Training method of intention recognition model and content recommendation method based on consumption intention

Abstract

The invention provides an intention recognition model training method and a consumption intention-based content recommendation method, and relates to the technical field of computers, in particular to the technical fields of deep learning, information flow, model training and the like. According to the specific implementation scheme, the method comprises the steps of obtaining subsequent interaction data associated with a click behavior of an entry resource; determining a first interaction record generated by executing the first resource recommendation strategy on the information flow and a second interaction record generated by executing the second resource recommendation strategy from the subsequent interaction data; marking resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain index to obtain a positive and negative sample pair; and training the first intention recognition model according to the positive and negative sample pairs and the advantage and disadvantage sorting relationship thereof to obtain a trained second intention recognition model. According to the technical scheme of the embodiment of the invention, the model can better learn and understand the consumption intention by comparing the value gains of different recommendation strategies.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to technical fields such as deep learning, information flow, and model training. Background Art

[0002] With the explosive growth of internet information, recommendation systems have become an indispensable core component of modern web applications. By analyzing users' historical behavior and preferences, they proactively filter massive amounts of information and push content (such as products, news, and videos) that may be of interest to users, significantly improving their efficiency and experience in obtaining information.

[0003] Information feeds, exemplified by short videos and graphic updates, are becoming the mainstream model for content consumption. In this model, users sequentially consume content provided by a recommendation system by continuously swiping or clicking. Research has found that after consuming a piece of content of interest (hereinafter referred to as a "portal"), users often develop a desire to continue consuming related or similar content. This phenomenon can be termed "extended consumption." Accurately identifying and addressing users' extended consumption intentions is crucial for improving user engagement and usage time. Summary of the Invention

[0004] The present disclosure provides a method for training an intent recognition model, a method, an apparatus, a device, and a storage medium for a content recommendation method based on consumption intent.

[0005] According to one aspect of the present disclosure, a method for training an intent recognition model is provided, comprising:

[0006] Obtain subsequent interaction data associated with click behavior on the entry resource;

[0007] Determining, from the subsequent interaction data, a first interaction record generated by executing the first resource recommendation strategy on the information flow, and a second interaction record generated by executing the second resource recommendation strategy;

[0008] Based on a preset value-gain indicator, the resource consumption behaviors in the first interaction record and the second interaction record are annotated to obtain positive and negative sample pairs;

[0009] The first intent recognition model is trained according to the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model.

[0010] According to another aspect of the present disclosure, a content recommendation method based on consumption intention is provided, comprising:

[0011] In response to a click operation on any entry resource, determining a target feature of the entry resource;

[0012] Input the target features into the extended consumption intention recognition model to obtain the target object's extended consumption intention intensity score for the entry resource;

[0013] Determine subsequent recommended content for the current session based on the extended consumption intention strength score;

[0014] Among them, the extended consumption intention recognition model is trained according to the training method of the intention recognition model provided in any embodiment of the present disclosure.

[0015] According to another aspect of the present disclosure, a training apparatus for an intent recognition model is provided, comprising:

[0016] An acquisition module, used to obtain subsequent interaction data associated with the click behavior of the entry resource;

[0017] a record determination module, configured to determine, from subsequent interaction data, a first interaction record generated by executing a first resource recommendation strategy on an information flow, and a second interaction record generated by executing a second resource recommendation strategy;

[0018] A sample construction module is used to annotate the resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain indicator to obtain positive and negative sample pairs;

[0019] The training module is used to train the first intent recognition model based on the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model.

[0020] According to another aspect of the present disclosure, a content recommendation device based on consumption intention is provided, comprising:

[0021] a feature determination module, configured to determine target features of an entry resource in response to a click operation on any entry resource;

[0022] An intention determination module is used to input the target features into the extended consumption intention recognition model to obtain the target object's extended consumption intention intensity score for the entry resource;

[0023] The recommendation module is used to determine subsequent recommended content for the current session based on the extended consumption intention strength score;

[0024] Among them, the extended consumption intention recognition model is trained according to the training method of the intention recognition model provided in any embodiment of the present disclosure.

[0025] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0026] at least one processor; and

[0027] a memory communicatively connected to the at least one processor; wherein,

[0028] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any method in the embodiments of the present disclosure.

[0029] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute any method according to the embodiments of the present disclosure.

[0030] According to another aspect of the present disclosure, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the computer program implements any one of the methods according to the embodiments of the present disclosure.

[0031] The technical solutions of the embodiments of the present disclosure can better learn and understand consumer intentions by comparing the value gains of different recommendation strategies.

[0032] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0034] Figure 1 is a flowchart of a method for training an intent recognition model according to an embodiment of the present disclosure;

[0035] Figure 2 is a schematic diagram of an APP interface provided according to an embodiment of the present disclosure;

[0036] Figure 3 This is a schematic diagram of the process of dividing subsequent interactive data provided by an embodiment of the present disclosure;

[0037] Figure 4 According to an embodiment of the present disclosure, an architectural diagram of a network structure of an intent recognition model is provided;

[0038] Figure 5 1 is a flowchart of a content recommendation method based on consumption intention provided according to an embodiment of the present disclosure;

[0039] Figure 6 1 is a schematic structural diagram of a training device for an intent recognition model according to an embodiment of the present disclosure;

[0040] Figure 7 is a structural diagram of a content recommendation device based on consumption intention provided according to an embodiment of the present disclosure;

[0041] Figure 8 is a block diagram of an electronic device for implementing embodiments of the present disclosure. DETAILED DESCRIPTION

[0042] Exemplary embodiments of the present disclosure are described herein below with reference to the accompanying drawings, in which various details of embodiments of the present disclosure are set forth to facilitate an understanding, and should be considered in a descriptive sense. A person skilled in the art will readily recognize that various alternative embodiments of the present disclosure maybe practiced without departing from the scope of the present disclosure. Similarly, it will be appreciated that, in the interest of clarity and brevity, descriptions of all features of the present disclosure are not necessarily disclosed herein.

[0043] In related art, in order to predict the behavior intention of a user, a recommendation system usually builds a complex deep learning model. A common technical path is to use a pointwise modeling method. Specifically, the model predicts a specific interaction indicator for the next content to be recommended in the information stream. For example, a binary classification model is built to predict whether a user will click on the content, or a regression model is built to predict the viewing time of the user on the content. The system sorts the candidate contents according to the prediction scores output by the model, and presents the content with the highest score to the user.

[0044] In the training process of the above model, feature engineering is a key link that determines the performance of the model. Existing technologies usually use features of multiple dimensions to describe a recommendation scenario. These features generally include: user portrait features that describe the long-term preferences of a user; content features that describe the attributes of the content itself (such as category, label, keyword, etc.); and context features that describe the environment in which the current recommendation occurs (such as time, geographic location, device type, etc.). These features are collectively input into the model to learn the association between the features and the user behavior (such as clicking).

[0045] However, the consumption behavior of a user is a continuous and dynamically evolving process. Simply predicting the single-point interaction behavior of a user on the next independent content may not fully capture the deep intention and satisfaction of the user in the entire consumption session. Therefore, how to more deeply understand and model the extended consumption mindset of a user in an information stream scenario, and how to more comprehensively measure the overall value gain brought to a user by a recommendation service, have become important directions for continuous exploration and optimization in the field of recommendation technology.

[0046] To at least partially solve one or more of the above problems and other potential problems, embodiments of the present disclosure provide a training method of an intention recognition model. By comparing the value gains of different recommendation strategies, the technical solutions of the embodiments of the present disclosure can better learn and understand the consumption intention.

[0047] Figure 1 FIG. 1 is a flow chart of a method for training an intent recognition model according to an embodiment of the present disclosure. Figure 1 As shown, the method comprises at least the following steps:

[0048] S110: Obtain subsequent interaction data associated with the click behavior of the entry resource.

[0049] S120: Determine, from the subsequent interaction data, a first interaction record generated by executing the first resource recommendation strategy on the information flow, and a second interaction record generated by executing the second resource recommendation strategy.

[0050] S130: Based on a preset value gain indicator, the resource consumption behaviors in the first interaction record and the second interaction record are marked to obtain positive and negative sample pairs.

[0051] S140: Training the first intent recognition model according to the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model.

[0052] In this embodiment, the target object can be understood as a terminal user who obtains and consumes recommended resources through an application on a terminal device. The entry resource can be understood as the initial media content that the target object clicks on in the information flow formed by the recommended resources or in the resource list, such as a short video or a graphic article, or it can also be resources such as commodities, songs, news information, and articles. Subsequent interaction data can refer to a series of behavioral data that occurs after the target object clicks on the entry resource, usually including recorded data generated by interactive behaviors such as playing, liking, and commenting on the recommended resource.

[0053] This embodiment can collect comparative data by distinguishing the interaction records generated by two recommendation strategies (one is a special first resource recommendation strategy, and the other is a conventional second resource recommendation strategy). Subsequently, these records are annotated using a value-gain indicator (a quantitative indicator that measures the improvement of user experience or consumer value) to form positive and negative sample pairs for model training. Finally, the ranking relationship between the sample pairs (i.e., which sample is better) is used to train an initial first intent recognition model. After training and optimization, a second intent recognition model with accurate prediction capabilities is obtained.

[0054] According to the solution of the embodiment of the present disclosure, a model can be trained to better understand the user's consumption intention. The model not only learns whether the user will click on the recommended resources, but also learns which recommendation method can obtain recommended resources that can bring higher user value.

[0055] Figure 2It is a schematic diagram of an application program (APP) interface provided according to an embodiment of the present disclosure.

[0056] like Figure 2 As shown, in an application scenario example, the APP can be a search engine APP. The page that the user enters after opening the APP can be called a list page. The list page can display a search box and multiple contents for users to choose from. The user can enter the recommendation page by clicking on a certain content.

[0057] The content that users click on on a list page is called a portal resource. When a user clicks on this portal resource, they are redirected to a fusion page, which can incorporate multiple content styles. The fusion page initially displays the portal resource, and users can refresh the recommended resources displayed on the fusion page by swiping up.

[0058] In one possible implementation, subsequent interaction data is in sessions. Each session includes a series of interaction records during a continuous use of the target object. The interaction records include at least:

[0059] Records clicks on any entry resource in the resource list page.

[0060] A record of resource consumption activities that occur on the resource integration page related to the entry resource. The session starts when the user navigates to the resource integration page and ends when the user returns to the resource list page.

[0061] In this embodiment, a session refers to a continuous period of interaction between a target object and an application. Figure 2 As shown, a preferred definition is that a session for analyzing extended consumption begins when a target user clicks on a portal resource from a general resource list page (such as the "Recommendations" or "Home" information flow of an app) and is redirected to a resource fusion page specifically presented for that resource (a page that aggregates playback, comment interactions, and related recommendations for the portal resource). The session ends when the target user returns from the fusion page to the previous list page or directly exits the app.

[0062] According to the solution of the embodiment of the present disclosure, by binding the session with the page jump behavior, all behavioral data related to an extended consumption intention can be accurately captured, forming a data chain with clear logic and clear cause and effect (click on the entry resource - enter the fusion page - consume on the fusion page - leave the fusion page), which greatly improves the accuracy of subsequent data analysis and sample annotation.

[0063] In a possible implementation, the first resource recommendation strategy is used to extract resource content from an extended acceptance queue related to an entry resource, and present it to a target object as a subsequent resource in an information flow.

[0064] The second resource recommendation strategy is used to recall subsequent resources from the personalized recommendation pool based on the profile characteristics and / or historical behavior characteristics of the target object.

[0065] In this embodiment, the first resource recommendation strategy is that when the system determines that the target object may have extended consumption intentions, it will select content from a pre-built extended acceptance queue. The content in the queue is highly relevant to the entry resource, such as the same theme, the same creator, or highly similar content style. It is also called an extended acceptance strategy. The second resource recommendation strategy is a conventional personalized recommendation strategy. It mainly recalls content that may be of interest to the target object from a huge personalized recommendation pool, but is not necessarily directly related to the entry resource, based on the long-term portrait features of the target object and recent historical behavior features (such as recently liked and viewed content).

[0066] In order to more clearly understand the two recommendation strategies defined in this embodiment, a typical short video information flow application scenario will be used as an example for explanation.

[0067] Assume that the basic profile of a target object (username: Xiao Zhang) is: female, white-collar worker, and has a high interest in travel and coffee culture.

[0068] Scenario Example 1: Executing the First Resource Recommendation Strategy (Extension and Succession Strategy)

[0069] Xiao Zhang was attracted by a video titled "Learn How to Make Pour-Over Coffee in One Minute" in the app's "Recommended" feed. She clicked on it and watched it in its entirety. This video became the entry point for this interaction.

[0070] The recommendation system's backend intent recognition model receives the characteristics of this interaction (including the entry resource ID, user Xiao Zhang's ID, 100% viewing completion, and the current leisure time), and the system decides to implement the first resource recommendation strategy. It temporarily sets aside the conventional personalized recommendation logic and instead searches for the next content to recommend from an extended queue related to the topic of "hand-poured coffee." This queue may include several highly relevant videos such as "How to Choose Your First Coffee Bean" and "V60 Filter Brewing Tutorial."

[0071] The system selects the highest-ranked video from the queue, such as "V60 Filter Cup Brewing Tutorial", and puts it as the next video in the information flow to present to Xiao Zhang.

[0072] After watching the introductory video, Xiao Zhang immediately moved on to the advanced tutorial. Her interest was captured and sustained, and she was likely to continue watching, significantly increasing her engagement with the "coffee" theme and the duration of each session. The system recorded this complete interaction, from "Learn Hand-Poured Coffee in One Minute" to "V60 Filter Brewing Tutorial," as the first interaction record for subsequent analysis and model training.

[0073] Scenario Example 2: Executing the Second Resource Recommendation Strategy (Conventional Personalization Strategy)

[0074] Similarly, user Xiao Zhang watched the entire video "Learn to Make Hand-Poured Coffee in One Minute".

[0075] In another hypothetical parallel scenario, the system decides to implement the second resource recommendation strategy. It activates the standard personalized recommendation logic and begins analyzing Xiao Zhang's comprehensive interests. Based on her profile and historical behavior, the system discovers that, in addition to coffee, she has recently shown a strong interest in "travel to Japan" and has watched several related vlogs.

[0076] After sorting through a vast pool of personalized recommendations, the system determined that a Kyoto travel video, "Wandering Through Kyoto's Quiet Alleys," would better suit Xiao Zhang's overall interests. The system then placed this travel video as the next piece of content in her feed.

[0077] Xiao Zhang then sees a travel video she might also be interested in. This is still an effective personalized recommendation. However, it interrupts the user's immediate flow on the topic of "coffee." The system records this complete interaction from "Learn to Pour Over Coffee in One Minute" to "Wandering Through the Quiet Alleys of Kyoto" as a second interaction record.

[0078] By comparing these two scenario examples, it can be clearly seen that this embodiment can distinguish and collect interaction data under two different strategies, providing a high-quality, comparable data foundation for subsequently measuring the pros and cons of the two strategies through value-gain indicators, and ultimately training the model to learn when to execute the first resource recommendation strategy.

[0079] This example clearly defines two recommendation strategies, creating the necessary conditions for subsequent comparative analysis (A / B testing) and model training. By comparing user behavior and value gains under the two strategies, the model can learn when extended recommendation is a better option than regular recommendation.

[0080] In one possible implementation, S130 labels the resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain indicator to obtain positive and negative sample pairs, further comprising the steps of:

[0081] S131: Calculate the subsequent consumption time length of the target object in the preset physical time length bucket in the first interaction record and the second interaction record, respectively, as a value gain indicator.

[0082] S132: For the same target object or the same type of target object, compare the subsequent consumption time length of the first interaction record with the subsequent consumption time length of the second interaction record. And

[0083] S133: According to the comparison result, determine the superior-inferior relationship order of the first interaction record and the second interaction record.

[0084] In the embodiments of the present disclosure, the value gain indicator is embodied as the subsequent consumption time length. In order to ensure the fairness of the comparison, the concept of preset physical time length bucket is introduced. The physical time length bucket is a category divided according to the video or content time length of the recommended subsequent resource itself. For example, the following buckets can be set in advance:

[0085] Short content bucket: resources with a time length of 0-60 seconds.

[0086] Medium content bucket: resources with a time length of 60-300 seconds.

[0087] Long content bucket: resources with a time length of more than 300 seconds.

[0088] In the S131 step, when the value gain indicator (i.e. the subsequent consumption time length) of an interaction record needs to be calculated, it will first check the time length of the recommended subsequent resource itself in the record, and put it into the corresponding physical time length bucket. For example, a first interaction record recommends a 45-second extended video, which will be put into the short content bucket.

[0089] When comparing in the S132 step, the system will follow the principle of same-bucket comparison. That is, the first interaction record described above which recommends a 45-second video, its calculated "subsequent consumption time length" will only be compared with the subsequent consumption time length of a second interaction record which also recommends short content (such as a 30-second video). It will not be compared with a record which recommends a 10-minute video, because such comparison is unfair.

[0090] In the S133 step, based on the fair comparison result within the same bucket, the system finally determines the superior-inferior relationship order of the two interaction records.

[0091] According to the solution of the embodiment of the present disclosure, by introducing a physical duration bucketing mechanism based on the duration of the resource itself, the natural impact of the recommended material's own attributes (duration) on the user's subsequent behavior is isolated. This allows subsequent value gain comparisons to more purely and accurately measure the pros and cons of the recommendation strategy itself (first strategy vs. second strategy) rather than the pros and cons of the content duration. This provides model training with more reliable sample labels with a higher signal-to-noise ratio, thereby significantly improving the accuracy and reliability of the ultimately trained intent recognition model.

[0092] In a possible implementation, S132 compares the subsequent consumption duration of the first interaction record with the subsequent consumption duration of the second interaction record for the same target object or the same type of target objects, including:

[0093] S132a: Determine, from the first interaction record, a target consumption behavior record that effectively interacts with the extended resource content.

[0094] S132b: Compare the subsequent consumption duration of the target consumption behavior record with the subsequent consumption duration of the second interaction record.

[0095] In this embodiment, effective interaction can be understood as consumption behavior that truly reflects the interests of the target object, rather than accidental touch or unintentional browsing. One or more thresholds can be set for judgment, for example: video viewing time exceeding 10 seconds, video completion rate exceeding 50%, or strong interactive behaviors such as likes / comments / shares. Before performing the comparison, the first interaction record is first filtered to retain only those records with effective interactions, that is, the target consumption behavior records. The subsequent consumption duration of these high-quality records is then compared with the subsequent consumption duration of the second interaction record.

[0096] The target consumption behavior record can be considered as an effective interactive behavior for extended consumption. For an interactive behavior to be judged as a successful extended consumption, the following conditions must be met in sequence:

[0097] 1. Triggering a session: A user clicks on and consumes a piece of content (video, image, text, etc.) in a regular information flow (such as a recommendation feed or short video flow).

[0098] 2. Recommendation System Intervention and Implementation: If the recommendation system (through intent recognition models or heuristics) determines that the current time is a good time to "extend consumption," it will change the default recommendation logic to an extended consumption strategy. At this point, it will retrieve the next content or batch of content from a specially prepared extended consumption queue related to the entry resource (for example, videos on the same theme, other works by the same author, or collections of content of similar style).

[0099] The system takes the content out of the extended continuation queue and presents it to the user as the next content.

[0100] 3. User response:

[0101] The user sees the relevant content that is "taken over" by the system. The user generates effective positive consumption behavior for this content, such as:

[0102] Watched the video in its entirety or to a high degree of completion.

[0103] The image or text remains on screen for longer than the preset time.

[0104] There was strong interaction such as likes, comments, and shares.

[0105] Only when all three steps are completed can the entire process be defined as a successful interactive behavior for extended consumption. For example, if Xiao Zhang engaged in positive consumption of the "V60 Filter Brewing Tutorial," this would be considered a record of targeted consumption behavior that effectively interacted with the extended resource content.

[0106] According to the solution of the embodiment of the present disclosure, the accuracy and signal-to-noise ratio of the strategy pros and cons judgment are improved. This ensures that the pros and cons relationship used for training is based on the user's true intention, thereby improving the reliability of the final trained model.

[0107] In one possible implementation, the model training method of the embodiment of the present disclosure further includes:

[0108] S310: Mark the second consumption behavior record that has no effective interaction with the extended resource content as a disadvantageous sample.

[0109] S320: Taking any second interaction record or any target consumption behavior record as a dominant sample to form a positive and negative sample pair.

[0110] In the disclosed embodiment, the second consumption behavior record specifically refers to the behavior record in which the target object does not have effective interaction after the system executes the first resource recommendation strategy (extended succession), that is, the behavior record of invalid consumption after succession. The most typical example is that Xiao Zhang swipes away immediately after the video appears in the application and starts playing. Still taking the above scenario example 1 as an example, if Xiao Zhang does not make effective consumption of the content taken over by the system, it is considered that there is no effective consumption after succession, thus forming a second consumption behavior record.

[0111] Second, consumption behavior records can be used as high-quality and low-quality samples. In the framework of pairwise learning, a low-quality sample cannot exist alone. It must be paired with a high-quality sample so that the model can learn the relationship between them.

[0112] Advantage samples can come from two sources: Strategy A is to pair them with ordinary regular recommendations, that is, with any second interaction record.

[0113] Sample pair example:

[0114] P (Advantage): User C watched a "latte art" video -> The system performed a regular recommendation and pushed a "weekend travel" video -> User C watched it for dozens of seconds, and the session duration met the baseline level.

[0115] N (disadvantage): User C watched the "latte art" video -> the system executed the follow-up strategy and pushed a "hand-brew kettle recommendation" video -> User C immediately swiped away.

[0116] This pairing approach aims to teach the model that a failed extension attempt is worth less than a typical action. The model learns to assign a very low score to scenarios in which it lacks confidence. This score not only indicates the likelihood of a successful extension, but also indicates that the expected benefit of executing the offer is lower than the average benefit of doing nothing (implementing the typical recommendation). This helps the model make more conservative and safe decisions in uncertain situations, avoiding sacrificing the user's current experience for an uncertain extension opportunity.

[0117] Strategy B:

[0118] Paired with a successful extension follow-up, that is, with any target consumer behavior record, this is intended to teach the model to finely distinguish between successful extensions and failed extensions.

[0119] According to the solution of the embodiment of the present disclosure, the user's negative feedback signals are fully utilized to construct high-quality negative samples, so that the model not only learns what is good, but also learns more deeply what is clearly bad, thereby significantly improving the decision-making accuracy of the model and avoiding interference to users.

[0120] Sample pair example:

[0121] P (Advantage): User A watched the "Latte Art" video -> the system took over the "Coffee Bean Tasting" video -> User A watched it in its entirety, and the session duration increased.

[0122] N (disadvantage): User B watched the "latte art" video -> the system launched the "hand brewing kettle recommendation" video -> User B immediately swiped away.

[0123] The goal of model learning: This pairing approach allows the model to perform highly refined learning. Two examples may have very similar antecedents (users viewed similar entry content, and the system executed the follow-up strategy in both cases), but very different outcomes. The model is forced to learn these subtle differences. By learning these pairings, the model can more accurately predict the success probability of the extended follow-up action itself, thereby optimizing when, for whom, and with what content the extended follow-up strategy should be executed.

[0124] Figure 3 This is a schematic diagram of the process of dividing the subsequent interactive data provided by the embodiment of the present disclosure. Figure 3 As shown in the figure, the subsequent interaction data is divided into three parts:

[0125] ●First division: based on whether the system intervenes

[0126] All data is first split into two parts by a key system behavior label:

[0127] A. Historical interactive behaviors without extended consumption

[0128] In this part of the data, after the user clicks on the entry resource, the system does not implement a special extended consumption strategy, but instead pushes the next content to the user according to the conventional personalized recommendation logic.

[0129] This data can be used as a metric to calculate value gain. It can also be used to calculate baseline follow-up consumption duration, which is how long users typically stay in the app without special intervention.

[0130] B. Interaction records of the extended consumption acceptance strategy

[0131] In this part of the data, the system proactively intervened and provided subsequent content from a dedicated, relevant extension queue after the user clicked on the entry resource.

[0132] ●Second division: based on whether the user responds

[0133] After the system has implemented the takeover strategy, it needs to be divided according to user response:

[0134] C. Interactive records of effective consumption behavior

[0135] Users have generated positive and quantifiable consumption behaviors towards the content received by the system. For example:

[0136] Video viewing time > 10 seconds

[0137] Video viewing completion rate > 50%

[0138] Strong interactions such as likes, comments, and shares occurred

[0139] Image and text dwell time > average page reading time

[0140] This data is a subset of interaction records that implement the extended consumption strategy, namely the target consumption behavior records. Only the data in this subset will be used to determine the value gain in the next step.

[0141] D. No effective consumption after taking over

[0142] Users do not effectively consume the content taken over by the system.

[0143] It is also a subset of the interaction records that implement the extended consumption acceptance strategy, that is, the second consumption behavior record, and is mutually exclusive with the interaction records of effective consumption behavior.

[0144] ●The third division: whether value gain is achieved

[0145] For those records of users who have generated effective consumption, they are divided according to value gain:

[0146] E. There is consumption and positive value gain

[0147] Not only did the users effectively consume the subsequent content, but their subsequent total consumption time was significantly higher than the baseline subsequent consumption time calculated in the first step.

[0148] F. Consumption but no value gain

[0149] Although the user consumed the transferred content, their subsequent total consumption time was equal to or lower than the baseline. This means that the system's transfer operation was useless or even negative in terms of user time value.

[0150] In one possible implementation, S140 trains the first intent recognition model based on the positive and negative sample pairs and their ranking order to obtain a trained second intent recognition model, including:

[0151] S141: Input the superior sample resource features and the inferior sample resource features in the positive and negative sample pairs into the first intent recognition model respectively.

[0152] S142: Obtain a first prediction score and a second prediction score corresponding to the superior sample and the inferior sample according to the prediction result of the first intention recognition model.

[0153] S143: Determine a loss value for measuring the difference between the prediction result and the ranking relationship between the positive and negative sample pairs, the first prediction score, and the second prediction score.

[0154] S144: According to the loss value, adjust the model parameters of the first intent recognition model through the back propagation algorithm to obtain a trained second intent recognition model.

[0155] In the disclosed embodiment, the features of a sample pair (comprising an advantageous sample A and an inferior sample B) are fed into the intent recognition model, and the model predicts A and B respectively, and outputs two scores (a first prediction score and a second prediction score). The actual order of merit is A>B, and the model will calculate a loss value based on this relationship and the prediction score. If the model prediction is wrong (the first prediction score ≤ the second prediction score), the loss value will be very large. This loss value will be used through the back-propagation algorithm to guide the model on how to adjust its internal parameters, and the direction of adjustment is to make the first prediction score larger and the second prediction score smaller. Through multiple rounds of iterative training, the trained second intent recognition model is finally obtained after the loss function converges or the number of training times reaches a threshold.

[0156] According to the solution of the disclosed embodiments, through this relative relationship-based training mechanism, the model's goal is no longer to fit an absolute score, but to learn how to correctly rank different recommendation opportunities. This ensures that the model's prediction results are highly consistent with the actual business goal (selecting the optimal recommendation strategy), and the resulting second intent recognition model has strong generalization capabilities and decision-making accuracy.

[0157] Figure 4 This is an architectural diagram of a network structure of an intent recognition model according to an embodiment of the present disclosure. Figure 4 As shown in the figure, this model is called Intention Net. The network adopts a twin network structure, which mainly includes symmetrical left tower (Intention Net) and right tower (Intention Net), and the top-level differential comparison and loss calculation module.

[0158] The left and right towers of the model have exactly the same network topology and hierarchy, and share the same set of model parameters (weights) during training.

[0159] During training, the features of the dominant sample in a positive-negative pair (e.g., the first interaction record with higher value gain) are input to the left tower. Meanwhile, the features of the disadvantaged sample (e.g., the second interaction record with lower value) are input to the right tower.

[0160] The left tower and the right tower process the input sample features respectively, and finally output a prediction score (i.e., the first prediction score and the second prediction score) respectively.

[0161] At the top level of the model, the output score of the left tower is subtracted from the output score of the right tower to get a score difference.

[0162] The difference is then fed into a loss function (e.g., logistic regression loss or Hinge loss) to calculate the final loss value for backpropagation, thereby driving the parameter update of the entire model. The final output is the label for classification.

[0163] In one possible implementation, S141 inputs the superior sample resource features and the inferior sample resource features in the positive and negative sample pairs into the first intent recognition model, including:

[0164] S141a: Obtain entry features based on at least one of the entry resource identification feature, entry resource author identification feature, category feature, basic information feature, genre feature, multimodal feature, and refined multi-objective Q value.

[0165] S141b: Obtain the resource characteristics of the dominant sample based on at least one of the entry characteristics and target object characteristics, scenario consumption characteristics, list page click representation, fusion page satisfaction representation, fusion page consumption characteristics, and list page consumption characteristics corresponding to the dominant sample.

[0166] S141c: Obtain the disadvantaged sample resource characteristics based on at least one of the entry characteristics and target object characteristics, scenario consumption characteristics, list page click representation, fusion page satisfaction representation, fusion page consumption characteristics, and list page consumption characteristics corresponding to the disadvantaged sample.

[0167] S141d: Inputting the superior sample resource features and the inferior sample resource features into the first intent recognition model respectively.

[0168] In the embodiment of the present disclosure, Figure 4 As shown, entry features can be understood as a set of features that describe the properties of the entry resource itself. S141b and S141c define the complete feature vector (advantage / disadvantage sample resource features) of the final input model. This vector is composed of two parts: one is the entry feature, and the other is features that describe the context and user status. For example, these include target object features (user profile), consumption scenario features (time and location), list page click representation (what else the user clicked on the list page), and fusion page satisfaction / consumption features (real-time behavioral feedback in the current session).

[0169] According to the solution of the disclosed embodiments, by building a multi-dimensional, information-rich feature system, the model can fully understand an interaction from multiple perspectives, including the user, content, real-time scenario, and historical sequence. This fundamental guarantee for significantly improving model accuracy, allowing model decisions to no longer be based on one-sided information, but on a comprehensive assessment of the overall situation.

[0170] In a possible implementation, the multi-modal feature is obtained by encoding at least one of the title, cover, picture or sound of the entry resource.

[0171] In this embodiment, the multi-modal feature refers to a deep feature extracted directly from the original form of the content beyond a simple text label. In specific implementation, the multi-modal feature includes the following aspects:

[0172] Title: The title can be encoded into a title Embedding vector containing semantic information by a natural language processing model such as BERT.

[0173] Cover: The cover can be encoded into a cover Embedding vector representing the image content and style by a convolutional neural network (CNN) such as ResNet.

[0174] Picture: Key frames can be extracted from the video and encoded by using a CNN.

[0175] Sound: The feature vector of the audio waveform of the background music or human voice can be extracted by analyzing the audio waveform.

[0176] According to the scheme of the embodiment of the present disclosure, the model can understand and listen to the content, rather than only reading the label. This makes the model understand the content more deeply, especially in judging the content similarity and style consistency, which is far beyond the traditional label-based method, so that a more human-intuitive extended recommendation can be made.

[0177] The embodiment of the present disclosure introduces the internal structure and feature acquisition of the single tower:

[0178] Since the left and right towers have the same structure, one of the towers is taken as an example to illustrate the internal data processing flow and feature acquisition method in detail.

[0179] S410: The feature acquisition and representation learning layer is responsible for processing the input raw data and generating various features required by the model. These features will be finally spliced together in the feature fusion layer.

[0180] S411: Obtain entry features

[0181] Concept introduction: This is a core feature set that describes the properties of the entry resource that triggers this prediction.

[0182] Acquisition method: As shown in the "entry feature-list page click" box in the figure, it is composed of multiple sub-features, and the specific implementation is as follows:

[0183] Multi-modal feature: As in the previous embodiment, the multi-modal feature is obtained by encoding the title, cover, picture, sound, etc. of the entry resource.

[0184] Category features (cate), genre features, and basic information features: structured information directly obtained from the content management system, such as "food", "Vlog", number of likes, number of collections, etc.

[0185] ID features (ID): include entry resource ID, author ID, etc., which are usually converted into low-dimensional dense vectors through an embedding layer.

[0186] Precision sorting multi-objective Q value (Q value): The estimated value of various a posteriori indicators of the entry resource (such as estimated duration, estimated completion rate, etc.) obtained from the precision sorting model responsible for the main information flow.

[0187] S412: Constructing Sequence Click / Consumption Representations

[0188] List page click representation:

[0189] Concept introduction: This feature aims to capture the user's short-term interests before entering the current session (i.e., clicking on the entry resource).

[0190] Acquisition method: As shown in the "List Page Click Point Set" module in the lower left corner of the figure, the system obtains a series of items (Xnid2, ..., Xnid20) that the user clicked on the list page before clicking the entry resource. The features of these items are fed into an encoder (for example, using a Transformer or Attention mechanism). The encoder outputs a context vector Hnid that aggregates information about all clicked items, which serves as the final "List Page Click Representation."

[0191] Integration page satisfaction indicators:

[0192] Concept introduction: This feature aims to dynamically capture the evolution trend of user satisfaction in the current session.

[0193] Acquisition method: As shown in the "Fusion Page Satisfaction Sequence" module in the lower right corner of the figure, the system obtains a series of content (Snid1, Snid2, ..., Snid20) that the user has consumed sequentially on the current fusion page and inputs their features into another encoder. This encoder captures the contextual information of the sequence, such as whether the user becomes increasingly engaged or gradually loses interest. The output context vector Cnid is the "Fusion Page Satisfaction Representation."

[0194] S413: Obtaining other context and statistical features

[0195] User characteristics: long-term static or slowly changing features obtained from the user portrait system, such as the user's age group, gender, long-term interest tags, etc.

[0196] Scenario consumption features: describe the macro-environmental features at the time of the current interaction, such as the current time (whether it is a weekend, holiday), network type (Wi-Fi / 5G), geographic location, etc.

[0197] List page / fusion page consumption features: refer to some real-time, simple statistics-based cross features or ratio features, such as the overall click rate of the user on the current list page, the average viewing time that has occurred on the current fusion page, etc.

[0198] S420: Feature fusion and prediction layer

[0199] S421: Feature concatenation

[0200] Concatenate all the feature vectors obtained or generated in the above steps S411 to S413 in dimension, to form a very high-dimensional and very rich fusion feature vector.

[0201] S422: Multilayer Perceptron (MLP) processing

[0202] Input the concatenated fusion feature vector into an MLP network. The MLP network usually includes multiple fully connected layers with nonlinear activation functions (such as ReLU).

[0203] S423: Output prediction score

[0204] The last layer of the MLP network is usually a single neuron without activation function or using linear activation, and its output is the final prediction score for the current sample (for example, the first prediction score output by the left tower).

[0205] The intent network model architecture shown in this embodiment has the following significant technical effects:

[0206] 1. The twin network structure and top layer difference design support the Pairwise learning paradigm based on the superiority-inferiority relationship, making the model training target highly unified with the ranking task target.

[0207] 2. By introducing an encoder with a Transformer or similar structure, the model can deeply mine the user's short-term dynamic interest and satisfaction changes from the unsorted click set (list page click set) and the time-ordered consumption sequence (fusion page satisfaction sequence), far exceeding traditional static features.

[0208] 3. The model integrates the multi-dimensional features of the content itself (entry features), long-term user portraits, real-time contextual scenarios, and dynamic sequential behavior representations, achieving a panoramic portrayal of a recommendation interaction and providing a solid foundation for accurate prediction.

[0209] 4. The design of shared parameters between the left and right towers greatly reduces the total number of model parameters, improves training efficiency, and helps the model learn a more generalizable metric function.

[0210] Figure 5 FIG is a flow chart of a content recommendation method based on consumption intention according to an embodiment of the present disclosure. Figure 5 As shown, the method includes the following steps:

[0211] S510: In response to a click operation on any entry resource, determine a target feature of the entry resource.

[0212] S520: Inputting the target features into the extended consumption intention recognition model to obtain the extended consumption intention intensity score of the target object for the entry resource.

[0213] S530: Determine subsequent recommended content for the current session based on the extended consumption intention strength score.

[0214] The extended consumption intention recognition model is trained according to any one of the methods in claims 1 to 9.

[0215] This example describes the model's online inference process. When a real target user clicks on an entry resource online, the recommendation system collects the target features of this scenario in real time. These features are fed into a trained secondary intent recognition model, which outputs a score: the strength of the extended consumption intention. A higher score indicates that the model believes the expected value of executing an extended consumption strategy is higher at this point. The recommendation system uses this score to determine whether to execute an extended consumption strategy or a regular recommendation.

[0216] According to the solution of the embodiment of the present disclosure, the offline trained model is applied to the online recommendation scenario that requires real-time and fast response, which can better improve the user experience when recommending resources based on the user's consumption intention.

[0217] In one possible implementation, S530 determines subsequent recommended content for the current session based on the extended consumption intention strength score, including:

[0218] S531: When the extended consumption intention strength score reaches a threshold, obtain extended content related to the entry resource.

[0219] S532: The extended content is placed as recommended content in the resource display position or information flow of the current session.

[0220] In the disclosed embodiment, a threshold is set in advance. The determination of this threshold usually requires finding an optimal balance point through offline backtesting and online A / B testing to weigh the accuracy and recall rate of extended recommendations. When the extended consumption intention intensity score output by the model in S520 is greater than or equal to this threshold, it is determined that this is a high-value extension opportunity, and relevant content is obtained from the extension acceptance queue. In S532, these contents will be inserted into the next position of the information flow and presented to the user. If the score is lower than the threshold, the system will give up this extension opportunity and continue to perform regular recommendations.

[0221] The technical effect of this embodiment is to provide a simple, efficient and easy-to-control online decision switch. By adjusting the threshold, the radicalism or conservatism of the extension recommendation strategy can be flexibly controlled to meet the business needs of different periods.

[0222] In one possible implementation, the target feature includes at least one of the following:

[0223] At least one of the identification feature of the entry resource, the identification feature of the author of the entry resource, the category feature, the basic information feature, the genre feature, the multimodal feature, and the refined multi-objective Q value is obtained.

[0224] The multimodal feature is obtained by encoding at least one of the title, cover, picture or sound of the entry resource.

[0225] The multi-objective Q value is obtained based on at least one of the estimated consumption time Q value, the estimated fast sliding rate Q value, the estimated completion rate Q value, and the estimated sliding rate Q value determined by the entry resource in the refined ranking model.

[0226] Figure 6 Schematic diagram of the structure of the training device of the intention recognition model provided in accordance with an embodiment of the present disclosure. Figure 6 As shown, the training device 600 includes:

[0227] The acquisition module 601 is used to acquire subsequent interaction data associated with the click behavior of the entry resource.

[0228] The record determination module 602 is configured to determine, from subsequent interaction data, a first interaction record generated by executing a first resource recommendation strategy on an information flow, and a second interaction record generated by executing a second resource recommendation strategy.

[0229] The sample construction module 603 is configured to mark the resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain indicator to obtain positive and negative sample pairs.

[0230] The training module 604 is configured to train the first intent recognition model according to the positive and negative sample pairs and the superior-inferior order relationship thereof, to obtain the trained second intent recognition model.

[0231] In a possible implementation, the subsequent interaction data is in units of sessions, and each session includes a series of interaction records of the target object in a continuous use process, and the interaction records at least include:

[0232] a click behavior record of any entry resource in the resource list page; and

[0233] a series of resource consumption behavior records occurring in the resource fusion page related to the entry resource; wherein the session starts when the session is adjusted to the resource fusion page, and ends when the session returns to the resource list page.

[0234] In a possible implementation, the first resource recommendation strategy is used to extract resource content from the extension queue related to the entry resource, as a subsequent resource in the information stream presented to the target object;

[0235] The second resource recommendation strategy is used to recall subsequent resources from the personalized recommendation pool according to the portrait features and / or historical behavior features of the target object.

[0236] In a possible implementation, the sample construction module 603 is configured to:

[0237] respectively calculate the subsequent consumption time length of the target object in a preset physical time length bucket in the first interaction record and the second interaction record as the value gain indicator;

[0238] for the same target object or the same type of target object, compare the subsequent consumption time length of the first interaction record with the subsequent consumption time length of the second interaction record; and

[0239] determine the superior-inferior order relationship of the first interaction record and the second interaction record according to the comparison result.

[0240] In a possible implementation, the sample construction module 603 is configured to:

[0241] determine, from the first interaction record, a target consumption behavior record that has an effective interaction with the extension resource content;

[0242] compare the subsequent consumption time length of the target consumption behavior record with the subsequent consumption time length of the second interaction record.

[0243] In a possible implementation, the apparatus further includes a labeling module configured to:

[0244] label a second consumption behavior record that does not have an effective interaction with the extension resource content as a disadvantage sample.

[0245] Any second interaction record or any target consumption behavior record is used as a dominant sample to form a positive and negative sample pair.

[0246] In one possible implementation, the training module 604 is configured to:

[0247] Inputting the superior sample resource features and the inferior sample resource features in the positive and negative sample pairs into the first intent recognition model respectively;

[0248] Obtaining a first prediction score and a second prediction score corresponding to the superior sample and the inferior sample according to the prediction result of the first intention recognition model;

[0249] Determine a loss value for measuring the difference between the prediction result and the ranking relationship of the positive and negative sample pairs, the first prediction score, and the second prediction score;

[0250] According to the loss value, the model parameters of the first intent recognition model are adjusted through the back propagation algorithm to obtain a trained second intent recognition model.

[0251] In one possible implementation, the training module 604 is configured to:

[0252] Obtaining entry features based on at least one of an entry resource identification feature, an entry resource author identification feature, a category feature, a basic information feature, a genre feature, a multimodal feature, and a refined ranking multi-objective Q value;

[0253] Obtaining the resource characteristics of the dominant sample based on at least one of the entry characteristics and target object characteristics, scenario consumption characteristics, list page click characteristics, fusion page satisfaction characteristics, fusion page consumption characteristics, and list page consumption characteristics corresponding to the dominant sample;

[0254] Obtaining resource features of the disadvantaged sample based on at least one of the entry features and target object features, scenario consumption features, list page click representation, fusion page satisfaction representation, fusion page consumption features, and list page consumption features corresponding to the disadvantaged sample;

[0255] The superior sample resource features and the inferior sample resource features are input into the first intent recognition model respectively.

[0256] In a possible implementation, the multimodal feature is obtained by encoding at least one of the title, cover, image, or sound of the entry resource.

[0257] Figure 7 Schematic diagram of the structure of a content recommendation device based on consumption intention according to an embodiment of the present disclosure. Figure 7 As shown, the content recommendation device 700 includes:

[0258] The feature determination module 701 is configured to determine target features of an entry resource in response to a click operation on any entry resource.

[0259] The intention determination module 702 is used to input the target features into the extended consumption intention recognition model to obtain the extended consumption intention intensity score of the target object for the entry resource.

[0260] The recommendation module 703 is configured to determine subsequent recommended content for the current session based on the extended consumption intention strength score.

[0261] Among them, the extended consumption intention recognition model is obtained by training the intention recognition model training method provided by any embodiment of the present disclosure.

[0262] In a possible implementation, the recommendation module 703 is configured to:

[0263] When the extended consumption intention intensity score reaches a threshold, the extended content related to the entry resource is obtained;

[0264] Place the extended content as recommended content in the resource display position or information flow of the current session.

[0265] In one possible implementation, the target feature includes at least one of the following:

[0266] At least one of the following is obtained: an identification feature of an entry resource, an identification feature of an entry resource author, a category feature, a basic information feature, a genre feature, a multimodal feature, and a refined multi-objective Q value;

[0267] The multimodal feature is obtained by encoding at least one of the title, cover, image or sound of the entry resource;

[0268] The multi-objective Q value is obtained based on at least one of the estimated consumption time Q value, the estimated fast sliding rate Q value, the estimated completion rate Q value, and the estimated sliding rate Q value determined by the entry resource in the refined ranking model.

[0269] For the description of specific functions and examples of each module and submodule of the device in the embodiment of the present disclosure, please refer to the relevant description of the corresponding steps in the above method embodiment, which will not be repeated here.

[0270] In the technical solutions disclosed herein, the acquisition, storage, and application of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0271] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0272] Figure 8A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0273] like Figure 8 As shown, the device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. Various programs and data required for the operation of the device 800 can also be stored in the RAM 803. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0274] Various components in device 800 are connected to I / O interface 805, including an input unit 806, such as a keyboard, mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, optical disk, etc.; and a communication unit 809, such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows device 800 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0275] The computing unit 801 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above, such as the training method of the intent recognition model or the content recommendation method based on consumption intent. For example, in some embodiments, the training method of the intent recognition model or the content recommendation method based on consumption intent can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the training method of the intent recognition model or the content recommendation method based on consumption intent described above can be performed. Alternatively, in other embodiments, the computing unit 801 may be configured in any other appropriate manner (eg, by means of firmware) to execute a training method for an intent recognition model or a content recommendation method based on consumption intent.

[0276] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0277] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0278] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0279] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0280] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0281] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0282] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0283] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A method for training an intent recognition model, comprising: Obtain subsequent interaction data associated with click behavior on the entry resource; Determining, from the subsequent interaction data, a first interaction record generated by executing a first resource recommendation strategy on the information flow, and a second interaction record generated by executing a second resource recommendation strategy; Based on a preset value gain indicator, the resource consumption behaviors in the first interaction record and the second interaction record are marked to obtain positive and negative sample pairs; The first intent recognition model is trained according to the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model.

2. The method according to claim 1, wherein The subsequent interaction data is in sessions. Each session includes a series of interaction records during a continuous use of the target object. The interaction records at least include: Record the click behavior of any entry resource on the resource list page; as well as A series of resource consumption behavior records that occur in the resource fusion page related to the entry resource; wherein, the session starts when adjusting to the resource fusion page and ends when returning to the resource list page.

3. The method according to claim 1, wherein The first resource recommendation strategy is used to extract resource content from the extended acceptance queue related to the entry resource and present it to the target object as a subsequent resource in the information flow; The second resource recommendation strategy is used to recall subsequent resources from the personalized recommendation pool based on the profile features and / or historical behavior features of the target object.

4. The method according to claim 1, wherein The tagging of resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain indicator to obtain positive and negative sample pairs includes: Calculating the subsequent consumption duration of the target object in the first interaction record and the second interaction record within the preset physical duration buckets respectively as the value gain indicator; For the same target object or the same type of target object, comparing the subsequent consumption duration of the first interaction record with the subsequent consumption duration of the second interaction record; and According to the comparison result, a ranking of the first interaction record and the second interaction record is determined.

5. The method according to claim 4, wherein The comparing, for the same target object or the same type of target object, the subsequent consumption duration of the first interaction record with the subsequent consumption duration of the second interaction record includes: Determining, from the first interaction record, a target consumer behavior record that effectively interacts with the extended resource content; Compare the subsequent consumption duration of the target consumption behavior record with the subsequent consumption duration of the second interaction record.

6. The method according to claim 5, further comprising: The second consumption behavior records that do not effectively interact with the extended resource content are marked as inferior samples; Any of the second interaction records or any of the target consumption behavior records is used as a dominant sample to form a positive and negative sample pair.

7. The method according to claim 1, wherein The first intent recognition model is trained according to the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model, including: Inputting the superior sample resource features and the inferior sample resource features in the positive and negative sample pairs into the first intent recognition model respectively; Obtaining a first prediction score and a second prediction score corresponding to the superior sample and the inferior sample according to the prediction result of the first intention recognition model; Determining a loss value for measuring the difference between the prediction result and the ranking relationship of the positive and negative sample pairs, the first prediction score, and the second prediction score; According to the loss value, the model parameters of the first intent recognition model are adjusted through a back propagation algorithm to obtain a trained second intent recognition model.

8. The method according to claim 7, wherein: The step of inputting the superior sample resource features and the inferior sample resource features in the positive and negative sample pairs into the first intent recognition model comprises: Obtaining entry features based on at least one of the entry resource identification feature, entry resource author identification feature, category feature, basic information feature, genre feature, multimodal feature, and refined ranking multi-objective Q value; Obtaining advantageous sample resource characteristics based on at least one of the entry characteristics and target object characteristics, scenario consumption characteristics, list page click characteristics, fusion page satisfaction characteristics, fusion page consumption characteristics, and list page consumption characteristics corresponding to the advantageous sample; Obtaining a resource feature of the disadvantaged sample based on at least one of the entry feature and target object feature, scenario consumption feature, list page click representation, fusion page satisfaction representation, fusion page consumption feature, and list page consumption feature corresponding to the disadvantaged sample; The advantageous sample resource features and the disadvantageous sample resource features are respectively input into the first intent recognition model.

9. The method according to claim 8, wherein The multimodal feature is obtained by encoding at least one of the title, cover, picture or sound of the entry resource.

10. A content recommendation method based on consumption intent, comprising: In response to a click operation on any entry resource, determining a target feature of the entry resource; Inputting the target feature into an extended consumption intention recognition model to obtain an extended consumption intention intensity score of the target object for the entry resource; Determining subsequent recommended content for the current session based on the extended consumption intention strength score; The extended consumption intention recognition model is trained according to the method described in any one of claims 1 to 9.

11. The method according to claim 10, wherein: Determining subsequent recommended content for the current session based on the extended consumption intention strength score includes: When the extended consumption intention strength score reaches a threshold, obtaining extended content related to the entry resource; The extended content is placed as recommended content in the resource display position or information flow of the current session.

12. The method according to claim 10, wherein: The target feature includes at least one of the following: Obtaining at least one of the identification feature of the entry resource, the identification feature of the author of the entry resource, the category feature, the basic information feature, the genre feature, the multimodal feature, and the refined multi-objective Q value; The multimodal feature is obtained by encoding at least one of the title, cover, image or sound of the entry resource; The multi-objective Q value is obtained based on at least one of the estimated consumption time Q value, the estimated fast sliding rate Q value, the estimated completion rate Q value, and the estimated sliding rate Q value determined by the entry resource in the refined ranking model.

13. A training device for an intent recognition model, comprising: An acquisition module, used to obtain subsequent interaction data associated with the click behavior of the entry resource; a record determination module, configured to determine, from the subsequent interaction data, a first interaction record generated by executing a first resource recommendation strategy on the information flow, and a second interaction record generated by executing a second resource recommendation strategy; a sample construction module, configured to annotate resource consumption behaviors in the first interaction record and the second interaction record based on a preset value gain indicator to obtain positive and negative sample pairs; The training module is used to train the first intent recognition model according to the positive and negative sample pairs and their quality ranking relationship to obtain a trained second intent recognition model.

14. A content recommendation device based on consumption intention, comprising: a feature determination module, configured to determine a target feature of any entry resource in response to a click operation on the entry resource; an intention determination module, configured to input the target features into an extended consumption intention recognition model to obtain a strength score of the target object's extended consumption intention for the entry resource; a recommendation module, configured to determine subsequent recommended content for the current session based on the extended consumption intention strength score; The extended consumption intention recognition model is trained according to the method described in any one of claims 1 to 9.

15. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 12.

16. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-12.

17. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 12.

Images