Recall model training method, object recall method, content recommendation method and content recommendation system

Abstract

The embodiment of the invention provides a recall model training method, an object recall method, a content recommendation method and a content recommendation system. The recall model training method comprises the steps of obtaining object features of sample objects and a plurality of historical object sequences of multiple types of interaction behaviors performed by sample users; obtaining a plurality of sample interaction behavior vectors through an initial recall model at least based on the plurality of historical object sequences; based on the object features, obtaining a sample object vector through an initial recall model; determining a plurality of interaction behavior loss values based on vector similarities between the plurality of sample interaction behavior vectors and the sample object vectors; and based on the plurality of interactive behavior loss values, training the initial recall model to obtain a target recall model. The multi-objective optimization accuracy of the recall model is improved, the model performance of the recall model is improved, and the accuracy and diversity of follow-up object recall are improved.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of machine learning, and in particular to a method for training a recall model, object recall, content recommendation, and a content recommendation system. Background Art

[0002] With the development of machine learning technology, object recall based on a recall model has gradually replaced object recall based on simple rules in multiple business fields and has been more widely applied. For example, in an advertising recommendation system, based on a recall model, the most likely ads that can arouse users' interest are screened out from a large number of ads for display, which is more accurate than traditional rule-based advertising recommendations (such as rotation by time, targeting by region, etc.), and can significantly improve the click-through rate and conversion rate of ads. Another example is that in a search engine, based on a recall model, web pages relevant to a user's query can be quickly found from a huge web page library, which is more intelligent than traditional keyword-matching search engines and can provide more relevant and personalized search results. Still another example is that in a note sharing platform, based on a recall model, the notes most likely to meet users' needs are screened out from a large number of notes for recommendation, which is more personalized than traditional click-through rate-based recommendation methods and can help users find valuable content faster, improving user satisfaction and platform activity.

[0003] Currently, the recall model relies on the vector similarity between the interaction behavior vector and the object vector to achieve more accurate object recall. Correspondingly, by designing a loss value based on the vector similarity between the interaction behavior vector and the sample object vector to train the recall model, the model performance of the recall model can be effectively improved, and thus the accuracy of object recall can be improved.

[0004] However, as business requirements become increasingly complex, the recall model currently not only needs to complete object recall for a single target, but also needs to balance between multiple recall targets to complete high-accuracy multi-target recall. How to train a recall model suitable for multi-target recall to balance multiple business targets and improve the overall recommendation effect has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the embodiments of this specification provide a method for training a recall model. One or more embodiments of this specification also relate to an object recall method, a content recommendation method, a content recommendation system, a recall model training device, an object recall device, a content recommendation device, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects existing in the prior art.

[0006] One aspect of the embodiments of this specification provides a method for training a recall model, including:

[0007] Obtain the object features of the sample object and multiple historical object sequences of the sample user performing multiple types of interaction behaviors;

[0008] Based on at least the multiple historical object sequences, obtain multiple sample interaction behavior vectors via the initial recall model;

[0009] Based on the object features, obtain the sample object vector via the initial recall model;

[0010] Determine multiple interaction behavior loss values based on the vector similarity between the multiple sample interaction behavior vectors and the sample object vector;

[0011] Train the initial recall model based on the multiple interaction behavior loss values to obtain the target recall model.

[0012] In one embodiment of this specification, obtaining the object features of the sample object and multiple historical object sequences of the sample user performing multiple types of interaction behaviors provides comprehensive sample data support for subsequent model training. Based on at least the multiple historical object sequences, obtaining multiple sample interaction behavior vectors via the initial recall model, and based on the object features, obtaining the sample object vector via the initial recall model ensure that the initial recall model can make full use of multi-dimensional feature information. Determining multiple interaction behavior loss values based on the vector similarity between the multiple sample interaction behavior vectors and the sample object vector, and training the initial recall model based on the multiple interaction behavior loss values to obtain a target recall model that can complete multi-target recall. At the same time, it fully explores the correlation between a single interaction behavior and the object, realizes the optimization of a single recall target, improves the accuracy of the recall model for multi-target optimization, enhances the model performance of the recall model, and improves the accuracy and diversity of subsequent object recall. Description of the Drawings

[0013] Figure 1 is a structural schematic diagram of a recall model;

[0014] Figure 2 is a structural schematic diagram of the recall model in a recall model training method provided by one embodiment of this specification;

[0015] Figure 3 is a flowchart of a recall model training method provided by one embodiment of this specification;

[0016] Figure 4 is a process architecture diagram of a recall model training method provided by one embodiment of this specification;

[0017] Figure 5 is a flowchart of an object recall method provided by one embodiment of this specification;

[0018] Figure 6 It is a flowchart of an object recall method provided by an embodiment of this specification;

[0019] Figure 7 It is a flowchart of a content recommendation method provided by an embodiment of this specification;

[0020] Figure 8 It is a front-end schematic diagram of a content recommendation method provided by an embodiment of this specification;

[0021] Figure 9 It is a schematic structural diagram of a content recommendation system provided by an embodiment of this specification;

[0022] Figure 10 It is a schematic structural diagram of a recall model training device provided by an embodiment of this specification;

[0023] Figure 11 It is a schematic structural diagram of an object recall device provided by an embodiment of this specification;

[0024] Figure 12 It is a schematic structural diagram of a content recommendation device provided by an embodiment of this specification;

[0025] Figure 13 It is a structural block diagram of a computing device provided by an embodiment of this specification. Detailed implementation manners

[0026] Many specific details are set forth in the following description in order to provide a thorough understanding of this specification. However, this specification can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of this specification. Therefore, this specification is not limited by the specific implementations disclosed below.

[0027] The terms used in one or more embodiments of the present invention are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of the present invention. The singular forms "a", "the", and "said" used in one or more embodiments of the present invention are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of the present invention refers to and includes any or all possible combinations of one or more of the associated listed items.

[0028] It should be understood that although the terms first, second, etc. may be used in one or more embodiments of the present invention to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of one or more embodiments of the present invention, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to a determination".

[0029] In addition, it should be noted that the data involved in one or more embodiments of the present invention are all information and data that have been authorized by the user or fully authorized by all parties, and the statistics, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entrances are provided for the user to select authorization or rejection.

[0030] First, the noun terms involved in one or more embodiments of this specification are explained.

[0031] Recall: The entire link system of the recommendation system can generally be classified into four major processes: recall -> rough ranking -> fine ranking -> re-ranking. The purpose of recall is to select a set of objects that the user may be interested in from tens of millions or even hundreds of millions of object candidates. To ensure that users obtain rich and relevant recommendations, the recall stage needs to have a high recall rate, which directly affects the overall performance and upper limit of the recommendation system.

[0032] Rough ranking: After obtaining a large number of candidate objects in the recall stage, the task of the rough ranking stage is to perform a preliminary ranking on these candidate objects to reduce the scale of the candidate set. This stage usually uses relatively simple models (such as linear models, decision trees, etc.) to filter out obviously irrelevant or low-quality candidate objects while ensuring efficiency. The purpose of rough ranking is to quickly screen out a batch of relatively high-quality candidate objects to prepare for the subsequent fine ranking stage.

[0033] Fine ranking: Based on the rough ranking, a more refined ranking is performed on the selected high-quality candidate objects. This stage usually uses more complex models (such as deep neural networks, GBDT, etc.) to more accurately predict the degree of interest of the user in each candidate object. The purpose of the fine ranking stage is to generate a more accurate recommendation list to ensure the high relevance and personalization of the recommendation results.

[0034] Re-ranking: Based on the recommended list generated in the fine-tuning stage, perform final adjustments and optimizations. This stage usually takes into account more business rules and user experience factors, such as diversity, novelty, user satisfaction, etc. The purpose of re-ranking is to ensure that the final recommended results are not only highly relevant but also provide a good user experience. Common re-ranking strategies include, but are not limited to: diversity optimization, fresh content recommendation, user interest balance, etc.

[0035] Approximate Nearest Neighbor (ANN): Approximate Nearest Neighbor is a method for finding the data points closest to a given query point in a large-scale dataset. However, unlike exact nearest neighbor search, it allows the returned results to have a certain error in exchange for higher efficiency.

[0036] K-Nearest Neighbors (KNN): K-Nearest Neighbors is an instance-based learning method used for classification and regression analysis. In classification tasks, the output is based on the majority vote of the nearest neighbors; in regression tasks, it may be the average of the nearest neighbors. K refers to the number of nearest neighbors to be considered in the search.

[0037] Multilayer Perceptron (MLP): A multilayer perceptron is a feedforward deep neural network consisting of at least three layers of neurons: an input layer, one or more hidden layers, and an output layer. The connections between the layers are fully connected, i.e., each neuron in one layer is connected to all neurons in the next layer.

[0038] Deep Neural Networks (DNN): A machine learning model characterized by containing multiple hidden layers, usually no less than two. These hidden layers enable the network to learn and represent complex features and patterns in the data. Through layer-by-layer abstraction, deep neural networks can extract high-level feature vectors from the original input features.

[0039] Latent Variable Estimation (LVE): Latent variable estimation refers to the process of inferring unobservable variables (latent variables) from observed data. These variables usually represent abstract features or factors behind the data and have wide applications in statistical modeling, machine learning, and other fields.

[0040] MultiQuery: A retrieval and recall technique in database operations that allows simultaneous execution of retrievals and recalls for multiple objectives to meet richer business requirements.

[0041] Attention Mechanism: A feature aggregation method that mimics the human visual or auditory attention selection process. By assigning corresponding weights to different parts of the input, the model can focus more on relevant information, thereby improving performance. The calculation formula of the attention mechanism is shown in Formula 1:

[0042]

[0043] Among them, Q is the query vector, K is the key vector, and V is the value vector. d k is the scaling factor, which is used to prevent the problem of gradient vanishing or explosion caused by too large dot product results. softmax() is the normalization function, and the specific formula is shown in Formula 2:

[0044]

[0045] Among them, the input of the softmax() function is a K-dimensional vector z i =[z1, z2, ……, z K , and the output of the softmax() function is a K-dimensional vector p i =[p1, p2, ……, p K , and e is the base of the natural logarithm (approximately equal to 2.71828).

[0046] Cross Attention: It mainly refers to a target vector calculating attention scores for multiple source vectors respectively, and then weighted summing the multiple source vectors according to the attention scores to integrate them into a single vector. For example, the attention process in the multi-target recall model in one or more embodiments of this specification belongs to the cross-attention mechanism: using the Basic User Vector as the target vector, calculating attention scores for the pagetime / engage / videoend three vectors respectively, and then weighting the three vectors into a single vector according to the attention scores.

[0047] Auxiliary Loss: Auxiliary loss refers to one or more additional loss terms introduced outside the main task, aiming to improve certain specific aspects in the model training process, such as improving the generalization ability of the model, accelerating convergence, etc. The auxiliary loss is usually optimized together with the main task loss to form a multi-task learning framework.

[0048] Information Noise Contrastive Estimation (infoNCE for short): infoNCE is a loss function for unsupervised learning, aiming to maximize the mutual information between positive sample pairs while minimizing the mutual information between negative sample pairs. This method trains the model by comparing positive samples with a large number of negative samples, thus effectively learning the representation of the data.

[0049] Dense parameters: In a machine learning model, these are the parameters that usually exist in continuous value form and are frequently updated in each training step. These parameters typically exist in network structures such as fully connected layers and convolutional layers, and are used to capture complex relationships and patterns in the data.

[0050] Sparse parameters: In a machine learning model, these are the parameters that usually exist in discrete value form and are not necessarily updated in each training step. These parameters typically exist in structures such as embedding layers and sparse matrices, and are used to represent high-dimensional sparse data, such as user identifiers and object identifiers.

[0051] Currently, multi-object recall models usually construct vectors for multiple objectives separately during the training phase, weight them into a single vector to calculate the vector similarity, and determine the loss value to complete the corresponding training. During the recall phase, the multi-object vectors are also weighted into a single vector for recall. Figure 1 Shows a schematic diagram of the structure of a recall model:

[0052] During the training phase:

[0053] Via the user extraction layer, extract the user vector corresponding to the user features. Via the object extraction layer, extract the object vector corresponding to the object features.

[0054] Via multiple types of behavior extraction layers, extract the interaction behavior vectors corresponding to multiple historical object sequences:

[0055] Via the behavior extraction layer of type_1, extract the interaction behavior vector corresponding to the historical object sequence of type_1; via the behavior extraction layer of type_2, extract the interaction behavior vector corresponding to the historical object sequence of type_2; via the behavior extraction layer of type_3, extract the interaction behavior vector corresponding to the historical object sequence of type_3.

[0056] Weight the interaction behavior vectors of multiple types to obtain the weighted interaction behavior vector.

[0057] Based on the vector similarity between the object vector and the weighted interaction behavior vector, determine the behavior loss value. Based on the vector similarity between the user vector and the object vector, determine the auxiliary selection loss value.

[0058] Train the initial recall model based on the behavior loss value and the auxiliary selection loss value.

[0059] However, this method of simply weighting multiple vectors for recall cannot fully release the value of different objectives and has the following obvious limitations: 1. Information loss: Simply weighting multiple types of interaction behavior vectors into a single vector may weaken or lose certain types of information, affecting the recall effect of the model. 2. Limited generalization ability: The single-vector recall method has weak generalization ability when dealing with sparse data, easily leading to overfitting or underfitting of the model.

[0060] In view of the above limitations, a specific embodiment of this specification has made targeted optimizations. In the training stage, independent target towers (processing paths for interaction behavior vectors) corresponding to three types are constructed respectively. Multiple types of interaction behavior vectors are extracted from the three target towers, and on this basis, an auxiliary tower (processing path for object vectors) is constructed to supplement the weakened information and enhance the generalization ability of sparse features, further improving the model performance of the recall model. Figure 2 The following shows a schematic structural diagram of a recall model in a recall model training method provided by an embodiment of this specification:

[0061] In the training stage:

[0062] Via the user extraction layer, extract the user vector corresponding to the user feature. Via the object extraction layer, extract the object vector corresponding to the object feature.

[0063] Via multiple types of behavior extraction layers, extract the interaction behavior vectors corresponding to multiple historical object sequences:

[0064] Via the behavior extraction layer of type_1, extract the interaction behavior vector corresponding to the historical object sequence of type_1; via the behavior extraction layer of type_2, extract the interaction behavior vector corresponding to the historical object sequence of type_2; via the behavior extraction layer of type_3, extract the interaction behavior vector corresponding to the historical object sequence of type_3.

[0065] Based on the vector similarity between the object vector and the interaction behavior vector of type_1, determine the behavior loss value of type_1. Based on the vector similarity between the object vector and the interaction behavior vector of type_2, determine the behavior loss value of type_2. Based on the vector similarity between the object vector and the interaction behavior vector of type_3, determine the behavior loss value of type_3. Based on the vector similarity between the user vector and the object vector, determine the auxiliary selection loss value.

[0066] Train the initial recall model based on the behavior loss values of multiple types and the auxiliary selection loss value.

[0067] The above-mentioned trained recall model can directly use multiple interaction behavior vectors for multi-vector recall in the recall stage. While meeting the requirements of multi-object recall, it fully unleashes the value of single-object recall and overcomes the limitations of the above-mentioned information loss and limited generalization ability.

[0068] In this specification, a method for training a recall model is provided. This specification also relates to an object recall method, a content recommendation method, a content recommendation system, a recall model training device, an object recall device, a content recommendation device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail one by one in the following embodiments.

[0069] See Figure 3 , Figure 3 FIG. shows a flowchart of a method for training a recall model provided by an embodiment of this specification, including the following specific steps:

[0070] Step 302: Obtain the object features of the sample object and multiple historical object sequences of the sample user's multiple types of interaction behaviors.

[0071] The embodiments of this specification are applied to applications, platforms, or systems with model training capabilities. For example, a virtual task platform with model training capabilities, or an open-source application for large model training.

[0072] The object is a specific object entity to be recalled, with rich features, and the user's interest level in the object can be judged through these features. For example, advertisements, videos, news, notes, commodities, etc. Object features are feature data describing at least one aspect attribute of the object and are the basis for constructing object vectors. Through these features, the user's interest level can be more accurately matched. For example, for advertisements, object features include: advertisement identifier, merchant identifier, advertisement position, advertisement size, advertisement type, advertisement taxonomy, etc. For another example, for videos, object features include: video identifier, author identifier, video length, video resolution, video tags, video taxonomy, etc. For still another example, for news, object features include: news identifier, author identifier, news release date, news keywords, news source, news taxonomy, etc. For still another example, for notes, object features include: note identifier, author identifier, note release time, note title, note content, note tags, note taxonomy, etc. For still another example, for commodities, object features include: commodity identifier, merchant identifier, commodity price, commodity inventory, commodity details, commodity taxonomy, etc.

[0073] The sample object is a specific object entity that serves as the sample data for training the initial recall model. It has rich features, and based on these features, the interest level of the sample user in the sample object can be determined. The object features of the sample object are the feature data that describe the attributes of the sample object in multiple aspects and are the basis for constructing the sample object vector. Through these features, the interest level of the sample user can be more accurately matched.

[0074] A user is a person who needs to recall a specific object entity. The user can select an object or perform interaction behaviors after selecting an object. The user has rich features, and based on these features, the interest level of the user in the object can be determined. For example, advertising viewers, video viewers, news readers, note users, commodity users, etc. A sample user is a user who serves as the sample data for training the initial recall model. The sample user has rich features, and based on these features, the interest level of the sample user in the object can be determined.

[0075] Multiple types refer to various types of interaction behaviors that occur between a user and an object. Different types of interaction behaviors reflect different interests and preferences of the user in the object. By analyzing these different types of behaviors, the needs of the user can be more comprehensively understood. Multiple types respectively correspond to multiple recall targets.

[0076] The historical object sequence is a sequence that includes multiple historical objects. Multiple types of interaction behaviors have occurred between the sample user and these historical objects. It is the basis for constructing the interaction behavior vector. Through these sequences, the interest of the sample user in the object can be characterized. The historical object sequence includes but is not limited to: the historical object sequence of viewing behaviors that meet the viewing duration, the historical object sequence of interaction behaviors (including behaviors such as clicking, liking, commenting, sharing, etc.), the historical object sequence of complete playback behaviors (playback completion behaviors), and the historical object sequence of search behaviors. It can be understood that the sample user has performed interaction behaviors on the historical objects in the historical object sequence of interaction behaviors, and the sample user has performed complete playback behaviors on the historical objects in the historical object sequence of complete playback behaviors, and so on. In an optional embodiment of this specification, multiple types of interaction behaviors include viewing behaviors that meet the viewing duration, interaction behaviors, and complete playback behaviors. The historical object sequence of viewing behaviors that meet the viewing duration is the historical object sequence in which the sample user has had viewing behaviors that meet the viewing duration with multiple historical objects. The historical object sequence of interaction behaviors is the historical object sequence in which the sample user has had interaction behaviors with multiple historical objects. The historical object sequence of complete playback behaviors is the historical object sequence in which the sample user has had complete playback behaviors with multiple historical objects.

[0077] For example, in a video recommendation scenario, the historical video sequence of user A's viewing behavior that meets the viewing duration is [Video X: 10 minutes, Video Y: 5 minutes, Video Z: 8 minutes], indicating that user A has the longest viewing duration for Video X and may have a greater preference for this type of content. In a note recommendation scenario, the historical note sequence of user B's viewing behavior that meets the viewing duration is [Note A: 3 minutes, Note B: 6 minutes, Note C: 2 minutes], reflecting a relatively high interest in Note B.

[0078] For example, in a social platform scenario, the historical video sequence of user C's like behavior is [Video M: liked, Note N: liked, Product P: liked], the historical video sequence of the comment behavior is [Video M: commented "wonderful content", Note N: commented "worth collecting"], and the historical note sequence of the sharing behavior is [Note N: shared to the circle of friends]. These behaviors indicate that user C has a relatively high level of engagement with Video M and Note N. In an e-commerce scenario, the historical product sequence of user D's purchase behavior is [Product A: purchased 2 times, Product B: purchased 1 time], reflecting a strong preference for Product A.

[0079] For example, in a short video scenario, the historical video sequence of user E's complete play behavior is [Video X: completely played, Video Y: completely played, Video Z: only played the first 10 seconds], indicating that user E is more inclined to watch Videos X and Y completely. In an online course scenario, the historical course sequence of user F's complete play behavior is [Course A: completed all chapters, Course B: only learned the first chapter], reflecting user F's in-depth learning behavior for Course A.

[0080] Optionally, by statistically analyzing the viewing duration of the sequence (such as weighted average, sliding window), the stability of user interests can be captured. For example, the historical video sequence of user G's viewing behavior that meets the viewing duration recently is [Video A: 8 minutes (3 days ago), Video B: 12 minutes (1 day ago), Video C: 15 minutes (today)]. Combining with the time decay factor, it can be found that user G's interest is gradually shifting towards long videos.

[0081] Optionally, different interaction types (such as like, comment, share) can be assigned different weights. For example, the historical note sequence of user H's sharing behavior [Note P: shared to the community, Note Q: shared in private chat] can reflect user H's willingness to actively recommend more than a simple like behavior.

[0082] Optionally, the complete play behavior is strongly correlated with the content quality. For example, user I watched a video multiple times after completely playing it. The historical video sequence of the complete play behavior is [Video X: completely played 3 times, Video Y: completely played 1 time], which may indicate that user I recognizes the high quality of Video X or has repeated learning / entertainment needs.

[0083] In the embodiments of this specification, by obtaining the viewing duration behavior vector, interaction behavior vector, and complete viewing behavior vector, the behavior patterns of sample users for historical objects in terms of the three recall objectives of viewing behavior, interaction behavior, and complete viewing behavior that meet the viewing duration are recorded, which helps to more accurately capture the interest evolution trend of sample users. By analyzing these historical object sequences, during the training process of the initial recall model, it can better learn to capture user interests, more fully explore the correlation between a single interaction behavior vector and object features, and achieve the optimization of a single recall objective.

[0084] To obtain the object features of a sample object, one optional method is to obtain the object features of the sample object from an object database. Another optional method is to obtain the object features of the sample object from a log file. Another optional method is to obtain the object features of the sample object from a data call interface. There is no limitation here.

[0085] To obtain multiple historical object sequences of a sample user performing multiple types of interaction behaviors, one optional method is to obtain multiple historical object sequences of the sample user performing multiple types of interaction behaviors from a behavior database. Another optional method is to obtain multiple historical object sequences of the sample user performing multiple types of interaction behaviors from a log file. Another optional method is to obtain multiple historical object sequences of the sample user performing multiple types of interaction behaviors from a data call interface. There is no limitation here.

[0086] Exemplarily, on a virtual task platform with model training capabilities, the developer of a certain note sharing application needs to train the current version of the note recall model to complete the model optimization for the three recall objectives of viewing behavior, interaction behavior, and complete viewing behavior that meet the viewing duration. Obtain 2,000,000 samples from the log database (each sample contains a note, a user, and label information; the label information records the interaction situation between the user and the note, such as whether they click, whether they interact, viewing duration, etc.). There may be overlaps in the notes / users of different samples. For example, among the 2,000,000 samples, there are only 1,000 non-repeating users and 20,000 non-repeating notes. Obtain the features of 20,000 notes from the note database. Obtain the historical object sequences of 1,000 users from the behavior database: the historical object sequence of viewing behavior that meets the viewing duration Pagetime LastN, the historical object sequence of interaction behavior Engage LastN, and the historical object sequence of complete viewing behavior Videoend LastN.

[0087] In step 302, obtaining the object features of the sample object and multiple historical object sequences of the sample user performing multiple types of interaction behaviors provides comprehensive sample data support for subsequent model training.

[0088] Step 304: Obtain multiple sample interaction behavior vectors via an initial recall model based at least on multiple historical object sequences.

[0089] A recall model is a machine learning model used to screen out a small number of objects that a user may be interested in from a large number of candidate objects. Recall models are commonly used in object recommendation systems (such as advertising recommendation systems, video recommendation systems, news recommendation systems, note recommendation systems, product recommendation systems, etc.), search engines, etc. By initially filtering a large number of candidate objects, the data processing burden of subsequent sorting (coarse sorting, fine sorting, re - sorting, etc.) and recommendation is reduced. Recall models usually rely on high - accuracy vector extraction capabilities and are based on the vector similarity between high - accuracy interaction behavior vectors and object vectors to achieve high - accuracy object recall. Recall models include, but are not limited to: multi - layer perceptrons, deep neural networks, etc. An initial recall model is a recall model that has not undergone the training described later.

[0090] Multiple sample interaction behavior vectors are vectors obtained by performing deep feature encoding on multiple historical object sequences of multiple types of interaction behaviors. Multiple sample interaction behavior vectors are abstract high - dimensional feature encoding vectors obtained after fully mining the historical object sequences.

[0091] One optional way to obtain multiple sample interaction behavior vectors via an initial recall model based at least on multiple historical object sequences is: perform deep feature encoding via the initial recall model based at least on multiple historical object sequences to obtain multiple sample interaction behavior vectors.

[0092] Step 306: Obtain a sample object vector via the initial recall model based on object features.

[0093] A sample object vector is a vector obtained by performing deep feature encoding on the object features of a sample object. A sample object vector is an abstract high - dimensional feature encoding vector obtained after fully mining the object features of the sample object.

[0094] One optional way to obtain a sample object vector via the initial recall model based on object features is: perform deep feature encoding via the initial recall model based on object features to obtain a sample object vector.

[0095] Exemplarily, obtain the note recall model of the current version. Based on the note features NoteFeature (note identifier, author identifier, note release time, note title, note content, note tags, note taxonomy, etc.), the historical object sequence Pagetime LastN of viewing behaviors that meet the viewing duration, the historical object sequence Engage LastN of interaction behaviors, and the historical object sequence Videoend LastN of complete viewing behaviors in batches (Batch), perform deep feature encoding through the note recall model to obtain the sample note vector Item Vector and three types of sample interaction behavior vectors: the sample viewing duration vector Pagetime UserVector, the sample interaction behavior vector Engage User Vector, and the sample complete viewing behavior vector Videoend User Vector.

[0096] In steps 304 and 306, at least based on multiple historical object sequences, multiple sample interaction behavior vectors are obtained through the initial recall model, and the sample object vector is obtained through the initial recall model based on the object features, ensuring that the initial recall model can make full use of multi-dimensional feature information.

[0097] Step 308: Determine multiple interaction behavior loss values based on the vector similarity between multiple sample interaction behavior vectors and the sample object vector.

[0098] The vector similarity between multiple sample interaction behavior vectors and the sample object vector is used to measure the similarity degree between multiple sample interaction behavior vectors and the sample object vector. This vector similarity reflects the interest degree of the sample user in the sample object and the behavior pattern after selection. The vector similarity includes but is not limited to: cosine similarity, Manhattan distance, etc.

[0099] The interaction behavior loss value is the model loss value calculated based on the vector similarity between multiple sample interaction behavior vectors and the sample object vector during the training of the initial recall model. The interaction behavior loss value is used to evaluate the error of the initial recall model in predicting the interest degree of the user in the object and the behavior pattern after selection. Based on this loss value, the initial recall model is trained to improve the multi-objective recall ability of the initial recall model. The interaction behavior loss value includes but is not limited to: mean square error loss value, cross-entropy loss value, information noise contrast estimation.

[0100] Exemplarily, calculate the cosine similarity CosSim_1 between the sample note vector Item Vector and the sample viewing duration vector Pagetime UserVector. Calculate the cosine similarity CosSim_2 between the sample note vector Item Vector and the sample interaction behavior vector Engage User Vector. Calculate the cosine similarity CosSim_3 between the sample note vector Item Vector and the sample video completion behavior vector Videoend User Vector. Use information noise contrast estimation to calculate the cosine similarity CosSim_1 to obtain the interaction behavior loss value PagetimeLoss of the viewing behavior type that meets the viewing duration. Use information noise contrast estimation to calculate the cosine similarity CosSim_2 to obtain the interaction behavior loss value Engage Loss of the interaction behavior type. Use information noise contrast estimation to calculate the cosine similarity CosSim_3 to obtain the interaction behavior loss value Videoend Loss of the video completion behavior type.

[0101] In step 308, based on the vector similarity between multiple sample interaction behavior vectors and the sample object vector, multiple interaction behavior loss values are determined, providing data support for the training loss value for subsequent training of the initial recall model.

[0102] Step 310: Based on multiple interaction behavior loss values, train the initial recall model to obtain the target recall model.

[0103] The target recall model is the recall model that has completed training, and the target recall model is the recall model that can accurately complete multi-target recall. Compared with the initial recall model, the target recall model fully explores the correlation between a single interaction behavior and object features on the basis of multi-target recall, achieving the optimization of a single recall target. The target recall model can more accurately capture the trend of users' interest evolution and provide more personalized objects for users. The target recall model can learn the commonalities and differences between different types of historical object sequences, improving the learning efficiency and model capabilities of each type.

[0104] Based on multiple interaction behavior loss values, train an initial recall model to obtain a target recall model. An optional method is as follows: weight the multiple interaction behavior loss values to obtain a target loss value, and based on the target loss value, adjust the model parameters of the initial recall model to obtain the target recall model. Further, based on the target loss value, adjust the model parameters of the initial recall model. An optional adjustment method is as follows: based on the target loss value, adjust the model parameters of the initial recall model by the gradient update method. Further, based on the target loss value, adjust the model parameters of the initial recall model to obtain the target recall model. An optional training method is as follows: based on the target loss value, iteratively adjust the model parameters of the initial recall model. When the preset training end condition is reached, obtain the trained target recall model, where the preset training end condition includes but is not limited to: preset number of iteration rounds, preset loss value threshold, the change amplitude of the model parameters is less than the preset change amplitude threshold, the performance metrics on the validation set no longer improve, etc.

[0105] Exemplarily, weight three types of interaction behavior loss values (interaction behavior loss value Pagetime Loss, interaction behavior loss value Engage Loss, interaction behavior loss value Videoend Loss) to obtain the target loss value Loss. Based on the target loss value Loss, adjust the model parameters of the note recall model by the gradient update method. When the preset number of iteration rounds is reached, obtain the trained note recall model.

[0106] In the embodiments of this specification, obtain the object features of the sample object and multiple historical object sequences of the sample user performing multiple types of interaction behaviors, providing comprehensive sample data support for subsequent model training. At least based on the multiple historical object sequences, obtain multiple sample interaction behavior vectors through the initial recall model, and based on the object features, obtain the sample object vector through the initial recall model, ensuring that the initial recall model can make full use of multi-dimensional feature information. Based on the vector similarity between the multiple sample interaction behavior vectors and the sample object vector, determine multiple interaction behavior loss values, and based on the multiple interaction behavior loss values, train the initial recall model to obtain a target recall model that can complete multi-target recall. At the same time, fully explore the correlation between a single interaction behavior and the object, realize the optimization of a single recall target, improve the accuracy of the recall model for multi-target optimization, enhance the model performance of the recall model, and improve the accuracy and diversity of subsequent object recall.

[0107] In the training phase, the initial recall model respectively utilizes the processing paths of multiple types of interaction behavior vectors to correspondingly construct multiple independent target towers, completing the modeling process of multiple types of interaction behavior patterns. On this basis, an auxiliary tower can also be constructed, introducing user vectors to enhance multiple types of interaction behavior vectors, so as to enhance the generalization ability for sparse interaction behavior vectors and further improve the model performance of the target recall model obtained through training.

[0108] In an optional embodiment of this specification, before step 308, the following specific steps are further included: obtaining the user features of a sample user; based on the user features, obtaining the sample user vector through the initial recall model; based on the sample user vector and multiple sample interaction behavior vectors, obtaining the sample auxiliary selection vector of the sample user; based on the vector similarity between the sample auxiliary selection vector and the sample object vector, as well as the predetermined interaction relationship between the sample user and the sample object, determining the auxiliary selection loss value;

[0109] Correspondingly, step 308 includes the following specific steps: training the initial recall model based on the auxiliary selection loss value and multiple interaction behavior loss values to obtain the target recall model.

[0110] The user features of a user are feature data describing at least one aspect attribute of the user, which is the basis for constructing the user vector and helps to improve the personalization degree of object recall. User features include but are not limited to: user identification, age, gender, geographical location, hobbies, occupation, educational background, etc. The user features of a sample user are feature data describing at least one aspect attribute of the sample user, which is the basis for constructing the sample user vector. The sample user vector is a vector obtained by deeply encoding the user features of the sample user. After fully mining the user features of the sample user, the sample user vector is an abstract high-dimensional feature encoding vector.

[0111] The sample auxiliary selection vector is a feature encoding vector determined based on the sample user vector and multiple sample interaction behavior vectors. Since the sample user can only perform multiple types of interaction behaviors on a historical object after selection, the sample auxiliary selection vector determined based on the sample user vector and multiple sample interaction behavior vectors can more comprehensively reflect the selection tendency of the sample user for the object, especially in the case of scarce interaction data, enhancing the generalization ability of the model. The sample auxiliary selection vector not only contains the basic attribute information of the sample user, but also integrates various interaction behavior information between the sample user and the object. It is a comprehensive and abstract high-dimensional feature encoding vector, which can better capture the selection tendency of the user for the object and improve the accuracy and personalization degree of the recall model in multi-object recall.

[0112] The vector similarity between the sample auxiliary selection vector and the sample object vector is used to measure the similarity between the sample auxiliary selection vector and the sample object vector. This vector similarity reflects the selection tendency of the sample user for the sample object and is an auxiliary indicator for evaluating the matching degree between the user and the object. In the object recall stage, by calculating the similarity between the auxiliary selection vector and the object vector, in addition to the multi-objective vector similarity of the vector similarities between multiple types of interaction behavior vectors and the object vector, an auxiliary similarity can be provided to more accurately predict the user's interest and selection probability for a certain object, which helps to improve the accuracy of personalized recommendation of the recall model. Therefore, correspondingly, in the training stage, it is necessary to determine the auxiliary selection loss value based on this vector similarity to train the initial recall model. This vector similarity includes but is not limited to: cosine similarity, Manhattan distance, etc.

[0113] The predetermined interaction relationship between the sample user and the sample object is whether the user generates an explicit feedback behavior (such as click, purchase, favorite, etc.) for the object, which characterizes whether the sample user selects the sample object as the interest target. For example, in the click behavior scenario, the predetermined interaction relationship is a binary classification label (click = 1 / not click = 0), which is used to supervise the model to learn the user's click tendency; in the purchase scenario, the predetermined interaction relationship can be extended to the purchase frequency (such as marked as 1 when the number of purchases ≥ 1, otherwise 0), which is used to reflect the user's consumption decision result.

[0114] The auxiliary selection loss value is the model loss value calculated based on the vector similarity between the sample auxiliary selection vector and the sample object vector when training the initial recall model. The auxiliary selection loss value is used to evaluate the error of the initial recall model in predicting the user's selection tendency for the object. Based on this loss value and multiple interaction behavior loss values, the initial recall model is trained to assist the multiple interaction behavior loss values to improve the multi-objective recall ability of the recall model. The auxiliary selection loss value includes but is not limited to: mean square error loss value, cross-entropy loss value, information noise contrast estimation.

[0115] One optional way to obtain the user characteristics of the sample user is: obtain the user characteristics of the sample user from the user database. Another optional way is: obtain the user characteristics of the sample user from the log file. Another optional way is: obtain the user characteristics of the sample user from the data call interface, which is not limited here.

[0116] Based on the user characteristics, the sample user vector is obtained via the initial recall model. One optional way is: based on the user characteristics, the sample user vector is obtained through deep feature encoding via the initial recall model.

[0117] Based on the sample user vector and multiple sample interaction behavior vectors, a sample auxiliary selection vector of the sample user is obtained. An optional way is: perform feature aggregation on the sample user vector and multiple sample interaction behavior vectors to obtain the sample auxiliary selection vector of the sample user. Further, an optional way is: use the attention mechanism to perform feature aggregation on the sample user vector and multiple sample interaction behavior vectors to obtain the sample auxiliary selection vector of the sample user.

[0118] In an optional embodiment of this specification, based on the auxiliary selection loss value and multiple interaction behavior loss values, an initial recall model is trained to obtain a target recall model, including the following specific steps: weight the auxiliary selection loss value and multiple interaction behavior loss values to obtain a target loss value; based on the target loss value, adjust the model parameters of the initial recall model to obtain the target recall model.

[0119] The target loss value is the comprehensive model loss value obtained after weighting the auxiliary selection loss value and multiple interaction behavior loss values when training the initial recall model. The target loss value is used to evaluate the overall performance of the initial recall model in the multi-objective recall task. By minimizing the target loss value, the model parameters of the initial recall model can be optimized, and the multi-objective recall ability of the initial recall model can be improved. When multiple types of interaction behaviors include viewing behaviors that meet the viewing duration, interaction behaviors, and complete-play behaviors, a calculation formula for an optional target loss value is shown in Formula 3:

[0120] Loss=Auxiliary Click Loss+Pagetime Loss+Engage Loss+Videoend Loss Formula 3

[0121] Where, Loss is the target loss value, Auxiliary Click Loss is the auxiliary selection loss value, PagetimeLoss is the interaction behavior loss value of the viewing behavior type that meets the viewing duration, Engage Loss is the interaction behavior loss value of the interaction behavior type, and Videoend Loss is the interaction behavior loss value of the complete-play behavior type.

[0122] Where, based on the target loss value, adjust the model parameters of the initial recall model to obtain the target recall model. Please refer to the content of step 308 and will not be elaborated here.

[0123] Exemplarily, user features of 10,000 sample users, namely User Feature: user identification and user profile settings such as age, gender, geographical location, hobbies, occupation, educational background, etc., are obtained from the user database of the note sharing application. Based on the user features User Feature in batches (Batch), deep feature encoding is performed via the note recall model to obtain the sample user vector Basic UserVector. Using the attention mechanism, feature aggregation is performed on the sample user vector Basic User Vector and three types of sample interaction behavior vectors (sample viewing duration vector Pagetime User Vector, sample interaction behavior vector Engage UserVector, and sample completion behavior vector Videoend User Vector) to obtain the sample auxiliary click vector User ClickVector for the sample user with respect to the sample note. Calculate the cosine similarity CosSim_4 between the sample auxiliary click vector User ClickVector and the sample note vector Item Vector to determine the auxiliary click loss value Auxiliary Click Loss. Using the above formula 3, the auxiliary click loss value Auxiliary ClickLoss and three types of interaction behavior loss values (interaction behavior loss value Pagetime Loss, interaction behavior loss value Engage Loss, interaction behavior loss value Videoend Loss) are weighted to obtain the target loss value Loss. Based on the target loss value Loss, the model parameters of the note recall model are adjusted by the gradient update method. When the preset number of iteration rounds is reached, the trained target recall model is obtained as the updated version of the note recall model.

[0124] In the embodiments of this specification, on the basis of multi-objective recall, an auxiliary vector processing path is constructed to enhance the generalization ability of the model. Based on the vector similarity between the sample auxiliary selection vector and the sample object vector, the auxiliary selection loss value is determined, and together with multiple interaction behavior loss values, the model parameters are optimized, improving the accuracy and personalization degree of the recall model in multi-objective recall. For the problem that the network convergence is difficult and the sparse generalization ability is insufficient due to the sparse problems of different degrees in a single type of historical object sequence, the generalization ability of the recall model facing the sparse historical object sequence is significantly improved.

[0125] In an alternative embodiment of this specification, based on the sample user vector and multiple sample interaction behavior vectors, obtaining the sample auxiliary selection vector of the sample user includes the following specific steps: performing target attention calculation on the sample user vector and multiple sample interaction behavior vectors to obtain the sample interest vector of the sample user; wherein, the target attention calculation can be cross-attention calculation. Based on the sample interest vector, obtaining the sample auxiliary selection vector of the sample user.

[0126] The sample interest vector is a deep encoded vector obtained by performing target attention calculation on the sample user vector and multiple sample interaction behavior vectors. The sample interest vector is an abstract high-dimensional feature encoded vector obtained after fully exploring the correlation between user features and interaction behaviors. This vector reflects the tendency of the interaction behavior pattern after the sample user selects historical objects, captures the correlation between user features and interaction behavior patterns, and fully determines whether the sample user tends to select an object based on their own user characteristics and for what purpose of interaction behavior.

[0127] One alternative way to perform target attention calculation on the sample user vector and multiple sample interaction behavior vectors to obtain the sample interest vector of the sample user is: using the sample user vector as the query vector, using multiple sample interaction behavior vectors as the key vector and value vector, performing target attention calculation to obtain the sample interest vector of the sample user. Another alternative way is: using the sample user vector as the key vector and value vector, using multiple sample interaction behavior vectors as the query vector, performing target attention calculation to obtain the sample interest vector of the sample user. This is not limited here.

[0128] One alternative way to obtain the sample auxiliary selection vector of the sample user based on the sample interest vector is: directly obtaining the sample auxiliary selection vector of the sample user based on the sample interest vector. For example, linear transformation (such as fully connected processing, etc.), or non-linear connection (such as fully connected layer plus activation, residual connection, etc.). Another alternative way is: obtaining the sample auxiliary selection vector of the sample user based on the sample user vector and the sample interest vector. This is not limited here.

[0129] Exemplarily, using the sample user vector Basic User Vector as the query vector Q, and using three types of sample interaction behavior vectors (sample viewing duration vector Pagetime UserVector, sample interaction behavior vector Engage UserVector, and sample completion behavior vector Videoend User Vector) as the key vector K and the value vector V, the above formulas 1 and 2 are used to perform target attention calculation to obtain the sample interest vector AttentionVector of the sample user for the sample note. The sample interest vector AttentionVector is processed through a fully connected layer to obtain the sample auxiliary click vector User ClickVector of the sample user for the sample note.

[0130] In the embodiments of this specification, by introducing multiple sample interaction behavior vectors using the target attention mechanism, a more comprehensive, rich, and accurate sample interest vector is obtained, clearly giving the training optimization direction for the correlation between user characteristics and interaction behavior patterns, and specifically enhancing the ability of the recall model to capture the correlation between user characteristics and interaction behavior patterns to achieve multi-target recall. At the same time, an auxiliary vector processing path is constructed to enhance the generalization ability of the model.

[0131] In an alternative embodiment of this specification, based on the sample interest vector, the sample auxiliary selection vector of the sample user is obtained, including the following specific steps: Concatenate the sample user vector and the sample interest vector to obtain the sample auxiliary selection vector of the sample user.

[0132] Exemplarily, using the sample user vector Basic User Vector as the query vector Q, and using three types of sample interaction behavior vectors (sample viewing duration vector Pagetime UserVector, sample interaction behavior vector Engage UserVector, and sample completion behavior vector Videoend User Vector) as the key vector K and the value vector V, the above formulas 1 and 2 are used to perform target attention calculation to obtain the sample interest vector AttentionVector of the sample user for the sample note. Concatenate the sample user vector Basic UserVector and the sample interest vector AttentionVector to obtain the sample auxiliary click vector User ClickVector of the sample user for the sample note.

[0133] In the embodiments of this specification, by concatenating the sample user vector and the sample interest vector to obtain the sample auxiliary selection vector of the sample user, the auxiliary selection loss value is determined, and the accuracy of the loss value is improved.

[0134] In an optional embodiment of this specification, the initial recall model includes a user extraction layer, an object extraction layer, multiple behavior extraction layers, an attention calculation layer, and a splicing layer;

[0135] Based on at least multiple historical object sequences, multiple sample interaction behavior vectors are obtained via the initial recall model, including the following specific steps: Input at least the multiple historical object sequences into the multiple behavior extraction layers to obtain multiple sample interaction behavior vectors;

[0136] Based on object features, a sample object vector is obtained via the initial recall model, including the following specific steps: Input the object features into the object extraction layer to obtain the sample object vector;

[0137] Based on user features, a sample user vector is obtained via the initial recall model, including the following specific steps: Input the user features into the user extraction layer to obtain the sample user vector;

[0138] Perform target attention calculation on the sample user vector and the multiple sample interaction behavior vectors to obtain the sample interest vector of the sample user, including the following specific steps: Input the sample user vector and the multiple sample interaction behavior vectors into the attention calculation layer for target attention calculation to obtain the sample interest vector of the sample user;

[0139] Splice the sample user vector and the sample interest vector to obtain the sample auxiliary selection vector of the sample user, including the following specific steps: Input the sample user vector and the sample interest vector into the splicing layer to obtain the sample auxiliary selection vector of the sample user.

[0140] The user extraction layer is a model layer for extracting user vectors from user features. The user extraction layer encodes and converts user features into high-dimensional abstract depth-encoded vectors through deep feature encoding. The user extraction layer includes, but is not limited to: multi-layer perceptron, deep neural network. The object extraction layer is a model layer for extracting object vectors from object features. The object extraction layer encodes and converts object features into high-dimensional abstract depth-encoded vectors through deep feature encoding. The object extraction layer includes, but is not limited to: multi-layer perceptron, deep neural network. The multiple behavior extraction layers are multiple model layers for extracting interaction behavior vectors from historical object sequences. Each behavior extraction layer specifically encodes and converts the corresponding type of historical object sequence into a high-dimensional abstract depth-encoded vector through deep feature encoding. The behavior extraction layer includes, but is not limited to: multi-layer perceptron, deep neural network. The attention calculation layer is a model layer for performing target attention calculation on user vectors and multiple types of interaction behavior vectors. The attention calculation layer captures the correlation between user features and interaction behavior patterns through an attention mechanism and generates a sample interest vector. The attention calculation layer includes, but is not limited to: attention pooling layer. The splicing layer is a model layer for splicing depth-encoded vectors. The splicing layer includes, but is not limited to: Concat layer.

[0141] For the specific implementation manners of the steps in the embodiments of this specification, reference may be made to the above-mentioned embodiments of the specification, which will not be elaborated herein.

[0142] In the embodiments of this specification, through the collaborative work of the user extraction layer, object extraction layer, behavior extraction layers corresponding to multiple types, attention calculation layer, and splicing layer, deep feature encoding of user features and various interaction behaviors and target attention calculation are achieved. The sample interest vector and sample auxiliary selection vector are spliced, and a model structure combining multiple target towers and auxiliary towers is adopted, significantly improving the generalization ability of the recall model and the accuracy of multi-target recall.

[0143] Combining the above-mentioned embodiments of each specification, taking the multi-target recall of notes as an example, Figure 4 shows the flow architecture diagram of a recall model training method provided by an embodiment of this specification:

[0144] In the training stage:

[0145] The sample note feature set includes: sample note identifier, sample author identifier... sample note taxonomy. The sample user feature set includes: sample user identifier, sample user profile.

[0146] Perform feature embedding on the sample note features to obtain sample note feature embeddings. Perform feature embedding on the sample user features to obtain sample user feature embeddings. Respectively perform feature embedding on the historical object sequences of viewing behaviors, interaction behaviors, and complete playback behaviors that meet the viewing duration, and correspondingly obtain sample viewing duration feature embeddings, sample interaction behavior feature embeddings, and sample complete playback behavior feature embeddings. In addition, perform feature embedding on the sample user identifier to obtain sample user identifier embeddings.

[0147] Input the sample note feature embeddings into a deep neural network (DNN) layer to extract and obtain sample note vectors. Input the sample user feature embeddings into the deep neural network layer to extract and obtain sample user vectors. Respectively input the sample viewing duration feature embeddings, sample interaction behavior feature embeddings, and sample complete playback behavior feature embeddings, and after splicing with the sample user identifier embeddings into the splicing layer, input them into the corresponding deep neural network layer to extract and obtain sample interaction behavior vectors of viewing behavior types that meet the viewing duration, sample interaction behavior vectors of interaction behavior types, and sample interaction behavior vectors of complete playback behavior types.

[0148] Input the sample user vectors and the sample interaction behavior vectors of 3 types into the attention calculation layer for target attention calculation. After obtaining the sample interest vector, input the sample interest vector and the sample user vector into the splicing layer and then into the deep neural network layer to obtain the sample auxiliary selection vector.

[0149] Determine the auxiliary selection loss value based on the cosine similarity between the sample note vector and the sample auxiliary click vector. Determine the interaction behavior loss value of the viewing behavior type that meets the viewing duration based on the cosine similarity between the interaction behavior vector of the viewing behavior type that meets the viewing duration and the sample note vector. Determine the interaction behavior loss value of the interaction behavior type based on the cosine similarity between the interaction behavior vector of the interaction behavior type and the sample note vector. Determine the interaction behavior loss value of the complete playback behavior type based on the cosine similarity between the interaction behavior vector of the complete playback behavior type and the sample note vector.

[0150] As Figure 4 shown, the note side models the object recall targets of user duration, interaction, and complete playback respectively using the historical object sequences of duration, interaction, and complete playback. Each single-target tower only learns the sample data corresponding to the target behavior. To address the problem that different degrees of sparsity in single-target behaviors lead to difficulties in network convergence and poor generalization of sparse embeddings, an auxiliary selection tower is constructed to enhance the generalization of low-level sparse embeddings. During the construction of the auxiliary selection tower, the attention mechanism is used to introduce the correlation information between the user and the duration, interaction, and complete playback vectors. Through the attention pooling layer, the three vectors are aggregated into a single sample interest vector, enhancing the ability of the recall model to capture the correlation between user features and interaction behavior patterns to achieve multi-target recall.

[0151] See Figure 5 , Figure 5 shows a flowchart of an object recall method provided by an embodiment of this specification, including the following specific steps:

[0152] Step 502: Obtain the object features of multiple candidate objects and at least one historical object sequence of the target user performing at least one type of interaction behavior.

[0153] The embodiments of this specification are applied to applications, platforms, or systems with object recall functions, such as advertising recommendation systems, video applications, news applications, note sharing platforms, e-commerce platforms, etc.

[0154] The object is a specific object entity to be recalled, with rich features, and the degree of user interest in the object can be judged through these features. For example, advertisements, videos, news, notes, commodities, etc. Object features are feature data describing the attributes of the object in multiple aspects and are the basis for constructing the object vector. Through these features, the degree of user interest can be more accurately matched. For example, for advertisements, object features include: advertisement identifier, merchant identifier, advertisement position, advertisement size, advertisement type, advertisement taxonomy, etc. Another example is that for videos, object features include: video identifier, author identifier, video length, video resolution, video tags, video taxonomy, etc. Another example is that for news, object features include: news identifier, author identifier, news release date, news keywords, news source, news taxonomy, etc. Another example is that for notes, object features include: note identifier, author identifier, note release time, note title, note content, note tags, note taxonomy, etc. Another example is that for commodities, object features include: commodity identifier, merchant identifier, commodity price, commodity inventory, commodity details, commodity taxonomy, etc.

[0155] The candidate object is a specific object entity that can be recalled, with rich features. These candidate objects are potential recommended objects, and through the recall process of the model, the target objects that the user may be interested in are screened out from a large number of candidate objects. The object features of the candidate object are feature data describing the attributes of the candidate object in multiple aspects and are the basis for constructing the candidate object vector. Through these features, the degree of interest of the sample user can be more accurately matched. The reference object is a specific object entity that the target user has selected and had an interaction behavior with. The reference object can be one of the candidate objects or other objects, which is not limited here. By analyzing the historical object sequence that has occurred between the target user and the reference object, the degree of interest of the target user in the candidate object can be more accurately predicted.

[0156] The user is the user who needs to recall the specific object entity. The user can select the object and can also perform interaction behaviors after selecting the object. The user has rich features, and the degree of user interest in the object can be judged through these features. For example, advertisement viewers, video viewers, news readers, note users, commodity users, etc. The target user is the user who needs to recall the target object and has rich features, and the degree of interest of the target user in multiple candidate objects can be judged through these features.

[0157] At least one type includes various interaction behavior types that occur between a user and an object. Different interaction behavior types reflect different interests and preferences of the user for the object. By analyzing these different types of behaviors, the user's needs can be understood more comprehensively. Each of the at least one type corresponds to at least one recall target. The historical object sequence of the at least one type is the characteristic data of various interaction behaviors that occur between the target user and the reference object, and is the basis for constructing the interaction behavior vector. Through these characteristics, the degree of interest of the target user in the candidate object can be predicted more accurately.

[0158] In an optional embodiment of this specification, the multiple types of interaction behaviors include the viewing behavior that meets the viewing duration, the interaction behavior, and the complete viewing behavior. The historical object sequence of the viewing behavior that meets the viewing duration is the historical object sequence of the viewing behavior that meets the viewing duration that has occurred between the sample user and multiple historical objects. The historical object sequence of the interaction behavior is the historical object sequence of the interaction behavior that has occurred between the sample user and multiple historical objects. The historical object sequence of the complete viewing behavior is the historical object sequence of the complete viewing behavior that has occurred between the sample user and multiple historical objects.

[0159] Specifically, refer to the above Figure 3 embodiment of the specification, which will not be elaborated here.

[0160] In the embodiment of this specification, by obtaining the historical object sequence of the viewing duration type, the historical object sequence of the interaction behavior type, and the historical object sequence of the complete viewing behavior type, the behavior patterns of the sample user for the historical object in the three recall targets of the viewing behavior that meets the viewing duration, the interaction behavior, and the complete viewing behavior are recorded, which helps to capture the interest evolution trend of the sample user more accurately. By analyzing these historical sequence data, in the application process of the recall model, it can better learn to capture the user's interest, more fully explore the correlation between a single historical object sequence and the object characteristics, and achieve the accurate recall of a single recall target.

[0161] To obtain the object characteristics of multiple candidate objects, one optional way is: obtain the object characteristics of multiple candidate objects from the object database. Another optional way is: obtain the object characteristics of multiple candidate objects from the log file. Another optional way is: obtain the object characteristics of multiple candidate objects from the data call interface. It is not limited here.

[0162] Obtain the historical object sequence of at least one type after the target user selects a reference object. An optional way is to obtain the historical object sequence of at least one type after the target user selects a reference object from the behavior database. Another optional way is to obtain the historical object sequence of at least one type after the target user selects a reference object from the log file. Another optional way is to obtain the historical object sequence of at least one type after the target user selects a reference object from the data call interface. This is not limited here.

[0163] Exemplarily, on a note sharing application, filter out the notes published in the last 30 days, and filter the notes according to certain filtering rules (for example: filter un-reviewed notes, filter notes with review violations) to determine 2,000,000 current popular compliant notes as candidate notes. Obtain the note features of the 2,000,000 candidate notes: note identifier, author identifier, note publication time, note title, note content, note tags, note taxonomy, etc. From the log file of user A on the note sharing application, obtain the historical object sequence of 3 types after user A clicks on the reference note within the last 30 days: the historical object sequence of viewing behaviors Pagetime LastN that meet the viewing duration, the historical object sequence of interaction behaviors Engage LastN, and the historical object sequence of complete play behaviors Videoend LastN.

[0164] In step 502, obtaining the object features of multiple candidate objects and the historical object sequence of at least one type after the target user selects a reference object provides comprehensive data support for subsequent object recall.

[0165] Step 504: Based on at least one historical object sequence, obtain at least one interaction behavior vector via the target recall model, where the target recall model is trained according to the above recall model training method.

[0166] The target recall model is a recall model that has completed training, and the target recall model is a recall model that can accurately complete multi-target recall. The target recall model is trained according to the above Figure 3 recall model training method, which will not be elaborated here.

[0167] At least one interaction behavior vector is a vector that performs deep feature encoding on the historical object sequence of at least one type. At least one interaction behavior vector is an abstract high-dimensional feature encoding vector obtained after fully mining the historical object sequence of at least one type.

[0168] Step 506: Based on the object features of multiple candidate objects, obtain multiple object vectors via the target recall model.

[0169] The object vector is a vector obtained by performing deep feature encoding on the object features of candidate objects. The object vector is an abstract high-dimensional feature encoding vector obtained after fully mining the object features of candidate objects.

[0170] Exemplarily, based on the note features of 2,000,000 candidate notes and 3 types of historical object sequences (the historical object sequence of viewing behavior satisfying the viewing duration Pagetime LastN, the historical object sequence of interaction behavior Engage LastN, and the historical object sequence of video completion behavior Videoend LastN), deep feature encoding is performed via the target recall model to obtain 2,000,000 note vectors Item Vector and 3 types of interaction behavior vectors: the viewing duration vector Pagetime User Vector, the interaction behavior vector Engage User Vector, and the video completion behavior vector VideoendUser Vector.

[0171] In steps 504 and 506, based on at least one historical object sequence, at least one interaction behavior vector is obtained via the target recall model, ensuring that the recall model can make full use of multi-dimensional feature information.

[0172] Step 508: Recall the target object from multiple candidate objects based on the vector similarity between at least one interaction behavior vector and multiple object vectors.

[0173] The vector similarity between at least one interaction behavior vector and multiple object vectors is used to measure the similarity degree between at least one interaction behavior vector and multiple object vectors. This vector similarity predicts the interest degree of the target user in multiple candidate objects and the behavior pattern after selection. The vector similarity includes but is not limited to: cosine similarity, Manhattan distance, etc.

[0174] The target object is a specific object entity that is screened out from multiple candidate objects by the target recall model and arouses the user interest of the target user. The recall of the target object comprehensively considers at least one type of historical object sequence, thereby achieving multi-target recall. By fully mining the correlation between at least one type of historical object sequence and object features, the target object recalled by the target recall model can provide diverse recommendation results while ensuring high accuracy, meeting the diverse needs of users.

[0175] Based on the vector similarity between at least one interaction behavior vector and multiple object vectors, to recall a target object from multiple candidate objects, an optional way is: to weight the vector similarity between at least one interaction behavior vector and multiple object vectors to determine the target vector similarity, and based on the target vector similarity, recall the target object from multiple candidate objects. Another optional way is: to recall the target object from multiple candidate objects respectively based on the vector similarity between at least one interaction behavior vector and multiple object vectors. Further, to recall the target object from multiple candidate objects, an optional way is: to perform approximate nearest neighbor search on multiple candidate objects based on the vector similarity to obtain the target object, and another optional way is: to perform K-nearest neighbor search on multiple candidate objects based on the vector similarity to obtain the target object, which is not limited herein.

[0176] Exemplarily, calculate the cosine similarity CosSim_1 between 2,000,000 note vectors Item Vector and the viewing duration vector Pagetime UserVector. Calculate the cosine similarity CosSim_2 between 2,000,000 note vectors Item Vector and the interaction behavior vector Engage User Vector. Calculate the cosine similarity CosSim_3 between 2,000,000 note vectors Item Vector and the video completion behavior vector Videoend User Vector. Weight the vector similarity between 3 types of interaction behavior vectors and 2,000,000 note vectors Item Vector to determine 2,000,000 target vector similarities. Based on the 2,000,000 target vector similarities, perform approximate nearest neighbor search on 2,000,000 candidate note candidate objects to recall 4,000 target notes, and perform subsequent rough ranking, fine ranking, and re-ranking operations on the 4,000 target notes.

[0177] In the embodiments of this specification, obtaining the object features of multiple candidate objects and at least one historical object sequence of the target user performing at least one type of interaction behavior provides comprehensive data support for subsequent object recall. Based on at least one historical object sequence, at least one interaction behavior vector is obtained via the target recall model, and based on the object features of multiple candidate objects, multiple object vectors are obtained via the target recall model, ensuring that the target recall model can make full use of multi-dimensional feature information. Based on the vector similarity between at least one interaction behavior vector and multiple object vectors, the target object is recalled from multiple candidate objects. While completing multi-target recall, the relevance between a single historical object sequence and object features is fully mined, the optimization of a single recall target is achieved, and high-accuracy and diverse multi-target recall are completed.

[0178] In an alternative embodiment of the present specification, step 508 includes the following specific steps: For each type, based on the vector similarity between the interaction behavior vector of this type and multiple object vectors, recall the target object of this type from multiple candidate objects; integrate the target objects of at least one type to obtain the target object.

[0179] For each type, based on the vector similarity between the interaction behavior vector of this type and multiple object vectors, to recall the target object of this type from multiple candidate objects, an alternative way is: For each type, based on the vector similarity between the interaction behavior vector of this type and multiple object vectors, perform approximate nearest neighbor search on multiple candidate objects to recall the target object of this type from multiple candidate objects. Another alternative way is: For each type, based on the vector similarity between the interaction behavior vector of this type and multiple object vectors, perform K-nearest neighbor search on multiple candidate objects to recall the target object of this type from multiple candidate objects. No limitation is made here.

[0180] To integrate the target objects of at least one type to obtain the target object, an alternative way is: Dedup and merge the target objects of at least one type to obtain the target object. Another alternative way is: Cluster the target objects of at least one type and select the representative object in each cluster as the target object. Another alternative way is: Sort the target objects of at least one type and select the representative object in each cluster as the target object. No limitation is made here.

[0181] Exemplarily, based on the cosine similarities CosSim_1, 2 or 3 between the interaction behavior vectors of 3 types and 2,000,000 note vectors Item Vector, determine the target vector similarities Sim_1, 2 or 3 of each type. Based on the target vector similarities Sim_1, 2 or 3 of 3 types, perform approximate nearest neighbor search on 2,000,000 candidate notes to recall 7 target notes of 3 types. Dedup and merge the 7 target notes to obtain 4 target notes, and display the 4 target notes on the home page after user A logs in to the note sharing application.

[0182] In the embodiment of the present specification, for each type, based on the vector similarity between the interaction behavior vector of this type and multiple object vectors, recall the target object of this type from multiple candidate objects, and integrate the target objects of at least one type to obtain the target object, which improves the accuracy and personalization degree of the recall model in multi-target recall.

[0183] In an optional embodiment of this specification, recalling a target object from multiple candidate objects based on the vector similarity between at least one interaction behavior vector and multiple object vectors includes the following specific steps: For any interaction behavior vector, recalling candidate objects from multiple candidate objects whose similarity to the any interaction behavior vector meets a threshold; obtaining the target object based on the candidate objects recalled for at least one interaction behavior vector; wherein, the number of candidate objects recalled based on any interaction behavior vector is determined based on the interaction behavior weight corresponding to the interaction behavior vector.

[0184] The interaction behavior weight corresponding to the interaction behavior vector is a priority parameter defined according to business objectives or user behavior value, and is used to balance the recall contributions of different interaction behavior types. For example: The viewing duration weight reflects the user's preference for content depth, and the weight can be set based on the contribution of session duration to user stickiness; The interaction behavior weight reflects the user's emotional feedback on the content (such as likes, comments), and the weight can be set based on the promotion effect of the interaction behavior on platform activity; The completion behavior weight reflects the user's recognition of the content integrity, and the weight can be set based on the correlation between the completion rate and user satisfaction.

[0185] For any interaction behavior vector, an optional way to recall candidate objects from multiple candidate objects whose similarity to the any interaction behavior vector meets a threshold is: For any interaction behavior vector, directly recalling candidate objects from multiple candidate objects whose similarity to the any interaction behavior vector meets a threshold. Another optional way is: For any interaction behavior vector, setting a dynamic threshold based on the similarity recalled from multiple candidate objects to the any interaction behavior vector, and screening candidate objects whose similarity meets the dynamic threshold. For example, setting the similarity threshold θ = μ + 2σ (where μ is the similarity mean and σ is the standard deviation), and only retaining candidate objects with a cosine similarity higher than θ. Another optional way is: For any interaction behavior vector, clustering based on the similarity recalled from multiple candidate objects to the any interaction behavior vector, and screening candidate objects whose similarity meets a preset threshold from the clustering results.

[0186] Based on the candidate objects recalled for at least one interaction behavior vector, a target object is obtained. An optional method is as follows: According to the preset interaction behavior weights, recall quotas of various types are allocated proportionally. Candidate objects are recalled according to the recall quotas of various types to obtain the target object. For example, if the weight of viewing duration is 40%, the weight of interaction is 30%, and the weight of complete playback is 30%, and the total recall quantity is 1000 objects, the recall quotas of various types are allocated proportionally: 400 objects are recalled for the viewing duration type, 300 for the interaction type, and 300 for the complete playback type, to obtain the target object. Another optional method is as follows: A weighted fusion strategy is adopted to perform score weighting on the recall results of various types, and candidate objects are recalled according to the weighted scores to obtain the target object. For example, multiply the similarity score of the candidate object in the viewing duration type by 0.4 + the score in the interaction type by 0.3 + the score in the complete playback type by 0.3, and select the top-N candidate objects sorted by the total score as the target object.

[0187] Exemplarily, the preset threshold adopts a dynamic quantile strategy: For 2 million candidate notes, statistical analysis is performed on the similarity distributions of various types, and the top 5% quantile of each type is taken as the similarity threshold. In the note recommendation scenario, the interaction behavior weights of viewing duration, interaction, and complete playback configured for the target recall model are 50%, 30%, and 20% respectively. For the 3 interaction behavior vectors of user A, the following operations are performed respectively: 1) Calculate the cosine similarity between the viewing duration vector and the 2 million note vectors, and recall 5 candidate notes; 2) Recall 3 candidate notes for the interaction behavior vector; 3) Recall 2 candidate notes for the complete playback behavior vector. A total of 10 recalled candidate notes are used as the target notes.

[0188] In the embodiments of this specification, by combining the interaction behavior weights and the threshold recall strategy, accurate multi-target recall is maintained, and the recall efficiency is improved.

[0189] Combining the above embodiments of this specification, taking the multi-target recall of notes as an example, Figure 6 The flowchart of an object recall method provided by an embodiment of this specification is shown:

[0190] In the offline publishing and index building request stage of online recall:

[0191] Feature extraction is performed on the note features of multiple candidate notes to obtain note vectors. An index set of candidate notes is built based on the note vectors and published online.

[0192] In the online recall stage of online recall:

[0193] Obtain an online user request, and in response to the online user request, obtain three types of historical object sequences of the user for the reference note. Extract features from the three types of historical object sequences to obtain three types of interaction behavior vectors: the interaction behavior vector of the viewing behavior that meets the viewing duration, the interaction behavior vector of the interaction behavior, and the interaction behavior vector of the interaction behavior.

[0194] Adopt the above Figure 5 method of the embodiment of the specification, and recall the target note from the index set of candidate notes through the K-nearest neighbor classification algorithm.

[0195] As Figure 6 shown, in the online recall stage, directly use the user's duration vector, interaction vector, and complete play vector for multi-target object recall. The multi-target vectors are the interaction behavior vectors of duration, interaction, and complete play. At the same time, use multiple target vectors to retrieve the note index library, that is, multi-query, to obtain the multi-vector recall result. The recall result quantities of the three vectors need to be adjusted according to the actual business requirements to ensure that the final recall result takes into account the multi-dimensional behavior preferences of users and maximizes the value of each target.

[0196] It should be noted that Figure 5 and Figure 6 the technical solutions of the object recall method belong to the same concept as the technical solutions of the above recall model training method. For the details not described in detail, reference can be made to the description of the technical solutions of the above recall model training method.

[0197] Refer to Figure 7 , Figure 7 shows a flowchart of a content recommendation method provided by an embodiment of this specification, which is applied to a content recommendation system and includes the following specific steps:

[0198] Step 702: Obtain the content features of multiple candidate contents from the content database, and based on the content features of the multiple candidate contents, obtain multiple content vectors through the target recall model, where the target recall model is trained according to the above recall model training method.

[0199] Step 704: Receive a content recommendation request sent by the user terminal of the target user.

[0200] Step 706: In response to the content recommendation request, obtain at least one historical content sequence of the target user performing at least one type of interaction behavior from the behavior database.

[0201] Step 708: Based on at least one historical content sequence, obtain at least one interaction behavior vector through the target recall model.

[0202] Step 710: Recall target content from multiple candidate contents based on the vector similarity between at least one interaction behavior vector and multiple content vectors.

[0203] Step 712: Integrate the target content and feedback the integrated target content to the user terminal.

[0204] The content recommendation system is an application, platform or system with an object recall function. For example, an advertisement recommendation system, a video application, a news application, a note sharing platform, an e-commerce platform, etc.

[0205] The content is a specific content entity to be recalled, which has rich features, and the degree of user interest in this content can be judged through these features. For example, advertisements, videos, news, notes, commodities, etc. The candidate content is a specific content entity that can be recalled and has rich features. These candidate contents are potential recommended contents, and through the recall process of the model, the target content that the user may be interested in is screened out from a large number of candidate contents. The content feature is the feature data describing at least one aspect attribute of the content, which is the basis for constructing the content vector. Through these features, the degree of user interest can be more accurately matched. For example, for a video, the content features include: video identifier, author identifier, video length, video resolution, video tags, video taxonomy, etc.

[0206] The user terminal of the target user is the device used by the target user, such as a smart phone, a tablet computer, a personal computer, etc. Through these devices, the target user can send a content recommendation request.

[0207] The content recommendation request is a request sent by the target user through the user terminal, indicating the hope to obtain recommended content. The request carries some user features.

[0208] The historical content sequence is the historical content sequence under multiple types of interaction behaviors that occur between the target user and the candidate content. It is the basis for constructing the interaction behavior vector. Through these sequences, the degree of interest of the target user in the candidate content can be characterized. The historical content sequence includes but is not limited to: the historical content sequence of the viewing behavior that meets the viewing duration, the historical content sequence of the interaction behavior, the historical content sequence of the complete playback behavior, and the historical content sequence of the search behavior.

[0209] The content vector is a vector obtained by deeply feature encoding the content features of the candidate content. The content vector is an abstract high-dimensional feature encoding vector obtained after fully mining the content features of the candidate content.

[0210] The target content is a specific content entity screened out from multiple candidate contents through the target recall model and is the most likely to arouse the interest of the target user. The recall of the target content comprehensively considers the user features and the historical object sequence, so as to achieve multi-target recall.

[0211] Integrate the target content in the following specific ways: sort the target content in a rough, fine and re-rank order. The four necessary steps of recall, rough, fine and re-ranking form a multi-level funnel structure. For example, the recall step selects thousands of notes from tens of millions of candidate notes and inputs them into the rough sorting. The rough sorting step performs a "coarse-grained" scoring and sorting on the thousands of notes provided by the recall, selects the top 1,000 notes with the best quality, and inputs them into the fine sorting. The fine sorting step performs a "fine-grained" scoring and sorting on the 1,000 notes provided by the rough sorting, selects the top 100 notes with the best quality, and inputs them into the re-ranking. The re-ranking step adjusts the order of the 100 notes provided by the fine sorting according to certain rules, selects the last 50 notes, and displays them to the user in order.

[0212] In the embodiments of this specification, a multi-objective recall mechanism is designed to balance multiple key business indicators in the recommendation process. Traditional single-objective modeling methods usually focus on optimizing a specific indicator, but this may lead to a lack of diversity in recommendation results or affect user experience. By adopting a multi-objective recall strategy, the content recommendation system can not only comprehensively consider and optimize these different dimensions of goals, but also ensure the diversity and fairness of content display while improving user satisfaction, thereby creating greater commercial value.

[0213] Take the note sharing app as an example. Figure 8 A front-end schematic diagram of a content recommendation method provided by an embodiment of this specification is shown:

[0214] After the user logs in to the note sharing application, four target notes recalled from multiple candidate notes are displayed on the "Discover" channel on the homepage of the note sharing application, and the four target notes are roughly sorted, finely sorted, and re-sorted in turn.

[0215] At the top of the homepage, there are a details button, a "Follow" channel, a "Discover" channel, a "Nearby" channel, and a query button. In the middle of the homepage, there are note covers of four target notes in order, namely: "Create your autumn ever-changing OOTD", "Explore the hidden food corners in the city", "How to take movie-like photos with a mobile phone", and "Those domestic niche travel destinations that cannot be missed". At the bottom of the homepage, there are a "Home" button, a "Video" button, a "Post Note" button, a "Message" button, and an "I" button.

[0216] It should be noted that Figure 7 and Figure 8 The technical solution of the content recommendation method and the technical solution of the recall model training method mentioned above belong to the same concept. For details not described in detail, please refer to the description of the technical solution of the recall model training method mentioned above.

[0217] Corresponding to the above method embodiments, this specification also provides content recommendation system embodiments. Figure 9 It shows a schematic structural diagram of a content recommendation system provided by an embodiment of this specification. As Figure 9 shown, the content recommendation system 900 includes a request interface 902, an indexing unit 904, a recall unit 906, a content database 908, and a behavior database 910;

[0218] The indexing unit 904 is configured to obtain content features of multiple candidate contents from the content database 908, and based on the content features of the multiple candidate contents, obtain multiple content vectors via a target recall model, where the target recall model is trained according to the above recall model training method;

[0219] The request interface 902 is configured to receive a content recommendation request sent by a user terminal of a target user;

[0220] The recall unit 906 is configured to, in response to the content recommendation request, obtain at least one historical content sequence of the target user performing at least one type of interaction behavior from the behavior database 910, obtain at least one interaction behavior vector via the target recall model based on the at least one historical content sequence, recall target content from the multiple candidate contents based on the vector similarity between the at least one interaction behavior vector and the multiple content vectors, and integrate the target content;

[0221] The request interface 902 is further configured to feedback the integrated target content to the user terminal.

[0222] The content recommendation system 900 is a recommendation system integrating multiple functional modules, aiming to screen out the content most in line with the user's interests from a large number of candidate contents for recommendation by analyzing the user's behavior and preferences. The system 900 includes multiple functional unit modules such as a request interface 902, an indexing unit 904, and a recall unit 906, as well as multiple data storage modules such as a user database 908, a content database 908, and a behavior database 910, which cooperate together to provide an efficient and personalized content recommendation service.

[0223] The request interface 902 is responsible for receiving content recommendation requests from the user terminal. When the user expresses the need for recommended content through their device (such as a smartphone, tablet, etc.), the request interface 902 will capture this request and pass it to other components inside the system for processing, and at the same time, be responsible for feedbacking the recommended results to the user terminal after content recommendation.

[0224] The indexing unit 904 is responsible for collecting relevant information of candidate content from the content database 908, especially content features, and converting these features into content vectors using a trained target recall model. These vectors are deep feature encodings of the candidate content and are used to match user interests in the subsequent content recommendation process.

[0225] The recall unit 906 is the core part of the content recommendation system 900. It is responsible for extracting the features of the target user from the user database 908 according to the received content recommendation request, and obtaining the historical object sequence between the user and the reference content from the behavior database 910. Subsequently, the recall unit 906 inputs these features together with the content features of the candidate content into the target recall model to generate auxiliary selection vectors, and selects the target content that is most likely to attract the user's interest from the candidate content based on the similarity calculation between the vectors. In addition, the recall unit 906 is also responsible for integrating the selected target content to ensure the quality and relevance of the recommended content.

[0226] The content database 908 is a data warehouse that stores candidate content available for recommendation and its relevant information. This information may include metadata such as the theme, type, author, publication time, tags of the content, as well as the content vectors generated by the indexing unit 904, providing the basic data support for the recall unit 906 to perform content recommendation.

[0227] The behavior database 910 is a data warehouse that records the interaction behavior data of users with the content in the system 900, such as the user's click, favorite, share, comment and other behaviors. These data are very important for understanding user preferences and predicting the user's future behavior, and are one of the key factors considered by the recall unit 906 when generating the recommendation list.

[0228] In the embodiments of this specification, the content recommendation system designed with a multi-objective recall mechanism is to balance multiple key business indicators in the recommendation process. Traditional single-objective modeling methods usually focus on optimizing a specific indicator, but this may lead to a lack of diversity in the recommendation results or affect the user experience. By adopting a multi-objective recall strategy, the content recommendation system can not only comprehensively consider and optimize these different-dimensional objectives, but also ensure the diversity and fairness of content display while improving user satisfaction, thereby creating greater commercial value.

[0229] The above is a schematic solution of a content recommendation system according to this embodiment. It should be noted that the technical solution of this content recommendation system and the technical solution of the above content recommendation method belong to the same concept. For the details not described in detail in the technical solution of the content recommendation system, reference can be made to the description of the technical solution of the above content recommendation method.

[0230] Corresponding to the above method embodiments, this specification also provides embodiments of a recall model training apparatus. Figure 10 The following shows a schematic structural diagram of a recall model training apparatus provided by an embodiment of this specification. As Figure 10 shown, the apparatus includes:

[0231] A first acquisition module 1002, configured to acquire object features of a sample object and multiple historical object sequences of a sample user performing multiple types of interaction behaviors;

[0232] A first extraction module 1004, configured to obtain multiple sample interaction behavior vectors via an initial recall model based at least on the multiple historical object sequences, and obtain a sample object vector via the initial recall model based on the object features;

[0233] A first determination module 1006, configured to determine multiple interaction behavior loss values based on the vector similarity between the multiple sample interaction behavior vectors and the sample object vector;

[0234] A first training module 1008, configured to train the initial recall model based on the multiple interaction behavior loss values to obtain a target recall model.

[0235] Optionally, the apparatus further includes: a first auxiliary module, configured to acquire user features of the sample user; obtain a sample user vector via the initial recall model based on the user features; obtain a sample auxiliary selection vector of the sample user based on the sample user vector and the multiple sample interaction behavior vectors; determine an auxiliary selection loss value based on the vector similarity between the sample auxiliary selection vector and the sample object vector and a predetermined interaction relationship between the sample user and the sample object;

[0236] Correspondingly, the first training module 1008 is further configured to: train the initial recall model based on the auxiliary selection loss value and the multiple interaction behavior loss values to obtain a target recall model.

[0237] Optionally, the first auxiliary module is further configured to: perform target attention calculation on the sample user vector and the multiple sample interaction behavior vectors to obtain a sample interest vector of the sample user; obtain the sample auxiliary selection vector of the sample user based on the sample interest vector.

[0238] Optionally, the first auxiliary module is further configured to: splice the sample user vector and the sample interest vector to obtain the sample auxiliary selection vector of the sample user.

[0239] Optionally, the initial recall model includes a user extraction layer, an object extraction layer, multiple behavior extraction layers, an attention calculation layer, and a splicing layer;

[0240] Correspondingly, the first extraction module 1004 is further configured to: input at least a plurality of historical object sequences into a plurality of behavior extraction layers to obtain a plurality of sample interaction behavior vectors; input object features into an object extraction layer to obtain a sample object vector;

[0241] Correspondingly, the first auxiliary module is further configured to: input user features into a user extraction layer to obtain a sample user vector; input the sample user vector and a plurality of sample interaction behavior vectors into an attention calculation layer for target attention calculation to obtain a sample interest vector of the sample user; input the sample user vector and the sample interest vector into a splicing layer to obtain a sample auxiliary selection vector of the sample user.

[0242] Optionally, multiple types of interaction behaviors include a viewing behavior that meets the viewing duration, an interaction behavior, and a complete playback behavior. The historical object sequence of the viewing behavior that meets the viewing duration is the historical object sequence in which the sample user has had a viewing behavior that meets the viewing duration with a plurality of historical objects. The historical object sequence of the interaction behavior is the historical object sequence in which the sample user has had an interaction behavior with a plurality of historical objects. The historical object sequence of the complete playback behavior is the historical object sequence in which the sample user has had a complete playback behavior with a plurality of historical objects.

[0243] In the embodiments of this specification, obtaining the object features of the sample object and a plurality of historical object sequences of the sample user performing multiple types of interaction behaviors provides comprehensive sample data support for subsequent model training. At least based on a plurality of historical object sequences, a plurality of sample interaction behavior vectors are obtained through an initial recall model. Based on the object features, a sample object vector is obtained through the initial recall model, ensuring that the initial recall model can make full use of multi-dimensional feature information. Based on the vector similarity between a plurality of sample interaction behavior vectors and the sample object vector, a plurality of interaction behavior loss values are determined. Based on the plurality of interaction behavior loss values, the initial recall model is trained to obtain a target recall model that can complete multi-target recall. At the same time, the correlation between a single interaction behavior and an object is fully mined, the optimization of a single recall target is realized, the accuracy of the recall model for multi-target optimization is improved, the model performance of the recall model is improved, and the accuracy and diversity of subsequent object recall are improved.

[0244] The above is a schematic solution of a recall model training device according to this embodiment. It should be noted that the technical solution of this recall model training device and the technical solution of the above recall model training method belong to the same concept. For the details not described in detail in the technical solution of the recall model training device, reference can be made to the description of the technical solution of the above recall model training method.

[0245] Corresponding to the above method embodiments, this specification also provides embodiments of an object recall device. Figure 11The figure shows a schematic structural diagram of an object recall device provided by an embodiment of this specification. As Figure 11 shown, the device includes:

[0246] A second acquisition module 1102, configured to acquire object features of multiple candidate objects, and at least one historical object sequence of a target user performing at least one type of interaction behavior;

[0247] A second extraction module 1104, configured to obtain at least one interaction behavior vector via a target recall model based on at least one historical object sequence, and obtain multiple object vectors via the target recall model based on the object features of multiple candidate objects, where the target recall model is trained according to the above-mentioned recall model training method;

[0248] A second recall module 1106, configured to recall a target object from multiple candidate objects based on the vector similarity between at least one interaction behavior vector and multiple object vectors.

[0249] Optionally, the second recall module 1106 is further configured to: for each type, recall the target object of this type from multiple candidate objects based on the vector similarity between the interaction behavior vector of this type and multiple object vectors; integrate the target objects of at least one type to obtain the target object.

[0250] Optionally, the second recall module 1106 is further configured to: for any interaction behavior vector, recall candidate objects whose similarity with any interaction behavior vector meets the threshold from multiple candidate objects; obtain the target object based on the candidate objects recalled for at least one interaction behavior vector; where the number of candidate objects recalled based on any interaction behavior vector is determined based on the interaction behavior weight corresponding to this interaction behavior vector.

[0251] In the embodiment of this specification, acquiring object features of multiple candidate objects and at least one historical object sequence of a target user performing at least one type of interaction behavior provides comprehensive data support for subsequent object recall. Obtaining at least one interaction behavior vector via a target recall model based on at least one historical object sequence, and obtaining multiple object vectors via the target recall model based on the object features of multiple candidate objects ensure that the target recall model can make full use of multi-dimensional feature information. Recalling the target object from multiple candidate objects based on the vector similarity between at least one interaction behavior vector and multiple object vectors, while completing multi-target recall, fully excavates the correlation between a single historical object sequence and object features, realizes the optimization of a single recall target, and completes multi-target recall with high accuracy and diversity.

[0252] The above is a schematic solution of an object recall device according to this embodiment. It should be noted that the technical solution of this object recall device and the technical solution of the above object recall method belong to the same concept. For the details not described in detail in the technical solution of the object recall device, reference can be made to the description of the technical solution of the above object recall method.

[0253] Corresponding to the above method embodiment, this specification also provides an embodiment of a content recommendation device. Figure 12 The following shows a schematic structural diagram of a content recommendation device provided by an embodiment of this specification. As Figure 12 shown, this device is applied to a content recommendation system and includes:

[0254] A third acquisition and extraction module 1202, configured to acquire the content features of multiple candidate contents from a content database, and based on the content features of the multiple candidate contents, obtain multiple content vectors via a target recall model, where the target recall model is trained according to the above recall model training method;

[0255] A third receiving module 1204, configured to receive a content recommendation request sent by a user terminal of a target user;

[0256] A third acquisition module 1206, configured to, in response to the content recommendation request, acquire at least one historical content sequence of the target user's interaction behaviors of at least one type from a behavior database;

[0257] A third extraction module 1208, configured to obtain at least one interaction behavior vector via the target recall model based on the at least one historical content sequence;

[0258] A third recall module 1210, configured to recall target content from the multiple candidate contents based on the vector similarity between the at least one interaction behavior vector and the multiple content vectors;

[0259] A third feedback module 1212, configured to integrate the target content and feedback the integrated target content to the user terminal.

[0260] In the embodiment of this specification, by adopting a multi-target recall strategy, the content recommendation system can not only comprehensively consider and optimize these different-dimensional targets, but also ensure the diversity and fairness of content display while improving user satisfaction, thereby creating greater commercial value.

[0261] The above is a schematic solution of a content recommendation device according to this embodiment. It should be noted that the technical solution of this content recommendation device and the technical solution of the above content recommendation method belong to the same concept. For the details not described in detail in the technical solution of the content recommendation device, reference can be made to the description of the technical solution of the above content recommendation method.

[0262] Figure 13 The block diagram of a computing device provided by an embodiment of this specification is shown. The components of the computing device 1300 include, but are not limited to, a memory 1310 and a processor 1320. The processor 1320 is connected to the memory 1310 through a bus 1330, and a database 1350 is used to store data.

[0263] The computing device 1300 further includes an access device 1340, which enables the computing device 1300 to communicate via one or more networks 1360. Examples of these networks include the Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 1340 may include one or more of any type of wired or wireless network interfaces (e.g., a Network Interface Controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface.

[0264] In an embodiment of this specification, the above components of the computing device 1300 and Figure 13 other components not shown may also be connected to each other, for example, through a bus. It should be understood that Figure 13 the block diagram of the computing device shown is only for illustrative purposes and not a limitation on the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0265] The computing device 1300 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.) or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 1300 can also be a mobile or stationary server.

[0266] Wherein, the processor 1320 is configured to execute the following computer program / instructions, and when the computer program / instructions are executed by the processor, the steps of the above-mentioned recall model training method, object recall method or content recommendation method are implemented.

[0267] The above is a schematic solution of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solutions of the above-mentioned recall model training method, object recall method and content recommendation method belong to the same concept. For the details not described in detail in the technical solution of the computing device, reference can be made to the descriptions of the technical solutions of the above-mentioned recall model training method, object recall method or content recommendation method.

[0268] An embodiment of this specification also provides a computer-readable storage medium, which stores computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned recall model training method, object recall method or content recommendation method are implemented.

[0269] The above is a schematic solution of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solutions of the above-mentioned recall model training method, object recall method and content recommendation method belong to the same concept. For the details not described in detail in the technical solution of the storage medium, reference can be made to the descriptions of the technical solutions of the above-mentioned recall model training method, object recall method or content recommendation method.

[0270] An embodiment of this specification also provides a computer program product, including computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above-mentioned recall model training method, object recall method or content recommendation method are implemented.

[0271] The above is a schematic solution of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solutions of the above-mentioned recall model training method, object recall method, and content recommendation method belong to the same concept. For the details not described in the technical solution of the computer program product, reference can be made to the descriptions of the technical solutions of the above-mentioned recall model training method, object recall method, or content recommendation method.

[0272] The above has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0273] The computer instructions include computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM for short), random access memory (RAM for short), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0274] It should be noted that for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the embodiments of this specification are not limited by the described order of actions, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this specification.

[0275] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0276] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation manners. Obviously, according to the content of the embodiments of the present specification, many modifications and variations can be made. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and utilize the present specification. The present specification is only limited by the claims and their full scope and equivalents.

Claims

1. A method for training a recall model, characterized in that, Including: Obtaining the object features of a sample object and multiple historical object sequences of a sample user performing multiple types of interaction behaviors; Obtaining multiple sample interaction behavior vectors via an initial recall model based at least on the multiple historical object sequences; Obtaining a sample object vector via the initial recall model based on the object features; Determining multiple interaction behavior loss values based on the vector similarity between the multiple sample interaction behavior vectors and the sample object vector; Training the initial recall model based on the multiple interaction behavior loss values to obtain a target recall model.

2. The method according to claim 1, wherein Before the training the initial recall model based on the multiple interaction behavior loss values to obtain a target recall model, it further includes: Obtaining the user features of the sample user; Obtaining a sample user vector via the initial recall model based on the user features; Obtaining a sample auxiliary selection vector of the sample user based on the sample user vector and the multiple sample interaction behavior vectors; Determining an auxiliary selection loss value based on the vector similarity between the sample auxiliary selection vector and the sample object vector and a predetermined interaction relationship between the sample user and the sample object; The training the initial recall model based on the multiple interaction behavior loss values to obtain a target recall model includes: Training the initial recall model based on the auxiliary selection loss value and the multiple interaction behavior loss values to obtain a target recall model.

3. The method according to claim 2, characterized in that, The obtaining a sample auxiliary selection vector of the sample user based on the sample user vector and the multiple sample interaction behavior vectors includes: Performing target attention calculation on the sample user vector and the multiple sample interaction behavior vectors to obtain a sample interest vector of the sample user; Obtaining a sample auxiliary selection vector of the sample user based on the sample interest vector.

4. The method according to claim 3, wherein The obtaining a sample auxiliary selection vector of the sample user based on the sample interest vector includes: Concatenating the sample user vector and the sample interest vector to obtain a sample auxiliary selection vector of the sample user.

5. The method according to claim 4, characterized in that The initial recall model includes a user extraction layer, an object extraction layer, the multiple behavior extraction layers, an attention calculation layer, and a concatenation layer; The obtaining multiple sample interaction behavior vectors via an initial recall model based at least on the multiple historical object sequences includes: Inputting at least the multiple historical object sequences into the multiple behavior extraction layers to obtain multiple sample interaction behavior vectors; The obtaining a sample object vector via the initial recall model based on the object features includes: Inputting the object features into the object extraction layer to obtain a sample object vector; The obtaining a sample user vector via the initial recall model based on the user features includes: Inputting the user features into the user extraction layer to obtain a sample user vector; The performing target attention calculation on the sample user vector and the multiple sample interaction behavior vectors to obtain a sample interest vector of the sample user includes: Inputting the sample user vector and the multiple sample interaction behavior vectors into the attention calculation layer for target attention calculation to obtain a sample interest vector of the sample user; Splicing the sample user vector and the sample interest vector to obtain a sample auxiliary selection vector of the sample user includes: Inputting the sample user vector and the sample interest vector into the splicing layer to obtain a sample auxiliary selection vector of the sample user.

6. The method according to claim 1, characterized in that, The multiple types of interaction behaviors include a viewing behavior that meets a predetermined viewing duration, an interaction behavior, and a complete viewing behavior. The historical object sequence of the viewing behavior that meets the viewing duration is the historical object sequence in which the sample user has had a viewing behavior that meets the viewing duration with multiple historical objects. The historical object sequence of the interaction behavior is the historical object sequence in which the sample user has had an interaction behavior with multiple historical objects. The historical object sequence of the complete viewing behavior is the historical object sequence in which the sample user has had a complete viewing behavior with multiple historical objects.

7. An object recall method, characterized in that, It includes: Obtaining the object features of multiple candidate objects and at least one historical object sequence of the target user performing at least one type of interaction behavior; Based on at least one historical object sequence, obtaining at least one interaction behavior vector via a target recall model, where the target recall model is trained according to the recall model training method described in any one of claims 1-6; Based on the object features of the multiple candidate objects, obtaining multiple object vectors via the target recall model; Based on the vector similarity between the at least one interaction behavior vector and the multiple object vectors, recalling a target object from the multiple candidate objects.

8. The method according to claim 7, characterized in that, The recalling a target object from the multiple candidate objects based on the vector similarity between the at least one interaction behavior vector and the multiple object vectors includes: According to each type, recalling a target object of this type from the multiple candidate objects based on the vector similarity between the interaction behavior vector of this type and the multiple object vectors; Integrating the at least one type of target object to obtain a target object.

9. The method according to claim 7, characterized in that The recalling a target object from the multiple candidate objects based on the vector similarity between the at least one interaction behavior vector and the multiple object vectors includes: For any one interaction behavior vector, recalling candidate objects from the multiple candidate objects whose similarity with the any one interaction behavior vector meets a threshold; Based on the candidate objects recalled for the at least one interaction behavior vector, obtaining a target object; where the number of candidate objects recalled based on any one interaction behavior vector is determined based on the interaction behavior weight corresponding to this interaction behavior vector.

10. A content recommendation method, characterized in that, Applied to a content recommendation system, it includes: Obtaining the content features of multiple candidate contents from a content database, and based on the content features of the multiple candidate contents, obtaining multiple content vectors via a target recall model, where the target recall model is trained according to the recall model training method described in any one of claims 1-6; Receiving a content recommendation request sent by a user terminal of a target user; In response to the content recommendation request, obtaining at least one historical content sequence of the target user performing at least one type of interaction behavior from the behavior database; Based on at least one historical content sequence, obtaining at least one interaction behavior vector via a target recall model; Recall target content from the multiple candidate contents based on the vector similarity between the at least one interaction behavior vector and the multiple content vectors; Integrate the target content and feedback the integrated target content to the user terminal.

11. A content recommendation system, characterized in that, It includes a request interface, an indexing unit, a recall unit, a content database, and a behavior database; The indexing unit is configured to obtain the content features of multiple candidate contents from the content database, and based on the content features of the multiple candidate contents, obtain multiple content vectors via a target recall model, where the target recall model is trained according to the recall model training method described in any one of claims 1-6; The request interface is configured to receive a content recommendation request sent by the user terminal of the target user; The recall unit is configured to, in response to the content recommendation request, obtain at least one historical content sequence of the target user performing at least one type of interaction behavior from the behavior database, obtain at least one interaction behavior vector via a target recall model based on the at least one historical content sequence, recall target content from the multiple candidate contents based on the vector similarity between the at least one interaction behavior vector and the multiple content vectors, and integrate the target content; The request interface is further configured to feedback the integrated target content to the user terminal.

12. A computing device, characterized in that, It includes: A memory and a processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the method described in any one of claims 1 to 11 are implemented.

13. A computer-readable storage medium, characterized in that, It stores computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the method described in any one of claims 1 to 11 are implemented.

14. A computer program product, characterized in that, It includes computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the method described in any one of claims 1 to 11 are implemented.

Images