Material recall method, device, computer equipment and storage medium

Abstract

This application relates to a method, apparatus, computer equipment, and storage medium for material recall, and pertains to the field of Internet technology. The method includes: upon receiving a user request from a target user, acquiring multiple user features of the target user; inputting the multiple user features into a dual-tower recall model for feature extraction and weighting processing to obtain a target user vector containing an enhanced first feature vector; the first feature vector is a feature vector representing the effective long-tail low-frequency features of the target user; based on the similarity between the target material vector and the target user vector of each material to be recalled, determining the recall material from the materials to be recalled and performing the recall; the target material vector contains an enhanced second feature vector, which is a feature vector representing the effective long-tail low-frequency features of the material to be recalled. Through this method, sufficient learning of long-tail low-frequency features is achieved, thereby improving the accuracy of feature acquisition and thus improving the effectiveness of material recall.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a method, apparatus, computer equipment, and storage medium for material recall. Background Technology

[0002] In interest-based social networking scenarios, personalized recommendation systems are responsible for distributing various materials to users who are interested, increasing click-through rates and improving user experience. Personalized recommendation systems typically consist of two modules: recall and ranking. The recall module uses various recall algorithms to filter materials from a pool of millions of candidates, selecting materials that a thousand users might be interested in, forming a recall material pool. The ranking module uses ranking algorithms to score the recalled materials and selects the Top K materials that users are most interested in based on the scores, distributing them to the users.

[0003] In related technologies, a dual-tower recall model can be used to screen recalled materials. This model calculates the similarity between user vectors and material vectors to represent the degree of user interest in the materials, thereby recalling the top K materials with the highest similarity.

[0004] However, due to the large number of long-tailed low-frequency features in the recommendation field, the traditional dual-tower recall model is not sufficiently or reliably learned for long-tailed low-frequency features, resulting in low model accuracy. This leads to the recall and recommendation of popular materials to users in interest-based social scenarios, with low personalization, thus resulting in poor material push performance. Summary of the Invention

[0005] This application provides a method, apparatus, computer device, and storage medium for material recall. The resulting material and user vectors better reflect the personalized characteristics of the materials to be recalled and the target users, improving feature acquisition accuracy. This, in turn, makes the recalled materials selected based on the similarity between the material and user vectors more aligned with the interests and needs of the target users, thus improving the effectiveness of material recall. The technical solution is as follows:

[0006] On the one hand, a material recall method is provided, the method comprising:

[0007] Upon receiving a user request from a target user, multiple user features of the target user are acquired; the multiple user features include effective long-tail low-frequency features.

[0008] The multiple user features are input into the dual-tower recall model for feature extraction and weighting to obtain the target user vector of the target user; the target user vector contains an enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user.

[0009] Based on the similarity between the target material vector of each recall material and the target user vector, recall materials are determined from each recall material and recall is performed; the target material vector of the recall material is obtained by the dual-tower recall model after performing feature extraction and feature weighting processing on multiple material features of the input recall material; the multiple material features include effective long-tail low-frequency features; the target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the recall material.

[0010] On the other hand, a material recall device is provided, the device comprising:

[0011] The feature acquisition module is used to acquire multiple user features of the target user when a user request from the target user is received; the multiple user features include effective long-tail low-frequency features;

[0012] The vector acquisition module is used to input the multiple user features into the dual-tower recall model for feature extraction and feature weighting to obtain the target user vector of the target user; the target user vector contains an enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user.

[0013] The recall module is used to determine and recall materials from the materials to be recalled based on the similarity between the target material vector and the target user vector of each material to be recalled. The target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled. The multiple material features include effective long-tail low-frequency features. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled.

[0014] In one possible implementation, the dual-tower recall model includes a first feature extraction layer and a first compressed excitation network SENet+ layer;

[0015] The vector acquisition module includes:

[0016] The feature extraction submodule is used to perform feature extraction processing on the input multiple user features through the first feature extraction layer of the dual-tower recall model to obtain a user feature vector; the user feature vector contains multiple feature vectors that correspond one-to-one with the multiple user features;

[0017] The feature weighting submodule is used to perform feature weighting processing on the user feature vector through the first SENet+ layer of the dual-tower recall model to obtain the target user vector.

[0018] In one possible implementation, the plurality of user features also includes invalid long-tailed low-frequency features;

[0019] The feature weighting submodule includes:

[0020] The vector splitting unit is used to split the plurality of feature vectors in the user feature vectors respectively to obtain at least two user splitting vectors corresponding to each of the plurality of user features;

[0021] The compression processing unit is used to perform compression processing based on the at least two user split vectors corresponding to each user feature, so as to obtain the compressed vector corresponding to each user feature.

[0022] The excitation processing unit is used to perform excitation processing based on the compression vector corresponding to each user feature to obtain the weight corresponding to each user splitting vector; wherein, the weight value of the user splitting vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the user splitting vector corresponding to the invalid long-tail low-frequency feature.

[0023] The weighted processing unit is used to perform weighted processing on each user split vector based on the weights corresponding to each user split vector to obtain the target user vector.

[0024] In one possible implementation, the compression processing unit is configured to extract feature values ​​from the at least two user split vectors corresponding to the target user feature, thereby obtaining at least two feature value sets that correspond one-to-one with the at least two user split vectors, wherein the feature value sets contain at least two feature values; the target user feature is any one of the plurality of user features.

[0025] The feature values ​​of at least two feature value sets corresponding to the target user feature are concatenated to obtain the compressed vector corresponding to the target user feature.

[0026] In one possible implementation, the feature set includes a first feature value and a second feature value, wherein the first feature value is the maximum value in the user split vector and the second feature value is the average value in the user split vector.

[0027] In one possible implementation, the excitation processing unit is used to concatenate the compressed vectors corresponding to the plurality of user features to obtain a first concatenated vector.

[0028] The first concatenated vector is processed by a first fully connected neural network with an activation function to obtain a first processing result.

[0029] The first processing result is fully processed by a second fully connected neural network without an activation function to obtain the weights corresponding to each user split vector output by each neuron of the second fully connected neural network; wherein the number of neurons in the second fully connected neural network is the same as the number of user split vectors.

[0030] In one possible implementation, the weighted processing unit is used to perform a product operation on each of the user split vectors and the corresponding weights to obtain each product result;

[0031] The summation process is performed on each of the product results and the corresponding user split vector to obtain each summation result;

[0032] The vector formed by the summation results is taken as the target user vector of the target user.

[0033] In one possible implementation, the dual-tower recall model is obtained based on training samples; the training samples include sample user characteristics of sample users, material characteristics of positive sample materials, and material characteristics of negative sample materials; the positive sample materials include sample materials that have received positive feedback from the sample users, and the negative sample materials include sample materials that have not received feedback from the sample users or sample materials that have received negative feedback from the sample users.

[0034] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the above-described material recall method.

[0035] On the other hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer program, which is loaded and executed by a processor to implement the above-described material recall method.

[0036] On the other hand, a computer program product is provided, the computer program product comprising at least one computer program, the computer program being loaded and executed by a processor to implement the material recall method provided in the various alternative implementations described above.

[0037] The technical solution provided in this application may include the following beneficial effects:

[0038] The material recall method provided in this application, when extracting the material vector of the material to be recalled and the user vector of the target user, strengthens the feature vector corresponding to the effective long-tail low-frequency features in the material features of the material to be recalled and the feature vector corresponding to the effective long-tail low-frequency features in the user features of the target user through a dual-tower recall model. This achieves full learning of the long-tail low-frequency features in the material features and the long-tail low-frequency features in the user features, so that the obtained material vector and user vector can better reflect the personalized features of the material to be recalled and the target user, improve the feature acquisition accuracy, and thus make the recalled materials obtained by screening based on the similarity between the material vector and the user vector more in line with the interests and needs of the target user, thereby improving the effect of material recall.

[0039] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0041] Figure 1 A schematic diagram of an exemplary dual-tower recall model structure provided in this application is shown;

[0042] Figure 2 This invention provides a schematic diagram of the structure of an upgraded dual-tower recall model according to an exemplary embodiment of the present application.

[0043] Figure 3 A flowchart illustrating an exemplary embodiment of this application provides a method for recalling materials;

[0044] Figure 4 A flowchart illustrating an exemplary embodiment of this application provides a method for recalling materials;

[0045] Figure 5 This invention illustrates a schematic diagram of the computational logic of the SENet+ layer provided in an exemplary embodiment of this application.

[0046] Figure 6 A flowchart of a material recall method provided in an exemplary embodiment of this application is shown;

[0047] Figure 7 A block diagram of a material recall apparatus provided in an exemplary embodiment of this application is shown;

[0048] Figure 8 A structural block diagram of a computer device illustrated in an exemplary embodiment of this application is shown;

[0049] Figure 9A structural block diagram of a computer device illustrated in an exemplary embodiment of this application is shown. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0051] In interest-based social scenarios, personalized recommendation systems are responsible for distributing various materials to users who are interested, thereby increasing the click-through rate of materials and improving the user experience.

[0052] Personalized recommendation systems typically consist of two modules: recall and ranking. The dual-tower recall algorithm is a simple and efficient representation learning algorithm that has been widely applied in the recall modules of various personalized recommendation systems. Figure 1 A schematic diagram of an exemplary dual-tower recall model structure provided in this application is shown, such as... Figure 1 As shown, the dual-tower recall model includes a user tower 110 (user-side multilayer neural network) and a material tower 120 (material-side multilayer neural network). The user tower 110 represents the input user features as user feature vectors, and the material tower 120 represents the input material features as material feature vectors. Then, the similarity between the material feature vectors and user feature vectors is calculated to represent the user's level of interest in the materials. During online service, the dual-tower recall model calculates the similarity between the user feature vector and the material feature vectors of all materials in the material library, thereby identifying and recalling the Top K materials with the highest similarity, and adding them to the recall material library.

[0053] like Figure 1 As shown, user characteristics can include user-related information such as gender, age, interests, and city; material characteristics can include material-related information such as the number of reposts, comments, likes, first-level tags, second-level tags, and third-level tags.

[0054] Because the recommendation field contains a large number of long-tailed, low-frequency features, traditional dual-tower recall models are insufficient and unreliable in learning these features, resulting in low model accuracy. Consequently, in interest-based social scenarios, the materials recalled and recommended to users tend to be popular and lack personalization, leading to poor material push performance. To address the shortcomings of traditional dual-tower recall models, this application provides an upgraded dual-tower recall model that improves the learning ability of long-tailed, low-frequency features, thereby enhancing the effectiveness of material recall.

[0055] Figure 2 This application shows a schematic diagram of the structure of an upgraded dual-tower recall model provided in an exemplary embodiment, as follows: Figure 2 As shown, the upgraded dual-tower recall model adds an SENet+ (Squeeze-Excitation Network) layer 210 to both the user tower and the material tower, based on the traditional dual-tower recall model. This SENet+ layer 210 is positioned between the feature extraction layer and the MLP (Multilayer Perceptron) layer. The SENet+ layer can suppress noise or invalid long-tail low-frequency features through feature weighting, while strengthening effective long-tail low-frequency features. When the upgraded dual-tower recall model is applied to material recall, it can improve the reliability of feature extraction, pay attention to the information of effective long-tail low-frequency features, improve model performance, and enhance the personalization of recalled materials in interest-based social scenarios, thereby making the recalled materials more in line with user interests.

[0056] Based on such Figure 2 The upgraded dual-tower recall model shown is Figure 3 This application illustrates a flowchart of an exemplary embodiment of a material recall method provided. This material recall method can be executed by a computer device, which can be implemented as a server or a terminal, such as... Figure 3 As shown, the material recall method may include the following steps:

[0057] Step 310: When a user request from a target user is received, multiple user features of the target user are obtained; among the multiple user features are effective long-tail low-frequency features.

[0058] The user request can be a material acquisition request generated and sent based on the target user's action of opening the target application, or a material acquisition request generated and sent based on the target user's action of refreshing the target application.

[0059] Multiple user characteristics can include user-related information such as gender, age, interests, and city. The content included in user characteristics can be set based on relevant individuals, and this application does not impose any restrictions on this. Long-tail low-frequency characteristics refer to characteristics that have a low probability of appearing in a specific field or have a low influence in big data statistics, which can also reflect the personalization of user needs.

[0060] Among them, long-tail low-frequency features can also include effective long-tail low-frequency features, which can refer to long-tail low-frequency features that are strongly related to user interests.

[0061] Step 320: Input multiple user features into the dual-tower recall model for feature extraction and feature weighting to obtain the target user vector of the target user; the target user vector contains the enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the target user.

[0062] The dual-tower recall model is pre-trained, such as... Figure 2 The upgraded dual-tower recall model shown can perform feature extraction and feature weighting on the input features to obtain enhanced feature vectors.

[0063] In this embodiment of the application, the computer device can enhance the effective long-tail low-frequency features by assigning higher weights to the effective long-tail low-frequency features.

[0064] Step 330: Based on the similarity between the target material vector and the target user vector of each material to be recalled, determine the materials to be recalled from each material to be recalled and perform recall; the target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled; the multiple material features include effective long-tail low-frequency features; the target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled.

[0065] In this embodiment, when the computer device receives a user request from a target user online, the computer device can acquire the user characteristics of the target user who initiated the request in real time, and obtain the target user vector of the target user based on the user characteristics. When the computer device receives materials to be recalled and puts them into the warehouse, it generates a target material vector for the materials to be recalled. Optionally, the computer device can periodically acquire materials to be recalled in batches and update the materials to be recalled in the material warehouse at specified time intervals to ensure the timeliness of the materials to be recalled in the material warehouse. Since the time of receiving the user request from the target user and the time of putting the materials to be recalled into the warehouse may differ, the process of acquiring the target material vector of the materials to be recalled and the process of acquiring the target user vector of the target user can be performed synchronously or asynchronously.

[0066] Computer equipment can obtain the target material vector of the materials to be recalled in the following ways:

[0067] Upon receiving materials to be recalled, multiple material characteristics of the materials to be recalled are obtained; among these multiple material characteristics are effective long-tailed low-frequency characteristics.

[0068] Multiple material features are input into the dual-tower recall model for feature extraction and weighting to obtain the target material vector of the material to be recalled. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled.

[0069] Material characteristics can include material-related information such as the number of reposts, comments, likes, first-level tags, second-level tags, and third-level tags.

[0070] In this embodiment, the similarity between the target material vector of the material to be recalled and the target user vector of the target user is used to characterize the degree of interest of the target user in the material to be recalled. Optionally, the computer device can identify the top K materials to be recalled with high similarity as the recalled materials and recall them, where K is a positive integer. The value of K can be set by relevant personnel based on actual needs, and this application does not impose any restrictions on it.

[0071] In summary, the material recall method provided in this application, when extracting the material vector of the material to be recalled and the user vector of the target user, strengthens the feature vector corresponding to the effective long-tail low-frequency features in the material features of the material to be recalled and the feature vector corresponding to the effective long-tail low-frequency features in the user features of the target user through a dual-tower recall model. This achieves full learning of the long-tail low-frequency features in the material features and the long-tail low-frequency features in the user features, thereby making the obtained material vector and user vector more reflective of the personalized features of the material to be recalled and the target user, improving the feature acquisition accuracy, and thus making the recalled materials obtained by screening based on the similarity between the material vector and the user vector more in line with the interests and needs of the target user, thereby improving the effect of material recall.

[0072] Based on such Figure 2 The upgraded dual-tower recall model shown is Figure 4 This application illustrates a flowchart of an exemplary embodiment of a material recall method provided. This material recall method can be executed by a computer device, which can be implemented as a server or a terminal, such as... Figure 4 As shown, the material recall method may include the following steps:

[0073] Step 410: When a user request from a target user is received, multiple user features of the target user are obtained; among the multiple user features are effective long-tail low-frequency features.

[0074] In this application embodiment, multiple user features may also include invalid long-tail low-frequency features; invalid long-tail low-frequency features may refer to long-tail low-frequency features that are weakly related to user interests. The distinction between valid and invalid long-tail low-frequency features can be determined by the trained dual-tower recall model. That is, the upgraded dual-tower recall model can enhance the identified valid long-tail low-frequency features to obtain more effective user personalized information. This application does not limit the distinction between valid and invalid long-tail low-frequency features.

[0075] Step 420: The first feature extraction layer of the dual-tower recall model performs feature extraction processing on the input multiple user features to obtain a user feature vector; the user feature vector contains multiple feature vectors that correspond one-to-one with the multiple user features.

[0076] In this embodiment of the application, the computer device may be configured with, for example, Figure 2 The upgraded dual-tower recall model is shown. This dual-tower recall model includes a user tower and a material tower. Both the user tower and the material tower contain a feature extraction layer and an SENet+ layer. In this application, the feature extraction layer in the user tower is referred to as the first feature extraction layer, the SENet+ layer in the user tower is referred to as the first SENet+ layer, the feature extraction layer in the material tower is referred to as the second feature extraction layer, and the SENet+ layer in the material tower is referred to as the second SENet+ layer. Therefore, when the computer device receives a user request from a target user, it can process multiple user features of the target user through the first feature extraction layer to obtain the user feature vector of the target user. The user feature vector of the target user is a representation of the multiple user features of the target user.

[0077] Optionally, the dual-tower recall model is obtained based on training samples; the training samples include sample user characteristics, positive sample material characteristics, and negative sample material characteristics; positive sample materials include sample materials that have received positive feedback from sample users, and negative sample materials include sample materials that have not received feedback from sample users or sample materials that have received negative feedback from sample users.

[0078] Optionally, the user characteristics of the sample users include long-tailed low-frequency features, which include both valid and invalid long-tailed low-frequency features.

[0079] To further improve the training effect of the dual-tower recall model, the positive sample material can optionally be the sample material that is strongly correlated with the long-tail low-frequency features of the user among the sample materials that have received positive feedback from the sample user.

[0080] In this embodiment, the training process of the dual-tower recall model can be implemented by a model training device, which can be a server or a terminal. The model training device can train the dual-tower recall model on a large-scale training sample set. After each round of training, the model effect is evaluated on the validation set, and the model with the best effect is saved as the online service model. That is, the dual-tower recall model with the highest accuracy in distinguishing between positive and negative sample materials is used as the online service model.

[0081] In this embodiment, an SENet+ layer is added to both the user tower and the material tower on the basis of the traditional dual-tower recall model. This allows the first SENet+ layer in the user tower to strengthen the effective long-tail low-frequency features in user features and suppress noise or invalid low-frequency features when recalling materials based on recommended users. Similarly, the second SENet+ layer in the material tower strengthens the effective long-tail low-frequency features in material features and suppresses noise or invalid low-frequency features. This improves the reliability of the feature vectors recalled by the dual-tower recall model, thereby enhancing model performance and increasing the personalization of recalled materials in interest-based social scenarios, ultimately recalling materials that better match user interests.

[0082] Step 430: The user feature vector is processed by the first SENet+ layer of the dual-tower recall model to obtain the target user vector.

[0083] The target user vector contains the enhanced first feature vector from the user feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the target user.

[0084] In this embodiment of the application, the computer device can perform weighted processing on the user feature vector extracted by the feature extraction layer through the first SENet+ layer in the upgraded dual-tower recall model, thereby obtaining the target user vector of the target user.

[0085] In one possible implementation, the process of obtaining the target user vector by weighting the user feature vector through the first SENet+ layer of the dual-tower recall model can be implemented as follows:

[0086] S4301, perform vector splitting on multiple feature vectors in the user feature vector to obtain at least two user split vectors corresponding to each of the multiple user features.

[0087] In this embodiment, each feature has a corresponding feature vector representation. Vector splitting is performed on a per-feature-vector basis, and the number of user split vectors corresponding to each user feature can be the same or different. For example, a target user has age, gender, and geographic features, each with its own corresponding feature vector. Taking the example that the number of user split vectors corresponding to each user feature is the same, during vector splitting, the feature vector corresponding to the age feature can be split into two user split vectors, the feature vector corresponding to the gender feature can be split into two user split vectors, and the feature vector corresponding to the geographic feature can be split into two user split vectors.

[0088] S4302, compresses at least two user split vectors corresponding to each user feature to obtain the compressed vector corresponding to each user feature.

[0089] In this embodiment of the application, compressing the feature vector can reduce the dimension of the feature vector while retaining the feature information in the feature vector, thereby reducing the amount of data processing required by the computer device.

[0090] Taking the compression process of the feature vector corresponding to a target user feature among various user features as an example, in one possible implementation, the process of compressing the feature vector corresponding to each user feature based on at least two user split vectors can be implemented as follows:

[0091] Feature values ​​are extracted from at least two user split vectors corresponding to the target user feature to obtain at least two feature value sets that correspond one-to-one with the at least two user split vectors. Each feature value set contains at least two feature values. The target user feature is any one of multiple user features.

[0092] The feature values ​​from at least two feature value sets corresponding to the target user feature are concatenated to obtain the compressed vector corresponding to the target user feature.

[0093] Optionally, at least two features in the feature set may be obtained in different ways.

[0094] Schematic, the feature set can contain two feature values, namely, a first feature value and a second feature value, wherein the first feature value is obtained in a different way than the second feature value. In one possible implementation, the first feature value is the maximum value in the user split vector, and the second feature value is the average value in the user split vector. Alternatively, the first and second feature values ​​can also be set to other values ​​in the user split vector based on actual needs, such as the minimum value, the median, etc.

[0095] Taking the first feature value as the maximum value in the user split vector and the second feature value as the average value in the user split vector as an example, after dividing the feature vector corresponding to the target user feature into multiple user split vectors, two values ​​(i.e., the first feature value and the second feature value) are extracted from each user split vector to form a feature value set. The feature values ​​of all user split vectors of the target user feature are concatenated together to obtain the compressed vector of the target user feature. For example, if the feature vector of the target user feature has an 8-dimensional dimension, during vector splitting, the feature vector is split into two user split vectors, each with a 4-dimensional dimension. The maximum value and average value are taken from each 4-dimensional vector, and each 4-dimensional vector is compressed into a 2-dimensional vector. The two compressed 2-dimensional vectors of the two user split vectors are concatenated to obtain the compressed vector of the target user feature's feature vector; that is, an 8-dimensional feature vector is compressed into two vectors * 2 dimensions = a 4-dimensional vector. It should be noted that the above process of compressing feature vectors is only illustrative. Relevant personnel can set the number of user-split vectors obtained by vector splitting based on actual needs. At the same time, they can also set the type and number of feature values ​​included in the extracted feature value set. This application does not impose any restrictions on this.

[0096] In this embodiment of the application, the computer device can compress the user split vector corresponding to each user feature through the first compression (Squeeze) network in the first SENet+ layer of the user tower in the dual-tower recall model.

[0097] S4303, based on the compression vector corresponding to each user feature, excitation processing is performed to obtain the weights corresponding to each user split vector; among them, the weight value of the user split vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the user split vector corresponding to the ineffective long-tail low-frequency feature.

[0098] In this embodiment of the application, the computer device can implement the excitation processing based on the compression vector corresponding to each user feature through the first excitation network in the first SENet+ layer of the user tower in the dual-tower recall model; the excitation network is used to calculate the weight of each user split vector.

[0099] Optionally, the process of obtaining the weights corresponding to each user's split vector by performing excitation processing based on the compressed vector corresponding to each user's feature can be implemented as follows:

[0100] The compressed vectors corresponding to multiple user features are concatenated to obtain the first concatenated vector;

[0101] The first concatenated vector is processed by a first fully connected neural network with an activation function to obtain the first processing result.

[0102] The first processing result is fully processed by a second fully connected neural network without an activation function to obtain the weights corresponding to each user split vector output by each neuron of the second fully connected neural network; wherein the number of neurons in the second fully connected neural network is the same as the number of user split vectors.

[0103] When the first activation network processes the activation based on the compressed vectors corresponding to each user feature, it first concatenates the compressed vectors of all user features together and inputs them into a first fully connected neural network with an activation function to obtain a first processing result. Then, the first processing result output by the first fully connected neural network is input into a second fully connected neural network without a nonlinear activation function to obtain the weights of each user split vector output by the second fully connected neural network. Optionally, the activation function in the first fully connected neural network is the ReLU (Rectified Linear Unit) activation function, and the number of neurons in the second fully connected neural network is the same as the number of user split vectors to ensure that the second fully connected neural network can output the weights corresponding to each user split vector.

[0104] S4304, based on the weights corresponding to each user's split vector, performs weighted processing on each user split vector to obtain the target user vector.

[0105] In this embodiment of the application, after obtaining the weights corresponding to each user split vector, the user split vector and its corresponding weight can be multiplied to obtain the product results. In one possible case, the computer device can determine the vector composed of the product results as the target user vector of the target user. In another possible case, after obtaining the product results, the computer device can sum the product results and the corresponding user split vector to obtain the summation results, and use the vector composed of the summation results as the target user vector of the target user.

[0106] Optionally, if the user feature corresponding to the user split vector is an invalid long-tailed low-frequency feature or noise, the weight corresponding to the user split vector can be negative or 0, thereby achieving the suppression effect of invalid long-tailed low-frequency features or noise.

[0107] Optionally, the computer device can use the first weighting (Reweight) network in the first SENet+ layer of the user tower of the dual-tower recall model to weight each user segmentation vector. Illustratively, the weighting network can multiply the weight of each user segmentation vector by the corresponding user segmentation vector and then add the corresponding user segmentation vector to complete the weighting of the original user feature vector.

[0108] Step 440: Based on the similarity between the target material vector and the target user vector of each material to be recalled, the recalled materials are determined from the materials to be recalled and recalled. The target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled. The multiple material features include effective long-tail low-frequency features. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled.

[0109] In the embodiments of this application, the computer device can calculate the cosine similarity or other optional similarity between the target material vector and the target user vector of each material to be recalled, and this application does not impose any restrictions on this.

[0110] The process of obtaining the target material vector of the recalled material by computer equipment is similar to the process of obtaining the target user vector of the target user. The computer equipment can extract the target material vector of the recalled material through the material tower containing the second SENet+ layer in the dual-tower recall model; the following uses the process of obtaining the target material vector of a recalled material as an example to illustrate the process.

[0111] S4401, upon receiving materials to be recalled, acquires multiple material characteristics of the materials to be recalled; among the multiple material characteristics are effective long-tailed low-frequency characteristics.

[0112] S4402, through the second feature extraction layer of the dual-tower recall model, performs feature extraction processing on multiple input material features to obtain a material feature vector; the material feature vector contains multiple feature vectors that correspond one-to-one with multiple material features;

[0113] S4403 uses the second SENet+ layer of the dual-tower recall model to perform feature weighting on the material feature vector to obtain the target material vector.

[0114] The process of obtaining the target material vector by weighting the material feature vector through the second SENet+ layer of the dual-tower recall model can be implemented as follows:

[0115] S1, perform vector splitting on multiple feature vectors in the material feature vector to obtain at least two material split vectors corresponding to each of the multiple material features.

[0116] S2, compress the material based on at least two material splitting vectors corresponding to each material feature to obtain the compression vector corresponding to each material feature.

[0117] Taking a material feature of the material to be recalled as an example, optionally, the process of compressing at least two material splitting vectors corresponding to each material feature to obtain the compressed vector corresponding to each material feature can be implemented as follows:

[0118] Feature values ​​are extracted from at least two material splitting vectors corresponding to the target material feature to obtain at least two feature value sets that correspond one-to-one with the at least two material splitting vectors. Each feature value set contains at least two feature values. The target material feature is any one of multiple material features.

[0119] The feature values ​​in at least two feature value sets corresponding to the target material feature are concatenated to obtain the compressed vector corresponding to the target material feature.

[0120] Taking a feature set extracted from a material splitting vector containing two feature values ​​as an example, optionally, a feature set can include a first feature value and a second feature value; in this case, the first feature value extracted from the material splitting vector can be the maximum value in the material splitting vector, and the second feature value is the average value in the material splitting vector. Alternatively, the values ​​of the first and second feature values ​​extracted from the material splitting vector can also be set differently by relevant personnel based on actual needs.

[0121] The process of obtaining the compression vectors of each material feature described above can be implemented by the second compression network in the second SENet+ layer of the material tower.

[0122] S3, based on the compression vector corresponding to each material feature, excitation processing is performed to obtain the weight corresponding to each material splitting vector; among them, the weight value of the material splitting vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the material splitting vector corresponding to the ineffective long-tail low-frequency feature.

[0123] Optionally, the process of obtaining the weights corresponding to each material split vector can be implemented as follows:

[0124] The compression vectors corresponding to multiple material features are concatenated to obtain a second concatenated vector;

[0125] The second concatenated vector is processed by a third fully connected neural network with an activation function to obtain the second processing result;

[0126] The second processing result is fully processed by a fourth fully connected neural network without activation functions to obtain the weights corresponding to each material splitting vector output by each neuron of the fourth fully connected neural network; wherein the number of neurons in the fourth fully connected neural network is the same as the number of material splitting vectors.

[0127] The above-mentioned excitation process based on the compression vector corresponding to each material characteristic can be implemented by the second excitation network in the second SENet+ layer of the material tower.

[0128] S4. Based on the weights corresponding to each material splitting vector, the weighted processing of each material splitting vector is performed to obtain the target material vector.

[0129] The above weighting process can be implemented through the second weighting network in the second SENet+ layer of the material tower.

[0130] The process of obtaining the target material vector of the recalled materials by computer equipment is similar to the process of obtaining the target user vector of the target user. Therefore, the relevant content in the process of obtaining the target material vector of the recalled materials can refer to the relevant content in the process of obtaining the target user vector of the target user in steps 410 to 430, which will not be repeated here.

[0131] Taking any SENet+ layer as an example, Figure 5 This application illustrates a schematic diagram of the computational logic of the SENet+ layer provided in an exemplary embodiment, as shown below. Figure 5 As shown, after receiving the initial feature vector uploaded by the feature extraction layer 510 (in the user tower, this initial feature vector is the user feature vector; in the material tower, this initial feature vector is the material feature vector), the compression network 520 performs vector splitting and compression processing on the initial feature vector to obtain compressed vectors of multiple split vectors corresponding to the initial feature vector (the process of obtaining compressed vectors of multiple split vectors through the compression network can be found in [reference]). Figure 4 The relevant content of the illustrated embodiment will not be repeated here; the compressed vectors of multiple split vectors are concatenated and then input into the excitation network 530. The excitation network calculates the weights of each split vector to obtain the weights of each split vector (the process of calculating the weights of each split vector through the excitation network can be found in [reference]). Figure 4 The relevant content of the illustrated embodiment will not be repeated here; the weighted network 540 weights each split vector based on the weight of each split vector, and outputs a new feature vector (in the user tower, the new feature vector is a new user feature vector; in the material tower, the new feature vector is a new material feature vector). After processing by the corresponding MLP layer, the corresponding target feature vector is obtained; thus, the initial feature vector is weighted so that the subsequent similarity calculation can be performed based on the weighted target vector to determine the recalled materials.

[0132] In summary, the material recall method provided in this application, when extracting the material vector of the material to be recalled and the user vector of the target user, strengthens the feature vector corresponding to the effective long-tail low-frequency features in the material features of the material to be recalled and the feature vector corresponding to the effective long-tail low-frequency features in the user features of the target user through a dual-tower recall model. This achieves full learning of the long-tail low-frequency features in the material features and the long-tail low-frequency features in the user features, thereby making the obtained material vector and user vector more reflective of the personalized features of the material to be recalled and the target user, improving the feature acquisition accuracy, and thus making the recalled materials obtained based on the similarity between the material vector and the user vector more in line with the interests and needs of the target user, thereby improving the effect of material recall.

[0133] Figure 6 A flowchart of a material recall method provided in an exemplary embodiment of this application is shown. This material recall method can be executed by a computer device, which can be implemented as a server or a terminal; Figure 6 As shown, the material recall method may include the following steps:

[0134] Step 610: When a user request from a target user is received, the target user vector of the target user is obtained through the user tower in the dual-tower recall model; the user tower contains a first SENet+ layer, which is used to perform feature weighting processing on the user feature vector of the target user and output the target user vector; the target user vector contains the enhanced first feature vector in the user feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user.

[0135] Step 620: Upon receiving each material to be recalled, the target material vector of each material to be recalled is obtained through the material tower in the dual-tower recall model. The material tower contains a second SENet+ layer, which is used to perform feature weighting processing on the material feature vector of the material to be recalled and output the target material vector of the material to be recalled. The target material vector contains a second feature vector that has been enhanced from the material feature vector. The second feature vector is the feature vector corresponding to the effective long-tail low-frequency feature of the material to be recalled.

[0136] Step 630: Based on the similarity between the target material vector and the target user vector of each material to be recalled, determine the materials to be recalled from each material to be recalled and carry out the recall.

[0137] In summary, the material recall method provided in this application, when extracting the material vector of the material to be recalled and the user vector of the target user, strengthens the feature vector corresponding to the effective long-tail low-frequency features in the material features of the material to be recalled and the feature vector corresponding to the effective long-tail low-frequency features in the user features of the target user through a dual-tower recall model. This achieves full learning of the long-tail low-frequency features in the material features and the long-tail low-frequency features in the user features, thereby making the obtained material vector and user vector more reflective of the personalized features of the material to be recalled and the target user, improving the feature acquisition accuracy, and thus making the recalled materials obtained based on the similarity between the material vector and the user vector more in line with the interests and needs of the target user, thereby improving the effect of material recall.

[0138] Figure 7 This illustration shows a block diagram of a material recall apparatus provided in an exemplary embodiment of this application, which can perform actions such as Figure 3 or Figure 4 All or part of the steps in the illustrated embodiments, such as Figure 7 As shown, the material recall device includes:

[0139] The feature acquisition module 710 is used to acquire multiple user features of the target user when a user request from the target user is received; the multiple user features include effective long-tail low-frequency features;

[0140] The vector acquisition module 720 is used to input the multiple user features into the dual-tower recall model for feature extraction and feature weighting to obtain the target user vector of the target user; the target user vector includes an enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user.

[0141] The recall module 730 is used to determine and recall materials from the materials to be recalled based on the similarity between the target material vector of each material to be recalled and the target user vector. The target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled. The multiple material features include effective long-tail low-frequency features. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled.

[0142] In one possible implementation, the dual-tower recall model includes a first feature extraction layer and a first compressed excitation network SENet+ layer;

[0143] The vector acquisition module 720 includes:

[0144] The feature extraction submodule is used to perform feature extraction processing on the input multiple user features through the first feature extraction layer of the dual-tower recall model to obtain a user feature vector; the user feature vector contains multiple feature vectors that correspond one-to-one with the multiple user features;

[0145] The feature weighting submodule is used to perform feature weighting processing on the user feature vector through the first SENet+ layer of the dual-tower recall model to obtain the target user vector.

[0146] In one possible implementation, the plurality of user features also includes invalid long-tailed low-frequency features;

[0147] The feature weighting submodule includes:

[0148] The vector splitting unit is used to split the plurality of feature vectors in the user feature vectors respectively to obtain at least two user splitting vectors corresponding to each of the plurality of user features;

[0149] The compression processing unit is used to perform compression processing based on the at least two user split vectors corresponding to each user feature, so as to obtain the compressed vector corresponding to each user feature.

[0150] The excitation processing unit is used to perform excitation processing based on the compression vector corresponding to each user feature to obtain the weight corresponding to each user splitting vector; wherein, the weight value of the user splitting vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the user splitting vector corresponding to the invalid long-tail low-frequency feature.

[0151] The weighted processing unit is used to perform weighted processing on each user split vector based on the weights corresponding to each user split vector to obtain the target user vector.

[0152] In one possible implementation, the compression processing unit is configured to extract feature values ​​from the at least two user split vectors corresponding to the target user feature, thereby obtaining at least two feature value sets that correspond one-to-one with the at least two user split vectors, wherein the feature value sets contain at least two feature values; the target user feature is any one of the plurality of user features.

[0153] The feature values ​​of at least two feature value sets corresponding to the target user feature are concatenated to obtain the compressed vector corresponding to the target user feature.

[0154] In one possible implementation, the feature set includes a first feature value and a second feature value, wherein the first feature value is the maximum value in the user split vector and the second feature value is the average value in the user split vector.

[0155] In one possible implementation, the excitation processing unit is used to concatenate the compressed vectors corresponding to the plurality of user features to obtain a first concatenated vector.

[0156] The first concatenated vector is processed by a first fully connected neural network with an activation function to obtain a first processing result.

[0157] The first processing result is fully processed by a second fully connected neural network without an activation function to obtain the weights corresponding to each user split vector output by each neuron of the second fully connected neural network; wherein the number of neurons in the second fully connected neural network is the same as the number of user split vectors.

[0158] In one possible implementation, the weighted processing unit is used to perform a product operation on each of the user split vectors and the corresponding weights to obtain each product result;

[0159] The summation process is performed on each of the product results and the corresponding user split vector to obtain each summation result;

[0160] The vector formed by the summation results is taken as the target user vector of the target user.

[0161] In one possible implementation, the dual-tower recall model is obtained based on training samples; the training samples include sample user characteristics of sample users, material characteristics of positive sample materials, and material characteristics of negative sample materials; the positive sample materials include sample materials that have received positive feedback from the sample users, and the negative sample materials include sample materials that have not received feedback from the sample users or sample materials that have received negative feedback from the sample users.

[0162] In summary, the material recall device provided in this application, when extracting the material vector of the material to be recalled and the user vector of the target user, strengthens the feature vector corresponding to the effective long-tail low-frequency features in the material features of the material to be recalled and the feature vector corresponding to the effective long-tail low-frequency features in the user features of the target user through a dual-tower recall model. This achieves full learning of the long-tail low-frequency features in the material features and the long-tail low-frequency features in the user features, thereby making the obtained material vector and user vector more reflective of the personalized features of the material to be recalled and the target user, improving the feature acquisition accuracy, and thus making the recalled materials obtained based on the similarity between the material vector and the user vector more in line with the interests and needs of the target user, thereby improving the effect of material recall.

[0163] Figure 8 A structural block diagram of a computer device 800 illustrated in an exemplary embodiment of this application is shown. This computer device can be implemented as a server as described above in this application. The computer device 800 includes a Central Processing Unit (CPU) 801, a system memory 804 including Random Access Memory (RAM) 802 and Read-Only Memory (ROM) 803, and a system bus 805 connecting the system memory 804 and the CPU 801. The computer device 800 also includes a mass storage device 806 for storing an operating system 809, application programs 810, and other program modules 811.

[0164] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage media are not limited to the above-mentioned types. The system memory 804 and mass storage device 806 described above can be collectively referred to as memory.

[0165] According to various embodiments of this application, the computer device 800 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 800 can be connected to a network 808 via a network interface unit 807 connected to the system bus 805, or the network interface unit 807 can be used to connect to other types of networks or remote computer systems (not shown).

[0166] The memory also includes at least one instruction, at least one program, code set, or instruction set, which are stored in the memory. The central processing unit 801 executes the at least one instruction, at least one program, code set, or instruction set to implement all or part of the steps in the material recall method shown in the above embodiments.

[0167] Figure 9 A structural block diagram of a computer device 900 illustrating an exemplary embodiment of this application is shown. The computer device 900 can be implemented as the aforementioned terminal, such as a smartphone, tablet computer, laptop computer, desktop computer, smartwatch, and television. The computer device 900 may also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or other names.

[0168] Typically, computer device 900 includes a processor 901 and a memory 902.

[0169] In some embodiments, the computer device 900 may optionally include a peripheral device interface 903 and at least one peripheral device. The processor 901, memory 902, and peripheral device interface 903 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 903 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 904, a display screen 905, a camera assembly 906, an audio circuit 907, and a power supply 908.

[0170] In some embodiments, the computer device 900 further includes one or more sensors 909. The one or more sensors 909 include, but are not limited to, an accelerometer 910, a gyroscope 911, a pressure sensor 912, an optical sensor 913, and a proximity sensor 914.

[0171] Those skilled in the art will understand that Figure 9 The structure shown does not constitute a limitation on the computer device 900, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0172] In one exemplary embodiment, a computer-readable storage medium is also provided, which stores at least one computer program that is loaded and executed by a processor to implement all or part of the steps in the above-described material recall method. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0173] In one exemplary embodiment, a computer-readable storage medium is also provided, which stores at least one computer program that is loaded and executed by a processor to implement all or part of the steps in the above-described material recall method. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0174] In one exemplary embodiment, a computer program product is also provided, comprising at least one computer program that is loaded and executed by a processor. Figure 3 , Figure 4 or Figure 6 All or part of the steps of the material recall method shown in any embodiment.

[0175] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0176] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for recalling materials, characterized in that, The method includes: Upon receiving a user request from a target user, multiple user features of the target user are obtained; the multiple user features include effective long-tail low-frequency features and invalid long-tail low-frequency features, the effective long-tail low-frequency features refer to long-tail low-frequency features that are strongly correlated with the user's interests, and the invalid long-tail low-frequency features refer to long-tail low-frequency features that are weakly correlated with the user's interests. The multiple user features are input into the dual-tower recall model for feature extraction and weighting to obtain the target user vector of the target user; the target user vector contains an enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user. Based on the similarity between the target material vector and the target user vector of each material to be recalled, recall materials are determined from the materials to be recalled and recalled. The target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled. The multiple material features include effective long-tail low-frequency features. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled. The dual-tower recall model includes a first feature extraction layer and a first compressed excitation network SENet+ layer. The step of inputting the multiple user features into the dual-tower recall model for feature extraction and weighting to obtain the target user vector of the target user includes: The first feature extraction layer of the dual-tower recall model performs feature extraction processing on the input multiple user features to obtain a user feature vector; the user feature vector contains multiple feature vectors that correspond one-to-one with the multiple user features; The plurality of feature vectors in the user feature vector are split into vectors respectively to obtain at least two user split vectors corresponding to each of the plurality of user features; Compression processing is performed on the at least two user split vectors corresponding to each user feature to obtain the compressed vector corresponding to each user feature. The compression vectors corresponding to each user feature are used for excitation processing to obtain the weights corresponding to each user splitting vector; wherein, the weight value of the user splitting vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the user splitting vector corresponding to the ineffective long-tail low-frequency feature. Based on the weights corresponding to each user segmentation vector, the user segmentation vectors are weighted to obtain the target user vector. The method further includes: Periodically acquire materials to be recalled in batches and update the materials to be recalled in the material library at specified time intervals.

2. The method according to claim 1, characterized in that, The compression process based on the at least two user segmentation vectors corresponding to each user feature, to obtain the compressed vector corresponding to each user feature's feature vector, includes: Feature values ​​are extracted from the at least two user splitting vectors corresponding to the target user feature to obtain at least two feature value sets that correspond one-to-one with the at least two user splitting vectors, and the feature value sets contain at least two feature values; the target user feature is any one of the plurality of user features. The feature values ​​in the at least two feature value sets corresponding to the target user feature are concatenated to obtain the compressed vector corresponding to the target user feature.

3. The method according to claim 2, characterized in that, The feature set includes a first feature value and a second feature value, wherein the first feature value is the maximum value in the user split vector and the second feature value is the average value in the user split vector.

4. The method according to claim 1, characterized in that, The activation process based on the compression vector corresponding to each user feature to obtain the weights corresponding to each user segmentation vector includes: The compressed vectors corresponding to the multiple user features are concatenated to obtain a first concatenated vector; The first concatenated vector is processed by a first fully connected neural network with an activation function to obtain a first processing result. The first processing result is fully processed by a second fully connected neural network without an activation function to obtain the weights corresponding to each user split vector output by each neuron of the second fully connected neural network; wherein the number of neurons in the second fully connected neural network is the same as the number of user split vectors.

5. The method according to claim 1, characterized in that, The step of weighting each user segmentation vector based on its respective weight to obtain the target user vector includes: The product operation is performed on each of the user split vectors and their corresponding weights to obtain the product results; The summation process is performed on each of the product results and the corresponding user split vector to obtain each summation result; The vector formed by the summation results is taken as the target user vector of the target user.

6. The method according to claim 1, characterized in that, The dual-tower recall model is obtained based on training samples; the training samples include sample user characteristics, positive sample material characteristics, and negative sample material characteristics; the positive sample material includes sample material that has received positive feedback from the sample user, and the negative sample material includes sample material that has not received feedback from the sample user or sample material that has received negative feedback from the sample user.

7. A material recall device, characterized in that, The device includes: The feature acquisition module is used to acquire multiple user features of the target user when a user request from the target user is received; the multiple user features include effective long-tail low-frequency features and invalid long-tail low-frequency features, the effective long-tail low-frequency features refer to long-tail low-frequency features that are strongly correlated with the user's interests, and the invalid long-tail low-frequency features refer to long-tail low-frequency features that are weakly correlated with the user's interests. The vector acquisition module is used to input the multiple user features into the dual-tower recall model for feature extraction and feature weighting to obtain the target user vector of the target user; the target user vector contains an enhanced first feature vector, which is the feature vector corresponding to the effective long-tail low-frequency feature of the target user. The recall module is used to determine and recall materials from the materials to be recalled based on the similarity between the target material vector and the target user vector of each material to be recalled. The target material vector of the material to be recalled is obtained by the dual-tower recall model after performing feature extraction and feature weighting on multiple material features of the input material to be recalled. The multiple material features include effective long-tail low-frequency features. The target material vector contains an enhanced second feature vector, which is the feature vector corresponding to the effective long-tail low-frequency features of the material to be recalled. The dual-tower recall model includes a first feature extraction layer and a first compressed excitation network SENet+ layer; The vector acquisition module includes a feature extraction submodule, which is used to perform feature extraction processing on the input multiple user features through the first feature extraction layer of the dual-tower recall model to obtain a user feature vector; the user feature vector contains multiple feature vectors that correspond one-to-one with the multiple user features; The feature weighting submodule includes: The vector splitting unit is used to split the plurality of feature vectors in the user feature vectors respectively to obtain at least two user splitting vectors corresponding to each of the plurality of user features; The compression processing unit is used to perform compression processing based on the at least two user split vectors corresponding to each user feature, so as to obtain the compressed vector corresponding to each user feature. The excitation processing unit is used to perform excitation processing based on the compression vector corresponding to each user feature to obtain the weight corresponding to each user splitting vector; wherein, the weight value of the user splitting vector corresponding to the effective long-tail low-frequency feature is higher than the weight value of the user splitting vector corresponding to the invalid long-tail low-frequency feature. The weighted processing unit is used to perform weighted processing on each user split vector based on the weights corresponding to each user split vector to obtain the target user vector. The feature acquisition module is also used to periodically acquire materials to be recalled in batches and update the materials to be recalled in the material library every specified time period.

8. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, which is loaded and executed by the processor to implement the material recall method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the material recall method as described in any one of claims 1 to 6.

Images