Content recommendation method and related products

Abstract

The application discloses a content recommendation method and related products, which comprises the following steps: obtaining a current intention state distribution of a first object; predicting a predicted intention state distribution corresponding to the first object according to a historical intention state sequence of the first object and a current interaction environment feature; performing multi-dimensional confidence evaluation on the first object according to the current intention state distribution, the current interaction environment feature and the predicted intention state distribution to obtain a comprehensive confidence; obtaining a current intention state recommendation score and a future intention state recommendation score corresponding to a target candidate content; performing weighted fusion processing on the current intention state recommendation score and the future intention state recommendation score according to the comprehensive confidence to obtain a final recommendation score corresponding to the target candidate content; and using the final recommendation score to determine a ranking position of the target candidate content in a content recommendation set.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a content recommendation method and related products. Background Technology

[0002] Current content recommendation algorithms are generally designed based on the assumption that "user interests are relatively stable," treating all user interactions as equivalent without considering the differences in intent behind different interactions. This assumption holds true for mature users because they are relatively stable, and their intent states have tended to stabilize. However, this is not the case for new users. One of the core characteristics of new users is that their intent evolves rapidly and their preferences are not yet fixed. Their intent states may change rapidly, and in reality, users in different intent states have different needs for recommended content. Current technology cannot identify the user's current intent state, nor has it considered the impact of changes in the user's intent state on recommended content, resulting in recommendations that deviate from the user's actual needs and have poor recommendation performance. Summary of the Invention

[0003] This application provides a content recommendation method and related products, wherein the related products include content recommendation device electronic equipment, computer-readable storage medium and computer program products.

[0004] Firstly, a content recommendation method is provided, including: Obtain the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; Based on the historical intent state sequence of the first object and the characteristics of the current interaction environment, a prediction is made to obtain the predicted intent state distribution corresponding to the first object; the predicted intent state distribution refers to the probability distribution of the intent state of the first object in a future time step. Based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution, a multidimensional confidence assessment is performed on the first object to obtain a comprehensive confidence score. Obtain the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the future time step; The current intent state recommendation score and the future intent state recommendation score are weighted and fused based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

[0005] In conjunction with any embodiment of this application, obtaining the current intent state distribution of the first object includes: Obtain the target multidimensional feature sequence of the first object; the target multidimensional feature sequence contains at least two multidimensional features; the interaction timestamps corresponding to the at least two multidimensional features are all greater than the target timestamp; the multidimensional features are composed of behavioral features, preference features, and interaction environment features. The position weight of each multidimensional feature is determined based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence; the sequence position of each multidimensional feature is sorted based on the interaction timestamp. Each multidimensional feature is weighted according to its corresponding position weight to obtain the weighted target multidimensional feature sequence. The first object is subjected to intent recognition processing based on the weighted target multidimensional feature sequence to obtain the current intent state distribution of the first object.

[0006] In any embodiment of this application, the step of predicting the distribution of predicted intention states corresponding to the first object based on the historical intention state sequence and current interaction environment features of the first object includes: Based on the historical intent state sequence of the first object and the current interaction environment features, multi-step intent state prediction processing is performed to obtain K predicted intent state distributions corresponding to the first object; among the K predicted intent state distributions, the i-th predicted intent state distribution refers to the probability distribution of the intent state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K.

[0007] In conjunction with any embodiment of this application, the step of performing multi-step intent state prediction processing based on the historical intent state sequence of the first object and the current interaction environment features to obtain K predicted intent state distributions corresponding to the first object includes: The historical intent state sequence of the first object is subjected to temporal encoding to obtain the historical intent state feature vector; The feature vectors corresponding to the historical intent state feature vector and the current interaction environment feature vector are subjected to feature fusion processing to obtain the initial prediction input vector; Traverse K future time steps. In the intention state prediction processing for the x-th future time step, obtain the predicted intention state distribution corresponding to the (x-1)-th future time step; x is a positive integer less than or equal to K. The (x-1)th predicted input vector is updated according to the predicted intention state distribution corresponding to the (x-1)th future time step to obtain the xth predicted input vector; when x-1 is 1, the (x-1)th predicted input vector is the initial predicted input vector; The x-th predicted input vector is processed for intent recognition to obtain the predicted intent state distribution corresponding to the x-th future time step; When the K future time steps are completed, the distribution of the K predicted intention states corresponding to the first object is obtained.

[0008] In any embodiment of this application, the predicted intention state distribution includes K predicted intention state distributions; the i-th predicted intention state distribution refers to the probability distribution of the intention state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; The step of performing a multi-dimensional confidence assessment on the first object based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution to obtain a comprehensive confidence score includes: The predicted intention state transition trajectory is determined based on the current intention state distribution and the K predicted intention state distributions; the predicted intention state transition trajectory includes the current intention state and K predicted intention states; the current intention state refers to the intention state with the highest probability in the current intention state distribution; the y-th predicted intention state refers to the intention state with the highest probability in the y-th predicted intention state distribution; y is a positive integer less than or equal to K; The predicted intention state transition trajectory and the standard intention state transition trajectory are compared to obtain the trajectory similarity; the standard intention state transition trajectory is generated based on the target historical intention state transition trajectories corresponding to at least two sample objects; the target historical intention state transition trajectory refers to the intention state transition trajectory whose starting state is the current state. The stability of the current interactive environment is obtained by performing a stability evaluation on the current interactive environment features. A deterministic evaluation process is performed on the K predicted intention state distributions to obtain the intention state prediction certainty. The trajectory similarity, the stability of the current interaction environment, and the certainty of the intention state prediction are weighted and fused to obtain a comprehensive confidence score.

[0009] In conjunction with any embodiment of this application, the step of performing deterministic evaluation processing on the K predicted intention state distributions to obtain the intention state prediction certainty includes: Entropy values ​​are calculated for each of the K predicted intention state distributions to obtain K basic entropy values; The weighted total entropy value is obtained by weighting and summing the K basic entropy values ​​according to the time decay weights corresponding to the K predicted intention states. The weighted total entropy value is normalized to obtain the normalized weighted total entropy value; The difference between the deterministic threshold and the normalized weighted total entropy value is determined as the determination of the intention state prediction.

[0010] In any embodiment of this application, the future intent state recommendation score includes K future intent state recommendation scores; the i-th future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; The step of weighting and fusing the current intent state recommendation score and the future intent state recommendation score based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content includes: The minimum value between the overall confidence level and the maximum threshold of the future intention weight is determined as the basic weight of the future intention state; The difference between the total weight threshold and the basic weight of the future intention state is determined as the weight of the current intention state; Multiply the time decay weights corresponding to the K future intention state recommendation scores by the base weights of the future intention states to obtain the future intention state decay weights corresponding to the K future intention state recommendation scores. The final recommendation score corresponding to the target candidate content is obtained by adding the product between the current intent state weight and the current intent state recommendation score, and the product between the future intent state decay weight corresponding to each future intent state recommendation score and each future intent state recommendation score.

[0011] Secondly, a content recommendation device is provided, including: The first acquisition module is used to acquire the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; The prediction module is used to predict the predicted intention state distribution of the first object based on the historical intention state sequence and the current interaction environment features; the predicted intention state distribution refers to the probability distribution of the intention state of the first object in a future time step. The evaluation module is used to perform a multi-dimensional confidence evaluation on the first object based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution, to obtain a comprehensive confidence score. The second acquisition module is used to acquire the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the future time step; The weighting module is used to perform weighted fusion processing on the current intent state recommendation score and the future intent state recommendation score based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

[0012] In any embodiment of this application, the first acquisition module is specifically used for: Obtain the target multidimensional feature sequence of the first object; the target multidimensional feature sequence contains at least two multidimensional features; the interaction timestamps corresponding to the at least two multidimensional features are all greater than the target timestamp; the multidimensional features are composed of behavioral features, preference features, and interaction environment features. The position weight of each multidimensional feature is determined based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence; the sequence position of each multidimensional feature is sorted based on the interaction timestamp. Each multidimensional feature is weighted according to its corresponding position weight to obtain the weighted target multidimensional feature sequence. The first object is subjected to intent recognition processing based on the weighted target multidimensional feature sequence to obtain the current intent state distribution of the first object.

[0013] In conjunction with any embodiment of this application, the prediction module is specifically used for: Based on the historical intent state sequence of the first object and the current interaction environment features, multi-step intent state prediction processing is performed to obtain K predicted intent state distributions corresponding to the first object; among the K predicted intent state distributions, the i-th predicted intent state distribution refers to the probability distribution of the intent state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K.

[0014] In conjunction with any embodiment of this application, the prediction module is specifically used for: The historical intent state sequence of the first object is subjected to temporal encoding to obtain the historical intent state feature vector; The feature vectors corresponding to the historical intent state feature vector and the current interaction environment feature vector are subjected to feature fusion processing to obtain the initial prediction input vector; Traverse K future time steps. In the intention state prediction processing for the x-th future time step, obtain the predicted intention state distribution corresponding to the (x-1)-th future time step; x is a positive integer less than or equal to K. The (x-1)th predicted input vector is updated according to the predicted intention state distribution corresponding to the (x-1)th future time step to obtain the xth predicted input vector; when x-1 is 1, the (x-1)th predicted input vector is the initial predicted input vector; The x-th predicted input vector is processed for intent recognition to obtain the predicted intent state distribution corresponding to the x-th future time step; When the K future time steps are completed, the distribution of the K predicted intention states corresponding to the first object is obtained.

[0015] In any embodiment of this application, the predicted intention state distribution includes K predicted intention state distributions; the i-th predicted intention state distribution refers to the probability distribution of the intention state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; the evaluation module is specifically used for: The predicted intention state transition trajectory is determined based on the current intention state distribution and the K predicted intention state distributions; the predicted intention state transition trajectory includes the current intention state and K predicted intention states; the current intention state refers to the intention state with the highest probability in the current intention state distribution; the y-th predicted intention state refers to the intention state with the highest probability in the y-th predicted intention state distribution; y is a positive integer less than or equal to K; The predicted intention state transition trajectory and the standard intention state transition trajectory are compared to obtain the trajectory similarity; the standard intention state transition trajectory is generated based on the target historical intention state transition trajectories corresponding to at least two sample objects; the target historical intention state transition trajectory refers to the intention state transition trajectory whose starting state is the current state. The stability of the current interactive environment is obtained by performing a stability evaluation on the current interactive environment features. A deterministic evaluation process is performed on the K predicted intention state distributions to obtain the intention state prediction certainty. The trajectory similarity, the stability of the current interaction environment, and the certainty of the intention state prediction are weighted and fused to obtain a comprehensive confidence score.

[0016] In conjunction with any embodiment of this application, the evaluation module is specifically used for: Entropy values ​​are calculated for each of the K predicted intention state distributions to obtain K basic entropy values; The weighted total entropy value is obtained by weighting and summing the K basic entropy values ​​according to the time decay weights corresponding to the K predicted intention states. The weighted total entropy value is normalized to obtain the normalized weighted total entropy value; The difference between the deterministic threshold and the normalized weighted total entropy value is determined as the determination of the intention state prediction.

[0017] In any embodiment of this application, the future intent state recommendation score includes K future intent state recommendation scores; the i-th future intent state recommendation score is used to characterize the recommendation degree of the target candidate content for the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; the weighting module is specifically used for: The minimum value between the overall confidence level and the maximum threshold of the future intention weight is determined as the basic weight of the future intention state; The difference between the total weight threshold and the basic weight of the future intention state is determined as the weight of the current intention state; Multiply the time decay weights corresponding to the K future intention state recommendation scores by the base weights of the future intention states to obtain the future intention state decay weights corresponding to the K future intention state recommendation scores. The final recommendation score corresponding to the target candidate content is obtained by adding the product between the current intent state weight and the current intent state recommendation score, and the product between the future intent state decay weight corresponding to each future intent state recommendation score and each future intent state recommendation score.

[0018] Thirdly, an electronic device is provided, comprising: a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs as described in the first aspect and any of its embodiments.

[0019] Fourthly, another electronic device is provided, comprising: a processor, a transmitting device, an input device, an output device, and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs as described in the first aspect and any of its embodiments.

[0020] Fifthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, the computer program including program instructions that, when executed by a processor, cause the processor to perform the first aspect and any of its embodiments described above.

[0021] In a sixth aspect, a computer program product is provided, the computer program product comprising a computer program or instructions that, when the computer program or instructions are executed on a computer, cause the computer to perform the first aspect described above and any of its embodiments.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application.

[0023] In this application, the current intent state distribution of a first object can be obtained; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; based on the historical intent state sequence of the first object and the characteristics of the current interaction environment, a predicted intent state distribution corresponding to the first object is obtained; the predicted intent state distribution refers to the probability distribution of the intent state of the first object at a future time step; based on the current intent state distribution, the characteristics of the current interaction environment, and the predicted intent state distribution, a multi-dimensional confidence assessment of the first object is performed to obtain a comprehensive confidence score; the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content are obtained; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content for the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content for the first object at a future time step; the current intent state recommendation score and the future intent state recommendation score are weighted and fused based on the comprehensive confidence score to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set. The method provided in this application divides the object's interest cognition into multiple intent states. The current intent state of the first object can be quickly determined by the distribution of the object's current intent state. Furthermore, the future intent state of the first object can be predicted based on the historical intent state sequence of the first object and the characteristics of the current interaction environment. The reliability of the prediction is quantified by a comprehensive confidence metric. Finally, the recommendation score of the current intent state and the recommendation score of the future intent state are weighted and fused based on the comprehensive confidence metric. The final recommendation score not only considers the current intent of the first object but also the impact of the evolution of the first object's future intent on the recommended content, making the recommendation result closer to the real needs of the first object and thus improving the recommendation effect. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0025] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0026] Figure 1 This application provides a schematic diagram of a network architecture. Figure 2 A flowchart illustrating a content recommendation method provided in an embodiment of this application; Figure 3 A schematic diagram of an intent recognition network architecture provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a content recommendation device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0028] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0029] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0030] The embodiments of this application are described below with reference to the accompanying drawings.

[0031] Please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. Figure 1As shown, this network architecture may include a server 2000 and a terminal cluster. The terminal cluster may specifically include one or more terminals; the number of terminals in the terminal cluster is not limited here. Figure 1 As shown, the multiple terminals may specifically include terminal 3000a, terminal 3000b, terminal 3000c, ..., terminal 3000n; terminal 3000a, terminal 3000b, terminal 3000c, ..., terminal 3000n can be directly or indirectly connected to server 2000 via wired or wireless communication, so that each terminal can interact with server 2000 through the network connection. It is understood that the above network connection is not limited to a specific connection method; it can be directly or indirectly connected via wired communication, wireless communication, or other methods. This application does not impose any restrictions on this.

[0032] Each terminal in the terminal cluster can include: smartphones, tablets, laptops, desktop computers, smart voice interaction devices, smart home appliances (e.g., smart TVs), wearable devices, in-vehicle terminals, aircraft, and other smart terminals with data processing capabilities. It should be understood that, as... Figure 1 Each terminal in the terminal cluster shown can have an application client installed. When the application client runs on each terminal, it can interact with the aforementioned... Figure 1 Data exchange occurs between the 2000 servers shown.

[0033] Specifically, the application client may include in-vehicle clients, smart home clients, entertainment clients (e.g., game clients), multimedia clients (e.g., video clients), social clients, and information clients (e.g., news clients) that have data processing capabilities. In this embodiment, the application client may be integrated into a client (e.g., a social client), or it may be a standalone client (e.g., a news client). This embodiment does not limit the type of application client.

[0034] Among them, such as Figure 1 The server 2000 shown can be the server corresponding to the application client. The server 2000 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0035] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 One of the multiple terminals shown is selected as the first terminal. For example, in the embodiments of this application, a terminal can be selected as the first terminal. Figure 1 The terminal 3000a shown serves as the first terminal, which can be equipped with an application client capable of data processing. In this case, the first terminal can interact with the server 2000 via the application client. Furthermore, the user corresponding to the first terminal can be referred to as the first object.

[0036] like Figure 1 As shown, the first terminal can generate a content recommendation request through a running application client, and then send the content recommendation request to the server 2000. After receiving the content recommendation request, the server 2000 can first generate a content recommendation candidate set for the content recommendation request, and then sort the candidate content in the content recommendation candidate set according to the current intent state and future intent state changes of the first object, thereby obtaining the content recommendation set. Specifically, server 2000 obtains the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; then, based on the historical intent state sequence of the first object and the characteristics of the current interaction environment, it makes a prediction to obtain the predicted intent state distribution corresponding to the first object; the predicted intent state distribution refers to the probability distribution of the intent state of the first object at a future time step; then, based on the current intent state distribution, the characteristics of the current interaction environment, and the predicted intent state distribution, it performs a multi-dimensional confidence assessment on the first object to obtain a comprehensive confidence score; next, it obtains the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content for the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content for the first object at a future time step; finally, based on the comprehensive confidence score, it performs a weighted fusion process on the current intent state recommendation score and the future intent state recommendation score to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

[0037] Furthermore, this application provides a content recommendation method, wherein the execution subject of the content recommendation method is a content recommendation device, which can be any electronic device capable of executing the technical solutions disclosed in the embodiments of this application. Optionally, the content recommendation device can be one of the following: a computer, a server. It should be understood that the embodiments of this application can also be implemented by a processor executing computer program code. In the implementation of this application, the content includes: text, video, images, audio, and graphic notes.

[0038] Please see Figure 2 , Figure 2 This is a flowchart illustrating a content recommendation method provided in an embodiment of this application.

[0039] S101. Obtain the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step.

[0040] Specifically, the cognitive interest of the first object can be divided into L intention states, where L is a positive integer. The L intention states can be defined based on the characteristics of the business domain. For example, in the e-commerce domain, the L intention states could be "browse → compare → evaluate → decide → purchase"; in the content domain, they could be "discover → participate → read in depth → interact → follow"; and in the social domain, they could be "lurk → observe → participate → contribute → lead". Different business domains can customize the number of states and transition paths according to their business characteristics; this application does not impose any restrictions on this.

[0041] Specifically, based on user cognitive psychology theory and large-scale user behavior data analysis, this application also provides a feasible implementation of L intention states, abstracting the evolution of a new user's intention into 5 progressive levels: L1 - Random Exploration: Users have no clear goal and browse various types of content randomly, characterized by a high bounce rate (>70%), short dwell time (<30 seconds), and low interaction depth; L2 - Discovery: Users begin to pay attention to specific categories or topics, characterized by continuous browsing of similar content and moderate dwell time (30-90 seconds); L3 - Focusing: Users develop a clear preference for certain attributes, characterized by repeated viewing of similar content and increased deep interaction behavior; L4 - Confirmation: Users form stable preferences, characterized by confirmatory behaviors such as saving, comparing, and deep browsing of details; L5 - Decision Preparation: Users are ready to make a final choice, characterized by price sensitivity, attention to evaluation, and purchase-related operations. This progressively hierarchical 5 intention states are applicable to most application domains.

[0042] Specifically, a feasible implementation process for obtaining the current intent state distribution of the first object can be as follows: Obtain the target multidimensional feature sequence of the first object; the target multidimensional feature sequence contains at least two multidimensional features; the interaction timestamps corresponding to at least two multidimensional features are both greater than the target timestamp; the multidimensional features are composed of behavioral features, preference features, and interaction environment features; determine the position weight corresponding to each multidimensional feature based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence; the sequence position of each multidimensional feature is sorted based on the interaction timestamps; perform position weighting processing on each multidimensional feature according to its position weight to obtain a weighted target multidimensional feature sequence; perform intent recognition processing on the first object based on the weighted target multidimensional feature sequence to obtain the current intent state distribution of the first object.

[0043] Among them, behavioral features can be multi-dimensional numerical feature vectors used to characterize fine-grained operational behaviors during the interaction between the object and the page, directly reflecting the object's attention to the current content and its willingness to interact. Specific feature dimensions may include: current page dwell time, click location coordinates, scrolling speed, scrolling depth, operation sequence complexity, page bounce probability, and return behavior frequency, among other core operational indicators. Preference features, also known as macro-sequence features, are multi-dimensional statistical feature vectors used to characterize the overall evolutionary pattern of the object's continuous behavioral sequence, reflecting the dynamic development trend of the object's intent. Specific feature dimensions may include: thematic consistency of recently viewed content, interest drift trend, decision urgency score, and behavioral pattern stability. Interaction environment features, also known as contextual features, are multi-dimensional categorical or numerical hybrid feature vectors used to describe the external environmental information in which the object's behavior occurs, eliminating the interference of environmental factors on intent recognition. Specific feature dimensions may include: access time period, device type, entry source, network environment, and geographical location. Categorical features are converted into numerical vectors through one-hot encoding to adapt to the input requirements of deep learning networks. In a feasible embodiment, behavioral features can be 12-dimensional, preference features can be 8-dimensional, and interaction environment features can be 5-dimensional.

[0044] The process of determining the position weight of each multidimensional feature based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence can be illustrated by the following formula (1): position_weights = exp(position_index / sequence_length) formula (1) Where position_weights refers to the position weights of the multidimensional features, position_index is the sequence position of the multidimensional features, and sequence_length is the sequence length of the target multidimensional feature sequence.

[0045] The implementation process of intent recognition can be achieved by first capturing the temporal dependencies of multiple weighted multidimensional features in the weighted multidimensional feature sequence of the target through a Bi-LSTM (Bidirectional Long Short-Term Memory) network, and then outputting the probability distribution of each state through a classifier to obtain the current intent state distribution.

[0046] To better understand the generation process of the current intent state distribution described above, please refer to [the relevant documentation / reference]. Figure 3 , Figure 3 This is a schematic diagram of an intent recognition network architecture provided in an embodiment of this application. Figure 3 As shown, the entire intent recognition network mainly includes a multi-scale feature encoding layer, a position-weighted layer, a bidirectional long short-term memory network layer, and a classifier. Therefore, the process of determining the current intent state distribution of the first object can be as follows: First, the behavioral data of the first object at different interaction timestamps (e.g., object data 1 corresponding to interaction timestamp 1, ..., object data n corresponding to interaction timestamp n) is obtained as the raw input to the network. It can be understood that, to avoid excessive data volume, a target timestamp can be set, and only behavioral data older than this target timestamp is selected. Then, in the multi-scale feature encoding layer, three types of features are extracted for each object data: 12-dimensional behavioral features (e.g., dwell time, click frequency, etc.), 8-dimensional preference features (e.g., topic consistency, interest drift trend, etc.), and 5-dimensional interaction environment features (e.g., time, device, source, etc., e.g., based on object data 1, behavioral feature 1, preference feature 1, and interaction environment feature 1 are obtained; based on object data n, behavioral feature n, preference feature n, and interaction environment feature n are obtained). Further, these three types of features are concatenated to obtain multi-dimensional features, e.g., based on behavioral feature 1, preference feature 1, and interaction environment feature 1, multi-dimensional feature 1 is obtained; based on behavioral feature n, preference feature n, and interaction environment feature n, multi-dimensional feature n is obtained. Next, the various multidimensional features are input into a position-weighted layer. Weight allocation enhances the influence of multidimensional features corresponding to recent behaviors (i.e., making later behaviors contribute more to intent recognition), outputting weighted multidimensional features (e.g., weighted multidimensional feature 1, weighted multidimensional feature n). Finally, all weighted multidimensional features are input into a bidirectional long short-term memory network layer in chronological order, simultaneously capturing the forward and reverse temporal dependencies of the behavior sequence to analyze the intent evolution pattern. The output of the bidirectional long short-term memory network layer is then fed into a classifier to obtain the current intent state distribution of the first object.

[0047] Optionally, the current intent state distribution of the first object can be determined based on rule matching. Another feasible implementation process for obtaining the current intent state distribution of the first object is as follows: Perform multi-feature statistical processing on the behavioral data of the first object to obtain R feature statistical values. Match each of the R feature statistical values ​​with the state determination rule corresponding to each intent state to obtain L intent state probabilities. R is a positive integer. The state determination rule corresponding to each intent state can be defined based on expert experience and a large amount of data statistics. A state determination rule corresponding to an intent state can contain at least Q feature thresholds, where Q is a positive integer less than or equal to R. For ease of understanding, taking the intent state as browsing as an example, the matching features of the object should be short dwell time + high page jumps + random browsing. Therefore, its state determination rule can be set as dwell time less than 30 seconds, page jumps greater than 5, and browsing category diversity greater than 0.8. The corresponding values ​​can be found in the R feature statistical values ​​of the first object. For example, if the dwell time is 35 seconds, page jumps are 4, and browsing category diversity is 0.9, then the probability of the first object being in a browsing state is determined to be 66.66%.

[0048] S102. Based on the historical intent state sequence of the first object and the current interaction environment features, a prediction is made to obtain the predicted intent state distribution corresponding to the first object; the predicted intent state distribution refers to the probability distribution of the intent state of the first object in a future time step.

[0049] Specifically, a feasible implementation process for predicting the distribution of predicted intention states corresponding to the first object based on the historical intention state sequence and current interaction environment features can be as follows: perform multi-step intention state prediction processing based on the historical intention state sequence and current interaction environment features of the first object to obtain K predicted intention state distributions corresponding to the first object; among the K predicted intention state distributions, the i-th predicted intention state distribution contains the probability of the intention state corresponding to the L intention states at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K.

[0050] Specifically, a feasible implementation process for multi-step intent state prediction processing based on the historical intent state sequence of the first object and the features of the current interaction environment to obtain the K predicted intent state distributions corresponding to the first object can be as follows: Perform temporal encoding processing on the historical intent state sequence of the first object to obtain a historical intent state feature vector; perform feature fusion processing on the feature vectors corresponding to the historical intent state feature vector and the current interaction environment features to obtain an initial prediction input vector; traverse K future time steps, and in the intent state prediction processing for the x-th future time step, obtain the predicted intent state distribution corresponding to the (x-1)-th future time step; x is a positive integer less than or equal to K; perform state update processing on the (x-1)-th predicted input vector according to the predicted intent state distribution corresponding to the (x-1)-th future time step to obtain the x-th predicted input vector; when x-1 is 1, the (x-1)-th predicted input vector is the initial predicted input vector; perform intent recognition processing on the x-th predicted input vector to obtain the predicted intent state distribution corresponding to the x-th future time step; when traversing K future time steps is completed, the K predicted intent state distributions corresponding to the first object are obtained. The temporal coding process can be implemented using a Transformer encoder (a deep learning model based on self-attention mechanism). Feature fusion can be achieved through methods such as concatenation, weighted summation, and attention fusion, which are not limited in this application. A feasible implementation of updating the state of the (x-1)th predicted input vector based on the predicted intention state distribution corresponding to the (x-1)th future time step to obtain the (x)th predicted input vector can be: integrating the predicted intention state distribution corresponding to the (x-1)th future time step into the (x-1)th predicted input vector to obtain the (x)th predicted input vector. The intention recognition process can be described by the following formula (2): pred = softmax(predictor[x-1](current_input)) formula (2) Here, `predictor[x-1]` refers to the dedicated predictor for the x-th future time step (e.g., x-1=0 is the predictor for the 1st future time step, and x-1=1 is the predictor for the 2nd future time step). Essentially, it's a neural network layer (such as a linear layer) specifically responsible for predicting the intent for the x-th future time step. `current_input` refers to the (x-1)-th predicted input vector. The `softmax()` function performs a softmax transformation on the raw scores output by the predictor, converting the rangeless scores into probability values ​​between 0 and 1, with the sum of all probabilities being 1. Therefore, `pred` refers to the predicted intent state distribution corresponding to the x-th future time step.

[0051] S103. Based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution, perform a multi-dimensional confidence assessment on the first object to obtain a comprehensive confidence score.

[0052] Specifically, the predicted intention state distribution includes K predicted intention state distributions; the i-th predicted intention state distribution refers to the probability distribution of the intention state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K. A feasible implementation process for evaluating the multi-dimensional confidence of the first object based on the current intention state distribution, current interaction environment characteristics, and predicted intention state distribution to obtain the comprehensive confidence is as follows: Determine the predicted intention state transition trajectory based on the current intention state distribution and the K predicted intention state distributions; the predicted intention state transition trajectory includes the current intention state and the K predicted intention states; the current intention state refers to the intention state with the highest probability of corresponding intention state in the current intention state distribution; the y-th predicted intention state refers to the intention state at the y-th future time step... The intention state with the highest probability corresponding to the predicted intention state distribution is identified; y is a positive integer less than or equal to K; trajectory similarity is obtained by comparing the predicted intention state transition trajectory and the standard intention state transition trajectory; the standard intention state transition trajectory is generated based on the target historical intention state transition trajectories corresponding to at least two sample objects; the target historical intention state transition trajectory refers to the intention state transition trajectory whose starting state is the current state; stability is assessed for the current interaction environment features to obtain the current interaction environment stability; deterministic assessment is performed on the K predicted intention state distributions to obtain the intention state prediction certainty; and a weighted fusion of trajectory similarity, current interaction environment stability, and intention state prediction certainty is performed to obtain the comprehensive confidence score.

[0053] One feasible implementation process for trajectory comparison processing can be represented by the following formula (3): consistency = max(DTW_similarity(trajectory, pattern) for pattern instandard_patterns) Formula (3) Wherein, consistency refers to trajectory similarity, trajectory refers to the predicted intention state transition trajectory, standard_patterns is the standard intention state transition trajectory library, pattern is a standard intention state transition trajectory in the standard intention state transition trajectory library, and DTW_similarity function is the Dynamic Time Warping (DTW) similarity function, used to calculate the similarity between two time series of different lengths. It can be understood that the meaning of formula (3) is to traverse all referenceable standard intention state transition trajectories in the standard intention state transition trajectory library, calculate the DTW similarity between the predicted intention state transition trajectory and each referenceable standard intention state transition trajectory, and finally select the maximum similarity value as the trajectory similarity.

[0054] One feasible implementation process for stability assessment can be represented by the following formula (4): stability = mean([time_stability, device_stability, source_stability]) formula (4) Here, stability refers to the stability of the current interaction environment. `time_stability` is the time stability score, ranging from [0,1], representing the stability of the user's access to the recommendation system over time. It can be obtained by calculating the variance or similarity of the consecutive access timestamps of the first object. The smaller the variance (or the higher the similarity), the closer the score is to 1. `device_stability` is the device stability score, ranging from [0,1], representing the stability of the terminal device features of the first object accessing the recommendation system. If the first object continuously uses the same device, the score is close to 1; if it frequently switches devices (phone → tablet → computer), the score is close to 0. `source_stability` is the source stability score, ranging from [0,1], representing the stability of the entry source features of the first object accessing the recommendation system. Entry sources include the application homepage, public account jumps, third-party links, etc. The more singular the source, the closer the score is to 1. The `mean` function is the arithmetic mean function, which calculates the average of the stability scores of multiple dimensions of the input by weighting them equally.

[0055] One feasible implementation process for determining the certainty of intention state prediction by performing deterministic evaluation on the distribution of K predicted intention states can be as follows: calculate the entropy value of each of the K predicted intention state distributions to obtain K basic entropy values; perform weighted summation of the K basic entropy values ​​according to the time decay weights corresponding to the K predicted intention states to obtain a weighted total entropy value; normalize the weighted total entropy value to obtain a normalized weighted total entropy value; and determine the difference between the deterministic threshold and the normalized weighted total entropy value as the certainty of intention state prediction.

[0056] S104. Obtain the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content.

[0057] Specifically, the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at future time steps.

[0058] Specifically, the current intent state recommendation score can also be used to characterize the degree to which the target candidate content recommends the first object in the current intent state. The current intent state refers to the intent state with the highest probability at the current time step among L intent states. The future intent state recommendation score can also be used to characterize the degree to which the target candidate content recommends the first object in the future intent state; the future intent state refers to the intent state with the highest probability at the future time step among L intent states.

[0059] Specifically, the content described in this application refers to the collective term for various digital information resources that can be recommended to the first object in the content recommendation system. The content may include, but is not limited to, audio and video content, graphic and text information content, product display content, live streaming content, dynamic information content, and other information content that can be digitally pushed and viewed or interacted with by the first object.

[0060] Specifically, the target candidate content can be any candidate content in the content recommendation candidate set. This content recommendation candidate set can be generated when a content recommendation request is received from the first object. The content recommendation request can refer to a request initiated by the first object based on needs such as content browsing, content retrieval, and personalized push notifications, seeking target content. The content recommendation candidate set refers to the set of content selected from a massive full-content resource library through preset filtering, matching, and recall algorithms that match the first object's content recommendation request and conform to the first object's content preferences and related characteristics. It can be understood that the content recommendation candidate set is a rough screening of massive amounts of content; to attract the interest of the first object, the candidate content needs to be precisely ranked to obtain the content recommendation set sent to the first object.

[0061] Specifically, a feasible implementation process for obtaining the current intent state recommendation score corresponding to the target candidate content can be as follows: Obtain the current intent state features of the first object and the content features of the candidate content; perform recommendation prediction processing based on the current intent state features, content features, and the intent-content matching matrix to obtain the current intent state recommendation score. The intent-content matching matrix can be constructed based on massive historical data to quantify the degree of preference of different intent states for different content features. It can be understood that the process of obtaining the future intent state recommendation score can refer to the process of obtaining the current intent state recommendation score, and will not be elaborated upon here.

[0062] S105. Based on the comprehensive confidence level, the current intent state recommendation score and the future intent state recommendation score are weighted and fused to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

[0063] Specifically, the future intent state recommendation score comprises K future intent state recommendation scores; the i-th future intent state recommendation score is used to characterize the recommendation degree of the target candidate content for the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; a feasible implementation process for weighted fusion processing of the current intent state recommendation score and the future intent state recommendation score based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content can be as follows: the minimum value between the comprehensive confidence level and the maximum threshold of the future intent weight is determined as the basic weight of the future intent state; the difference between the total weight threshold and the basic weight of the future intent state is determined as the weight of the current intent state; the time decay weights corresponding to the K future intent state recommendation scores are multiplied by the basic weights of the future intent states respectively to obtain the future intent state decay weights corresponding to the K future intent state recommendation scores respectively; the product between the current intent state weight and the current intent state recommendation score, and the product between the future intent state decay weight corresponding to each future intent state recommendation score and each future intent state recommendation score are added together to obtain the final recommendation score corresponding to the target candidate content. The maximum threshold for future intent weight is a pre-set threshold, which can be 0.4 or adjusted based on actual circumstances. The total weight threshold is 1. For ease of understanding, assuming a comprehensive confidence level of 0.3 and a maximum threshold for future intent weight of 0.4, the base weight for the future intent state is determined to be 0.3, and the weight for the current intent state is 0.7. In the calculation, we determine two future intent state recommendation scores: the time decay weight for the first future intent state recommendation score is 0.6, and the time decay weight for the second future intent state recommendation score is 0.4. Therefore, the time decay weight for the first future intent state recommendation score is 0.3 × 0.6 = 0.18, and the time decay weight for the second future intent state recommendation score is 0.3 × 0.4 = 0.12. The final recommendation score = 0.7 × current intent state recommendation score + 0.18 × first future intent state recommendation score + 0.12 × two future intent state recommendation scores.

[0064] The method provided in this application divides the object's interest cognition into L intention states. This not only quickly determines the current intention state of the first object, but also predicts the future intention state of the first object in the next K time steps based on the historical intention state sequence and current interaction environment features. The reliability of the prediction is quantified by a comprehensive confidence metric. Finally, the recommendation score of the current intention state and the recommendation scores of the K future intention states are weighted and fused based on the comprehensive confidence metric. The final recommendation score not only considers the current intention of the first object, but also the influence of the evolution of the first object's future intention on the recommended content, making the recommendation result closer to the real needs of the first object, thereby improving the recommendation effect.

[0065] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0066] The methods of the embodiments of this application have been described in detail above, and the apparatus of the embodiments of this application is provided below.

[0067] Please see Figure 4 , Figure 4 This is a schematic diagram of a content recommendation device provided in an embodiment of this application. The content recommendation device 1 includes: a first acquisition module 11, a prediction module 12, an evaluation module 13, a second acquisition module 14, and a weighting module 15. Specifically: The first acquisition module 11 is used to acquire the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; The prediction module 12 is used to make predictions based on the historical intention state sequence of the first object and the current interaction environment features to obtain the predicted intention state distribution corresponding to the first object; the predicted intention state distribution refers to the probability distribution of the intention state of the first object in a future time step. Evaluation module 13 is used to perform multi-dimensional confidence evaluation on the first object based on the current intent state distribution, the current interaction environment features and the predicted intent state distribution, to obtain a comprehensive confidence score; The second acquisition module 14 is used to acquire the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the future time step; The weighting module 15 is used to perform weighted fusion processing on the current intent state recommendation score and the future intent state recommendation score according to the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

[0068] In any embodiment of this application, the first acquisition module 11 is specifically used for: Obtain the target multidimensional feature sequence of the first object; the target multidimensional feature sequence contains at least two multidimensional features; the interaction timestamps corresponding to the at least two multidimensional features are all greater than the target timestamp; the multidimensional features are composed of behavioral features, preference features, and interaction environment features. The position weight of each multidimensional feature is determined based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence; the sequence position of each multidimensional feature is sorted based on the interaction timestamp. Each multidimensional feature is weighted according to its corresponding position weight to obtain the weighted target multidimensional feature sequence. The first object is subjected to intent recognition processing based on the weighted target multidimensional feature sequence to obtain the current intent state distribution of the first object.

[0069] In any embodiment of this application, the prediction module 12 is specifically used for: Based on the historical intent state sequence of the first object and the current interaction environment features, multi-step intent state prediction processing is performed to obtain K predicted intent state distributions corresponding to the first object; among the K predicted intent state distributions, the i-th predicted intent state distribution refers to the probability distribution of the intent state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K.

[0070] In any embodiment of this application, the prediction module 12 is specifically used for: The historical intent state sequence of the first object is subjected to temporal encoding to obtain the historical intent state feature vector; The feature vectors corresponding to the historical intent state feature vector and the current interaction environment feature vector are subjected to feature fusion processing to obtain the initial prediction input vector; Traverse K future time steps. In the intention state prediction processing for the x-th future time step, obtain the predicted intention state distribution corresponding to the (x-1)-th future time step; x is a positive integer less than or equal to K. The (x-1)th predicted input vector is updated according to the predicted intention state distribution corresponding to the (x-1)th future time step to obtain the xth predicted input vector; when x-1 is 1, the (x-1)th predicted input vector is the initial predicted input vector; The x-th predicted input vector is processed for intent recognition to obtain the predicted intent state distribution corresponding to the x-th future time step; When the K future time steps are completed, the distribution of the K predicted intention states corresponding to the first object is obtained.

[0071] In any embodiment of this application, the predicted intention state distribution includes K predicted intention state distributions; the i-th predicted intention state distribution refers to the probability distribution of the intention state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; the evaluation module 13 is specifically used for: The predicted intention state transition trajectory is determined based on the current intention state distribution and the K predicted intention state distributions; the predicted intention state transition trajectory includes the current intention state and K predicted intention states; the current intention state refers to the intention state with the highest probability in the current intention state distribution; the y-th predicted intention state refers to the intention state with the highest probability in the y-th predicted intention state distribution; y is a positive integer less than or equal to K; The predicted intention state transition trajectory and the standard intention state transition trajectory are compared to obtain the trajectory similarity; the standard intention state transition trajectory is generated based on the target historical intention state transition trajectories corresponding to at least two sample objects; the target historical intention state transition trajectory refers to the intention state transition trajectory whose starting state is the current state. The stability of the current interactive environment is obtained by performing a stability evaluation on the current interactive environment features. A deterministic evaluation process is performed on the K predicted intention state distributions to obtain the intention state prediction certainty. The trajectory similarity, the stability of the current interaction environment, and the certainty of the intention state prediction are weighted and fused to obtain a comprehensive confidence score.

[0072] In any embodiment of this application, the evaluation module 13 is specifically used for: Entropy values ​​are calculated for each of the K predicted intention state distributions to obtain K basic entropy values; The weighted total entropy value is obtained by weighting and summing the K basic entropy values ​​according to the time decay weights corresponding to the K predicted intention states. The weighted total entropy value is normalized to obtain the normalized weighted total entropy value; The difference between the deterministic threshold and the normalized weighted total entropy value is determined as the determination of the intention state prediction.

[0073] In any embodiment of this application, the future intent state recommendation score includes K future intent state recommendation scores; the i-th future intent state recommendation score is used to characterize the recommendation degree of the target candidate content for the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; the weighting module 15 is specifically used for: The minimum value between the overall confidence level and the maximum threshold of the future intention weight is determined as the basic weight of the future intention state; The difference between the total weight threshold and the basic weight of the future intention state is determined as the weight of the current intention state; Multiply the time decay weights corresponding to the K future intention state recommendation scores by the base weights of the future intention states to obtain the future intention state decay weights corresponding to the K future intention state recommendation scores. The final recommendation score corresponding to the target candidate content is obtained by adding the product between the current intent state weight and the current intent state recommendation score, and the product between the future intent state decay weight corresponding to each future intent state recommendation score and each future intent state recommendation score.

[0074] Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device 4 includes a processor 41 and a memory 42. Optionally, the electronic device 4 also includes an input device 43 and an output device 44. The processor 41, memory 42, input device 43, and output device 44 are coupled together via connectors, which include various interfaces, transmission lines, or buses, etc., and are not limited in this embodiment. It should be understood that in the various embodiments of this application, coupling refers to mutual connection in a specific way, including direct connection or indirect connection through other devices, such as through various interfaces, transmission lines, buses, etc.

[0075] Processor 41 may include one or more processors, such as one or more central processing units (CPUs). If the processor is a CPU, it may be a single-core CPU or a multi-core CPU. Optionally, processor 41 may be a processor group consisting of multiple CPUs, with the multiple processors coupled to each other via one or more buses. Optionally, the processor may also be other types of processors, etc., which are not limited in this embodiment.

[0076] The memory 42 can be used to store computer program instructions, as well as various types of computer program code, including program code for executing the present application. Optionally, the memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM), which is used for related instructions and data.

[0077] Input device 43 is used to input data and / or signals, and output device 44 is used to output data and / or signals. Input device 43 and output device 44 can be independent devices or an integrated device.

[0078] It is understood that in this embodiment of the application, the memory 42 can be used not only to store related instructions, but also to store related data. This embodiment of the application does not limit the specific data stored in the memory.

[0079] Understandable Figure 5 This is merely a simplified design of an electronic device. In practical applications, the electronic device may also include other necessary components, including, but not limited to, any number of input / output devices, processors, memories, etc., and all electronic devices that can implement the embodiments of this application are within the protection scope of this application.

[0080] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0081] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will also readily understand that the various embodiments of this application have different focuses, and for the sake of convenience and brevity, the same or similar parts may not be repeated in different embodiments. Therefore, parts not described or not described in detail in one embodiment can be referred to the descriptions in other embodiments.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0085] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0086] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A content recommendation method, characterized in that, The method includes: Obtain the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; Based on the historical intent state sequence of the first object and the characteristics of the current interaction environment, a prediction is made to obtain the predicted intent state distribution corresponding to the first object; the predicted intent state distribution refers to the probability distribution of the intent state of the first object in a future time step. Based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution, a multidimensional confidence assessment is performed on the first object to obtain a comprehensive confidence score. Obtain the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the future time step; The current intent state recommendation score and the future intent state recommendation score are weighted and fused based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

2. The method according to claim 1, characterized in that, The step of obtaining the current intent state distribution of the first object includes: Obtain the target multidimensional feature sequence of the first object; the target multidimensional feature sequence contains at least two multidimensional features; the interaction timestamps corresponding to the at least two multidimensional features are all greater than the target timestamp; the multidimensional features are composed of behavioral features, preference features, and interaction environment features. The position weight of each multidimensional feature is determined based on its sequence position in the target multidimensional feature sequence and the sequence length of the target multidimensional feature sequence; the sequence position of each multidimensional feature is sorted based on the interaction timestamp. Each multidimensional feature is weighted according to its corresponding position weight to obtain the weighted target multidimensional feature sequence. The first object is subjected to intent recognition processing based on the weighted target multidimensional feature sequence to obtain the current intent state distribution of the first object.

3. The method according to claim 1, characterized in that, The step of predicting the distribution of predicted intention states corresponding to the first object based on the historical intention state sequence and current interaction environment features of the first object includes: Based on the historical intent state sequence of the first object and the current interaction environment features, multi-step intent state prediction processing is performed to obtain K predicted intent state distributions corresponding to the first object; among the K predicted intent state distributions, the i-th predicted intent state distribution refers to the probability distribution of the intent state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K.

4. The method according to claim 3, characterized in that, The step of performing multi-step intent state prediction processing based on the historical intent state sequence of the first object and the current interaction environment features to obtain K predicted intent state distributions corresponding to the first object includes: The historical intent state sequence of the first object is subjected to temporal encoding to obtain the historical intent state feature vector; The feature vectors corresponding to the historical intent state feature vector and the current interaction environment feature vector are subjected to feature fusion processing to obtain the initial prediction input vector; Traverse K future time steps. In the intention state prediction processing for the x-th future time step, obtain the predicted intention state distribution corresponding to the (x-1)-th future time step; x is a positive integer less than or equal to K. The (x-1)th predicted input vector is updated according to the predicted intention state distribution corresponding to the (x-1)th future time step to obtain the xth predicted input vector; when x-1 is 1, the (x-1)th predicted input vector is the initial predicted input vector; The x-th predicted input vector is processed for intent recognition to obtain the predicted intent state distribution corresponding to the x-th future time step; When the K future time steps are completed, the distribution of the K predicted intention states corresponding to the first object is obtained.

5. The method according to claim 1, characterized in that, The predicted intention state distribution includes K predicted intention state distributions; the i-th predicted intention state distribution refers to the probability distribution of the intention state of the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; The step of performing a multi-dimensional confidence assessment on the first object based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution to obtain a comprehensive confidence score includes: The predicted intention state transition trajectory is determined based on the current intention state distribution and the K predicted intention state distributions; the predicted intention state transition trajectory includes the current intention state and K predicted intention states; the current intention state refers to the intention state with the highest probability in the current intention state distribution; the y-th predicted intention state refers to the intention state with the highest probability in the y-th predicted intention state distribution; y is a positive integer less than or equal to K; The predicted intention state transition trajectory and the standard intention state transition trajectory are compared to obtain the trajectory similarity; the standard intention state transition trajectory is generated based on the target historical intention state transition trajectories corresponding to at least two sample objects; the target historical intention state transition trajectory refers to the intention state transition trajectory whose starting state is the current state. The stability of the current interactive environment is obtained by performing a stability evaluation on the current interactive environment features. A deterministic evaluation process is performed on the K predicted intention state distributions to obtain the intention state prediction certainty. The trajectory similarity, the stability of the current interaction environment, and the certainty of the intention state prediction are weighted and fused to obtain a comprehensive confidence score.

6. The method according to claim 5, characterized in that, The deterministic evaluation process of the K predicted intention state distributions to obtain the intention state prediction certainty includes: Entropy values ​​are calculated for each of the K predicted intention state distributions to obtain K basic entropy values; The weighted total entropy value is obtained by weighting and summing the K basic entropy values ​​according to the time decay weights corresponding to the K predicted intention states. The weighted total entropy value is normalized to obtain the normalized weighted total entropy value; The difference between the deterministic threshold and the normalized weighted total entropy value is determined as the determination of the intention state prediction.

7. The method according to claim 1, characterized in that, The future intention state recommendation score includes K future intention state recommendation scores; the i-th future intention state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the i-th future time step; K is a positive integer, and i is a positive integer less than or equal to K; The step of weighting and fusing the current intent state recommendation score and the future intent state recommendation score based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content includes: The minimum value between the overall confidence level and the maximum threshold of the future intention weight is determined as the basic weight of the future intention state; The difference between the total weight threshold and the basic weight of the future intention state is determined as the weight of the current intention state; Multiply the time decay weights corresponding to the K future intention state recommendation scores by the base weights of the future intention states to obtain the future intention state decay weights corresponding to the K future intention state recommendation scores. The final recommendation score corresponding to the target candidate content is obtained by adding the product between the current intent state weight and the current intent state recommendation score, and the product between the future intent state decay weight corresponding to each future intent state recommendation score and each future intent state recommendation score.

8. A content recommendation device, characterized in that, The device includes: The first acquisition module is used to acquire the current intent state distribution of the first object; the current intent state distribution refers to the probability distribution of the intent state of the first object at the current time step; The prediction module is used to predict the predicted intention state distribution of the first object based on the historical intention state sequence and the current interaction environment features; the predicted intention state distribution refers to the probability distribution of the intention state of the first object in a future time step. The evaluation module is used to perform a multi-dimensional confidence evaluation on the first object based on the current intent state distribution, the current interaction environment features, and the predicted intent state distribution, to obtain a comprehensive confidence score. The second acquisition module is used to acquire the current intent state recommendation score and the future intent state recommendation score corresponding to the target candidate content; the current intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the current time step; the future intent state recommendation score is used to characterize the degree of recommendation of the target candidate content to the first object at the future time step; The weighting module is used to perform weighted fusion processing on the current intent state recommendation score and the future intent state recommendation score based on the comprehensive confidence level to obtain the final recommendation score corresponding to the target candidate content; the final recommendation score is used to determine the ranking position of the target candidate content in the content recommendation set.

9. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program or instructions; when the computer program or instructions are executed on a computer, the computer causes the computer to perform the method according to any one of claims 1 to 7.

Images