Model training method, object recommendation method, device and apparatus
By dynamically adjusting the embedding length weights of users and objects, the problems of low model prediction accuracy and resource waste in existing technologies are solved, achieving more efficient model training and resource utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA CONSTRUCTION BANK
- Filing Date
- 2023-02-27
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, embedding of users or items based on a fixed embedding length results in low model prediction accuracy and wasted system resources, failing to effectively utilize the differences in the amount of information carried by each user or item.
By acquiring the first and second sample data, the initial model is trained to generate user feature weights and object feature weights. The model parameters are then updated to obtain the target model. The weight values for the dynamic embedding length are generated using the frequency of occurrence of users and objects.
It improves the accuracy of model predictions, makes better use of system resources, and optimizes the efficiency of embedding processing.
Smart Images
Figure CN116167458B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of computer technology, artificial intelligence and big data, and in particular to a model training method, an object recommendation method, an apparatus and device. Background Technology
[0002] In the field of intelligent recommendation, service providers can use deep learning models to predict a user's level of liking for a particular item, and then choose whether to recommend that item to the user based on the prediction results. For example, a deep learning model can be used to predict a user's click behavior on an item. If the prediction results show that the user will click on the item, it indicates that the user is interested in the item, and the service provider can then recommend the item to the user.
[0003] Deep learning models typically include an input layer, a hidden layer, and an output layer. In the input layer, the user or item needs to be embedded to obtain user input features and item input features. Then, the subsequent model calculations are performed based on the processed user input features and item input features.
[0004] In related technologies, users or items are typically embedded using a fixed embedding length. However, because the frequency of each user or item's occurrence varies, the amount of information carried by each user or item also varies, resulting in different embedding lengths for each user or item. For example, the more frequently a user appears, the longer their embedding length; conversely, the less frequently a user appears, the shorter their embedding length. Therefore, using a fixed embedding length for all users or items not only easily leads to low model prediction accuracy but also wastes system resources. Summary of the Invention
[0005] This application provides a model training method, an object recommendation method, an apparatus, and a device to improve the accuracy of model predictions and to make reasonable use of system resources.
[0006] In a first aspect, embodiments of this application provide a model training method, including:
[0007] Acquire first sample data and second sample data, wherein the first sample data and the second sample data respectively include information of sample users, information of sample objects, and sample interaction behavior of the sample users on the sample objects;
[0008] The initial model is trained based on the first sample data to obtain the user feature weights and object feature weights of the initial model. The user feature weights include target weight values for each embedding length corresponding to multiple users, and the object feature weights include target weight values for each embedding length corresponding to multiple objects.
[0009] The model parameters of the initial model are updated based on the second sample data, the user feature weights, and the object feature weights to obtain the target model corresponding to the initial model.
[0010] In one possible implementation, the initial model is trained based on the first sample data to obtain the user feature weights and object feature weights of the initial model, including:
[0011] Obtain the initial user weights and initial object weights of the initial model. The initial user weights include the initial weight values of each embedding length corresponding to the plurality of users, and the initial object weights include the initial weight values of each embedding length corresponding to the plurality of objects.
[0012] The initial model is trained based on the first sample data, the initial user weights, and the initial object weights to obtain the user feature weights and the object feature weights.
[0013] In one possible implementation, obtaining the initial user weights and initial object weights of the initial model includes:
[0014] Obtain the first occurrence count of the multiple users in the recommendation system and the maximum occurrence count of the user in the recommendation system, and generate the initial user weight based on the first occurrence count, the maximum occurrence count of the user, and a first length threshold;
[0015] Obtain the second occurrence count of the multiple objects in the recommendation system and the maximum occurrence count of the object in the recommendation system. Generate the initial object weights based on the second occurrence count, the maximum occurrence count of the object, and the second length threshold.
[0016] In one possible implementation, the initial user weight is generated based on the first occurrence count, the maximum user occurrence count, and the first length threshold, including:
[0017] For any one of the multiple users, the initial embedding length of the user is determined based on the first occurrence count, the maximum occurrence count of the user, and the first length threshold.
[0018] The initial weight value of the initial embedding length corresponding to the user is set to a first preset value, and the initial weight values of other embedding lengths corresponding to the user are set to zero, so as to obtain the initial weight value of the user in each embedding length.
[0019] In one possible implementation, the initial model is trained based on the first sample data, the initial user weights, and the initial object weights to obtain the user feature weights and the object feature weights, including:
[0020] Based on the initial user weights, determine the embedding lengths of multiple candidate users corresponding to the first sample data;
[0021] Based on the initial object weights, determine the embedding lengths of multiple candidate objects corresponding to the first sample data;
[0022] Based on the embedding lengths of the multiple candidate users and the information of the sample users in the first sample data, multiple candidate user input features are generated.
[0023] Based on the embedding lengths of the multiple candidate objects and the information of the sample objects in the first sample data, multiple candidate object input features are generated.
[0024] The initial model is trained based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data to obtain the user feature weights and the object feature weights.
[0025] In one possible implementation, the initial model is trained based on the plurality of candidate user input features, the plurality of candidate object input features, and sample interaction behaviors in the first sample data to obtain the user feature weights and the object feature weights, including:
[0026] Based on the multiple candidate user input features and the multiple candidate object input features, multiple input feature combinations are determined, wherein the input feature combination includes one candidate user input feature and one candidate object input feature;
[0027] Each input feature combination is processed by the initial model to obtain multiple first predictive interaction behaviors;
[0028] Based on the multiple first predictive interaction behaviors and sample interaction behaviors, determine the model error corresponding to each input feature combination;
[0029] Based on the model error corresponding to each input feature combination, a target input feature combination is determined from the plurality of input feature combinations, wherein the model error corresponding to the target input feature combination is the smallest.
[0030] The initial user weights are updated based on the lengths of the candidate user input features in the target input feature combination to obtain the user feature weights;
[0031] The initial object weights are updated based on the length of the candidate object input features in the target input feature combination to obtain the object feature weights.
[0032] In one possible implementation, updating the model parameters of the initial model based on the second sample data, the user feature weights, and the object feature weights to obtain the target model corresponding to the initial model includes:
[0033] Based on the user feature weights and the information of the sample users in the second sample data, the sample user input features are determined;
[0034] The input features of the sample objects are determined based on the object feature weights and the information of the sample objects in the second sample data;
[0035] The initial model is used to process the input features of the sample users and the input features of the sample objects to obtain the second predicted interaction behavior;
[0036] Based on the second predicted interaction behavior and the sample interaction behavior in the second sample data, the model parameters of the initial model are updated to obtain the target model.
[0037] Secondly, embodiments of this application provide an object recommendation method, including:
[0038] Obtain user information and object information;
[0039] Based on the user feature weights of the recommendation model, the length of the user feature corresponding to the user information is determined, and user features are generated based on the user feature length and the user information.
[0040] Based on the object feature weights of the recommendation model, the length of the object feature corresponding to the object information is determined, and object features are generated based on the object feature length and the object information.
[0041] The user features and object features are processed by the recommendation model to obtain the user's interaction behavior with the object;
[0042] The probability of recommending the object to the user is determined based on the interaction behavior.
[0043] Thirdly, embodiments of this application provide a model training apparatus, including a first acquisition module, a training module, and an update module, wherein...
[0044] The first acquisition module is used to acquire first sample data and second sample data, wherein the first sample data and the second sample data respectively include information of the sample user, information of the sample object, and sample interaction behavior of the sample user on the sample object;
[0045] The training module is used to train the initial model based on the first sample data to obtain the user feature weights and object feature weights of the initial model. The user feature weights include target weight values for each embedding length corresponding to multiple users, and the object feature weights include target weight values for each embedding length corresponding to multiple objects.
[0046] The update module is used to update the model parameters of the initial model according to the second sample data, the user feature weights, and the object feature weights, so as to obtain the target model corresponding to the initial model.
[0047] In one possible implementation, the training module is specifically used for:
[0048] Obtain the initial user weights and initial object weights of the initial model. The initial user weights include the initial weight values of each embedding length corresponding to the plurality of users, and the initial object weights include the initial weight values of each embedding length corresponding to the plurality of objects.
[0049] The initial model is trained based on the first sample data, the initial user weights, and the initial object weights to obtain the user feature weights and the object feature weights.
[0050] In one possible implementation, the training module is further configured to:
[0051] Obtain the first occurrence count of the multiple users in the recommendation system and the maximum occurrence count of the user in the recommendation system, and generate the initial user weight based on the first occurrence count, the maximum occurrence count of the user, and a first length threshold;
[0052] Obtain the second occurrence count of the multiple objects in the recommendation system and the maximum occurrence count of the object in the recommendation system. Generate the initial object weights based on the second occurrence count, the maximum occurrence count of the object, and the second length threshold.
[0053] In one possible implementation, the training module is further configured to:
[0054] For any one of the multiple users, the initial embedding length of the user is determined based on the first occurrence count, the maximum occurrence count of the user, and the first length threshold.
[0055] The initial weight value of the initial embedding length corresponding to the user is set to a first preset value, and the initial weight values of other embedding lengths corresponding to the user are set to zero, so as to obtain the initial weight value of the user in each embedding length.
[0056] In one possible implementation, the training module is further configured to:
[0057] Based on the initial user weights, determine the embedding lengths of multiple candidate users corresponding to the first sample data;
[0058] Based on the initial object weights, determine the embedding lengths of multiple candidate objects corresponding to the first sample data;
[0059] Based on the embedding lengths of the multiple candidate users and the information of the sample users in the first sample data, multiple candidate user input features are generated.
[0060] Based on the embedding lengths of the multiple candidate objects and the information of the sample objects in the first sample data, multiple candidate object input features are generated.
[0061] The initial model is trained based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data to obtain the user feature weights and the object feature weights.
[0062] In one possible implementation, the training module is further configured to:
[0063] Based on the multiple candidate user input features and the multiple candidate object input features, multiple input feature combinations are determined, wherein the input feature combination includes one candidate user input feature and one candidate object input feature;
[0064] Each input feature combination is processed by the initial model to obtain multiple first predictive interaction behaviors;
[0065] Based on the multiple first predictive interaction behaviors and sample interaction behaviors, determine the model error corresponding to each input feature combination;
[0066] Based on the model error corresponding to each input feature combination, a target input feature combination is determined from the plurality of input feature combinations, wherein the model error corresponding to the target input feature combination is the smallest.
[0067] The initial user weights are updated based on the lengths of the candidate user input features in the target input feature combination to obtain the user feature weights;
[0068] The initial object weights are updated based on the length of the candidate object input features in the target input feature combination to obtain the object feature weights.
[0069] In one possible implementation, the update module is specifically used for:
[0070] Based on the user feature weights and the information of the sample users in the second sample data, the sample user input features are determined;
[0071] The input features of the sample objects are determined based on the object feature weights and the information of the sample objects in the second sample data;
[0072] The initial model is used to process the input features of the sample users and the input features of the sample objects to obtain the second predicted interaction behavior;
[0073] Based on the second predicted interaction behavior and the sample interaction behavior in the second sample data, the model parameters of the initial model are updated to obtain the target model.
[0074] Fourthly, embodiments of this application provide an object recommendation apparatus, comprising: a second acquisition module, a first determination module, a second determination module, a processing module, and a third determination module, wherein:
[0075] The second acquisition module is used to acquire user information and object information of the object;
[0076] The first determining module is used to determine the length of the user feature corresponding to the user information based on the user feature weights of the recommendation model, and to generate user features based on the user feature length and the user information.
[0077] The second determining module is used to determine the object feature length corresponding to the object information based on the object feature weights of the recommendation model, and to generate object features based on the object feature length and the object information;
[0078] The processing module is used to process the user features and the object features through the recommendation model to obtain the user's interaction behavior with the object;
[0079] The third determining module is used to determine the probability of recommending the object to the user based on the interaction behavior.
[0080] Fifthly, embodiments of this application provide an electronic device, including: a processor and a memory;
[0081] The memory is used to store computer-executed instructions;
[0082] The processor is configured to execute computer execution instructions stored in the memory to implement the method as described in any of the first aspects.
[0083] Sixthly, embodiments of this application provide an electronic device, including: a processor and a memory;
[0084] The memory is used to store computer-executed instructions;
[0085] The processor is configured to execute computer execution instructions stored in the memory to implement the method as described in any of the second aspects.
[0086] In a seventh aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in any of the first aspects.
[0087] Eighthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, are used to implement the method described in any of the second aspects.
[0088] Ninthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method as described in any of the first aspects.
[0089] In a tenth aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the second aspects.
[0090] The model training method, object recommendation method, apparatus, and device provided in this application can acquire first sample data and second sample data, train an initial model based on the first sample data to obtain user feature weights and object feature weights of the initial model, and update the model parameters of the initial model based on the second sample data, the aforementioned user feature weights, and the aforementioned object feature weights to obtain a target model corresponding to the initial model. In the above process, by first training the initial model to obtain the corresponding user feature weights and object feature weights, and then training the initial model based on the obtained user feature weights and object feature weights to obtain the target model, not only is the accuracy of model prediction improved, but it also facilitates the rational use of system resources. Attached Figure Description
[0091] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0092] Figure 1 A schematic diagram illustrating the application scenarios provided in the embodiments of this application;
[0093] Figure 2 This is a schematic diagram of the structure of a length selector provided in an embodiment of this application;
[0094] Figure 3A schematic flowchart illustrating a model training method provided in an embodiment of this application;
[0095] Figure 4 A schematic diagram of a weight matrix proposed in an embodiment of the application;
[0096] Figure 5 A flowchart illustrating another model training method provided in an embodiment of this application;
[0097] Figure 6 A schematic diagram illustrating the process of determining user feature weights and object feature weights provided in this application embodiment;
[0098] Figure 7 A flowchart illustrating an object recommendation method provided in an embodiment of this application;
[0099] Figure 8 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application;
[0100] Figure 9 This is a schematic diagram of the structure of an object recommendation device provided in an embodiment of this application;
[0101] Figure 10 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.
[0102] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0103] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0104] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0105] Figure 1 This is a schematic diagram illustrating an application scenario provided in an embodiment of this application. Please refer to [link / reference]. Figure 1 This includes a recommendation model 101, which can be applied to the field of intelligent recommendation. The recommendation model 101 can predict the user's interaction behavior with objects, so that the service provider can recommend objects to the user based on the prediction results.
[0106] Recommendation model 101 can be a machine learning model, which includes two length selectors and one base learner. In recommendation model 101, the two length selectors have the same workflow, which will be discussed below. Figure 2 Taking the length selector 1 corresponding to user information as an example, the length selector of this application embodiment will be described.
[0107] Figure 2 This is a schematic diagram of a length selector provided in an embodiment of this application. Please refer to... Figure 2 The length selector takes user information as input and outputs user features. The specific working process of the length selector is as follows: the first layer is used to obtain user information; the second layer is used to obtain the target weight values for each embedding length corresponding to the user based on the user information; the third layer is used to select the embedding length corresponding to the largest target weight value as the user feature length; and the fourth layer is used to generate user features based on the user feature length and user information.
[0108] The base learner can be an information prediction model based on deep neural networks (DNNs). The base learner includes an input layer, a hidden layer, and an output layer. The base learner can predict the user's interaction behavior with objects by processing the user features output by length selector 1 and the object features output by length selector 2.
[0109] When training the recommendation model, the length selector and the base learner can be trained alternately based on the Differentiable Architecture Search (DARTS) algorithm until the recommendation model converges, so as to obtain more accurate model prediction results.
[0110] In related technologies, users or items are typically embedded using a fixed embedding length. However, because the frequency of each user or item's occurrence varies, the amount of information carried by each user or item also varies, resulting in different embedding lengths for each user or item. For example, the more frequently a user appears, the longer their embedding length; conversely, the less frequently a user appears, the shorter their embedding length. Therefore, using a fixed embedding length for all users or items not only easily leads to low model prediction accuracy but also wastes system resources.
[0111] In this embodiment, first sample data and second sample data can be obtained. The initial model is first trained using the first sample data to obtain the user feature weights and object feature weights corresponding to the initial model. Then, the target model is trained using the second sample data and the obtained user feature weights and object feature weights. This process not only improves the accuracy of model prediction but also facilitates the rational use of system resources.
[0112] The method described in this application will now be illustrated through specific embodiments. It should be noted that the following embodiments may exist independently or in combination with each other; identical or similar content will not be repeated in different embodiments.
[0113] Figure 3 This is a schematic flowchart illustrating a model training method provided in an embodiment of this application. Please refer to [link / reference]. Figure 3 The method may include:
[0114] S301. Obtain the first sample data and the second sample data.
[0115] The execution subject of this application embodiment can be an electronic device or a model training device installed in an electronic device. The model training device can be implemented by software or by a combination of software and hardware.
[0116] The first sample data and the second sample data respectively include information about sample users, information about sample objects, and sample user interaction behavior with sample objects.
[0117] Optionally, the first sample data and the second sample data may include multiple sample data, and the format of each sample data may be represented as (sample user information, sample object information, sample user's sample interaction behavior with the sample object).
[0118] Optionally, the sample user information may include the user's identifier, basic user attributes, and user behavior information. For example, basic user attributes may include age and city, while user behavior information may include the number of historical clicks.
[0119] Optionally, the information of the sample object may include the object's identifier, the object's basic attribute information, and the object's behavioral information.
[0120] Optionally, the sample interaction behavior can be the click behavior of the sample user on the sample object, which can have two results: click or no click. Optionally, these two results can be marked with 0 or 1. For example, 0 can be used to mark that the sample user did not click the sample object, and 1 can be used to mark that the sample user clicked the sample object.
[0121] S302. Train the initial model based on the first sample data to obtain the user feature weights and object feature weights of the initial model.
[0122] User feature weights include target weight values for each embedding length corresponding to multiple users. Object feature weights include target weight values for each embedding length corresponding to multiple objects.
[0123] Optionally, in this embodiment, a length threshold K can be preset, where K is any positive integer greater than 1, and the embedding length corresponding to each user or object can include any length between 1 and K.
[0124] Optionally, user feature weights or object feature weights can be stored in the database as a weight matrix. The row coordinates of this weight matrix can be user identifiers of multiple users or object identifiers of multiple objects in the recommendation system, and the column coordinates can be K embedding lengths between 1 and K. Below, we take the weight matrix corresponding to user feature weights as an example, combined with... Figure 4 The weight matrix will be explained.
[0125] Figure 4 This is a schematic diagram of a weight matrix proposed in an embodiment of the application. Please refer to [link / reference]. Figure 4 In the weight matrix, each user can correspond to K embedding lengths and the target weight values for these K embedding lengths.
[0126] Optionally, the user feature weights and object feature weights of the initial model can be obtained in the following way: obtain the initial user weights and initial object weights of the initial model, wherein the initial user weights include the initial weight values of each embedding length corresponding to multiple users, and the initial object weights include the initial weight values of each embedding length corresponding to multiple objects; train the initial model based on the first sample data, the initial user weights, and the initial object weights to obtain the user feature weights and object feature weights.
[0127] S303. Update the model parameters of the initial model based on the second sample data, user feature weights, and object feature weights to obtain the target model corresponding to the initial model.
[0128] Optionally, gradient descent can be used to update the model parameters of the initial model to obtain the target model corresponding to the initial model. During the training of the initial model using the first and second sample data, steps S302 to S303 can be repeated until the initial model converges and the target model is obtained.
[0129] The model training method provided in this application can acquire first sample data and second sample data. The initial model is first trained using the first sample data to obtain the user feature weights and object feature weights corresponding to the initial model. Then, the initial model is trained again using the second sample data and the obtained user feature weights and object feature weights to obtain the target model. This process not only improves the accuracy of model prediction but also facilitates the rational use of system resources.
[0130] Based on any of the above embodiments, the following, in conjunction with Figure 5 This paper describes the detailed process of a model training method.
[0131] Figure 5 This is a flowchart illustrating another model training method provided in an embodiment of this application. Please refer to... Figure 5 The method may include:
[0132] S501, Obtain the first sample data and the second sample data.
[0133] It should be noted that the execution process of S501 can be found in the execution process of S301, and will not be repeated here.
[0134] Optionally, after obtaining the first sample data, multiple candidate user input features corresponding to each sample user can be determined based on the information of the sample users in the first sample data, and multiple candidate object input features corresponding to each sample object can be determined based on the information of the sample objects in the first sample data; then, the initial model can be trained based on the first sample data, multiple candidate user input features, and multiple candidate object input features to obtain user feature weights and object feature weights.
[0135] For any sample user in the first sample data, the process of determining the input features of multiple candidate users corresponding to that sample user will be introduced below, in conjunction with S502 to S504.
[0136] S502. Obtain the first occurrence count of multiple users in the recommendation system and the maximum occurrence count of a user in the recommendation system. Generate initial user weights based on the first occurrence count, the maximum occurrence count of a user, and the first length threshold.
[0137] Optionally, the initial user weights can be generated as follows: for any one of the multiple users, determine the initial embedding length of the user based on the first occurrence count, the maximum number of user occurrences, and the first length threshold; set the initial weight value of the initial embedding length corresponding to the user to the first preset value, and set the initial weight values of the other embedding lengths corresponding to the user to zero, thereby obtaining the initial weight values of the user at each embedding length.
[0138] Optionally, the initial embedding length for each user can be determined using the following formula:
[0139]
[0140] Where i is the initial embedding length of each user, i∈[1,K]; the first occurrence count is the number of times each user appears in the recommendation system; the maximum user count is the maximum number of times a single user appears in the recommendation system; and K is a preset first length threshold.
[0141] For example, in a recommendation system with 100 users, user 5 appears most frequently (500 times). Therefore, the maximum number of times a user appears is 500. Assuming the first length threshold K is 100 and user 1 appears 100 times, the initial embedding length for user 1 is 20. In the initial user weights, the initial weight value at embedding length 20 corresponding to user 1 can be set to a first preset value (e.g., 0.01), while the initial weight values at other embedding lengths corresponding to user 1 can be set to 0.
[0142] S503. Determine the embedding length of multiple candidate users corresponding to the first sample data based on the initial user weights.
[0143] Optionally, for any user, the initial embedding length d can be determined based on the first preset value in the initial user weights; and using Auto Machine Learning (AutoML) technology, based on the preset interval parameter n, multiple candidate user embedding lengths are automatically searched from the initial embedding length d, sequentially obtained between 1 and the first length threshold K, namely d, d-1, d+1, d-2, d+2, ..., dn, d+n, where:
[0144]
[0145] S504. Generate multiple candidate user input features based on the embedding lengths of multiple candidate users and the information of sample users in the first sample data.
[0146] Optionally, K initial embedding tables can be preset based on user information from multiple users in the recommendation system and a first length threshold K. These K initial embedding tables correspond to embedding lengths 1, 2, ..., K, respectively. In practical applications, based on the sample user information and the embedding lengths of multiple candidate users searched in step S503 above, AutoML technology can be used to sequentially obtain the multiple candidate user input features (embeddings) corresponding to the sample user from the initial embedding tables corresponding to each candidate user embedding length.
[0147] Optionally, in the model training method provided in this application embodiment, AutoML can be used to automatically search for candidate embedding length and candidate user input features, and the click-through rate (CTR) can be combined with the prediction task during subsequent model prediction, which is beneficial to improving the training efficiency of the model.
[0148] For any sample object in the first sample data, the process of determining the input features of multiple candidate objects corresponding to that sample object will be introduced below, in conjunction with S505 to S507.
[0149] S505. Obtain the second occurrence count of multiple objects in the recommendation system and the maximum occurrence count of objects in the recommendation system. Generate initial object weights based on the second occurrence count, the maximum occurrence count of objects, and the second length threshold.
[0150] Optionally, the initial object weights can be generated as follows: for any one of the multiple objects, determine the initial embedding length of the object based on the second occurrence count, the maximum occurrence count of the object, and the second length threshold; set the initial weight value of the initial embedding length corresponding to the object to the first preset value, and set the initial weight values of the other embedding lengths corresponding to the object to zero, thereby obtaining the initial weight values of the object at each embedding length.
[0151] S506. Determine the embedding length of multiple candidate objects corresponding to the first sample data based on the initial object weights.
[0152] S507. Generate multiple candidate object input features based on the embedding length of multiple candidate objects and the information of sample objects in the first sample data.
[0153] It should be noted that the execution process of S505 to S507 can be found in the execution process of S502 to S504, and will not be repeated here.
[0154] S508. Based on the input features of multiple candidate users, the input features of multiple candidate objects, and the sample interaction behavior in the first sample data, train the initial model to obtain the user feature weights and object feature weights.
[0155] Below, in conjunction with Figure 6 The process of determining user feature weights and object feature weights is explained in detail.
[0156] Figure 6 This is a schematic diagram illustrating a process for determining user feature weights and object feature weights, provided as an embodiment of this application. Please refer to... Figure 6 The process includes:
[0157] S601. Determine multiple combinations of input features based on multiple candidate user input features and multiple candidate object input features.
[0158] The input feature combination includes a candidate user input feature and a candidate object input feature.
[0159] Optionally, assuming there are M candidate user input features and N candidate object input features, there are M×N possible combinations of input features, where M and N are both positive integers greater than 1.
[0160] For example, assuming there are 4 candidate user input features and 4 candidate user input features, then there will be 16 possible combinations of input features, as shown in Table 1:
[0161] Table 1
[0162]
[0163]
[0164] S602. Each input feature combination is processed using the initial model to obtain multiple first predictive interaction behaviors.
[0165] S603. Based on multiple first prediction interaction behaviors and sample interaction behaviors, determine the model error corresponding to each input feature combination.
[0166] Optionally, the model error corresponding to each input feature combination can be determined based on the difference between the first predicted interaction behavior and the sample interaction behavior.
[0167] S604. Based on the model error corresponding to each input feature combination, determine the target input feature combination from multiple input feature combinations.
[0168] The model with the smallest error corresponds to the combination of target input features.
[0169] S605. Update the initial user weights according to the length of the candidate user input features in the target input feature combination to obtain the user feature weights.
[0170] Optionally, in the initial user weights, the initial weight value of the length of the candidate user input feature can be updated according to a second preset value to obtain the target weight value. For example, if the second preset value is 0.1, and the initial weight value at the length of the candidate user input feature is 0, the updated target weight value corresponding to the length of the candidate user input feature is 0 + 0.1 = 0.1. When updating the initial user weights, the updated target weight value can also be normalized so that the target weight value in the user feature weights is between 0 and 1.
[0171] S606. Update the initial object weights according to the length of the candidate object input features in the target input feature combination to obtain the object feature weights.
[0172] The process of determining object feature weights is similar to that of user feature weights, and will not be elaborated here.
[0173] After obtaining the user feature weights and object feature weights, you can also obtain the sample user input features and sample object input features by executing S509 to S510.
[0174] S509. Determine the input features of the sample users based on the user feature weights and the information of the sample users in the second sample data.
[0175] Optionally, the sample user feature length corresponding to the sample user can be determined based on the user feature weight, and the target weight value corresponding to the sample user feature length is the largest; then, the sample user input features can be determined based on the sample user feature length and the sample user information.
[0176] S510. Determine the input features of the sample objects based on the object feature weights and the information of the sample objects in the second sample data.
[0177] The process of determining the input features of the sample object is similar to the process of determining the input features of the sample user, and will not be described in detail here.
[0178] S511. The second predicted interaction behavior is obtained by processing the input features of sample users and sample objects through the initial model.
[0179] S512. Based on the second predicted interaction behavior and the sample interaction behavior in the second sample data, update the model parameters of the initial model to obtain the target model.
[0180] Optionally, the loss value of the initial model can be calculated based on the second predicted interaction behavior and the sample interaction behavior in the first sample data, and then the model parameters of the initial model can be updated based on the loss value to obtain the target model.
[0181] The model training method provided in this application can acquire first sample data and second sample data. The initial model is first trained using the first sample data to obtain the user feature weights and object feature weights corresponding to the initial model. Then, the initial model is trained again using the second sample data and the obtained user feature weights and object feature weights to obtain the target model. This process not only improves the accuracy of model prediction but also facilitates the rational use of system resources.
[0182] Figure 7 This is a flowchart illustrating an object recommendation method provided in an embodiment of this application. Please refer to [link / reference]. Figure 7 The method may include:
[0183] S701. Obtain user information and object information.
[0184] The execution subject of this application embodiment can be an electronic device or an object recommendation device installed in an electronic device. The object recommendation device can be implemented by software or by a combination of software and hardware.
[0185] Optionally, user information can be user identifiers, and object information can be object identifiers.
[0186] S702. Based on the user feature weights of the recommendation model, determine the length of the user features corresponding to the user information, and generate user features based on the user feature length and user information.
[0187] Optionally, for any user identifier, the embedding length of the target weight value corresponding to that user identifier can be obtained from the user feature weight. This embedding length is the user feature length.
[0188] Optionally, the recommendation model includes a pre-defined embedding table corresponding to each embedding length. This embedding table can be used to generate user features based on the user feature length and the user identifier. After obtaining the user feature length, the embedding table corresponding to that user feature length can be queried based on the user identifier to generate the user feature.
[0189] S703. Based on the object feature weights of the recommendation model, determine the object feature length corresponding to the object information, and generate object features based on the object feature length and object information.
[0190] It should be noted that the execution process of S703 can be found in the execution process of S702, and will not be repeated here.
[0191] S704. By processing user features and object features through a recommendation model, the user's interaction behavior with the object is obtained.
[0192] Optionally, user interaction with an object can include two outcomes: the first outcome is interaction between the user and the object; the second outcome is no interaction between the user and the object.
[0193] S705. Determine the probability of recommending objects to users based on interaction behavior.
[0194] Optionally, if the interaction behavior shows the first result mentioned above, then an object can be recommended to the user; if the interaction behavior shows the second result mentioned above, then the probability of recommending an object to the user is relatively low.
[0195] The object recommendation method provided in this application can, after obtaining user information, determine the length of user features corresponding to the user information based on the user feature weights of the recommendation model, and generate user features based on the user feature length and user information; after obtaining object information, determine the length of object features corresponding to the object information based on the object feature weights of the recommendation model, and generate object features based on the object feature length and object information; then, process the user features and object features through the recommendation model to obtain the user's interaction behavior with the object, and then determine the probability of recommending the object to the user based on the interaction behavior. In the above process, the recommendation model can learn the length of user features corresponding to the user information and the length of object features corresponding to the user information, making the recommendation model more accurate in generating user features and object features and predicting user interaction behavior with objects, which is beneficial to improving the accuracy of object recommendation.
[0196] Figure 8 This is a schematic diagram of a model training device provided in an embodiment of this application. Please refer to... Figure 8 The model training device 10 includes: a first acquisition module 11, a training module 12, and an update module 13, wherein,
[0197] The first acquisition module 11 is used to acquire first sample data and second sample data, wherein the first sample data and the second sample data respectively include information of the sample user, information of the sample object, and sample interaction behavior of the sample user on the sample object;
[0198] The training module 12 is used to train the initial model based on the first sample data to obtain the user feature weights and object feature weights of the initial model. The user feature weights include target weight values for each embedding length corresponding to multiple users, and the object feature weights include target weight values for each embedding length corresponding to multiple objects.
[0199] The update module 13 is used to update the model parameters of the initial model according to the second sample data, the user feature weights and the object feature weights, so as to obtain the target model corresponding to the initial model.
[0200] The model training device 10 provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.
[0201] In one possible implementation, the training module 12 is specifically used for:
[0202] Obtain the initial user weights and initial object weights of the initial model. The initial user weights include the initial weight values of each embedding length corresponding to the plurality of users, and the initial object weights include the initial weight values of each embedding length corresponding to the plurality of objects.
[0203] The initial model is trained based on the first sample data, the initial user weights, and the initial object weights to obtain the user feature weights and the object feature weights.
[0204] In one possible implementation, the training module 12 is further configured to:
[0205] Obtain the first occurrence count of the multiple users in the recommendation system and the maximum occurrence count of the user in the recommendation system, and generate the initial user weight based on the first occurrence count, the maximum occurrence count of the user, and a first length threshold;
[0206] Obtain the second occurrence count of the multiple objects in the recommendation system and the maximum occurrence count of the object in the recommendation system. Generate the initial object weights based on the second occurrence count, the maximum occurrence count of the object, and the second length threshold.
[0207] In one possible implementation, the training module 12 is further configured to:
[0208] For any one of the multiple users, the initial embedding length of the user is determined based on the first occurrence count, the maximum occurrence count of the user, and the first length threshold.
[0209] The initial weight value of the initial embedding length corresponding to the user is set to a first preset value, and the initial weight values of other embedding lengths corresponding to the user are set to zero, so as to obtain the initial weight value of the user in each embedding length.
[0210] In one possible implementation, the training module 12 is further configured to:
[0211] Based on the initial user weights, determine the embedding lengths of multiple candidate users corresponding to the first sample data;
[0212] Based on the initial object weights, determine the embedding lengths of multiple candidate objects corresponding to the first sample data;
[0213] Based on the embedding lengths of the multiple candidate users and the information of the sample users in the first sample data, multiple candidate user input features are generated.
[0214] Based on the embedding lengths of the multiple candidate objects and the information of the sample objects in the first sample data, multiple candidate object input features are generated.
[0215] The initial model is trained based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data to obtain the user feature weights and the object feature weights.
[0216] In one possible implementation, the training module 12 is further configured to:
[0217] Based on the multiple candidate user input features and the multiple candidate object input features, multiple input feature combinations are determined, wherein the input feature combination includes one candidate user input feature and one candidate object input feature;
[0218] Each input feature combination is processed by the initial model to obtain multiple first predictive interaction behaviors;
[0219] Based on the multiple first predictive interaction behaviors and sample interaction behaviors, determine the model error corresponding to each input feature combination;
[0220] Based on the model error corresponding to each input feature combination, a target input feature combination is determined from the plurality of input feature combinations, wherein the model error corresponding to the target input feature combination is the smallest.
[0221] The initial user weights are updated based on the lengths of the candidate user input features in the target input feature combination to obtain the user feature weights;
[0222] The initial object weights are updated based on the length of the candidate object input features in the target input feature combination to obtain the object feature weights.
[0223] In one possible implementation, the update module 13 is specifically used for:
[0224] Based on the user feature weights and the information of the sample users in the second sample data, the sample user input features are determined;
[0225] The input features of the sample objects are determined based on the object feature weights and the information of the sample objects in the second sample data;
[0226] The initial model is used to process the input features of the sample users and the input features of the sample objects to obtain the second predicted interaction behavior;
[0227] Based on the second predicted interaction behavior and the sample interaction behavior in the second sample data, the model parameters of the initial model are updated to obtain the target model.
[0228] Figure 9 This is a schematic diagram of an object recommendation device provided in an embodiment of this application. Please refer to... Figure 9 The object recommendation device 20 includes: a second acquisition module 21, a first determination module 22, a second determination module 23, a processing module 24, and a third determination module 25, wherein...
[0229] The second acquisition module 21 is used to acquire user information and object information of the object;
[0230] The first determining module 22 is used to determine the length of the user feature corresponding to the user information based on the user feature weights of the recommendation model, and to generate user features based on the length of the user feature and the user information.
[0231] The second determining module 23 is used to determine the object feature length corresponding to the object information based on the object feature weights of the recommendation model, and generate object features based on the object feature length and the object information;
[0232] The processing module 24 is used to process the user features and the object features through the recommendation model to obtain the user's interaction behavior with the object;
[0233] The third determining module 25 is used to determine the probability of recommending the object to the user based on the interaction behavior.
[0234] The object recommendation device 20 provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.
[0235] Figure 10 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application. Please refer to... Figure 10 The electronic device 30 may include a processor 31 and a memory 32. The processor 31 and the memory 32 can communicate; for example, the processor 31 and the memory 32 communicate via a communication bus 33.
[0236] The memory 32 is used to store computer execution instructions;
[0237] The processor 31 is used to execute computer execution instructions stored in the memory 32, causing the processor 31 to execute the model training method or object recommendation method as shown in the above method embodiments.
[0238] Optionally, the electronic device 30 may also include a communication interface, which may include a transmitter and / or a receiver.
[0239] Optionally, the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0240] This application provides a computer-readable storage medium storing computer-executable instructions; the computer-executable instructions are used to implement the model training method as described in any of the above embodiments.
[0241] This application provides a computer-readable storage medium storing computer-executable instructions; the computer-executable instructions are used to implement the object recommendation method as described in any of the above embodiments.
[0242] This application provides a computer program product, which includes a computer program that, when executed, causes a computer to perform the above-described model training method.
[0243] This application provides a computer program product, which includes a computer program that, when executed, causes a computer to perform the above-described object recommendation method.
[0244] All or part of the steps in the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable memory. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof.
[0245] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processing unit of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0246] It is understood that the various numerical designations used in the embodiments of this application are merely for descriptive convenience and are not intended to limit the scope of the embodiments of this application.
[0247] In this application, the term "comprising" and its variations can refer to non-limiting inclusion; the term "or" and its variations can refer to "and / or". The terms "first", "second", etc., in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.
[0248] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.
Claims
1. A model training method, characterized in that, include: Acquire first sample data and second sample data, wherein the first sample data and the second sample data respectively include information of sample users, information of sample objects, and sample interaction behavior of the sample users on the sample objects; The initial model is trained based on the first sample data to obtain the user feature weights and object feature weights of the initial model. The user feature weights include target weight values for each embedding length corresponding to multiple users, and the object feature weights include target weight values for each embedding length corresponding to multiple objects. The model parameters of the initial model are updated based on the second sample data, the user feature weights, and the object feature weights to obtain the target model corresponding to the initial model; The initial model is trained based on the first sample data to obtain the user feature weights and object feature weights of the initial model, including: Obtain the first occurrence count of the multiple users in the recommendation system and the maximum occurrence count of the user in the recommendation system. For any one of the multiple users, determine the initial embedding length of the user based on the first occurrence count, the maximum occurrence count of the user, and a first length threshold. The initial weight value of the initial embedding length corresponding to the user is set to a first preset value, and the initial weight values of other embedding lengths corresponding to the user are set to zero, to obtain the initial weight value of the user at each embedding length; the initial user weight includes the initial weight value of each embedding length corresponding to the multiple users; Obtain the second occurrence count of the multiple objects in the recommendation system and the maximum occurrence count of the object in the recommendation system. Based on the second occurrence count, the maximum occurrence count of the object, and the second length threshold, generate initial object weights, which include initial weight values for each embedding length corresponding to the multiple objects. Based on the initial user weights, determine the embedding lengths of multiple candidate users corresponding to the first sample data; based on the initial object weights, determine the embedding lengths of multiple candidate objects corresponding to the first sample data; based on the embedding lengths of multiple candidate users and the information of sample users in the first sample data, generate multiple candidate user input features; based on the embedding lengths of multiple candidate objects and the information of sample objects in the first sample data, generate multiple candidate object input features; based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data, train the initial model to obtain the user feature weights and the object feature weights.
2. The method according to claim 1, characterized in that, Based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data, the initial model is trained to obtain the user feature weights and the object feature weights, including: Based on the multiple candidate user input features and the multiple candidate object input features, multiple input feature combinations are determined, wherein the input feature combination includes one candidate user input feature and one candidate object input feature; Each input feature combination is processed by the initial model to obtain multiple first predictive interaction behaviors; Based on the multiple first predictive interaction behaviors and sample interaction behaviors, determine the model error corresponding to each input feature combination; Based on the model error corresponding to each input feature combination, a target input feature combination is determined from the plurality of input feature combinations, wherein the model error corresponding to the target input feature combination is the smallest. The initial user weights are updated based on the lengths of the candidate user input features in the target input feature combination to obtain the user feature weights; The initial object weights are updated based on the length of the candidate object input features in the target input feature combination to obtain the object feature weights.
3. The method according to any one of claims 1-2, characterized in that, The model parameters of the initial model are updated based on the second sample data, the user feature weights, and the object feature weights to obtain the target model corresponding to the initial model, including: Based on the user feature weights and the information of the sample users in the second sample data, the sample user input features are determined; The input features of the sample objects are determined based on the object feature weights and the information of the sample objects in the second sample data; The initial model is used to process the input features of the sample users and the input features of the sample objects to obtain the second predicted interaction behavior; Based on the second predicted interaction behavior and the sample interaction behavior in the second sample data, the model parameters of the initial model are updated to obtain the target model.
4. An object recommendation method, characterized in that, The method employs a recommendation model, which is the target model obtained by the method described in any one of claims 1-3, comprising: Obtain user information and object information; Based on the user feature weights of the recommendation model, the length of the user feature corresponding to the user information is determined, and user features are generated based on the user feature length and the user information. Based on the object feature weights of the recommendation model, the length of the object feature corresponding to the object information is determined, and object features are generated based on the object feature length and the object information. The user features and object features are processed by the recommendation model to obtain the user's interaction behavior with the object; The probability of recommending the object to the user is determined based on the interaction behavior.
5. A model training device, characterized in that, It includes a first acquisition module, a training module, and an update module, among which, The first acquisition module is used to acquire first sample data and second sample data, wherein the first sample data and the second sample data respectively include information of the sample user, information of the sample object, and sample interaction behavior of the sample user on the sample object; The training module is used to train the initial model based on the first sample data to obtain the user feature weights and object feature weights of the initial model. The user feature weights include target weight values for each embedding length corresponding to multiple users, and the object feature weights include target weight values for each embedding length corresponding to multiple objects. The update module is used to update the model parameters of the initial model according to the second sample data, the user feature weights and the object feature weights, so as to obtain the target model corresponding to the initial model; The training module is specifically used to: obtain the first occurrence count of the multiple users in the recommendation system and the maximum occurrence count of the user in the recommendation system; for any one of the multiple users, determine the initial embedding length of the user based on the first occurrence count, the maximum occurrence count of the user, and a first length threshold; set the initial weight value of the initial embedding length corresponding to the user to a first preset value, and set the initial weight values of the other embedding lengths corresponding to the user to zero, to obtain the initial weight value of the user at each embedding length; the initial user weight includes the initial weight value of each embedding length corresponding to the multiple users; obtain the second occurrence count of the multiple objects in the recommendation system and the maximum occurrence count of the object in the recommendation system, and determine the initial embedding length of the user based on the second occurrence count, the maximum occurrence count of the object, and a second length threshold. The initial object weights are generated by taking the initial user weights and determining the embedding lengths of multiple candidate users corresponding to the first sample data. Based on the initial user weights, the embedding lengths of multiple candidate objects corresponding to the first sample data are determined. Based on the initial object weights, the embedding lengths of multiple candidate objects corresponding to the first sample data are determined. Based on the embedding lengths of multiple candidate users and the information of sample users in the first sample data, multiple candidate user input features are generated. Based on the embedding lengths of multiple candidate objects and the information of sample objects in the first sample data, multiple candidate object input features are generated. Based on the multiple candidate user input features, the multiple candidate object input features, and the sample interaction behavior in the first sample data, the initial model is trained to obtain the user feature weights and the object feature weights.
6. An object recommendation device, characterized in that, The device employs a recommended model, which is the target model obtained by the method of any one of claims 1-3, and includes: a second acquisition module, a first determination module, a second determination module, a processing module, and a third determination module, wherein: The second acquisition module is used to acquire user information and object information of the object; The first determining module is used to determine the length of the user feature corresponding to the user information based on the user feature weights of the recommendation model, and to generate user features based on the user feature length and the user information. The second determining module is used to determine the object feature length corresponding to the object information based on the object feature weights of the recommendation model, and to generate object features based on the object feature length and the object information; The processing module is used to process the user features and the object features through the recommendation model to obtain the user's interaction behavior with the object; The third determining module is used to determine the probability of recommending the object to the user based on the interaction behavior.
7. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer-executed instructions; The processor is configured to execute computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 3 or claim 4.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when executed by a processor, are used to implement the method described in any one of claims 1 to 3 or claim 4.
9. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 3 or claim 4.