Model training method, resource allocation method, device and computer equipment

Abstract

Embodiments of the present specification disclose a model training method, a resource allocation method, a device and computer equipment. The method comprises: obtaining a plurality of sample data, the sample data comprising user data, resource data and a relationship label, the relationship label indicating whether the user data and the resource data have an association relationship; counting a first distribution of the user data and a second distribution of the resource data; fusing the first distribution of a plurality of user data and fusing the second distribution of a plurality of resource data; and determining model parameters of a model according to the user data, the resource data, the fused first distribution, the fused second distribution and the relationship label, the model being used to predict the association relationship between the user data and the resource data. Embodiments of the present specification can correct the sample data and improve the model training effect.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of computer technology, and in particular to a model training method, resource allocation method, apparatus, and computer equipment. Background Technology

[0002] Currently, to increase user engagement, resources can be distributed to users for their use. These resources may include red envelopes, coupons, etc. To this end, a model can be trained to distribute resources to users. Summary of the Invention

[0003] This specification provides a model training method to improve training effectiveness. Additionally, this specification also provides a resource allocation method to improve the accuracy of resource allocation.

[0004] A first aspect of the embodiments of this specification provides a model training method, including:

[0005] Acquire a number of sample data, including user data, resource data, and relationship tags, wherein the relationship tags indicate whether user data and resource data are related.

[0006] The system calculates a first distribution of user data and a second distribution of resource data. The first distribution represents the association between user data and multiple resource data, and the second distribution represents the association between resource data and multiple user data.

[0007] The first distribution of multiple user data is merged, and the second distribution of multiple resource data is merged.

[0008] Based on user data, resource data, the fused first distribution, the fused second distribution, and relationship labels, the model parameters are determined, and the model is used to predict the association between user data and resource data.

[0009] A second aspect of the embodiments of this specification provides a resource allocation method, including:

[0010] Receive resource allocation requests;

[0011] Based on user data, the historical resource distribution of user data, the resource data in the resource dataset, and the historical user distribution of resource data, the model predicts the correlation between user data and the resource data in the resource dataset.

[0012] Based on the prediction results, target resource data is selected from the resource dataset;

[0013] Based on the target resource data, resources are allocated to the users corresponding to the user data.

[0014] A third aspect of the embodiments of this specification provides a model training apparatus, comprising:

[0015] An acquisition unit is used to acquire a number of sample data, the sample data including user data, resource data and relationship tags, the relationship tags indicating whether user data and resource data have a correlation relationship;

[0016] The statistical unit is used to statistically analyze the first distribution of user data and the second distribution of resource data. The first distribution represents the relationship between user data and multiple resource data, and the second distribution represents the relationship between resource data and multiple user data.

[0017] The fusion unit is used to fuse the first distribution of multiple user data and the second distribution of multiple resource data.

[0018] The determining unit is used to determine the model parameters of the model based on user data, resource data, the fused first distribution, the fused second distribution, and relationship labels. The model is used to predict the association relationship between user data and resource data.

[0019] A fourth aspect of the embodiments of this specification provides a resource allocation apparatus, comprising:

[0020] The receiving unit is used to receive resource allocation requests;

[0021] The prediction unit is used to predict the relationship between user data and resource data in the resource dataset based on user data, the historical resource distribution of user data, resource data in the resource dataset, and the historical user distribution of resource data.

[0022] The selection unit is used to select target resource data from the resource dataset based on the prediction results;

[0023] The allocation unit is used to allocate resources to the users corresponding to the user data based on the target resource data.

[0024] A fifth aspect of the embodiments of this specification provides a computer device, including:

[0025] At least one processor;

[0026] A memory storing program instructions configured to be executed by the at least one processor, the program instructions including instructions for performing the method of the first or second aspect.

[0027] The technical solutions provided in the embodiments of this specification can acquire a number of sample data, including user data, resource data, and relationship tags; they can statistically analyze a first distribution of user data and a second distribution of resource data; they can fuse the first distributions of multiple user data sets and the second distributions of multiple resource data sets; and they can determine the model parameters based on the user data, resource data, the fused first distribution, the fused second distribution, and the relationship tags. The fused first and second distributions are less affected by external factors. Causal intervention can be performed using the fused first and second distributions to correct biases in the sample data and improve model training performance. Furthermore, the technical solutions provided in the embodiments of this specification can also accurately recommend resources to users through the model. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1a This is a user distribution diagram for small-amount coupons related to the technology.

[0030] Figure 1b This is a user distribution diagram for large-value coupons in related technologies;

[0031] Figure 2a This is a diagram illustrating the distribution of coupons among high-API users in related technologies.

[0032] Figure 2b This is a diagram illustrating the distribution of coupons for low-API users in related technologies.

[0033] Figure 3 For cause-effect graphs in related technologies;

[0034] Figure 4 The causal graph after intervening in the causal dry cleaning of nodes D and U, and the causal relationship between nodes K and I;

[0035] Figure 5 This specification includes a flowchart illustrating the model training method in the embodiments.

[0036] Figure 6 This specification contains a schematic diagram of the functional structure of the model in the embodiments.

[0037] Figure 7 This specification includes flowcharts illustrating the resource allocation methods in the embodiments.

[0038] Figure 8 This specification provides a schematic diagram of the resource request page in the embodiments.

[0039] Figure 9 This specification includes schematic diagrams illustrating the prompting information in the embodiments.

[0040] Figure 10 This is a schematic diagram of the model training device in the embodiments of this specification;

[0041] Figure 11 This is a schematic diagram of the resource allocation device in the embodiments of this specification;

[0042] Figure 12 This is a functional structure diagram of the computer device in the embodiments of this specification. Detailed Implementation

[0043] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. The specific embodiments described herein are only used to explain this disclosure, and not to limit this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure are within the scope of protection of this disclosure. In addition, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0044] In related technologies, sample data can be constructed based on users and the resources they receive. Models can then be trained using this sample data and used to distribute resources to users. However, the resources available to users are often influenced by external factors such as marketing strategies. This means that the sample data is not truly real data, but rather biased data influenced by external factors. Models trained on biased data have low confidence levels and cannot accurately recommend resources to users.

[0045] For example, in a marketing campaign, coupons of varying amounts can be issued to users for their use. The conversion rate represents the probability of a coupon being used. A higher conversion rate indicates a greater probability of coupon use, while a lower conversion rate indicates a lower probability. Coupons have an monetary attribute that affects their conversion rate; that is, the higher the amount, the higher the conversion rate. Users also have an API (Activity Participation Intention) attribute that affects coupon conversion rates; that is, the higher the API, the higher the conversion rate. In practice, coupon issuance incurs costs. To maximize Return on Investment (ROI) within a limited budget, coupons can be issued to users through marketing strategies. These strategies may include: prioritizing the issuance of smaller coupons to high-API users to save costs; and prioritizing the issuance of larger coupons to low-API users to increase DAU (Daily Active Users). This results in a distribution of users receiving smaller coupons (see [link to relevant documentation]). Figure 1a In the data, high-API users account for a larger proportion, while low-API users account for a smaller proportion; the distribution of users receiving large-value coupons (see [link to relevant documentation]). Figure 1b In the data, low-API users account for a larger proportion, while high-API users account for a smaller proportion; the coupon distribution for high-API users (see [link to relevant documentation]). Figure 2a In the data, small-amount coupons account for a larger proportion, while large-amount coupons account for a smaller proportion; the coupon distribution for low-API users (see [link to relevant documentation]). Figure 2b In this context, high-value coupons account for a larger proportion, while low-value coupons account for a smaller proportion. This indicates that the coupons users receive are not entirely dependent on the user's API or the coupon's value, but are also influenced by marketing strategies.

[0046] In the aforementioned related technologies, the influence of external factors on the model training process can be explained using a causal graph. The causal graph can be a directed graph, specifically including nodes and directed edges. Nodes represent subjects, directed edges represent causal relationships between subjects, and the direction of the directed edges represents the direction of the causal relationship. For example, the direction of the directed edge can be from the source node to the target node. The source node can be the cause, and the target node can be the result.

[0047] exist Figure 3In the causal graph shown, node U represents user data; node I represents resource data; node D represents the historical resource distribution of users; node K represents the historical user distribution of resources; node T represents the joint distribution of users and resources, which is used to represent the degree of correlation between users and resources; node L represents external factors; and node Y represents the conversion rate of resources. The directed edge U→Y represents the influence of user attributes (e.g., user API) on Y; the directed edge I→Y represents the influence of resource attributes (e.g., coupon amount) on Y; the directed edge D→U represents the influence of historical resource distribution on U; the directed edge K→I represents the influence of historical user distribution on I; the directed edges D, U, K, I→T represent the combined influence of user data, historical resource distribution, resource data, and historical user distribution on T; the directed edge T→Y represents the influence of the joint distribution of users and resources on Y; the directed edge L→D represents the influence of external factors on D; and the directed edge L→K represents the influence of external factors on K. The dashed lines indicate that the influence of node L on nodes D and K is unobservable but real.

[0048] according to Figure 3 As shown in the causal graph, node D can influence node Y not only through node T, but also through node U. D→U→Y constitutes a backdoor path. Similarly, node K can influence node Y not only through node T, but also through node I. K→I→Y constitutes a backdoor path. Due to the existence of these backdoor paths, a falsely high value of Y appears between a specific user and a specific resource. The specific user is a user with a high percentage in the historical user distribution of the resource, and the specific resource is a resource with a high percentage in the historical resource distribution of the user. The high percentage of a specific user is due to the influence of external factors in addition to the user's own attributes. The high percentage of a specific resource is due to the influence of external factors in addition to the influence of the resource's own attributes.

[0049] In the aforementioned related technologies, the influence of external factors on the model training process can also be understood through the following formula.

[0050]

[0051]

[0052] Equations (1a), (1b), and (1c) can be obtained from the law of total probability. Equations (1d) and (1e) can be obtained from Bayes' theorem. In equation (1f), the summation of the numerator with respect to t has a value only under the conditions of D = d, U = u, K = k, and I = i. Therefore, the summation with respect to t can be eliminated. P(T|D u K i Uu I i = 1. In formula (1g), the sum of the denominators is 1. In formula (1h), when U is given, D is fixed. P(D) u |U u = 1. Similarly, when I is given, K is also fixed. P(K) i |I i = 1. Therefore, the summation of d and k can be eliminated. Furthermore, formula (1h) shows that the model training process is through D... u and K i To find T(D) u K i U, I), U, I). This allows external factors to be influenced by D. u and K i Effects on T(D) u K i ,U,I),U,I), which in turn affects Y.

[0053] Please see Figure 4 If the causal relationship between nodes D and U, and between nodes K and I, can be intervened through backdoor adjustments, the backdoor path can be blocked, mitigating the impact of external factors on the sample data. This would allow for data correction and improve the confidence of the trained model. For further information, please refer to [link to relevant documentation]. Figure 5 This specification provides a model training method. The model training method can be executed by a computer device. The computer device can include any apparatus, equipment, platform, or cluster of devices with computing power. The method can include the following steps.

[0054] Step S11: Obtain a number of sample data, including user data, resource data, and relationship tags, wherein the relationship tags indicate whether user data and resource data are related.

[0055] In some embodiments, the plurality of sample data may include any number of sample data. The sample data may include user data, resource data, and relationship tags. The user data describes a user and may include one or more sub-data. The sub-data may include basic data, wealth data, transaction data, etc. The basic data may include age, gender, occupation, API (activity participation intention), etc. The wealth data may include income, purchasing power, the quantity of resources received, the amount of resources received, etc. The transaction data may include transfer amount, number of transactions, etc. The resource data describes resources and may include one or more sub-data. The resources may include virtual resources and physical resources; virtual resources may include red envelopes, coupons, recharge cards, etc., and physical resources may include physical products, etc. The sub-data may include the amount of the resource, the discount of the resource, the identifier of the resource, etc. The relationship tags are used to indicate whether user data and resource data have a relationship. For example, the relationship tags may include 0 and 1, where 0 indicates that user data and resource data have no relationship, and 1 indicates that user data and resource data have a relationship.

[0056] The association relationship can include an interaction relationship. The association relationship can be used to indicate that a user corresponding to user data has performed a specific interactive behavior on a resource corresponding to resource data. The specific interactive behavior can include clicking, browsing, transaction, verification, adding to cart, etc. In practical applications, the association relationship can vary depending on the application scenario of the embodiments of this specification. For example, the application scenario of the embodiments of this specification can be a virtual resource distribution scenario, and the association relationship can include a verification relationship. The verification relationship is used to indicate whether a user has used virtual resources (e.g., whether a user has used a coupon). Another example is the application scenario of the embodiments of this specification can be a physical resource transaction scenario, and the association relationship can include a transaction relationship. The purchase relationship is used to indicate whether a user has purchased physical resources. Yet another example is the application scenario of the embodiments of this specification can be a virtual resource pricing scenario, where resource data can include virtual resource pricing data, and the association relationship can include the correspondence between user data and pricing data. The correspondence relationship is used to indicate whether a price has been set for a user's virtual resources (e.g., whether a certain price has been set for a user to activate a membership card).

[0057] In some embodiments, the user data and resource data in different sample datasets may be the same or different. For example, the sample datasets may include three sample datasets: data1, data2, and data3. Sample data1 may include user data (user1), resource data (Item1), and relationship label 0. Sample data2 may include user data (user1), resource data (Item2), and relationship label 1. Sample data3 may include user data (user2), resource data (Item1), and relationship label 1.

[0058] In some embodiments, user data may belong to multiple user data groups within the plurality of sample data. Each user data group includes at least one user data group. Each user data group may correspond to an activity level. The plurality of user data groups may correspond to multiple activity levels. The activity level is related to the user's intention to participate in the activity. A higher activity level indicates a higher intention to participate in the activity. A lower activity level indicates a lower intention to participate in the activity.

[0059] In some embodiments, resource data may belong to multiple resource data groups within the plurality of sample data. Each resource data group includes at least one resource data. Each resource data group corresponds to a cost tier. The plurality of resource data groups may correspond to multiple cost tiers. The cost tier is related to resource cost (e.g., coupon cost). The higher the cost tier, the greater the cost required to use the resource. The lower the cost tier, the lower the cost required to use the resource.

[0060] In some embodiments, a number of sample data can be collected. Alternatively, a number of sample data can be received from other devices. Alternatively, a sample dataset can be provided; the sample data in the sample dataset can be divided into several batches; and a batch of sample data can be obtained. The sample dataset may include multiple sample data.

[0061] Step S13: Calculate the first distribution of user data and the second distribution of resource data. The first distribution represents the association between user data and multiple resource data, and the second distribution represents the association between resource data and multiple user data.

[0062] In some embodiments, a first distribution can be statistically analyzed for each user data point. This first distribution represents the association between the user data and multiple resource data points. The multiple resource data points may include all or part of the resource data from the plurality of sample data points. The first distribution can serve as the historical resource distribution corresponding to the user data.

[0063] Additionally, a second distribution can be calculated for each resource data point. This second distribution represents the association between the resource data and multiple user data points. The multiple user data points may include all or a portion of the user data from the plurality of sample data points. The second distribution can serve as the historical user distribution corresponding to the resource data.

[0064] In some embodiments, the first distribution may include the association probability between user data and multiple resource data. The association probability can be understood as the probability that user data and resource data are associated. The first distribution may be a vector or matrix, etc. The number of associations between user data and each resource data among the multiple resource data can be counted as a first association count; the first distribution can be determined based on the first association count. Specifically, the association probability between user data and that resource data can be obtained by dividing the first association count corresponding to each resource data by the sum of the first association counts corresponding to the multiple resource data.

[0065] For example, the number of sample data may include four sample data such as data1, data2, data3, and data4.

[0066] Table 1

[0067] data1 user1 item1 0 data2 user2 item1 1 data3 user2 item2 0 data4 user1 item2 1

[0068] In Table 1, relationship label 0 indicates that user data and resource data are not related, and relationship label 1 indicates that user data and resource data are related. Based on Table 1, the first association counts between user data (user1) and resource data (item1) and (item2) are 0 and 1, respectively. The association probabilities between user data (user1) and resource data (item1) and (item2) are respectively... The first distribution corresponding to user data user1 can be a vector [0, 1]. According to Table 1, the first association counts between user data user2 and resource data item1 and item2 are 1 and 0, respectively. The association probabilities between user data user2 and resource data item1 and item2 are respectively... Then the first distribution corresponding to user data user2 can be the vector [1, 0].

[0069] The second distribution may include the association probability between resource data and multiple user data. Association probability can be understood as the probability that user data and resource data are related. The second distribution can be a vector or matrix, etc. The number of associations between resource data and each user data in the multiple user data sets can be counted as a second association count; the second distribution can be determined based on the second association count. Specifically, the association probability between resource data and that user data can be obtained by dividing the second association count corresponding to each user data set by the sum of the second association counts corresponding to the multiple user data sets.

[0070] In some embodiments, the multiple resource data may belong to multiple resource data groups. The first distribution may include the association probability between user data and multiple resource data groups. The association probability can be understood as the probability that user data and resource data groups have an association relationship. The first distribution may be a vector or matrix, etc. The number of associations between user data and each resource data group in the multiple resource data groups can be counted as a third association count; the first distribution can be determined based on the third association count. Specifically, the association probability between user data and that resource data group can be obtained by dividing the third association count corresponding to each resource data group by the sum of the third association counts corresponding to the multiple resource data groups. The third association count may include the sum of the association counts between user data and each resource data in the resource data group.

[0071] For example, the multiple resource data groups may include three resource data groups: group1, group2, and group3. Resource data group1 includes two resource data items: item1 and item2. Resource data group2 includes two resource data items: item3 and item4. Resource data group3 includes two resource data items: item5 and item6.

[0072] Table 2

[0073] data1 user1 item1 0 data2 user2 item2 1 data3 user2 item3 0 data4 user1 item2 1 data5 user1 item4 0 data6 user2 item5 1 data7 user3 item6 0 data8 user3 item6 1 data9 user2 item5 0 data10 user1 item4 1 data11 user3 item1 1

[0074] In Table 2, relationship label 0 indicates that user data and resource data are not related, and relationship label 1 indicates that user data and resource data are related. Based on Table 2, the number of third associations between user data `user1` and resource data groups `group1`, `group2`, and `group3` are 1, 1, and 0, respectively. The association probabilities of user data `user1` with resource data groups `group1`, `group2`, and `group3` are respectively... The first distribution corresponding to user data user1 can be a vector [0.5, 0.5, 0]. According to Table 2, the third association counts between user data user2 and resource data groups group1, group2, and group3 are 1, 0, and 1, respectively. The association probabilities between user data user2 and resource data groups group1, group2, and group3 are respectively... The first distribution corresponding to user data user2 can be a vector [0.5, 0, 0.5]. According to Table 2, the third association counts between user data user3 and resource data groups group1, group2, and group3 are 1, 0, and 1, respectively. The association probabilities between user data user3 and resource data groups group1, group2, and group3 are respectively... Then the first distribution corresponding to the user data user3 can be the vector [0.5, 0, 0.5].

[0075] The multiple user data sets may belong to multiple user data groups. The second distribution may include the association probability between resource data and multiple user data groups. The association probability can be understood as the probability that resource data and user data groups have an association relationship. The second distribution may be a vector or matrix, etc. The number of associations between resource data and each user data group in the multiple user data groups can be counted as the fourth association count; the second distribution can be determined based on the fourth association count. Specifically, the association probability between resource data and that user data group can be obtained by dividing the fourth association count corresponding to each user data group by the sum of the fourth association counts corresponding to the multiple user data groups. The fourth association count may include the sum of the association counts between resource data and each user data set within the user data group.

[0076] Step S15: Merge the first distribution of multiple user data and merge the second distribution of multiple resource data.

[0077] In some embodiments, the first distribution of the plurality of user data may include the first distribution of all user data or the first distribution of a portion of the user data in the plurality of sample data. This can be determined according to the formula ∑ d∈Dp(d)d represents the fusion of the first distributions of multiple user data sets. D represents the set of the first distributions of the multiple user data sets, d represents the first distribution in set D, and p(d) represents the probability of d appearing in set D. It should be noted that the first distributions of different user data sets can be the same or different. This allows one or more first distributions in set D to be identical. The probability of d appearing in set D can be obtained by dividing the number of d by the number of first distributions in set D. Of course, other methods can also be used to fuse the first distributions of multiple user data sets. For example, the fusion can be achieved by performing mathematical operations such as addition, subtraction, multiplication, and division on the first distributions of multiple user data sets.

[0078] The fused first distribution is obtained by fusing the first distributions of multiple user data sets. This makes the fused first distribution less susceptible to external factors and closer to reality compared to the first distribution of individual user data sets. Therefore, the fused first distribution can be used to intervene in the causal relationship between the first distribution and user data.

[0079] In some embodiments, the second distribution of the plurality of resource data may include the second distribution of all resource data or the second distribution of some resource data in the plurality of sample data. This can be determined according to the formula ∑ k∈K p(k)k merges the second distributions of multiple resource data. K represents the set formed by the second distributions of the multiple resource data, k represents the second distribution in set K, and p(k) represents the probability of k appearing in set K. It should be noted that the second distributions of different resource data can be the same or different. This allows one or more second distributions in set K to be identical. The probability of k appearing in set K can be obtained by dividing the number of k by the number of second distributions in set K. Of course, other methods can also be used to merge the second distributions of multiple resource data. For example, the second distributions of multiple resource data can be merged by performing mathematical operations such as addition, subtraction, multiplication, and division.

[0080] The fused second distribution is obtained by fusing the second distributions of multiple resource data. This makes the fused second distribution less susceptible to external factors and closer to reality compared to the second distribution of a single resource data point. Therefore, the fused second distribution can be used to intervene in the causal relationship between the second distribution and the resource data.

[0081] Step S17: Determine the model parameters based on user data, resource data, the fused first distribution, the fused second distribution, and the relationship labels. The model is used to predict the association between user data and resource data.

[0082] In some embodiments, as previously mentioned, the fused first distribution can be used to intervene in the causal relationship between the first distribution and user data, and the fused second distribution can be used to intervene in the causal relationship between the second distribution and resource data. That is, the model can utilize the fused first and second distributions to intervene in the causal relationship through backdoor adjustments, thereby correcting the bias of the sample data. In some scenario examples, the causal intervention can be implemented based on Do-Calculus. The specific causal intervention process can be understood through the following formula.

[0083]

[0084] In formulas (2c) and (2d), This represents the expectation. In formula (2e), ∑ d∈D p(D u )D u This can be understood as the first distribution of fusion. Compared to a single D u , ∑ d∈D p(D u )D u Less affected by external factors. k∈K p(K i )K i This can be understood as a second distribution after fusion. Compared to a single K... i , ∑ k∈k p(K i )K i It is less affected by external factors. (Through ∑) d∈D p(D u )D u and ∑ k∈K p(K i )K i , can be used for D u The causal relationship between U and K i Intervene in the causal relationship between I and Y. Reduce the influence of external factors on T, thereby reducing the influence of external factors on Y. It should be noted that ∑... d∈D p(D u )D u As the first distribution of fusion and ∑ k∈K p(K i )K i As a second distribution after fusion, it is only for the purpose of facilitating the understanding of the technical solutions of the embodiments in this specification, and does not constitute an improper limitation on the embodiments in this specification.

[0085] In some embodiments, the output of the model can be determined based on user data, resource data, the fused first distribution, and the fused second distribution; the model parameters can be determined based on the model output and relation labels.

[0086] Please see Figure 6 The model may include an embedding layer, an association layer, and a prediction layer. The embedding layer is used to obtain data representations. The association layer is used to determine the degree of association, which represents the degree of association between user data and resource data. The association layer may include factorization machines (FM), deep neural networks (DNN), multi-gate mixture-of-experts (MMOE), etc. The prediction layer is used to predict the association between user data and resource data. The prediction layer may include a classifier model, which may include a neural network model, etc.

[0087] The model can determine user representations, resource representations, first distribution representations, and second distribution representations based on user data, resource data, a fused first distribution, and a fused second distribution, respectively. It can also determine a correlation degree representation based on these representations, indicating the degree of correlation between user data and resource data. Finally, it can determine the model's output based on these representations. The user representation, resource representation, first distribution representation, second distribution representation, and correlation degree representation can be numerical values, vectors, or matrices. The model's output can be a score, representing the probability that user data and resource data are correlated.

[0088] In practical applications, user data, resource data, the fused first distribution, and the fused second distribution can be input into the embedding layer to obtain user representations, resource representations, first distribution representations, and second distribution representations. These representations can then be input into the correlation layer to obtain a correlation degree representation. Finally, they can be input into the prediction layer to obtain the model's output. The correlation layer can obtain the correlation degree representation by cross-referencing the user representation, resource representation, first distribution representation, and second distribution representation. For example, the correlation layer can be a factorization machine model. The factorization machine model can be expressed using the formula... Determine the characterization of the degree of association. V represents the number of model parameters in the factorization machine model, ω aω can be obtained by concatenating the a-th model parameter from the user representation, resource representation, first distribution representation, second distribution representation, and factorization machine model. a It can be obtained by concatenating the b-th model parameter in the user representation, resource representation, first distribution representation, second distribution representation, and factorization machine model, where ⊙ represents the Hadamard product.

[0089] Loss information can be calculated using a loss function based on the model's output and relation labels; model parameters can then be determined based on this loss information. The loss function can include cross-entropy loss, maximum likelihood loss (MLE), etc. Specifically, loss information can be obtained by substituting the model's output and relation labels into the loss function. For example, the loss information can be expressed as L = loss(f(u,i,T),y). ui u and i represent user representation and resource representation, respectively; T represents the degree of correlation between user data and resource data; f(u,i,T) represents the output of the model; y ui These labels represent the relationship between user data and resource data. Based on the loss information, the backpropagation mechanism can be used to optimize model parameters. For example, the gradient of the model parameters can be calculated using backpropagation; the model parameters can then be adjusted based on the gradient. In practical applications, all model parameters can be adjusted. Alternatively, some model parameters can be kept constant while others are adjusted. For example, the model parameters of the embedding and association layers can be kept constant while the model parameters of the prediction layer are adjusted.

[0090] In some embodiments, the model is used to predict the association between user data and resource data.

[0091] As previously mentioned, the relationships described in the embodiments of this specification may differ depending on the application scenario. Consequently, the models may also differ depending on the application scenario of the embodiments of this specification. For example, the application scenario of the embodiments of this specification may be a virtual resource distribution scenario, and the model may be a redemption probability prediction model. Another example is that the application scenario of the embodiments of this specification may be a physical resource trading scenario, and the model may be a trading probability prediction model. Yet another example is that the application scenario of the embodiments of this specification may be a virtual resource pricing scenario, and the model may be a virtual resource pricing prediction model.

[0092] In some embodiments, the model training method may include multiple iterations. Specifically, steps S11-S17 may be executed iteratively until a termination condition is met. The termination condition can be set according to actual needs. For example, the termination condition may be: the number of iterations reaches a threshold. As another example, the sample data in the sample dataset can be divided into several batches; in step S11, one batch of sample data can be obtained, and then this batch of sample data can be used to train the model. Therefore, the termination condition could also be: all batches of sample data have been used.

[0093] In some embodiments, the model training method may include multiple iterations. To simplify the fusion step of the first distribution, the first distribution of user data in the current iteration can be fused with the first distribution fused in the previous iteration. For example, it can be fused using the formula d×α+d′×(1-α), where d represents the first distribution of user data in the current iteration, d′ represents the first distribution fused in the previous iteration, and α represents the hyperparameter. To simplify the fusion step of the second distribution, the second distribution of resource data in the current iteration can be fused with the second distribution fused in the previous iteration. For example, it can be fused using the formula k×β+k′×(1-β), where k represents the second distribution of resource data in the current iteration, k′ represents the second distribution fused in the previous iteration, and β represents the hyperparameter.

[0094] The model training method described in this specification can acquire a number of sample data, including user data, resource data, and relationship labels; it can statistically analyze a first distribution of user data and a second distribution of resource data; it can fuse multiple first distributions of user data and multiple second distributions of resource data; and it can determine the model parameters based on the user data, resource data, the fused first distribution, the fused second distribution, and the relationship labels. The fused first and second distributions are less affected by external factors. Causal intervention can be performed using the fused first and second distributions to correct biases in the sample data and improve the model training effect.

[0095] Please see Figure 7 This specification also provides a resource allocation method. The resource allocation method can be executed by a computer device. The computer device can include any apparatus, device, platform, device cluster, etc., with computing power. The resource allocation method can include the following steps.

[0096] Step S21: Receive resource allocation request.

[0097] In some embodiments, a resource allocation request sent by a terminal device can be received. The terminal device includes, but is not limited to, smartphones, tablet computers, portable computers, and personal computers. The resource allocation request may include a user identifier. The user identifier is used to identify the user and may specifically be the user's mobile phone number, user ID, or email address. The resource allocation request is used to request the allocation of resources to the user. The resources may include virtual resources and physical resources. Virtual resources may include red envelopes, coupons, recharge cards, etc., and physical resources may include physical products, etc.

[0098] In some embodiments, the terminal device may display a resource request page. Users can perform operations on the resource request page to send instructions to the terminal device. Upon receiving an instruction, the terminal device can send a resource allocation request. See also... Figure 8 The resource request page may include a "Tap to Grab Coupons" button control. Users can click the "Tap to Grab Coupons" button control. After detecting a click operation on the "Tap to Grab Coupons" button control, the terminal device can send a resource allocation request. The resource allocation request may include the user's user identifier.

[0099] Step S23: Based on user data, the historical resource distribution of user data, the resource data in the resource dataset, and the historical user distribution of resource data, predict the correlation between user data and the resource data in the resource dataset using a model.

[0100] In some embodiments, corresponding user data can be obtained based on a user identifier. The user data may include one or more sub-data. The sub-data may include basic data, wealth data, transaction data, etc. Historical resource distribution is used to represent the association relationship between user data and multiple resource data. Specifically, historical resource distribution may include the association probability between user data and multiple resource data. The multiple resource data may include resource data in a resource dataset. Alternatively, historical resource distribution may also include the association probability between user data and multiple resource data groups. The resource data groups may include resource data in a resource dataset. Historical resource distribution can be a preset distribution. The preset distribution can be flexibly set according to actual needs. For example, historical resource distribution can be a uniform distribution [0.25, 0.25, 0.25, 0.25].

[0101] The resource dataset may include at least one resource data. The resource data may include one or more sub-data. The sub-data may include the resource's amount, discount, identifier, etc. The resource data in the resource dataset may correspond to a historical user distribution. The historical user distribution is used to represent the association relationship between resource data and multiple user data. Specifically, the historical user distribution may include the association probability between resource data and multiple user data. Alternatively, the historical user distribution may also include the association probability between resource data and a group of user data. The historical user distribution can be a preset distribution. The preset distribution can be flexibly set according to actual needs. For example, the historical user distribution can be a uniform distribution [0.25, 0.25, 0.25, 0.25].

[0102] In some embodiments, the model can be trained based on the foregoing embodiments.

[0103] In some embodiments, for each resource data in the resource dataset, user data, the historical resource distribution of the user data, the resource data itself, and the historical user distribution of the resource data can be input into the model to predict the association between the user data and the resource data. It should be noted that in practical applications, the model's output can be controlled by setting the historical resource distribution and historical user distribution, thereby controlling resource recommendations. For example, the historical resource distribution and historical user distribution can be set to a uniform distribution, allowing the model to recommend resources according to this uniform distribution.

[0104] Step S25: Select target resource data from the resource dataset based on the prediction results.

[0105] In some embodiments, after step S23, the resource data in the resource dataset may correspond to prediction results. The prediction results may include scores. The scores represent the probability that user data and resource data are correlated. The higher the score, the greater the probability that user data and resource data are correlated.

[0106] In some embodiments, target resource data can be selected directly from the resource dataset. For example, the resource data with the largest prediction result can be selected from the resource dataset as the target resource data. Another example is that the resource data may correspond to cost data, which represents the resource cost (e.g., coupon cost). The cost data may include the amount of the red envelope, the amount of the coupon, the discount of the coupon, etc. The prediction result corresponding to the resource data can be multiplied by the cost data to obtain the gain of the resource data; the resource data with the largest gain can be selected from the resource dataset as the target resource data.

[0107] In some embodiments, the resource dataset includes multiple resource data groups. Each resource data group includes at least one resource data. Each resource data group corresponds to a cost level. The multiple resource data groups may correspond to multiple cost levels. The cost level is related to the resource cost. The higher the cost level, the greater the resource cost. The lower the cost level, the smaller the resource cost.

[0108] In this way, the activity level corresponding to user data can be determined; a target resource data group can be selected from the resource dataset based on the activity level; the association between user data and resource data in the target resource data group can be predicted by the model based on user data, the historical resource distribution of user data, resource data in the target resource data group, and the historical user distribution of resource data; and target resource data can be selected from the target resource data group based on the prediction results.

[0109] Based on user data, an algorithm can be used to calculate the activity level corresponding to the user data. There is a correspondence between activity level and cost level. This correspondence can be flexibly set according to actual needs. For example, high activity level can correspond to a low-cost level, and low activity level can correspond to a high-cost level. Based on the activity level, a target resource data group can be selected from the resource dataset. The cost level corresponding to the target resource data group matches the activity level. For each resource data in the target resource data group, user data, the historical resource distribution of user data, resource data, and the historical user distribution of resource data can be input into the model to predict the association between user data and resource data. The process of selecting target resource data from the target resource data group is similar to the process of selecting target resource data from the resource dataset, and will not be described in detail here.

[0110] Step S27: Allocate resources to the users corresponding to the user data based on the target resource data.

[0111] In some embodiments, the resources described by the target resource data can be assigned to users. See also Figure 9 It can also send a notification message to the terminal device to inform the user of the resources they have received. The notification message may include the resource's name, identifier, and description. The terminal device can receive and display the notification message.

[0112] The resource allocation method described in this specification can receive resource allocation requests; predict the association between user data and resource data in the resource dataset using a model based on user data, the historical resource distribution of user data, resource data in the resource dataset, and the historical user distribution of resource data; select target resource data from the resource dataset based on the prediction results; and allocate resources to the user corresponding to the user data based on the target resource data. In this way, resources can be accurately recommended to users through the model.

[0113] Please see Figure 10 The embodiments of this specification also provide a model training device, including the following units.

[0114] The acquisition unit 31 is used to acquire a number of sample data, the sample data including user data, resource data and relationship tags, the relationship tags indicating whether user data and resource data have an association relationship;

[0115] Statistical unit 33 is used to statistically analyze a first distribution of user data and a second distribution of resource data. The first distribution represents the association between user data and multiple resource data, and the second distribution represents the association between resource data and multiple user data.

[0116] The fusion unit 35 is used to fuse the first distribution of multiple user data and the second distribution of multiple resource data;

[0117] The determining unit 37 is used to determine the model parameters of the model based on user data, resource data, the fused first distribution, the fused second distribution, and the relationship labels. The model is used to predict the association relationship between user data and resource data.

[0118] Please see Figure 11 This specification also provides a resource allocation device, which includes the following units.

[0119] Receiving unit 41 is used to receive resource allocation requests;

[0120] Prediction unit 43 is used to predict the relationship between user data and resource data in the resource dataset based on user data, historical resource distribution of user data, resource data in the resource dataset, and historical user distribution of resource data.

[0121] Selection unit 45 is used to select target resource data from the resource dataset based on the prediction results;

[0122] The allocation unit 47 is used to allocate resources to the user corresponding to the user data based on the target resource data.

[0123] The following describes an embodiment of the computer device described in this manual. Figure 12 This is a schematic diagram of the hardware structure of the computer device in this embodiment. For example... Figure 12 As shown, the computer device may include one or more (only one is shown in the figure) processors, memory, and transmission modules. Of course, those skilled in the art will understand that... Figure 12 The hardware structure shown is for illustrative purposes only and does not limit the hardware structure of the computer device described above. In practice, the computer device may also include more... Figure 12 Showing more or fewer component units; or, having the same as Figure 12 The different configurations shown.

[0124] The memory may include high-speed random access memory; or it may include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. Of course, the memory may also include remotely accessible network memory. The memory can be used to store program instructions or modules of application software, such as those described in this specification. Figure 5 or Figure 7 The program instructions or modules corresponding to the embodiments.

[0125] The processor can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. The processor can read and execute program instructions or modules in the memory.

[0126] The transmission module can be used to transmit data via a network, such as the Internet, corporate intranet, local area network, or mobile communication network.

[0127] This specification also provides an embodiment of a computer storage medium. The computer storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), cache, hard disk drive (HDD), memory card, etc. The computer storage medium stores computer program instructions. When the computer program instructions are executed, they implement: this specification. Figure 5 or Figure 7 The program instructions or modules corresponding to the embodiments.

[0128] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0129] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. A computer can be a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0130] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification 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.

[0131] Those skilled in the art will understand that the descriptions of the various embodiments have different focuses, and parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. Furthermore, it is understood that those skilled in the art, after reading this specification, can conceive of any combination of some or all of the embodiments listed in this specification without creative effort, and such combinations are also within the scope of disclosure and protection of this specification.

[0132] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations and modifications are possible with respect to this specification, and it is intended that the appended claims cover such variations and modifications without departing from the spirit of this specification.

Claims

1. A model training method, comprising: Acquire a number of sample data, including user data, resource data, and relationship tags. The relationship tags indicate whether user data and resource data are related. The user data describes users and includes at least one of basic data, wealth data, and transaction data. The basic data includes at least one of age, gender, occupation, and activity participation intention. The wealth data includes at least one of income, purchasing power, the quantity of resources received, and the amount of resources received. The transaction data includes at least one of transfer amount and number of transactions. The resource data describes resources and includes at least one of the amount of resources, resource discounts, and resource identifiers. The relationship indicates that the user described by the user data has performed a specific interactive behavior on the resource described by the resource data. The specific interactive behavior includes at least one of click behavior, browsing behavior, transaction behavior, verification behavior, and adding to cart behavior. The system calculates a first distribution of user data and a second distribution of resource data. The first distribution represents the association between user data and multiple resource data, and the second distribution represents the association between resource data and multiple user data. The first distribution of multiple user data is merged, and the second distribution of multiple resource data is merged. Based on user data, resource data, the fused first distribution, the fused second distribution, and relationship labels, the model parameters are determined, and the model is used to predict the association between user data and resource data.

2. The method according to claim 1, wherein the first distribution includes the association probability between user data and multiple resource data; the step of statistically analyzing the first distribution includes: The number of associations between user data and each of the multiple resource data is counted and used as the first association count; The first distribution is determined based on the number of first associations; The second distribution includes the probability of association between resource data and multiple user data; The steps involved in statistically analyzing the second distribution include: The number of associations between the statistical resource data and each user data in the plurality of user data is used as the second association count; The second distribution is determined based on the second association frequency.

3. The method according to claim 1, wherein the plurality of resource data belongs to a plurality of resource data groups, and the first distribution includes the association probability between user data and the plurality of resource data groups; the step of statistically analyzing the first distribution includes: The number of associations between user data and each resource data group in multiple resource data groups is counted as the third association count, which includes the sum of the number of associations between user data and each resource data in the resource data group; The first distribution is determined based on the third association frequency; The multiple user data belong to multiple user data groups, and the second distribution includes the association probability between resource data and multiple user data groups; The steps involved in statistically analyzing the second distribution include: The number of associations between the resource data and each user data group in multiple user data groups is used as the fourth association count, which includes the sum of the association counts between the resource data and each user data in the user data group; The second distribution is determined based on the fourth association number.

4. The method according to claim 1, wherein the step of fusing the first distribution includes: According to the formula The first distributions of multiple user data are merged, where D represents the set formed by the first distributions of the multiple user data, and d represents the first distribution in the set. This represents the probability of d appearing in the set; The step of fusing the second distribution includes: According to the formula The second distributions of multiple resource data are merged, where K represents the set formed by the second distributions of the multiple resource data, and k represents the second distribution in the set. This represents the probability that k appears in the set.

5. The method according to claim 1, wherein the step of fusing the first distribution includes: The first distribution of user data is merged with the first distribution fused in the previous iteration. The step of fusing the second distribution includes: The second distribution of resource data is merged with the second distribution fused in the previous iteration.

6. The method according to claim 1, wherein the step of determining the model parameters includes: The model output is determined based on user data, resource data, the first fusion distribution, and the second fusion distribution; Based on the model's output and relation labels, determine the model parameters.

7. The method of claim 6, wherein the step of determining the output of the model includes: Based on user data, resource data, the fused first distribution, and the fused second distribution, user representation, resource representation, first distribution representation, and second distribution representation are determined respectively. Based on user representation, resource representation, first distribution representation, and second distribution representation, a correlation degree representation is determined, which is used to represent the degree of correlation between user data and resource data. The model output is determined based on user representation, resource representation, and correlation degree representation.

8. The method according to claim 6, wherein the step of determining the model parameters includes: Based on the model's output and relation labels, loss information is calculated using a loss function. Based on the loss information, determine the model parameters.

9. A resource allocation method, comprising: Receive resource allocation requests; Based on user data, the historical resource distribution of user data, resource data in the resource dataset, and the historical user distribution of resource data, a model predicts the association between user data and resource data in the resource dataset. The user data describes users and includes at least one of basic data, wealth data, and transaction data. The basic data includes at least one of age, gender, occupation, and activity participation intention. The wealth data includes at least one of income, purchasing power, the quantity of resources received, and the amount of resources received. The transaction data includes at least one of transfer amount and number of transactions. The resource data describes resources and includes at least one of the amount of resources, resource discounts, and resource identifiers. The association indicates that the user described by the user data has performed a specific interactive behavior on the resource described by the resource data. The specific interactive behavior includes at least one of click behavior, browsing behavior, transaction behavior, verification behavior, and adding to cart behavior. Based on the prediction results, target resource data is selected from the resource dataset; Based on the target resource data, resources are allocated to the users corresponding to the user data.

10. The method according to claim 9, wherein the model is trained according to any one of claims 1-8.

11. The method according to claim 9, wherein the resource dataset comprises multiple resource data groups; The steps involved in predicting association relationships include: Select the target resource data group from the resource dataset based on the activity level corresponding to the user data. Based on user data, the historical resource distribution of user data, resource data in the target resource data group, and the historical user distribution of resource data, the model predicts the correlation between user data and resource data in the target resource data group. The steps involved in selecting target resource data include: Based on the prediction results, target resource data is selected from the target resource data group.

12. A model training device, comprising: The acquisition unit is used to acquire a number of sample data, including user data, resource data, and relationship tags. The relationship tags indicate whether the user data and resource data are related. The user data describes the user and includes at least one of basic data, wealth data, and transaction data. The basic data includes at least one of age, gender, occupation, and activity participation intention. The wealth data includes at least one of income, purchasing power, the quantity of resources received, and the amount of resources received. The transaction data includes at least one of transfer amount and number of transactions. The resource data describes the resource and includes at least one of the resource amount, resource discount, and resource identifier. The relationship indicates that the user described by the user data has performed a specific interactive behavior on the resource described by the resource data. The specific interactive behavior includes at least one of click behavior, browsing behavior, transaction behavior, verification behavior, and adding to cart behavior. The statistical unit is used to statistically analyze the first distribution of user data and the second distribution of resource data. The first distribution represents the relationship between user data and multiple resource data, and the second distribution represents the relationship between resource data and multiple user data. The fusion unit is used to fuse the first distribution of multiple user data and the second distribution of multiple resource data. The determining unit is used to determine the model parameters of the model based on user data, resource data, the fused first distribution, the fused second distribution, and relationship labels. The model is used to predict the association relationship between user data and resource data.

13. A resource allocation device, comprising: The receiving unit is used to receive resource allocation requests; The prediction unit is used to predict the association between user data and resource data in the resource dataset based on user data, the historical resource distribution of user data, resource data in the resource dataset, and the historical user distribution of resource data. The user data describes the user and includes at least one of basic data, wealth data, and transaction data. The basic data includes at least one of age, gender, occupation, and activity participation intention. The wealth data includes at least one of income, purchasing power, the quantity of resources received, and the amount of resources received. The transaction data includes at least one of transfer amount and number of transactions. The resource data describes the resource and includes at least one of the resource amount, resource discount, and resource identifier. The association indicates that the user described by the user data has performed a specific interactive behavior on the resource described by the resource data. The specific interactive behavior includes at least one of click behavior, browsing behavior, transaction behavior, redemption behavior, and adding to cart behavior. The selection unit is used to select target resource data from the resource dataset based on the prediction results; The allocation unit is used to allocate resources to the users corresponding to the user data based on the target resource data.

14. A computer device, comprising: At least one processor; A memory storing program instructions configured to be executed by the at least one processor, the program instructions including instructions for performing the method according to any one of claims 1-11.

Images