the network, including hidden layers and output layers. It is the first The weighted input of the k-th neuron in the layer; It is the connection of the first The j-th neuron in layer 1 and the j-th neuron The weight of the k-th neuron in the layer; It is the first The activation output of the k-th neuron in layer 1; It is the first Layer bias terms; It is the front Layer-wise linear weighted summation; yes The number of neurons in the layer, It is the first The activation function of the layer.

[0062] In step S5202, the loss value of the target Wide&Deep neural network model is calculated using the cross-entropy loss function based on the predicted score of the click pass rate and the actual score of the sample.

[0063] For example, loss calculation quantifies the difference between the model's predicted values ​​and the true labels. It calculates a loss value (L), which is a scalar used to adjust the parameters. The formula is as follows: Formula 6 Where L is the total loss across the entire dataset; N is the number of samples; This is the true label; in binary classification, it typically takes the value 0 or 1. It is the prediction score, and its value is between 0 and 1.

[0064] In step S5203, based on the loss value, the gradients of each parameter of the target Wide&Deep neural network model are calculated using the backpropagation algorithm.

[0065] For example, backpropagation uses the chain rule to decompose the output layer (L) of the neural network layer by layer in reverse, calculating the gradient (denoted as δ) of the loss function with respect to each parameter in the network (the weights w and biases b of each layer). The formula is as follows: Formula 7 in, It is the output layer of the neural network. It is the error of the output layer; It is the activation value of the j-th neuron in the L-th layer; It is the real label corresponding to the j-th output neuron.

[0066] Formula 8 in, It is a layer of a neural network. It is the first The error term of the i-th neuron in the layer; It refers to the connection weight, indicating the first... The i-th neuron in layer i and the th +1 The connection weights between the j-th neurons in layer j, where the index order ji indicates "from the j-th neuron". layer i to the first +1 layer j”; It is the first The error term of the j-th neuron in layer +1 can be regarded as the error signal “feedback” from the next layer; It is the first +1 floor all The error contributions of each neuron (after weighting) are summed up; It is the first The weighted input of the i-th neuron in the layer; It is the derivative of the activation function.

[0067] In step S5204, an adaptive optimization algorithm is used to update the weights and bias parameters of the target Wide&Deep neural network model according to the gradient until the target Wide&Deep neural network model converges, thus obtaining the trained target Wide&Deep neural network model.

[0068] For example, the AdaGrad optimization algorithm can be used to update parameters, adjusting all weights and biases in the model based on the gradient direction and the learning rate (η).

[0069] Formula 9 in, It is a layer of a neural network. It is the first The connection weights from the i-th neuron to the j-th neuron in the layer; It is the learning rate; It is the first Error term of the j-th neuron in layer +1; It is the first The activation value of the i-th neuron in the layer.

[0070] Formula 10 in, It is any layer in the network. It is the first The bias term of the i-th neuron in the layer is a constant that is independent of the input and is used to adjust the activation threshold of the neuron. It's the learning rate. It is the first The error term of the i-th neuron in the layer.

[0071] In this embodiment, the multi-dimensional feature vectors of the training samples are input into the model, and forward propagation is performed to obtain the click-through rate prediction score for each sample. The predicted score is compared with the actual score of the sample, and the cross-entropy loss function is used to accurately calculate the model's loss value, quantifying the model's prediction bias. Based on the loss value, the gradient of each parameter of the model is derived layer by layer through the backpropagation algorithm to clarify the direction of parameter adjustment. The AdaGrad adaptive optimization algorithm is used to dynamically update the model's weights and bias parameters based on the gradient information, iterating repeatedly until the model converges, resulting in a trained model. This rigorous training process continuously optimizes the model parameters, ensuring stable and accurate prediction performance after convergence, providing a core guarantee for the high-quality development of recommendation services.

[0072] Figure 7 This is a block diagram of an electronic device provided in one embodiment of this application. For example... Figure 7 As shown, the electronic device 700 may include a processor 701 and a memory 702. The electronic device 700 may also include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.

[0073] The processor 701 controls the overall operation of the electronic device 700 to complete all or part of the steps in the recommended method based on the Wide&Deep model or to complete all or part of the steps in the training method of the Wide&Deep model. The memory 702 stores various types of data to support the operation of the electronic device 700. This data may include, for example, instructions for any application or method operating on the electronic device 700, and application-related data. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 703 may include a screen and audio components. I / O interface 704 provides an interface between processor 701 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical. Communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, and is not limited herein. Therefore, the corresponding communication component 705 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0074] In an exemplary embodiment, the electronic device 700 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the recommended method based on the Wide&Deep model described above or the training method for the Wide&Deep model described above.

[0075] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the recommendation method based on the Wide&Deep model described above. For example, the computer-readable storage medium may be the memory 702 including the program instructions described above, which may be executed by the processor 701 of the electronic device 700 to complete the recommendation method based on the Wide&Deep model described above or to complete the training method of the Wide&Deep model described above.

[0076] In another exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program executable by a programmable device, the computer program having a code portion for executing the above-described recommended method based on the Wide&Deep model or a code portion for executing the above-described training method of the Wide&Deep model when executed by the programmable device.

[0077] The preferred embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this application, various simple modifications can be made to the technical solution of this application, and these simple modifications all fall within the protection scope of this application.

[0078] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this application will not describe the various possible combinations separately.

[0079] Furthermore, various different implementations of this application can be combined in any way, as long as they do not violate the spirit of this application, they should also be regarded as the content disclosed in this application.

Claims

1. A recommendation method based on the Wide & Deep model, characterized in that, The recommendation method includes: A multi-dimensional feature vector of the target recommended product is obtained. The multi-dimensional feature vector includes the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product. The multidimensional feature vector is input into a target Wide&Deep neural network model to obtain the click-through rate prediction score of the target user for the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform non-linear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score. The multidimensional feature vectors corresponding to different target recommended products are input into the target Wide&Deep neural network model to obtain the click pass rate prediction scores of multiple target recommended products; Based on the click-through rate prediction scores of the multiple target recommended products, products are recommended to the target user.

2. The recommended method according to claim 1, characterized in that, The user profile feature vector includes at least one of the following: user profile data features, communication behavior data features, package usage features, system service data features, user interaction features, and contract status features; The agent feature vector includes at least one of the following: agent skill features, historical conversion rate features, average call duration features, and customer satisfaction rating features. The context feature vector includes at least one of time features and session context features; The product feature vector includes at least one of the following: product tag, product discount, product category, and product lifecycle stage.

3. The recommended method according to claim 1, characterized in that, The step of inputting the multidimensional feature vector into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target user for the target recommended product includes: Configure a first weight vector for the sparse binary feature vector corresponding to the sparse binary feature, and perform a weighted summation on the sparse binary feature vector based on the first weight vector to obtain the first score; Configure a second weight vector for the combined feature vector corresponding to the category feature and the continuous feature, and perform a weighted summation on the combined feature vector based on the second weight vector to obtain the second score; The sum of the first score and the second score is input into an activation function to obtain the click-through rate prediction score of the target recommended product.

4. The recommended method according to claim 1, characterized in that, The step of recommending products to the target user based on the click-through rate prediction scores of the multiple target recommended products includes: The click pass rate prediction scores of the multiple target recommended products are sorted to obtain the recommended product ranking results; Based on the ranking results of the recommended products, a list of recommended products and corresponding recommendation messages for each product are generated.

5. The recommended method according to claim 4, characterized in that, The step of recommending products to the target user based on the click-through rate prediction scores of the multiple target recommended products further includes: The recommended sales scripts for each product are sent to the customer service agent so that the agent can recommend the products to the target user.

6. A training method for a Wide&Deep model, characterized in that, The training method includes: Multiple multidimensional feature vectors of the target recommended product are obtained. The multidimensional feature vectors include the product feature vector of the target recommended product, the user profile feature vector of the target user of the target recommended product, the agent feature vector of the customer service agent who recommends the target recommended product to the target user, and the context feature vector of the customer service agent when recommending the target recommended product to the target user. The agent feature vector is used to characterize the skill expertise and business capabilities of the customer service agent. The context feature vector is used to characterize the time and environmental characteristics of the customer service agent when recommending the target recommended product to the target user. The product feature vector is used to characterize the tag features and pricing features of the target recommended product. Each multidimensional feature vector is input into the target Wide&Deep neural network model to obtain the click-through rate prediction score of the target recommended product. The target Wide&Deep neural network model includes a Wide sub-network layer, a Deep sub-network layer, and an output layer. The Wide sub-network layer is used to perform linear processing on the sparse binary features of the multidimensional feature vector to obtain a first score. The Deep sub-network layer is used to perform non-linear processing on the categorical features and continuous features in the multidimensional feature vector to obtain a second score. The output layer is used to perform weighted fusion on the first score and the second score to obtain the click-through rate prediction score. The target Wide&Deep neural network model is trained using multiple multidimensional feature vectors as samples and the click pass rate prediction score as labels.

7. The training method according to claim 6, characterized in that, The step of training the target Wide&Deep neural network model using multiple multidimensional feature vectors as samples and the click pass rate prediction score as a label includes: The multidimensional feature vector is input into the target Wide&Deep neural network model, and forward propagation calculation is performed to obtain the click pass rate prediction score; Based on the predicted score of the click pass rate and the actual score of the sample, the loss value of the target Wide&Deep neural network model is calculated using the cross-entropy loss function; Based on the loss value, the gradients of each parameter of the target Wide&Deep neural network model are calculated using the backpropagation algorithm; An adaptive optimization algorithm is used to update the weights and bias parameters of the target Wide&Deep neural network model according to the gradient until the target Wide&Deep neural network model converges, thus obtaining the trained target Wide&Deep neural network model.

8. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method of any one of claims 1-5 or the steps of the method of any one of claims 6-7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5 or the steps of the method according to any one of claims 6-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-5 or the steps of the method according to any one of claims 6-7.

Images