A pre-estimation ranking method and device, a computer device and a storage medium

Abstract

The application discloses a kind of estimation ranking method, device, computer equipment and storage medium, belong to artificial intelligence technical field.The application is based on the exposure event carrying first behavior and the exposure event not carrying first behavior to construct first sample set and based on the exposure event carrying second behavior and the exposure event not carrying second behavior to construct second sample set, based on first sample set and second sample set fitting training to multi-task learning model, obtain estimation ranking model, import user data into estimation ranking model and obtain information estimation ranking result, complete information push based on information estimation ranking result.In addition, the application also relates to blockchain technology, and user data can be stored in blockchain.The application is fitted by first behavior feature and second behavior feature in multi-task learning model, solves sample selection bias problem, improves model generalization ability, improves prediction accuracy and user experience.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a predictive sorting method, apparatus, computer equipment, and storage medium. Background Technology

[0002] In information retrieval scenarios such as search and recommendation, the most basic goal is the user's CTR (Click-Through-Rate), which represents whether a user will click to read a piece of content after seeing it. However, CTR only represents the click-to-read behavior, while user behavior on a product can be diverse, such as liking, saving, sharing, commenting, and conversion (purchase). Such behaviors need to be measured by the model using a unified paradigm standard.

[0003] While it's possible to optimize a user's CTR (Click-Through Rate) for a single objective, this approach is somewhat one-sided and prone to producing clickbait headlines, leading to negative consequences. Content recommended based solely on user clicks may not be entirely satisfactory to the user. While user readership may increase, other behaviors (likes, favorites, shares, etc.) may decrease. This decline indicates that users are receiving too much irrelevant information. Therefore, it's necessary to integrate reading, liking, favorites, sharing, conversion (purchase), and other behaviors into a single model for learning—a multi-objective learning recommendation system. Multi-objective ranking utilizes multiple objective functions to find a comprehensive ranking method that optimizes all objectives overall.

[0004] Currently, the latest multi-objective learning recommendation systems are based on the multi-task learning framework MMOE, which evolved from OMOE (One-gate Mixture of Experts). MMOE is actually a variant and optimization of MOE for multi-task learning. Compared to the structure of OMOE where all tasks share a single gating network, the structure of MMOE is optimized so that each task uses a separate gating network. This improvement allows for different Experts weights to be obtained for different tasks, thereby enabling selective utilization of Experts. The gating networks corresponding to different tasks can learn different combinations of Experts, making it easier for the model to capture the correlation and differences between sub-tasks.

[0005] However, the aforementioned multi-objective learning recommendation models still have some shortcomings in CVR (Conversion Rate) prediction. First, multi-task recommendation models like MMOE only use samples with click behavior to train the CVR prediction model, but the trained model is used to make inferences and predictions across the entire sample space. Since click events are much fewer than impression events, they represent only a small subset of the sample space. This phenomenon—where training samples are extracted from a small subset of the overall sample space, while the trained model needs to make inferences and predictions based on samples from the entire sample space—is called sample selection bias. Sample selection bias can harm the generalization performance of the learned model. Summary of the Invention

[0006] The purpose of this application is to propose a prediction ranking method, apparatus, computer device, and storage medium to solve the problem that existing solutions have sample selection bias when training multi-objective learning recommendation models. The sample selection bias problem leads to low generalization ability and prediction accuracy of the trained recommendation model, which affects the user experience.

[0007] To address the aforementioned technical problems, this application provides a method for predicting and ranking, employing the following technical solution:

[0008] A ranking prediction method, applied to a ranking prediction model for multi-task learning, the method comprising:

[0009] Obtain exposure events carrying the first behavior and exposure events not carrying the first behavior, and construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior;

[0010] Obtain exposure events with and without the second behavior, and construct a second sample set based on the exposure events with and without the second behavior;

[0011] The preset initial multi-task learning model is fitted and trained based on the first sample set and the second sample set to obtain the predicted ranking model.

[0012] The system receives information recommendation instructions, obtains user data of the target users, imports the user data into the prediction ranking model, obtains the information prediction ranking result, and completes information push based on the information prediction ranking result.

[0013] Further, the first action is a user click action, and the step of obtaining exposure events carrying the first action and exposure events not carrying the first action, and constructing a first sample set based on the exposure events carrying the first action and exposure events not carrying the first action, specifically includes:

[0014] Exposure events with user clicks are taken as the first positive sample, and exposure events without user clicks are taken as the first negative sample.

[0015] Combine the first positive sample and the first negative sample to construct the first sample set.

[0016] Furthermore, the second behavior is a user purchase behavior. The process of acquiring exposure events carrying the second behavior and exposure events not carrying the second behavior, and constructing a second sample set based on these events, specifically includes:

[0017] Exposure events with user purchase behavior are obtained as the second positive sample, and exposure events without user purchase behavior are obtained as the second negative sample;

[0018] Combine the second positive sample and the second negative sample to construct the second sample set.

[0019] Furthermore, the predicted ranking model also includes a multi-head attention network. The step of fitting and training a preset initial multi-task learning model based on the first sample set and the second sample set to obtain the predicted ranking model specifically includes:

[0020] Based on the multi-head attention network, feature extraction is performed on the first positive sample to obtain the first sample features;

[0021] Based on the multi-head attention network, feature extraction is performed on the second positive sample to obtain the second sample features;

[0022] The first action is predicted based on the features of the first sample using the initial multi-task learning model to obtain the prediction result of the first action;

[0023] The second behavior is predicted based on the features of the second sample using the initial multi-task learning model, and the prediction result of the second behavior is obtained.

[0024] The first behavior prediction result and the second behavior prediction result are fitted to obtain the multi-task prediction fitting result;

[0025] The prediction error between the multi-task prediction fitting result and the preset standard result is calculated based on the loss function of the initial multi-task learning model.

[0026] The initial multi-task learning model is iteratively updated based on the prediction error to obtain the estimated ranking model.

[0027] Furthermore, the first behavior prediction result and the second behavior prediction result are fitted based on the following formula:

[0028] p(y=1,z=1|x)=p(y=1|x1)*p(z=1|y=1,x2)

[0029] In the formula, p(y=1,z=1|x) is the multi-task prediction fitting result, p(y=1|x1) is the first row of prediction results, p(z=1|y=1,x2) is the second row of prediction results, x represents the sample features, x includes the first sample feature x1 and the second sample feature x2, y represents the first row, and z represents the second row.

[0030] Furthermore, the initial multi-task learning model includes a first learning network and a second learning network. The first learning network is used to predict the first action, and the second learning network is used to predict the second action. The loss function of the initial multi-task learning model is expressed as follows:

[0031]

[0032] In the formula, L is the cross-entropy loss function of the initial multi-task learning model, θ1 represents the network parameters of the first learning network, θ2 represents the network parameters of the second learning network, N is the total number of samples, and f(x) i ,θ1)=p(y=1|x i ), f(x) i ,θ1)*f(x i ,θ2)=p(z=1|y=1,x i ).

[0033] Furthermore, the iterative update of the initial multi-task learning model based on the prediction error to obtain the estimated ranking model specifically includes:

[0034] The prediction error is propagated through each network layer of the initial multi-task learning model;

[0035] Obtain the prediction error of each network layer;

[0036] The prediction error of each network layer is compared with a preset error threshold. When the prediction error of any network layer is greater than the error threshold, the initial multi-task learning model is iteratively updated until the prediction error of all network layers of the initial multi-task learning model is less than or equal to the error threshold, and the estimated ranking model is output.

[0037] To address the aforementioned technical problems, this application also provides a predictive sorting device, which employs the following technical solution:

[0038] A predictive sorting device, comprising:

[0039] The first sample construction module is used to obtain exposure events carrying the first behavior and exposure events not carrying the first behavior, and to construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior;

[0040] The second sample construction module is used to obtain exposure events carrying the second behavior and exposure events without the second behavior, and to construct a second sample set based on the exposure events carrying the second behavior and exposure events without the second behavior;

[0041] The multi-task model training module is used to fit and train a preset initial multi-task learning model based on the first sample set and the second sample set to obtain a predicted ranking model.

[0042] The information prediction and ranking module is used to receive information recommendation instructions, obtain user data of the target users, import the user data into the prediction and ranking model, obtain the information prediction and ranking result, and complete the information push based on the information prediction and ranking result.

[0043] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution:

[0044] A computer device includes a memory and a processor, the memory storing computer-readable instructions, the processor executing the computer-readable instructions to implement the steps of the estimated sorting method as described in any of the preceding claims.

[0045] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below:

[0046] A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, implement the steps of the estimated sorting method as described in any one of the preceding descriptions.

[0047] Compared with the prior art, the embodiments of this application have the following main advantages:

[0048] This application discloses a method, apparatus, computer device, and storage medium for predicting ranking, belonging to the field of artificial intelligence technology. This application acquires exposure events carrying a first behavior and exposure events without carrying the first behavior, and constructs a first sample set based on these events. It then acquires exposure events carrying a second behavior and exposure events without carrying the second behavior, and constructs a second sample set based on these events. A preset initial multi-task learning model is fitted and trained using the first and second sample sets to obtain a predictive ranking model. An information recommendation instruction is received, user data of the target user is acquired, and the user data is imported into the predictive ranking model to obtain the information prediction ranking result. Information is then pushed based on the information prediction ranking result. This application constructs sample sets of first behavior events and second behavior events respectively, and uses these sample sets to fit and train a multi-task learning model, enabling the first behavior features and second behavior features to be fitted, thus solving the sample selection bias problem, improving the model's generalization ability, and enhancing prediction accuracy and user experience. Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model. Attached Figure Description

[0049] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 An exemplary system architecture diagram is shown, in which this application can be applied;

[0051] Figure 2 A structural diagram of the OMOE network in an embodiment of this application is shown;

[0052] Figure 3 A structural diagram of the MMOE network in an embodiment of this application is shown;

[0053] Figure 4 A structural diagram of a multi-task prediction ranking model according to the prediction ranking method of this application is shown;

[0054] Figure 5 A flowchart of one embodiment of the pre-sorting method according to this application is shown;

[0055] Figure 6 A schematic diagram of one embodiment of the predictive sorting apparatus according to this application is shown;

[0056] Figure 7 A schematic diagram of the structure of one embodiment of a computer device according to this application is shown. Detailed Implementation

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0058] 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.

[0059] 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.

[0060] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0061] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0062] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0063] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103. The server can be a standalone server 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, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0064] It should be noted that the estimated sorting method provided in the embodiments of this application is generally executed by the server, and correspondingly, the estimated sorting device is generally set in the server.

[0065] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0066] Currently, the latest multi-objective learning recommendation systems are based on the multi-task learning framework MMOE. For example, the paper "Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts" proposes a novel multi-task learning framework MMOE, which is evolved from OMOE.

[0067] OMOE (One-gate Mixture of Experts) consists of many networks of "experts", such as Figure 2As shown, each "expert" network has a simple feedforward neural network and a trainable gating network. This gating network selects a sparse combination of the "expert" networks to process each input. It can automatically assign parameters to capture information shared across multiple tasks or task-specific information without adding many new parameters for each task. Furthermore, all parts of the network can be trained together via backpropagation. Because there is only one gate network, this structure is also called OMOE (One-gate Mixture-of-Experts) to correspond with MMOE. The MOE model can be formally represented as:

[0068]

[0069] Among them, f i (i = 1, ..., n) are n expert networks (an expert network can be considered as a neural network), and g is a gating network that combines the results of the experts. Specifically, g generates probability distributions over n experts, and the final output is the weighted sum of all experts. Clearly, MOE can be viewed as an ensemble method based on multiple independent models. Note that MOE only corresponds to a portion of the diagram above; the weighted results can be input into the tower networks corresponding to the subtasks for learning.

[0070] MMOE is actually a variant and optimization of MOE for multi-task learning. The MMOE structure diagram is as follows: Figure 3 As shown, compared to the OMOE structure where all tasks share a single gating network, the MMOE structure optimizes this by using a separate gating network for each task. This improvement allows for different Experts weights for different tasks, enabling selective utilization of Experts. The gating networks corresponding to different tasks can learn different Experts combination patterns, thus making it easier for the model to capture the correlations and differences between subtasks.

[0071] Multi-task recommendation models like MMOE only use samples with click behavior to train the CVR prediction model, but the trained model is then used to make inferences and predictions across the entire sample space. This leads to sample selection bias, resulting in low generalization ability and prediction accuracy, negatively impacting the user experience. Furthermore, each "expert" network in MMOE is a simple feedforward neural network that can only extract basic features and cannot capture dependencies within user behavior sequences.

[0072] To address the aforementioned technical problems, this application discloses a multi-task learning prediction ranking model based on the above-mentioned MMOE structure. The structure of the multi-task learning prediction ranking model of this application is as follows: Figure 4 As shown, based on the traditional multi-task learning ranking model, a result fitting layer is added. After completing CTR and CVR predictions, the CTR and CVR results are fitted, enabling the model to learn fitting features of various user behaviors, solving the sample selection bias problem, and improving the model's generalization ability and prediction accuracy. Specifically, the left sub-network is used to fit pCTR, and the right sub-network is used to fit pCVR. Furthermore, the multiplication of the outputs of the two sub-networks yields the fitted pCTR / CVR.

[0073] This application also discloses a prediction ranking method for the prediction ranking model applied to the above-mentioned multi-task learning. By constructing sample sets of first-action events and second-action events respectively, and using the constructed sample sets to fit and train the multi-task learning model, the first-action features and second-action features are fitted, thus solving the sample selection bias problem, improving the model's generalization ability, and enhancing prediction accuracy and user experience.

[0074] Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model.

[0075] Continue to refer to Figure 5 The diagram illustrates a flowchart of an embodiment of the prediction and sorting method according to this application. Embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0076] Foundational artificial intelligence technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning. The aforementioned ranking prediction method includes the following steps:

[0077] S501, obtain exposure events carrying the first behavior and exposure events not carrying the first behavior, and construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior.

[0078] The server's database stores a large number of pre-collected exposure events. Exposure events are interactive behaviors resulting from virtual clicks by users when a webpage loads, such as when product or activity recommendations are displayed. Simply put, an exposure event is an event where a specific element or area on the page is exposed (i.e., displayed at the front of the page). Exposure events in the database are pre-labeled with different tags based on different user behaviors, giving each exposure event a distinct behavioral label.

[0079] In this embodiment, the server retrieves exposure events carrying the first behavior and exposure events without the first behavior from the database, and constructs a first sample set based on the exposure events carrying the first behavior and exposure events without the first behavior.

[0080] Furthermore, the first action is the user click action. Exposure events carrying the first action and those not carrying the first action are acquired. A first sample set is constructed based on these two sets, specifically including:

[0081] Exposure events with user clicks are taken as the first positive sample, and exposure events without user clicks are taken as the first negative sample.

[0082] Combine the first positive sample and the first negative sample to construct the first sample set.

[0083] In one embodiment of this application, the first action is a user click action, that is, a user performs a click operation on the exposure page, including liking, favoriting, sharing, commenting, conversion (purchase), etc. The server obtains exposure events with user click actions as first positive samples and exposure events without user click actions as first negative samples, and combines the first positive samples and the first negative samples to construct a first sample set.

[0084] In the above embodiments, this application constructs a first sample set by using exposure events with user click behavior and exposure events without user click behavior, and uses the first sample set to predict user CTR.

[0085] S502, obtain exposure events with the second behavior and exposure events without the second behavior, and construct a second sample set based on the exposure events with the second behavior and exposure events without the second behavior.

[0086] In this embodiment, the server retrieves exposure events carrying the second behavior and exposure events without the second behavior from the database, and constructs a second sample set based on the exposure events carrying the second behavior and exposure events without the second behavior.

[0087] Furthermore, the second action is the user's purchase behavior. Exposure events carrying the second action and those not carrying the second action are acquired. A second sample set is constructed based on these two sets, specifically including:

[0088] Exposure events with user purchase behavior are obtained as the second positive sample, and exposure events without user purchase behavior are obtained as the second negative sample;

[0089] Combine the second positive sample and the second negative sample to construct the second sample set.

[0090] In one embodiment of this application, the second action is a user purchase action, i.e., a user makes a purchase on the exposure page. The server obtains exposure events with user purchase actions as second positive samples and exposure events without user purchase actions as second negative samples, and combines the second positive samples and the second negative samples to construct a second sample set.

[0091] In the above embodiments, this application constructs a second sample set by using exposure events with user purchase behavior and exposure events without user purchase behavior as second negative samples, and realizes user CVR prediction through the second sample set.

[0092] S503, based on the first and second sample sets, the preset initial multi-task learning model is fitted and trained to obtain the predicted ranking model.

[0093] In this embodiment, the server trains a preset initial multi-task learning model based on a first sample set and a second sample set to obtain a predicted ranking model. This application adds a result fitting layer to the traditional multi-task learning ranking model. After completing CTR and CVR predictions, the CTR and CVR results are fitted, enabling the model to learn fitting features of various user behaviors, thus solving the sample selection bias problem and improving the model's generalization ability and prediction accuracy.

[0094] Furthermore, the ranking prediction model also includes a multi-head attention network, which is used to fit and train a pre-defined initial multi-task learning model based on the first and second sample sets to obtain the ranking prediction model, specifically including:

[0095] The first positive sample is feature extracted using a multi-head attention network to obtain the features of the first sample.

[0096] The second positive sample is feature extracted using a multi-head attention network to obtain the features of the second sample.

[0097] The first action is predicted based on the features of the first sample using the initial multi-task learning model, and the prediction result of the first action is obtained.

[0098] The second action is predicted based on the features of the second sample using the initial multi-task learning model, and the prediction result of the second action is obtained.

[0099] The prediction results of the first row and the prediction results of the second row are fitted to obtain the multi-task prediction fitting results;

[0100] The prediction error between the multi-task prediction fitting result and the preset standard result is calculated based on the loss function of the initial multi-task learning model.

[0101] The initial multi-task learning model is iteratively updated based on the prediction error to obtain the predicted ranking model.

[0102] In this embodiment, features are extracted from the first positive sample using a multi-head attention network to obtain first sample features; features are also extracted from the second positive sample using the same multi-head attention network to obtain second sample features; an initial multi-task learning model is used to predict the first behavior based on the first sample features to obtain a first behavior prediction result; the same initial multi-task learning model is used to predict the second behavior based on the second sample features to obtain a second behavior prediction result; the first and second behavior prediction results are then fitted to obtain a multi-task prediction fitting result; the prediction error between the multi-task prediction fitting result and a preset standard result is calculated based on the loss function of the initial multi-task learning model, where the preset standard result is the behavior label corresponding to the exposure event; and the initial multi-task learning model is iteratively updated based on the prediction error to obtain a predicted ranking model.

[0103] In the above embodiments, this application learns the first behavioral feature through a sub-network of the initial multi-task learning model and outputs the first behavioral prediction result CTR. It learns the second behavioral feature through another sub-network of the initial multi-task learning model and outputs the second behavioral prediction result CVR. The CTR result and CVR result are fitted, so that the model can learn the fitting features of multiple user behaviors.

[0104] Furthermore, the prediction results for the first row and the second row are fitted based on the following formula:

[0105] p(y=1,z=1|x)=p(y=1|x1)*p(z=1|y=1,x2)

[0106] In the formula, p(y=1,z=1|x) is the multi-task prediction fitting result, p(y=1|x1) is the first row of prediction result, p(z=1|y=1,x2) is the second row of prediction result, x represents the sample feature (i.e., impression), x includes the first sample feature x1 and the second sample feature x2, y represents the first row, and z represents the second row.

[0107] In this embodiment, (x,y) is input into the CTR task to obtain the estimated value of CTR, and (x,z) is input into the CVR task to obtain the estimated value of CVR. The estimated values ​​of CTR and CVR are multiplied together to obtain the estimated value of the fitted CTCVR.

[0108] Furthermore, the initial multi-task learning model includes a first learning network and a second learning network. The first learning network is used to predict the first action, and the second learning network is used to predict the second action. The loss function of the initial multi-task learning model is expressed as follows:

[0109]

[0110] In the formula, L is the cross-entropy loss function of the initial multi-task learning model, θ1 represents the network parameters of the first learning network, θ2 represents the network parameters of the second learning network, N is the total number of samples, and f(x) i ,θ1)=p(y=1|x i ), f(x) i ,θ1)*f(x i ,θ2)=p(z=1|y=1,x i ).

[0111] In one specific embodiment of this application, exposure events with click behavior are used as positive samples, and exposure events without click behavior are used as negative samples to construct a first sample set for training the CTR prediction task. Exposure events with purchase behavior are used as positive samples, and exposure events without purchase behavior are used as negative samples to train the CVR prediction task. The inputs to the two tasks are actually the same, but the labels are different, therefore the extracted features are also different. The CTR task predicts the click-through rate, while the CVR task predicts the conversion rate. Multiplying the predicted CTR and CVR values ​​yields the fitted result CTCVR. The error between the fitted result CTCVR and the behavior label of the exposure event is calculated using the model loss function, and the calculated prediction error is used to iteratively update the model until the model is perfectly fitted. The model loss function is expressed as:

[0112]

[0113] In the formula, θ ctr and θ cvr These are the parameters for the CTR network and the CVR network, respectively.

[0114] Furthermore, the initial multi-task learning model is iteratively updated based on the prediction error to obtain the estimated ranking model, which specifically includes:

[0115] Propagate the prediction error through each network layer of the initial multi-task learning model;

[0116] Obtain the prediction error of each network layer;

[0117] The prediction error of each network layer is compared with a preset error threshold. When the prediction error of any network layer is greater than the error threshold, the initial multi-task learning model is iteratively updated until the prediction error of all network layers of the initial multi-task learning model is less than or equal to the error threshold, and the predicted ranking model is output.

[0118] In one specific embodiment of this application, prediction errors are propagated in each network layer of the initial multi-task learning model based on a preset backpropagation algorithm. The prediction errors of each network layer are obtained, and the prediction errors of each network layer are compared with preset error thresholds. When the prediction error of any network layer is greater than the error threshold, the initial multi-task learning model is iteratively updated until the prediction errors of all network layers of the initial multi-task learning model are less than or equal to the error threshold, and the estimated ranking model is output.

[0119] In the above embodiments, this application uses the backpropagation algorithm to achieve iterative training of the multi-task learning model, thereby improving the model training efficiency and the model's prediction accuracy.

[0120] S504 receives the information recommendation instruction, obtains the user data of the target users, imports the user data into the prediction ranking model, obtains the information prediction ranking result, and completes the information push based on the information prediction ranking result.

[0121] In this embodiment, the server receives an information recommendation instruction, obtains user data of the target user, including the historical exposure events of the target user, labels the historical exposure events of the target user, imports the labeled user data into the prediction ranking model, obtains the information prediction ranking result, and completes the information push based on the information prediction ranking result.

[0122] In this embodiment, the prediction and sorting method runs on an electronic device (e.g., Figure 1 The server shown can receive information recommendation instructions via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future wireless connection methods.

[0123] This application discloses a ranking prediction method, belonging to the field of artificial intelligence technology. This application acquires exposure events carrying a first behavior and exposure events without carrying the first behavior, and constructs a first sample set based on these events. It then acquires exposure events carrying a second behavior and exposure events without carrying the second behavior, and constructs a second sample set based on these events. A pre-set initial multi-task learning model is fitted and trained using the first and second sample sets to obtain a ranking prediction model. Upon receiving an information recommendation instruction, user data of the target user is acquired, and the user data is imported into the ranking prediction model to obtain the information ranking prediction result. Information is then pushed based on the ranking prediction result. This application constructs sample sets of first-behavior events and second-behavior events respectively, and uses these sample sets to fit and train a multi-task learning model. This allows the first-behavior features and second-behavior features to be fitted, solving the sample selection bias problem, improving the model's generalization ability, and enhancing prediction accuracy and user experience. Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model.

[0124] It should be emphasized that, to further ensure the privacy and security of the aforementioned user data, the user data can also be stored in a blockchain node.

[0125] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0126] 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 instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium, and when executed, they can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0127] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0128] Further reference Figure 6 As a response to the above Figure 5 To implement the method shown, this application provides an embodiment of a predictive sorting device, which is similar to... Figure 5 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0129] like Figure 6 As shown, the estimated sorting device 600 described in this embodiment includes:

[0130] The first sample construction module 601 is used to acquire exposure events carrying the first behavior and exposure events not carrying the first behavior, and to construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior.

[0131] The second sample construction module 602 is used to obtain exposure events carrying the second behavior and exposure events not carrying the second behavior, and to construct a second sample set based on the exposure events carrying the second behavior and exposure events not carrying the second behavior;

[0132] The multi-task model training module 603 is used to fit and train a preset initial multi-task learning model based on a first sample set and a second sample set to obtain a predicted ranking model.

[0133] The information prediction and ranking module 604 is used to receive information recommendation instructions, obtain user data of the target users, import the user data into the prediction and ranking model, obtain the information prediction and ranking results, and complete the information push based on the information prediction and ranking results.

[0134] Furthermore, the first line represents the user's click behavior, and the first sample construction module 601 specifically includes:

[0135] The first event acquisition unit is used to acquire exposure events with user click behavior as the first positive sample, and to acquire exposure events without user click behavior as the first negative sample.

[0136] The first sample construction unit is used to combine the first positive sample and the first negative sample to construct the first sample set.

[0137] Furthermore, the second action is the user's purchase behavior, and the second sample construction module 602 specifically includes:

[0138] The second event acquisition unit is used to acquire exposure events with user purchase behavior as the second positive sample, and to acquire exposure events without user purchase behavior as the second negative sample.

[0139] The second sample construction unit is used to combine the second positive sample and the second negative sample to construct the second sample set.

[0140] Furthermore, the ranking prediction model also includes a multi-head attention network, and the multi-task model training module 603 specifically includes:

[0141] The first feature extraction unit is used to extract features from the first positive sample based on a multi-head attention network to obtain the features of the first sample.

[0142] The second feature extraction unit is used to extract features from the second positive sample based on a multi-head attention network to obtain the features of the second sample.

[0143] The first action prediction unit is used to predict the first action based on the features of the first sample using the initial multi-task learning model, and obtain the prediction result of the first action.

[0144] The second behavior prediction unit is used to predict the second behavior based on the features of the second sample using the initial multi-task learning model, and obtain the prediction result of the second behavior.

[0145] The behavior prediction fitting unit is used to fit the first behavior prediction result and the second behavior prediction result to obtain the multi-task prediction fitting result.

[0146] The prediction error calculation unit is used to calculate the prediction error between the multi-task prediction fitting result and the preset standard result based on the loss function of the initial multi-task learning model.

[0147] The model iterative update unit is used to iteratively update the initial multi-task learning model based on the prediction error to obtain the estimated ranking model.

[0148] Furthermore, the prediction results for the first row and the second row are fitted based on the following formula:

[0149] p(y=1,z=1|x)=p(y=1|x1)*p(z=1|y=1,x2)

[0150] In the formula, p(y=1,z=1|x) is the multi-task prediction fitting result, p(y=1|x1) is the first row of prediction result, p(z=1|y=1,x2) is the second row of prediction result, x represents the sample feature, x includes the first sample feature x1 and the second sample feature x2, y represents the first row, and z represents the second row.

[0151] Furthermore, the initial multi-task learning model includes a first learning network and a second learning network. The first learning network is used to predict the first action, and the second learning network is used to predict the second action. The loss function of the initial multi-task learning model is expressed as follows:

[0152]

[0153] In the formula, L is the cross-entropy loss function of the initial multi-task learning model, θ1 represents the network parameters of the first learning network, θ2 represents the network parameters of the second learning network, N is the total number of samples, and f(x) i ,θ1)=p(y=1|x i ), f(x) i ,θ1)*f(x i ,θ2)=p(z=1|y=1,x i ).

[0154] Furthermore, the model iterative update unit specifically includes:

[0155] The prediction error propagation subunit is used to propagate the prediction error through the various network layers of the initial multi-task learning model.

[0156] The network layer error acquisition subunit is used to acquire the prediction error of each network layer.

[0157] The model iterative update subunit is used to compare the prediction error of each network layer with a preset error threshold. When the prediction error of any network layer is greater than the error threshold, the initial multi-task learning model is iteratively updated until the prediction error of all network layers of the initial multi-task learning model is less than or equal to the error threshold, and the estimated ranking model is output.

[0158] This application discloses a predictive ranking device, belonging to the field of artificial intelligence technology. This application acquires exposure events carrying a first behavior and exposure events without carrying the first behavior, and constructs a first sample set based on these events. It then acquires exposure events carrying a second behavior and exposure events without carrying the second behavior, and constructs a second sample set based on these events. A preset initial multi-task learning model is fitted and trained using the first and second sample sets to obtain a predictive ranking model. The device receives information recommendation instructions, acquires user data of the target users, imports the user data into the predictive ranking model, obtains the information prediction ranking result, and completes information push based on the prediction ranking result. This application constructs sample sets of first behavior events and second behavior events respectively, and uses these sample sets to fit and train a multi-task learning model, enabling the first behavior features and second behavior features to be fitted, thus solving the sample selection bias problem, improving the model's generalization ability, and enhancing prediction accuracy and user experience. Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model.

[0159] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 7 , Figure 7 This is a basic structural block diagram of the computer device in this embodiment.

[0160] The computer device 7 includes a memory 71, a processor 72, and a network interface 73 that are interconnected via a system bus. It should be noted that only the computer device 7 with components 71-73 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0161] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0162] The memory 71 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 71 may be an internal storage unit of the computer device 7, such as the hard disk or memory of the computer device 7. In other embodiments, the memory 71 may also be an external storage device of the computer device 7, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 7. Of course, the memory 71 may include both the internal storage unit and its external storage device of the computer device 7. In this embodiment, the memory 71 is typically used to store the operating system and various application software installed on the computer device 7, such as computer-readable instructions for predictive sorting methods. In addition, the memory 71 can also be used to temporarily store various types of data that have been output or will be output.

[0163] In some embodiments, the processor 72 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 72 is typically used to control the overall operation of the computer device 7. In this embodiment, the processor 72 is used to execute computer-readable instructions stored in the memory 71 or to process data, for example, to execute computer-readable instructions for the estimated sorting method.

[0164] The network interface 73 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 7 and other electronic devices.

[0165] This application discloses a computer device belonging to the field of artificial intelligence technology. This application acquires exposure events carrying a first behavior and exposure events without carrying the first behavior, and constructs a first sample set based on these events. It then acquires exposure events carrying a second behavior and exposure events without carrying the second behavior, and constructs a second sample set based on these events. A preset initial multi-task learning model is fitted and trained using the first and second sample sets to obtain a predicted ranking model. The device receives information recommendation instructions, acquires user data of the target users, imports the user data into the predicted ranking model, obtains the predicted ranking result, and completes information push based on the predicted ranking result. This application constructs sample sets of first behavior events and second behavior events respectively, and uses these sample sets to fit and train a multi-task learning model, enabling a proper fit between the first and second behavior features. This solves the sample selection bias problem, improves the model's generalization ability, and enhances prediction accuracy and user experience. Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model.

[0166] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the estimated sorting method described above.

[0167] This application discloses a storage medium belonging to the field of artificial intelligence technology. This application acquires exposure events carrying a first behavior and exposure events not carrying the first behavior, and constructs a first sample set based on these events. It then acquires exposure events carrying a second behavior and exposure events not carrying the second behavior, and constructs a second sample set based on these events. A preset initial multi-task learning model is fitted and trained using the first and second sample sets to obtain a predicted ranking model. Upon receiving an information recommendation instruction, user data of the target user is acquired, and the user data is imported into the predicted ranking model to obtain the predicted ranking result. Information is then pushed based on the predicted ranking result. This application constructs sample sets of first-behavior events and second-behavior events respectively, and uses these sample sets to fit and train a multi-task learning model. This allows the first-behavior features and second-behavior features to be fitted, solving the sample selection bias problem, improving the model's generalization ability, and enhancing prediction accuracy and user experience. Furthermore, this application introduces a multi-head attention network into the multi-task learning model to obtain the dependencies between behavioral features, form user behavior sequence information, and further improve the prediction accuracy of the model.

[0168] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0169] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0170] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A predictive sorting method, characterized in that, The predicted ranking method is applied to a predicted ranking model in multi-task learning, and the method includes: Obtain exposure events carrying the first behavior and exposure events not carrying the first behavior, and construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior. Among them, exposure events with user click behavior are obtained as the first positive sample, and exposure events without user click behavior are obtained as the first negative sample. The first positive sample and the first negative sample are combined to construct the first sample set. Obtain exposure events carrying the second behavior and exposure events not carrying the second behavior, and construct a second sample set based on the exposure events carrying the second behavior and exposure events not carrying the second behavior. Among them, exposure events with user purchase behavior are obtained as the second positive sample, and exposure events without user purchase behavior are obtained as the second negative sample. Combine the second positive sample and the second negative sample to construct the second sample set. The preset initial multi-task learning model is fitted and trained based on the first sample set and the second sample set to obtain the predicted ranking model. Receive information recommendation instructions, obtain user data of the target users, import the user data into the prediction ranking model, obtain information prediction ranking results, and complete information push based on the information prediction ranking results; The predicted ranking model further includes a multi-head attention network. The process of fitting and training a preset initial multi-task learning model based on the first and second sample sets to obtain the predicted ranking model specifically includes: Based on the multi-head attention network, feature extraction is performed on the first positive sample to obtain the first sample features; Based on the multi-head attention network, feature extraction is performed on the second positive sample to obtain the second sample features; The first action is predicted based on the features of the first sample using the initial multi-task learning model to obtain the prediction result of the first action; The second behavior is predicted based on the features of the second sample using the initial multi-task learning model, and the prediction result of the second behavior is obtained. The first behavior prediction result and the second behavior prediction result are fitted to obtain the multi-task prediction fitting result; The prediction error between the multi-task prediction fitting result and the preset standard result is calculated based on the loss function of the initial multi-task learning model. The initial multi-task learning model is iteratively updated based on the prediction error to obtain the estimated ranking model.

2. The prediction and sorting method as described in claim 1, characterized in that, The first behavior prediction result and the second behavior prediction result are fitted based on the following formula: In the formula, The results are the fitting results for multi-task prediction. The first row represents the prediction result. The second row represents the prediction result. Indicates sample features, Including the first sample features Second sample features , Indicates the first action. Indicates the second action.

3. The prediction and sorting method as described in claim 2, characterized in that, The initial multi-task learning model includes a first learning network and a second learning network. The first learning network is used to predict the first action, and the second learning network is used to predict the second action. The loss function of the initial multi-task learning model is expressed as follows: In the formula, It is the cross-entropy loss function of the initial multi-task learning model. This represents the network parameters of the first learning network. This represents the network parameters of the second learning network. The total number of samples, , = .

4. The prediction and sorting method as described in claim 1, characterized in that, The iterative update of the initial multi-task learning model based on the prediction error to obtain the estimated ranking model specifically includes: The prediction error is propagated through each network layer of the initial multi-task learning model; Obtain the prediction error of each network layer; The prediction error of each network layer is compared with a preset error threshold. When the prediction error of any network layer is greater than the error threshold, the initial multi-task learning model is iteratively updated until the prediction error of all network layers of the initial multi-task learning model is less than or equal to the error threshold, and the estimated ranking model is output.

5. A predictive sorting device, characterized in that, include: The first sample construction module is used to obtain exposure events carrying a first behavior and exposure events not carrying a first behavior, and to construct a first sample set based on the exposure events carrying the first behavior and exposure events not carrying the first behavior. Specifically, exposure events with user click behavior are obtained as first positive samples, and exposure events without user click behavior are obtained as first negative samples. The first positive samples and the first negative samples are combined to construct the first sample set. The second sample construction module is used to obtain exposure events carrying the second behavior and exposure events not carrying the second behavior, and to construct a second sample set based on the exposure events carrying the second behavior and exposure events not carrying the second behavior. Specifically, exposure events with user purchase behavior are obtained as second positive samples, and exposure events without user purchase behavior are obtained as second negative samples. The second positive samples and the second negative samples are combined to construct the second sample set. The multi-task model training module is used to fit and train a preset initial multi-task learning model based on the first sample set and the second sample set to obtain a predicted ranking model. The information prediction and ranking module is used to receive information recommendation instructions, obtain user data of the target users, import the user data into the prediction and ranking model, obtain information prediction and ranking results, and complete information push based on the information prediction and ranking results. The prediction ranking model also includes a multi-head attention network, and the multi-task model training module specifically includes: The first feature extraction unit is used to extract features from the first positive sample based on the multi-head attention network to obtain the first sample features; The second feature extraction unit is used to extract features from the second positive sample based on the multi-head attention network to obtain the features of the second sample; The first behavior prediction unit is used to predict the first behavior based on the features of the first sample using the initial multi-task learning model, and obtain the first behavior prediction result. The second behavior prediction unit is used to predict the second behavior based on the features of the second sample using the initial multi-task learning model, and obtain the second behavior prediction result. A behavior prediction fitting unit is used to fit the first behavior prediction result and the second behavior prediction result to obtain a multi-task prediction fitting result; The prediction error calculation unit is used to calculate the prediction error between the multi-task prediction fitting result and the preset standard result based on the loss function of the initial multi-task learning model. The model iterative update unit is used to iteratively update the initial multi-task learning model based on the prediction error to obtain the estimated ranking model.

6. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the prediction sorting method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions that, when executed by a processor, implement the steps of the estimated sorting method as described in any one of claims 1 to 4.

Images