Model training method and device based on deep target optimization, and information recommendation method and device

Abstract

Embodiments of the present specification provide a model training method, an information recommendation method and device based on deep target optimization. In the model training method based on deep target optimization, each training sample used in the model training process includes feature data and label data matched with a direct training target and a deep training target based thereon, the feature data includes user features, item features and historical interaction features; according to the feature data, a conditional probability value between each training target and a target prediction value corresponding to each training target and a corresponding conditional prediction value are obtained; then a deep-direct target causal strength value corresponding to each training sample is determined, and a deep target prediction value corresponding to each training sample is determined; then a loss value is determined according to a difference between the deep target prediction value of each current training sample and the corresponding label data; the model parameters of the current prediction model are adjusted according to the loss value until a training end condition is met.

Description

Technical Field

[0001] The embodiments in this specification generally relate to the field of computer technology, and in particular to model training methods, information recommendation methods and apparatus based on deep target optimization. Background Technology

[0002] With the rapid development of internet technology, information recommendation technology based on multi-objective combination optimization has been increasingly widely applied. However, due to the widespread existence of data bias in recommendation systems—for example, the interaction data on which the recommendation system relies is inevitably affected by exposure mechanisms and user choices, recommendation data exhibits long-tail characteristics, and feedback loops in recommendation data exacerbate the bias—it is of great significance to learn how to debias and thus build more effective multi-objective recommendation systems. Summary of the Invention

[0003] In view of the above, embodiments of this specification provide a model training method, an information recommendation method, and an apparatus based on deep objective optimization. Using this method and apparatus, a more efficient multi-objective recommendation system can be implemented.

[0004] According to one aspect of an embodiment of this specification, a model training method based on deep target optimization is provided, comprising: cyclically executing the following model training process using a training sample set until a training termination condition is met, wherein each training sample in the training sample set includes feature data and label data matching a training target, wherein the training target includes a direct training target and a deep training target based on the direct training target, and the feature data includes user features, item features, and historical interaction features, wherein the historical interaction features are used to indicate whether there has been an interaction between a user and an item that matches the training target in the past period; for each current training sample in the current training sample set, obtaining conditional probability values ​​between each training target based on the user features and historical interaction features of the current training sample; and retrieving the feature data of the current training sample... The current prediction model is provided with target prediction values ​​corresponding to each training objective, wherein the target prediction values ​​corresponding to the training objectives include conditional prediction values ​​involving other training objectives with causal relationships; the deep-direct target causal strength value corresponding to the current training sample is determined based on the obtained conditional probability value and the corresponding conditional prediction value; the deep target prediction value corresponding to the current training sample is determined based on the obtained deep-direct target causal strength value and the corresponding target prediction value; the loss value of the current model training process is determined based on the difference between the deep target prediction values ​​and the corresponding label data of each current training sample; and in response to the failure to meet the training termination condition, the model parameters of the current prediction model are adjusted according to the loss value, wherein the prediction model after model parameter adjustment serves as the current prediction model for the next model training process.

[0005] According to another aspect of the embodiments of this specification, an information recommendation method is provided, comprising: generating a user-information pair feature set based on relevant information of a target user and a set of information to be recommended, wherein each user-information pair feature in the user-information pair feature set includes user features, item features, and historical interaction features generated based on the relevant information of the target user and the information to be recommended; providing the user-information pair feature set to a prediction model to obtain at least one predicted value corresponding to each user-information pair feature, wherein the prediction model is trained by the model training method based on deep training objective optimization as described above, and each predicted value corresponds to a direct training objective or a deep training objective based on a direct training objective; determining target recommendation information from the set of information to be recommended based on the obtained at least one predicted value; and providing the target recommendation information to the target user.

[0006] According to another aspect of the embodiments of this specification, a model training apparatus based on deep target optimization is provided. The apparatus is configured such that a training unit cyclically executes a model training process using a training sample set until a training termination condition is met. Each training sample in the training sample set includes feature data and label data matching a training target. The training target includes a direct training target and a deep training target based on the direct training target. The feature data includes user features, item features, and historical interaction features. The historical interaction features indicate whether there has been an interaction between a user and an item that matches the training target within a past period. The training unit includes: a conditional probability determination module configured to obtain conditional probability values ​​between each training target based on the user features and historical interaction features of the current training sample; and an initial prediction module configured to provide the feature data of the current training sample to a current prediction model to obtain conditional probability values ​​between each training target. The apparatus includes: a target prediction value corresponding to the target, wherein the target prediction value corresponding to the training target includes conditional prediction values ​​involving other training targets with causal relationships; a causal strength determination module configured to determine the deep-direct target causal strength value corresponding to the current training sample based on the obtained conditional probability value and the corresponding conditional prediction value; a deep target prediction module configured to determine the deep target prediction value corresponding to the current training sample based on the obtained deep-direct target causal strength value and the corresponding target prediction value; a loss value determination module configured to determine the loss value of the current model training process based on the difference between the deep target prediction values ​​of each current training sample and the corresponding label data; and the apparatus further includes: a parameter adjustment unit configured to adjust the model parameters of the current prediction model according to the loss value in response to the failure to meet the training termination condition, wherein the prediction model after model parameter adjustment serves as the current prediction model for the next model training process.

[0007] According to another aspect of the embodiments of this specification, an information recommendation apparatus is provided, comprising: a feature extraction unit configured to generate a user-information pair feature set based on relevant information of a target user and a set of information to be recommended, wherein each user-information pair feature in the user-information pair feature set includes user features, item features, and historical interaction features generated based on the relevant information of the target user and the information to be recommended; a probability prediction unit configured to provide the user-information pair feature set to a prediction model to obtain at least one predicted value corresponding to each user-information pair feature, wherein the prediction model is trained by a model training device based on deep training objective optimization as described above, and each predicted value corresponds to a direct training objective or a deep training objective based on a direct training objective; and an information recommendation unit configured to determine target recommendation information from the set of information to be recommended based on the obtained at least one predicted value; and to provide the target recommendation information to the target user.

[0008] According to another aspect of the embodiments of this specification, a model training apparatus based on deep objective optimization is provided, comprising: at least one processor, and a memory coupled to the at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the model training method based on deep objective optimization as described above.

[0009] According to another aspect of the embodiments of this specification, an information recommendation apparatus is provided, comprising: at least one processor, and a memory coupled to the at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the information recommendation method as described above.

[0010] According to another aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, implements the model training method and / or information recommendation method based on deep objective optimization as described above.

[0011] According to another aspect of the embodiments of this specification, a computer program product is provided, including a computer program that is executed by a processor to implement the model training method and / or information recommendation method based on deep objective optimization as described above. Attached Figure Description

[0012] A further understanding of the nature and advantages of this specification can be achieved by referring to the following figures. In the figures, similar components or features may have the same reference numerals.

[0013] Figure 1An exemplary architecture of a model training method, information recommendation method, and apparatus based on deep objective optimization according to embodiments of this specification is shown.

[0014] Figure 2 A flowchart illustrating an example of a model training method based on deep objective optimization according to an embodiment of this specification is shown.

[0015] Figure 3 A flowchart illustrating an example of a process for determining a target predicted value according to an embodiment of this specification is shown.

[0016] Figure 4 A flowchart illustrating an example of a process for determining the causal strength value of a deep-direct target according to an embodiment of this specification is shown.

[0017] Figure 5 A schematic diagram illustrating an example of a path for transforming a direct training objective into a deep training objective according to an embodiment of this specification is shown.

[0018] Figure 6 A flowchart illustrating an example of the process for determining the loss value in the current model training process according to an embodiment of this specification is provided.

[0019] Figure 7 A schematic diagram illustrating an example application scenario of the model training method based on deep objective optimization according to an embodiment of this specification is shown.

[0020] Figure 8 A flowchart illustrating an example of an information recommendation method according to an embodiment of this specification is shown.

[0021] Figure 9 A block diagram of an example of a model training apparatus based on deep objective optimization according to an embodiment of this specification is shown.

[0022] Figure 10 A block diagram illustrating an example of an information recommendation device according to an embodiment of this specification is shown.

[0023] Figure 11 A block diagram of an example of a model training apparatus based on deep objective optimization according to an embodiment of this specification is shown.

[0024] Figure 12 A block diagram illustrating an example of an information recommendation device according to an embodiment of this specification is shown. Detailed Implementation

[0025] The subject matter described herein will be discussed below with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of the embodiments described herein. Various processes or components may be omitted, substituted, or added as needed in the various examples. Furthermore, features described in some examples may be combined in other examples.

[0026] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0027] The following will describe in detail, with reference to the accompanying drawings, the model training method, information recommendation method, and apparatus based on deep target optimization according to embodiments of this specification.

[0028] Figure 1 An exemplary architecture 100 of a model training method, information recommendation method, and apparatus based on deep objective optimization according to embodiments of this specification is shown.

[0029] exist Figure 1 In this context, network 110 is used to interconnect terminal device 120 and application server 130.

[0030] Network 110 can be any type of network capable of interconnecting network entities. Network 110 can be a single network or a combination of various networks. In terms of coverage, network 110 can be a local area network (LAN), a wide area network (WAN), etc. In terms of the carrying medium, network 110 can be a wired network, a wireless network, etc. In terms of data switching technology, network 110 can be a circuit-switched network, a packet-switched network, etc.

[0031] Terminal device 120 can be any type of electronic computing device capable of connecting to network 110, accessing servers or websites on network 110, processing data or signals, etc. For example, terminal device 120 can be a desktop computer, laptop computer, tablet computer, smartphone, etc. Although in Figure 1 Only one terminal device is shown in the diagram, but it should be understood that a different number of terminal devices may be connected to network 110.

[0032] In one implementation, terminal device 120 can be used by a user. Terminal device 120 may include an application client (e.g., application client 121) that can provide various services to the user. In some cases, application client 121 can interact with application server 130. For example, application client 121 can transmit user-inputted messages to application server 130 and receive responses associated with those messages from application server 130. However, it should be understood that, given sufficient computing and storage capabilities, application client 121 may also generate responses to user-inputted messages locally, instead of interacting with application server 130. In this document, "message" can refer to any input information, such as a user's click on a page.

[0033] Application server 130 can store a prediction model to determine a recommendation score for recommending specific information to a specific user. Application server 130 can connect to a database of information to be recommended 140. The database of information to be recommended 140 can include multiple pieces of information to be recommended. The information to be recommended can include various forms, such as product information, news, video information, music information, etc. Application server 130 can also connect to a model training server 150. The model training server 150 can be used to train the aforementioned prediction model. However, it should be understood that in other cases, application server 130 can also store the information to be recommended and train the aforementioned prediction model locally, instead of interacting with the database of information to be recommended 140 and the model training server 150.

[0034] It should be understood that Figure 1 All network entities shown are exemplary, and any other network entities may be involved in Architecture 100 depending on the specific application requirements.

[0035] Figure 2 A flowchart of a model training method 200 based on deep target optimization according to an embodiment of this specification is shown.

[0036] like Figure 2 As shown in 210, the following model training process 220-270 is executed iteratively using the training sample set until the training termination condition is met.

[0037] In this embodiment, each training sample in the training sample set may include feature data and label data matching the training objective. The training objective may include a direct training objective and a deep training objective based on the direct training objective. The feature data may include user features, item features, and historical interaction features. The historical interaction features can be used to indicate whether there have been interactions between the user and items that match the training objective over a past period of time.

[0038] In one example, a user-item pair can be obtained based on whether a user has previously recommended an item to that user. This allows for the acquisition of relevant information about the user and the item, such as user-authorized access information (e.g., city, preferences, interaction history) and item attribute information (e.g., category, price range). The interaction history can include, for example, login habits, average session duration, and recent purchases. Various feature extraction methods can then be used to obtain corresponding user characteristics, item characteristics, and historical interaction features. For instance, historical interaction features could include information indicating whether the user has clicked on and / or purchased the item within the past month.

[0039] In one example, a direct training objective can be used to indicate a metric strongly correlated with the recommendation system. A deep training objective can be used to indicate a metric that is somewhat correlated with the recommendation system but is also influenced by other factors. Furthermore, the evaluation of a deep training objective often depends on at least one of the aforementioned direct training objectives. In one example, a direct training objective might include, for example, clicks or conversions. A deep training objective might include, for example, user activity or user retention.

[0040] Accordingly, for a feature data, the label data that matches the training objective may include information such as whether the user clicked on the item when it was recommended, whether the user purchased the item, whether the user is an active user, and whether the user is a retained user.

[0041] For each current training sample in the current training sample set, perform the following model training process 220-250 to obtain the deep target prediction value corresponding to each current training sample.

[0042] In step 220, the conditional probability values ​​between each training objective are obtained based on the user characteristics and historical interaction characteristics of the current training sample.

[0043] In this embodiment, based on the user characteristics and historical interaction characteristics of each current training sample, various conditional probability determination methods can be used to obtain the conditional probability values ​​between each training objective. In one example, the conditional dependencies between training objectives can typically be predetermined based on their actual meaning. For example, the probability of satisfying a deeper training objective under the condition of satisfying a direct training objective is calculated, rather than the probability of satisfying a direct training objective under the condition of satisfying a deeper training objective. Another example is calculating the probability of conversion under a click condition, rather than the probability of clicking under a conversion condition.

[0044] Optionally, the user features and historical interaction features of the current training sample can be first provided to the encoding structure in the autoencoder to obtain the encoding result. The autoencoder can be trained based on the prediction differences of the historical interaction features. Then, a matrix indicating the conditional probability values ​​between various training objectives can be generated based on the encoding result.

[0045] In one example, the input feature (e.g., "X") can include 10 features, where features 1-6 are user features, and features 7-10 (e.g., "1,0,1,1" represents clicked, no conversion, active user, and retained user) are historical interaction features. During training, these input features are provided to the current autoencoder to obtain an output containing 10 features (e.g., ...). Then, based on a predetermined loss function, the difference between the 7th to 10th features in the output (e.g., "0.6, 0.3, 0.7, 0.8") and its own 7th to 10th features (e.g., "1, 0, 1, 1" above) is used, rather than the input feature (e.g., "X") and the output (e.g., ...). The loss value is calculated based on the differences between the 10 features included. The parameters of the current autoencoder can then be adjusted based on the calculated loss value to obtain the final autoencoder.

[0046] In one example, the user features and historical interaction features of the current training sample are provided to the encoding structure of the autoencoder trained in the above manner to obtain the encoding result. In one example, the above encoding structure may include two nonlinear transformation functions, such as f1 and f2. In one example, the user features and historical interaction features can be provided together to f1 to obtain the corresponding transformation result. Then, the transformation result of the historical interaction features and the historical interaction features are provided together to f2 to obtain the secondary transformation result of the historical interaction features. The transformation result of the user features and the secondary transformation result of the historical interaction features can be combined to form the above encoding result (for example, it can be represented by "Z"). In one example, the feature used to indicate the input (e.g., represented by "X") can be obtained by multiplying the encoded result itself with its transpose. This is a matrix (e.g., 10×10 dimensions) showing the conditional probability values ​​between the features in the training dataset. A portion (e.g., 4×4 dimensions) indicating the conditional probability values ​​between the various training objectives (e.g., click-through rate, conversion rate, user activity rate, user retention rate) can be extracted from this matrix. In one example, the element in the i-th row and j-th column (j>i) of the matrix can be used to indicate the probability that the training objective in the j-th column (e.g., the user is an active user) is satisfied given that the training objective in the i-th row (e.g., a click) is satisfied.

[0047] Based on this, by introducing an autoencoder, we can more effectively mine the conditional probability information between various training objectives hidden in various user features and historical interaction features, thereby providing a basis for determining subsequent causal relationships.

[0048] At 230, the feature data of the current training sample is provided to the current prediction model to obtain the target prediction value corresponding to each training objective.

[0049] In this embodiment, the current prediction model described above can be implemented as the structure of various multi-objective learning models, such as the MMoE model, the PLE (Progressive Layered Extraction) model, etc. The target prediction value corresponding to the training objective can include conditional prediction values ​​involving other causally related training objectives. In one example, the training objectives can include clicks and conversions. The target prediction value corresponding to the training objective can include click probability prediction values ​​and conversion probability prediction values. The conditional prediction values ​​involving other causally related training objectives can include conversion probability prediction values ​​based on clicks.

[0050] Optionally, continue to refer to Figure 3 , Figure 3 A flowchart illustrating an example of a process 300 for determining a target predicted value according to an embodiment of this specification is shown.

[0051] In this embodiment, the feature data of each training sample in the training sample set also includes recommendation domain features. The current prediction model may include a current domain awareness model and a current prediction value generation model. In one example, recommendation domain features can be used to represent the presentation location of items corresponding to the user. In one example, recommendation domain features may include, but are not limited to, at least one of the following: the name of the section to which the presentation location belongs (e.g., trending topics, card groups, recommendation feed), the number of clicks on the presentation location over a period of time, and whether the presentation location is highlighted (e.g., a red dot is added to the icon, characters are displayed with a background, etc.).

[0052] like Figure 3 As shown in step 310, the recommended domain features of the current training sample are provided to the current domain-aware model to obtain the domain-aware features of the current training sample.

[0053] In this embodiment, the current domain-aware model can be implemented as various artificial neural networks for feature extraction, such as convolutional neural networks and multilayer perceptrons. The aforementioned domain-aware features can be viewed as a higher-dimensional abstraction of the recommendation domain features.

[0054] In step 320, the obtained domain-aware features and corresponding user features, item features, and historical interaction features are provided to the current prediction value generation model to obtain the initial target prediction value corresponding to each training objective.

[0055] In this embodiment, the obtained domain-aware features can be combined with corresponding user features, item features, and historical interaction features, and the combined features are provided to the current prediction value generation model. The aforementioned current prediction value generation model can be implemented as the structure of various multi-objective learning models, such as the MMoE model, PLE model, etc. In one example, the current prediction value generation model may include a shared feature network and multiple prediction networks. Each prediction network can obtain a corresponding initial target prediction value based on the features extracted by the shared feature network. In one example, the prediction network may include, but is not limited to, at least one of the following: a network for predicting click probability, a network for predicting conversion probability, a network for predicting the probability of conversion based on a click, and a network for predicting the probability that the user is an active user based on both a click and a conversion.

[0056] In step 330, the initial target prediction value corresponding to the current training sample is corrected based on the initial target prediction values ​​corresponding to other current training samples that match the recommendation domain features corresponding to the current training sample, and the corrected initial target prediction value is used as the corresponding target prediction value.

[0057] In one example, the above correction could refer to normalizing the initial target prediction values ​​for each current training sample that matches the recommendation domain feature corresponding to the current training sample. In another example, if the recommendation domain feature of the current training sample indicates that the recommendation domain is a card group, then all current training samples whose recommendation domain feature indicates that the recommendation domain is a card group can be selected from the current training sample set to form a card group subset. Then, the initial target prediction values ​​(e.g., initial click-through rate prediction values) for each current training sample in the card group subset can be normalized (e.g., using Z-score normalization, linear normalization, etc.) to obtain the target prediction values ​​(e.g., click probability prediction values) for each current training sample. Similarly, the corresponding conversion probability prediction values ​​and the conversion probability prediction values ​​based on clicks can be corrected. It is understandable that the initial target prediction values ​​for each current training sample whose recommendation domain feature indicates that the recommendation domain is a hot search list or a recommendation feed can also be corrected.

[0058] In one example, the above correction can also be achieved by adding batch normalization (BN) layers after each prediction network. It should be noted that the normalization operation of each BN layer is based on the matching recommendation domain features (i.e., different subsets of samples from the current training sample set), rather than uniformly based on all training samples in the current training sample set.

[0059] It should be noted that matching the recommendation domain features does not have to be based on the section name to which the presentation location belongs, but can be based on at least one of the following: the number of clicks on the presentation location over a period of time, or whether the presentation location is highlighted (e.g., a red dot is added to the icon, or the characters are displayed with a background). There are no restrictions here.

[0060] Based on this, this solution can fully explore the influence of different recommendation domains on the final target prediction value from two aspects: providing the extracted domain-aware features along with the corresponding user features, item features, and historical interaction features to the current prediction value generation model, and correcting the corresponding target prediction value output by the model based on the matching domain-aware features. This makes the obtained target prediction value more considerate of the inherent characteristics of different recommendation domains, thereby improving the accuracy of cross-domain recommendation of items using a single prediction model.

[0061] In 240, the deep-direct target causal strength value corresponding to the current training sample is determined based on the obtained conditional probability value and the corresponding conditional prediction value.

[0062] In this embodiment, the deep-to-direct-target causal strength value can be used to indicate the extent to which the achievement of the deep training objective is attributable to the direct training objective.

[0063] Optionally, refer to the following: Figure 4 , Figure 4 A flowchart illustrating an example of a process 400 for determining the causal strength value of a deep-direct target according to an embodiment of this specification is shown.

[0064] In this embodiment, the direct training objective may include a first objective and a second objective. The deep training objective may include a third objective based on the first and second objectives. The conditional probability values ​​may include a first conditional probability value and a second conditional probability value representing the conditional probability between the direct training objective and the deep training objective. The target prediction values ​​corresponding to each training objective include the conditional prediction values ​​between the first and second objectives.

[0065] In one example, the first, second, and third objectives could be clicks, conversions, and user engagement, respectively. The first conditional probability value can be the probability that the user is an active user given a click (e.g., using causal probability).actv|clk (Representation). The second conditional probability value can be the probability that the user is an active user under the condition of conversion (for example, it can be expressed as causal). actv|conv (This can be represented as a conditional prediction between the first and second objectives, which can be the predicted probability of conversion under the condition of a click (e.g., it can be represented as pred(cvr|ctr)).

[0066] like Figure 4 As shown in Figure 410, based on the obtained conditional probability value, target prediction value, and the corresponding path from direct training target to deep training target, the causal conditional weight value corresponding to the current training sample is determined.

[0067] In this embodiment, for a deep training objective, the causal conditional weight value corresponding to the current training sample can be obtained based on the path from each direct training objective to the deep training objective and the corresponding conditional probability value and / or target prediction value during the transformation process.

[0068] Optionally, continue to refer to Figure 5 , Figure 5 A schematic diagram illustrating an example of a path 500 for transforming a direct training objective into a deep training objective according to an embodiment of this specification is shown. Figure 5 As shown, the path from direct training target clicks to deep training target user activity can include: click → user activity, click → conversion → user activity. Optionally, if the deep training target also includes user retention, the path from direct training target click-through rate to deep training target user retention rate can include: click → user retention, click → user activity → user retention, click → conversion → user activity → user retention, click → conversion → user retention.

[0069] In one example, the causal conditional weights between user activity, clicks, and conversions corresponding to the current training sample can be represented as w = causal actv|clk *(pred(cvr|ctr)+causal actv|conv The meanings of the relevant symbols can be found above. It is evident that the conditional probability values ​​and / or target predicted values ​​corresponding to the same path can be added together, while the conditional probability values ​​and / or target predicted values ​​between different paths can be multiplied together to obtain the corresponding causal conditional weight values. It can be understood that a similar method can be used to obtain the causal conditional weight values ​​between user retention, clicks, and conversions corresponding to the current training sample.

[0070] Back Figure 4 In 420, based on the relationship between the determined causal condition weight value and 1, the deep-direct target causal strength value corresponding to the current training sample is determined.

[0071] In one example, the deep-direct target causal strength value corresponding to the current training sample (e.g., w) can be used. condition The value is determined to be the smaller of the determined causal condition weight value and 1.

[0072] Based on this, this scheme can determine the causal condition weight value corresponding to each current training sample according to the path of transformation from direct training target to deep training target and the corresponding conditional probability value and / or target prediction value, so that the deep-direct target causal strength value used to characterize the causal relationship between deep training target and direct training target is closer to the real situation.

[0073] Back Figure 2 At 250, based on the obtained deep-direct target causal strength value and the corresponding target prediction value, the deep target prediction value corresponding to the current training sample is determined.

[0074] In this embodiment, the aforementioned deep target prediction value can be used to represent the predicted probability of satisfying deep training after considering various factors that are causally related to it. In one example, when the deep training target is user activity, the aforementioned deep target prediction value can be used to represent the predicted probability that the user is an active user after considering factors such as clicks and conversions.

[0075] Optionally, in some implementations of this embodiment, the target prediction value corresponding to each training objective further includes a first target prediction value, a conditional prediction value that satisfies the deep training objective while satisfying the direct training objective, and a conditional prediction value that still satisfies the deep training objective even if the direct training objective is not satisfied. The deep target prediction value corresponding to the current training sample can be calculated based on the obtained deep-direct target causal strength value and the corresponding target prediction value using the full probability expansion formula.

[0076] In one example, the first target prediction value can be the click probability prediction value (e.g., represented by pred(ctr)). The conditional prediction value that satisfies the deep training objective in addition to the direct training objective can be the predicted probability of user activity based on clicks and conversions (e.g., represented by pred(actv|clk,conv)). The conditional prediction value that still satisfies the deep training objective despite not satisfying the direct training objective can be the predicted probability of user activity without clicks or conversions (e.g., represented by pred(actv|nclk,nconv)). The predicted probability of a user being an active user, calculated based on the full probability expansion formula and considering click and conversion factors, can be expressed as:

[0077] in, This can be used to represent the joint probability of clicks and conversions. In one example, The meanings of other symbols can be found above.

[0078] In step 260, the loss value for the current model training process is determined based on the difference between the deep target prediction value and the corresponding label data for each current training sample.

[0079] In one example, the deep target prediction value could be the predicted probability that the user is an active user, and the corresponding label data could be a label indicating whether the user is an active user (e.g., 0 or 1). In another example, the deep target prediction value could be the predicted probability that the user is a retained user, and the corresponding label data could be a label indicating whether the user is a retained user (e.g., 0 or 1). A predetermined loss function can be used to determine the loss value for each current training sample corresponding to the deep training target. Then, in one example, the loss value for the current model training process can be obtained by summing the loss values ​​for each current training sample in the current training sample set corresponding to the deep training target.

[0080] Optionally, refer to the following: Figure 6 , Figure 6 A flowchart illustrating an example of a loss value determination process 600 in the current model training process according to an embodiment of this specification is shown.

[0081] like Figure 6 As shown in Figure 610, the deep target loss value is determined based on the difference between the deep target prediction value and the corresponding label data of each current training sample.

[0082] In one example, the deep target loss value can be obtained by summing the loss values ​​corresponding to the deep training objective for each current training sample in the current training sample set. In one example, the deep target loss value can be expressed as... in, This can be used to represent the predicted probability that the user corresponding to the i-th current training sample is an active user. (actv) label_i B can be used to represent the label indicating whether the user corresponding to the i-th current training sample is an active user. B can be used to represent the number of current training samples in the current training sample set. loss(·) can be used to represent the loss function.

[0083] In step 620, the direct target loss value is determined based on the difference between the predicted value and the corresponding label data of the direct training target on which the deep training target of each current training sample is based.

[0084] In this embodiment, the direct target loss value corresponding to the direct training target on which the deep training target is based can be determined by referring to the calculation method of the deep target loss value described above. In one example, the direct target loss value can be represented as loss. Direct =loss CTR +loss CVR Among them, loss CTR and loss CVR These can be used to represent click loss and conversion loss, respectively. In one example, loss CTR and loss CVR These can be represented as loss respectively. CTR =loss(pred(ctr),clk label ) and loss CVR =loss(pred(cvr),cvr label ), where pred(ctr) and pred(cvr) can be used to represent the predicted click probability and conversion probability, respectively. Correspondingly, clk label and CVR label These can be used to indicate whether a user clicks and converts, respectively.

[0085] At 630, the loss value for the current model training process is determined based on the determined deep target loss value and direct target loss value.

[0086] In one example, the determined deep target loss value and direct target loss value can be weighted and summed to obtain the loss value of the current model training process.

[0087] Based on this, the proposed solution can comprehensively consider both deep training objectives and direct training objectives during the training process, thereby achieving explicit optimization of both deep training objectives and direct training objectives, which can improve the effectiveness of the trained prediction model.

[0088] Back Figure 2 At 270, determine whether the training termination condition is met.

[0089] In one example, the training termination condition can be determined by checking whether the number of iterations has reached the preset number, whether the training duration has reached the preset duration, and whether the loss value has converged.

[0090] At 280, in response to the failure to meet the training termination condition, the model parameters of the current prediction model are adjusted based on the loss value.

[0091] In this embodiment, the prediction model after model parameter adjustment serves as the current prediction model for the next model training process. Then, the current training sample set can be redefined using the aforementioned training sample set, and model training processes 220-270 can continue until the training termination condition is met.

[0092] In response to the completion of training conditions, the current prediction model is identified as the training-completed prediction model. This allows the use of the trained prediction model to obtain deep target prediction values ​​corresponding to user-item pair features, thus providing a more effective reference for item recommendation.

[0093] The following is for reference. Figure 7 , Figure 7 A schematic diagram illustrating an example application scenario 700 of the model training method based on deep target optimization according to an embodiment of this specification is shown.

[0094] like Figure 7 As shown, conditional probability values ​​between various training objectives (e.g., clicks, conversions, activity, retention) can be obtained based on user characteristics and historical interaction characteristics of the current training sample. User characteristics, historical interaction characteristics, and item characteristics can also be provided to the current prediction model to obtain target prediction values ​​corresponding to each training objective, such as click probability prediction, conversion probability prediction under click conditions, activity probability prediction under click and conversion conditions, and activity probability prediction under no click and no conversion conditions. Then, the causal strength value between user activity and clicks / conversions can be calculated based on the probability that the user is an active user under click conditions, the probability that the user is an active user under conversion conditions, and the predicted probability of conversion under click conditions. Based on the obtained causal strength value and the corresponding target prediction value, the predicted probability that the user corresponding to the current training sample is an active user is determined. Then, based on the predicted probability that each user corresponding to each current training sample is an active user... The loss value is calculated based on the difference between the label indicating whether the user corresponding to the current training sample is an active user, and then the parameters of the prediction model are adjusted based on the loss value.

[0095] Optionally, the recommended domain features can be provided to the current domain-aware model to obtain domain-aware features. Then, the domain-aware features, user features, historical interaction features, and item features can be provided to the current prediction value generation model to obtain target prediction values ​​corresponding to each training objective.

[0096] use Figures 1-7The model training method based on deep target optimization disclosed herein explores the causal dependency between deep training targets and direct training targets by calculating the conditional probability values ​​between multiple training targets. Then, it optimizes the deep training targets based on the above causal dependency and can also optimize the direct training targets to a certain extent with the help of the above causal dependency, thereby effectively improving the performance of the prediction model.

[0097] Figure 8 A flowchart of an information recommendation method 800 according to an embodiment of this specification is shown.

[0098] like Figure 8 As shown in Figure 810, a user-information pair feature set is generated based on the relevant information of the target user and the set of information to be recommended.

[0099] In this embodiment, the features of each user-information pair in the user-information pair feature set may include user features, item features, and historical interaction features generated based on the relevant information of the target user and the information to be recommended. The information to be recommended can be in various forms, such as news, products, music, videos, etc.

[0100] It should be noted that the specific user characteristics, item characteristics, and historical interaction characteristics mentioned above can be found in the preceding text. Figure 2 The relevant description of step 210 in the embodiment.

[0101] In step 820, the user-information pair feature dataset is provided to the prediction model to obtain at least one predicted value corresponding to each user-information pair feature.

[0102] In this embodiment, the above prediction model can be achieved through, for example... Figures 1-7 The model is trained using the deep training objective optimization method described in the embodiments. Each predicted value corresponds to a direct training objective or a deep training objective based on the direct training objective. In one example, the predicted values ​​may be those described above... Figure 2 The predicted data corresponding to the tag data in step 210 of the embodiment. For example, the predicted click-through rate, predicted conversion rate, user activity probability, user retention probability, etc.

[0103] In step 830, target recommendation information is determined from the set of information to be recommended based on at least one predicted value obtained.

[0104] In one example, the recommended information whose predicted value for a specific training objective is greater than a predetermined threshold can be selected as the target recommended information. In another example, the recommended information with the highest predicted value for a specific training objective can be selected as the target recommended information. In yet another example, the predicted values ​​for multiple training objectives can be weighted and summed to obtain a comprehensive score, and the recommended information with the highest comprehensive score can be selected as the target recommended information.

[0105] In step 840, target recommendation information is provided to the target user.

[0106] In this embodiment, the aforementioned target recommendation information can be provided to the user in various forms. For example, the target recommendation information can be arranged in a list format, ordered from largest to smallest predicted value.

[0107] It should be noted that the user to whom the target recommendation information is provided can be the same user as the user described in step 810 above, or a user using the same client as the user described in step 810 above; there is no limitation here.

[0108] Based on the above, a method is provided to apply the deep objective optimization-based model to the field of information recommendation, so as to improve the recommendation effect of information recommendation systems.

[0109] Figure 9 A block diagram of an example of a deep objective optimization-based model training apparatus 900 according to an embodiment of this specification is shown. This apparatus embodiment can be used with... Figures 2-7 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0110] like Figure 9 As shown, the model training device 900 based on deep target optimization may include a training unit 910 and a parameter adjustment unit 920. The device 1300 is configured such that the training unit 910 cyclically executes the model training process using a training sample set until the training termination condition is met. Each training sample in the training sample set includes feature data and label data matching the training target. The training target includes a direct training target and a deep training target based on the direct training target. The feature data includes user features, item features, and historical interaction features, where the historical interaction features indicate whether there has been an interaction between the user and items that matches the training target over a past period. The training unit 910 includes: a conditional probability determination module 911, an initial prediction module 912, a causal strength determination module 913, a deep target prediction module 914, and a loss value determination module 915.

[0111] The conditional probability determination module 911 is configured to obtain conditional probability values ​​between various training objectives based on the user characteristics and historical interaction characteristics of the current training sample. The operation of the conditional probability determination module 911 can be referenced above. Figure 2 The operation described in section 220.

[0112] In one example, the conditional probability determination module 911 can be further configured to: provide the user features and historical interaction features of the current training sample to the encoding structure in the autoencoder to obtain an encoding result, wherein the autoencoder is trained based on the prediction differences of the historical interaction features; and generate a matrix indicating the conditional probability values ​​between the various training objectives based on the encoding result. The operation of the conditional probability determination module 911 can be referenced above. Figure 2 The operation of the 220 optional implementations described.

[0113] The initial prediction module 912 is configured to provide the feature data of the current training sample to the current prediction model to obtain target prediction values ​​corresponding to each training objective. These target prediction values ​​include conditional prediction values ​​related to other causally correlated training objectives. The operation of the initial prediction module 912 can be referenced above. Figure 2 The operation described in section 230.

[0114] In one example, the feature data of each training sample in the training sample set further includes recommendation domain features, and the current prediction model includes a current domain-aware model and a current prediction value generation model. The initial prediction module 912 can be further configured to: provide the recommendation domain features of the current training sample to the current domain-aware model to obtain the domain-aware features of the current training sample; provide the obtained domain-aware features and corresponding user features, item features, and historical interaction features to the current prediction value generation model to obtain initial target prediction values ​​corresponding to each training objective; and correct the initial target prediction values ​​corresponding to the current training sample based on the initial target prediction values ​​of other current training samples that match the recommendation domain features of the current training sample, and use the corrected initial target prediction values ​​as the corresponding target prediction values. The operation of the initial prediction module 912 can be referred to above. Figure 3 The operation described in 310-330.

[0115] The causal strength determination module 913 is configured to determine the deep-direct target causal strength value corresponding to the current training sample based on the obtained conditional probability value and the corresponding conditional prediction value. The operation of the causal strength determination module 913 can be referred to above. Figure 2 The operation described in section 240.

[0116] In one example, the direct training objective includes a first objective and a second objective, and the deep training objective includes a third objective based on the first and second objectives. The conditional probability values ​​include a first conditional probability value and a second conditional probability value characterizing the conditional probability between the direct training objective and the deep training objective. The target prediction values ​​corresponding to each training objective include the conditional prediction values ​​between the first objective and the second objective. The causal strength determination module 913 can be further configured to: determine the causal conditional weight value corresponding to the current training sample based on the obtained conditional probability values, the target prediction values, and the corresponding path from the direct training objective to the deep training objective; and determine the deep-direct objective causal strength value corresponding to the current training sample based on the relationship between the determined causal conditional weight value and 1. The operation of the causal strength determination module 913 can be referred to above. Figure 4 The operation described in 410-420

[0117] The deep target prediction module 914 is configured to determine the deep target prediction value corresponding to the current training sample based on the obtained deep-direct target causal strength value and the corresponding target prediction value. The operation of the deep target prediction module 914 can be referred to above. Figure 2 The operation described is 250.

[0118] In one example, the target prediction values ​​corresponding to each training objective further include a first target prediction value, a conditional prediction value that satisfies the deep training objective while satisfying the direct training objective, and a conditional prediction value that still satisfies the deep training objective even if the direct training objective is not satisfied. The deep target prediction module 914 can be further configured to: calculate the deep target prediction value corresponding to the current training sample based on the obtained deep-direct target causal strength value and the corresponding target prediction value using the full probability expansion formula. The operation of the deep target prediction module 914 can be referred to above. Figure 2 The operation of the 250 optional implementations is described.

[0119] The loss determination module 915 is configured to determine the loss value for the current model training process based on the difference between the deep target prediction value and the corresponding label data for each current training sample. The operation of the loss determination module 915 can be referenced above. Figure 2 The operation described in section 260.

[0120] In one example, the loss determination module 915 can be further configured to determine a deep target loss value based on the difference between the deep target prediction value and the corresponding label data of each current training sample; determine a direct target loss value based on the difference between the prediction value and the corresponding label data of the direct training target on which the deep training target of each current training sample is based; and determine the loss value of the current model training process based on the determined deep target loss value and the direct target loss value. The operation of the loss determination module 915 can be referenced above. Figure 6 The operation described in section 610-630.

[0121] The parameter adjustment unit 920 is configured to adjust the model parameters of the current prediction model based on the loss value in response to the failure to meet the training termination condition. The prediction model after parameter adjustment serves as the current prediction model for the next model training process. The operation of the parameter adjustment unit 920 can be referred to above. Figure 2 The operation described in section 280.

[0122] The following is for reference. Figure 10 , Figure 10 A block diagram illustrating an example of an information recommendation device 1000 according to an embodiment of this specification is shown. This device embodiment can be used with... Figure 8 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0123] like Figure 10 As shown, the information recommendation device 1000 may include a feature extraction unit 1010, a probability prediction unit 1020, and an information recommendation unit 1030.

[0124] The feature extraction unit 1010 is configured to generate a user-information pair feature set based on the relevant information of the target user and the set of information to be recommended. Each user-information pair feature in the feature set includes user features, item features, and historical interaction features generated based on the relevant information of the target user and the information to be recommended. The operation of the feature extraction unit 1010 can be referred to above. Figure 8 The operation of 810 is described.

[0125] The probability prediction unit 1020 is configured to provide the user-information pair feature set to the prediction model to obtain at least one predicted value corresponding to each user-information pair feature. The prediction model is trained using the model training device based on deep training objective optimization as described above, and each predicted value corresponds to a direct training objective or a deep training objective based on the direct training objective. The operation of the probability prediction unit 1020 can be referred to above. Figure 8 The operation of 820 is described.

[0126] The information recommendation unit 1030 is configured to determine target recommendation information from the set of information to be recommended based on at least one obtained prediction value; and to provide the target recommendation information to the target user. The operation of the information recommendation unit 1030 can be referred to the above. Figure 8 The operation described in section 830-840.

[0127] Reference above Figures 1 to 10 This specification describes embodiments of a model training method and apparatus based on deep target optimization, as well as an information recommendation method and apparatus.

[0128] The model training device and information recommendation device based on deep objective optimization described in this specification can be implemented in hardware, software, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of its host device reading the corresponding computer program instructions from the memory into memory and executing them. In the embodiments of this specification, the model training device and information recommendation device based on deep objective optimization can, for example, be implemented using electronic devices.

[0129] Figure 11 A schematic diagram of a model training apparatus 1100 based on deep target optimization according to an embodiment of this specification is shown.

[0130] like Figure 11 As shown, the deep target optimization-based model training apparatus 1100 may include at least one processor 1110, a memory (e.g., non-volatile memory) 1120, a RAM 1130, and a communication interface 1140, and the at least one processor 1110, memory 1120, RAM 1130, and communication interface 1140 are connected together via a bus 1150. The at least one processor 1110 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0131] In one embodiment, computer-executable instructions are stored in memory, which, when executed, cause at least one processor 1110 to: cyclically execute the following model training process using a training sample set until a training termination condition is met, wherein each training sample in the training sample set includes feature data and label data matching a training objective, wherein the training objective includes a direct training objective and a deep training objective based on the direct training objective, and the feature data includes user features, item features, and historical interaction features, wherein the historical interaction features are used to indicate whether there has been an interaction between the user and an item that matches the training objective in the past period; for each current training sample in the current training sample set, based on the user features and historical interaction features of the current training sample, obtain conditional probability values ​​between each training objective; and then, based on the user features and historical interaction features of the current training sample, obtain conditional probability values ​​between each training objective; and then, based on the user features and historical interaction features of the current training sample, obtain conditional probability values ​​between each training objective; and then, based on the user features and historical interaction features of the current training sample, obtain conditional probability values ​​between each training objective. The training data is provided to the current prediction model to obtain target prediction values ​​corresponding to each training objective, wherein the target prediction values ​​corresponding to the training objectives include conditional prediction values ​​involving other training objectives with causal relationships; the deep-direct target causal strength value corresponding to the current training sample is determined based on the obtained conditional probability value and the corresponding conditional prediction value; the deep target prediction value corresponding to the current training sample is determined based on the obtained deep-direct target causal strength value and the corresponding target prediction value; the loss value of the current model training process is determined based on the difference between the deep target prediction values ​​and the corresponding label data of each current training sample; and in response to the failure to meet the training termination condition, the model parameters of the current prediction model are adjusted according to the loss value, wherein the prediction model after the model parameter adjustment serves as the current prediction model for the next model training process.

[0132] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 1110 to perform the above-described combinations in the various embodiments of this specification. Figures 1-7 The description includes various operations and functions.

[0133] Figure 12 A schematic diagram of an information recommendation device 1200 according to an embodiment of this specification is shown.

[0134] like Figure 12 As shown, the information recommendation device 1200 may include at least one processor 1210, a memory (e.g., non-volatile memory) 1220, a RAM 1230, and a communication interface 1240, and the at least one processor 1210, memory 1220, RAM 1230, and communication interface 1240 are connected together via a bus 1250. At least one processor 1210 executes at least one computer-readable instruction (i.e., the elements implemented in software described above) stored or encoded in the memory.

[0135] In one embodiment, computer-executable instructions are stored in memory, which, when executed, cause at least one processor 1210 to: generate a user-information pair feature set based on relevant information of a target user and a set of information to be recommended, wherein each user-information pair feature in the user-information pair feature set includes user features, item features, and historical interaction features generated based on the relevant information of the target user and the information to be recommended; provide the user-information pair feature set to a prediction model to obtain at least one predicted value corresponding to each user-information pair feature, wherein the prediction model is trained by the model training method based on deep training objective optimization as described above, and each predicted value corresponds to a direct training objective or a deep training objective based on a direct training objective; determine target recommendation information from the set of information to be recommended based on the obtained at least one predicted value; and provide the target recommendation information to the target user.

[0136] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor 1210 to perform the above-described combinations in the various embodiments of this specification. Figure 8 The description includes various operations and functions.

[0137] According to one embodiment, a program product, such as a computer-readable medium, is provided. The computer-readable medium may have instructions (i.e., the elements implemented in software as described above), which, when executed by a computer, cause the computer to perform the above-described combinations of the various embodiments of this specification. Figures 1-8 The description includes various operations and functions.

[0138] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0139] In this case, the program code itself, which can be read from a readable medium, can perform the functions of any of the above embodiments. Therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0140] The computer program code required for the operation of each part of this manual can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB, .NET, and Python; conventional procedural programming languages ​​such as C, Visual Basic 2003, Perl, COBOL 2002, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service, such as Software as a Service (SaaS).

[0141] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0142] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0143] Not all steps and units in the above process and system structure diagrams are mandatory; some steps or units can be omitted as needed. The execution order of each step is not fixed and can be determined as required. The device structure described in the above embodiments can be a physical structure or a logical structure. That is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0144] The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" over other embodiments. Detailed descriptions are included for the purpose of providing an understanding of the described techniques. However, these techniques may be practiced without these detailed descriptions. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0145] The optional embodiments of the present specification have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present specification are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present specification, various simple modifications can be made to the technical solutions of the embodiments of the present specification, and these simple modifications all fall within the protection scope of the embodiments of the present specification.

[0146] The foregoing description of this specification is provided to enable any person skilled in the art to implement or use the content of this specification. Various modifications to the content of this specification will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of protection of this specification. Therefore, this specification is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for training a model based on deep target optimization, comprising: performing a model training process cyclically using a training sample set until a training end condition is met, each training sample in the training sample set comprising feature data and label data matching a training target, wherein the training target comprises a direct training target and a deep training target based on the direct training target, the feature data comprising user features, item features, and historical interaction features indicating whether there is an interaction matching the training target between a user and an item in a past period of time; for each current training sample in the training sample set, obtaining conditional probability values between the training targets based on the user features and the historical interaction features of the current training sample, the training targets having a conditional dependency relationship, the conditional probability values comprising probabilities of satisfying the deep training target under the condition of satisfying the direct training target; providing the feature data of the current training sample to a current prediction model to obtain target prediction values corresponding to the training targets, wherein the target prediction values corresponding to the training targets comprise conditional prediction values involving other training targets having a causal association; determining a deep-direct target causal strength value corresponding to the current training sample based on the obtained conditional probability values and the corresponding conditional prediction values, the deep-direct target causal strength value indicating to what extent the implementation of the deep training target is attributed to the direct training target; determining a deep target prediction value corresponding to the current training sample based on the obtained deep-direct target causal strength value and the corresponding target prediction values; determining a loss value of the current model training process based on differences between the deep target prediction values of each current training sample and the corresponding label data; and in response to the training end condition not being met, adjusting model parameters of the current prediction model based on the loss value, wherein the prediction model after the model parameter adjustment serves as a current prediction model of a next model training process.

2. The method of claim 1, wherein, the obtaining of the conditional probability values between the training targets based on the user features and the historical interaction features of the current training sample comprises: providing the user features and the historical interaction features of the current training sample to an encoding structure in an autoencoder to obtain an encoding result, wherein the autoencoder is obtained by training based on prediction differences of the historical interaction features; and generating a matrix indicating the conditional probability values between the training targets based on the encoding result.

3. The method of claim 1, wherein, the feature data of each training sample in the training sample set further comprises recommendation domain features, and the current prediction model comprises a current domain perception model and a current prediction value generation model, the providing of the feature data of the current training sample to the current prediction model to obtain the target prediction values corresponding to the training targets comprises: providing the recommendation domain features of the current training sample to the current domain perception model to obtain domain perception features of the current training sample; and provide the obtained domain-aware feature and corresponding user feature, item feature and historical interaction feature to a current prediction value generation model to obtain initial target prediction values corresponding to each training target; and correct the initial target prediction value corresponding to the current training sample according to each initial target prediction value corresponding to other current training samples matched with the recommendation domain feature corresponding to the current training sample, and take the corrected initial target prediction value as the corresponding target prediction value.

4. The method of claim 1, wherein, the direct training target includes a first target and a second target, the deep training target includes a third target based on the first target and the second target, the conditional probability value includes a first conditional probability value and a second conditional probability value for representing the conditional probability between the direct training target and the deep training target, the target prediction value corresponding to each training target includes a conditional prediction value between the first target and the second target, the determining of the deep-direct target causal strength value corresponding to the current training sample according to the obtained conditional probability value and the corresponding target prediction value includes: determining a causal condition weight value corresponding to the current training sample according to the obtained conditional probability value and the target prediction value and a path of the direct training target to the deep training target; and determining the deep-direct target causal strength value corresponding to the current training sample according to the size relationship between the determined causal condition weight value and 1.

5. The method of claim 4, wherein, the target prediction value corresponding to each training target further includes a first target prediction value, a conditional prediction value meeting the deep training target on the basis of meeting the direct training target, and a conditional prediction value meeting the deep training target on the basis of not meeting the direct training target, the determining of the deep target prediction value corresponding to the current training sample according to the obtained deep-direct target causal strength value and the corresponding target prediction value includes: calculating the deep target prediction value corresponding to the current training sample according to the obtained deep-direct target causal strength value and the corresponding target prediction value based on a total probability expansion formula.

6. The method of any one of claims 1 to 5, wherein, the determining of the loss value of the current model training process according to the difference between the deep target prediction value of each current training sample and the corresponding label data includes: determining a deep target loss value according to the difference between the deep target prediction value of each current training sample and the corresponding label data; determining a direct target loss value according to the difference between the prediction value of the direct training target based on which the deep training target of each current training sample is and the corresponding label data; and determining the loss value of the current model training process according to the determined deep target loss value and direct target loss value.

7. An information recommendation method, comprising: generating a user-information pair feature set according to relevant information of a target user and a set of to-be-recommended information, wherein each user-information pair feature in the user-information pair feature set includes a user feature, an item feature and a historical interaction feature generated based on the relevant information of the target user and the to-be-recommended information; provide the user-information pair feature set to a prediction model to obtain at least one prediction value corresponding to each user-information pair feature, wherein the prediction model is trained by the model training method based on deep training target optimization according to any one of claims 1 to 6, and each prediction value corresponds to a direct training target or a deep training target based on the direct training target; determine target recommendation information from the set of to-be-recommended information according to the obtained at least one prediction value; and provide the target recommendation information to the target user.

8. A model training device based on deep target optimization, the device being configured to perform a model training process by a training unit with a training sample set cyclically until a training end condition is satisfied, each training sample in the training sample set comprising feature data and label data matching a training target, wherein, The training target includes a direct training target and a deep training target based on the direct training target, the feature data includes user features, item features, and historical interaction features, the historical interaction features are used to indicate whether there is an interaction between the user and the item within a past period of time that matches the training target, and the training unit includes: a conditional probability determination module configured to obtain a conditional probability value between each training target according to user features and historical interaction features of a current training sample, the training targets having a conditional dependency relationship, and the conditional probability value including a probability of satisfying the deep training target under the condition of satisfying the direct training target; an initial prediction module configured to provide feature data of the current training sample to a current prediction model to obtain target prediction values corresponding to each training target, wherein the target prediction values corresponding to the training targets include conditional prediction values involving other training targets having a causal association; a causal strength determination module configured to determine a deep-direct target causal strength value corresponding to the current training sample according to the obtained conditional probability value and the corresponding conditional prediction value, the deep-direct target causal strength value being used to indicate to what extent the implementation of the deep training target is attributed to the direct training target; a deep target prediction module configured to determine a deep target prediction value corresponding to the current training sample according to the obtained deep-direct target causal strength value and the corresponding target prediction value; a loss value determination module configured to determine a loss value of the current model training process according to a difference between the deep target prediction value of each current training sample and corresponding label data; and The device further includes: a parameter adjustment unit configured to adjust model parameters of the current prediction model according to the loss value in response to not satisfying the training end condition, wherein the prediction model after the model parameter adjustment serves as the current prediction model of the next model training process.

9. An information recommendation device, comprising: a feature extraction unit configured to generate a user-information pair feature set according to relevant information of a target user and a set of to-be-recommended information, wherein each user-information pair feature in the user-information pair feature set includes user features, item features, and historical interaction features generated based on the relevant information of the target user and the to-be-recommended information; a probability prediction unit configured to provide the user-information pair feature set to a prediction model to obtain at least one prediction value corresponding to each user-information pair feature, wherein the prediction model is trained by the model training apparatus based on deep training target optimization as claimed in claim 8, and each prediction value corresponds to a direct training target or a deep training target based on the direct training target; an information recommendation unit configured to determine target recommendation information from the set of to-be-recommended information according to the obtained at least one prediction value, and provide the target recommendation information to the target user.

10. A model training apparatus based on deep target optimization, comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored on the memory, wherein the at least one processor executes the computer program to implement the method as claimed in any one of claims 1 to 6.

11. An information recommendation apparatus comprising: at least one processor, a memory coupled to the at least one processor, and a computer program stored on the memory, wherein the at least one processor executes the computer program to implement the information recommendation method as claimed in claim 7.

12. A computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the model training method based on deep target optimization as claimed in any one of claims 1 to 6 and / or the information recommendation method as claimed in claim 7.

13. A computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the model training method based on deep target optimization as claimed in any one of claims 1 to 6 and / or the information recommendation method as claimed in claim 7.

Images