Information processing method and device, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention discloses an information processing method and device, electronic equipment, a storage medium and a program product. The information processing method can comprise the steps of obtaining multiple pieces of feature information of attribute information of a to-be-processed object; determining the feature information amount of the feature information, and performing grouping processing on the plurality of pieces of feature information according to the feature information amount to obtain a plurality of feature groups; multiple times of accumulation and combination are carried out on the multiple feature groups, multiple hierarchical structures arranged according to hierarchies are obtained, and in every two adjacent hierarchical structures of the multiple hierarchical structures, the next hierarchical structure comprises the feature group in the previous hierarchical structure; and carrying out embedded representation learning based on the hierarchical structure to obtain embedded representation of the attribute information of the to-be-processed object, and carrying out content recommendation on the to-be-processed object based on the embedded representation. According to the scheme, the learning effect of embedded representation learning can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of machine learning, specifically to an information processing method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] Embedding is a crucial concept in machine learning, primarily used to transform high-dimensional discrete variables into low-dimensional continuous vector representations. This transformation is essential for improving model performance because it helps models better understand and utilize the structured information in the data, thus leading to its widespread application in recommender systems. Embedding representation learning is a key step in achieving this goal.

[0003] However, when learning embedding representations of features in related technologies, a fixed embedding dimension is often used. For features with less information and lower complexity, high-dimensional embedding does not bring significant performance improvement. Instead, it increases the storage pressure and computational cost during embedding learning, resulting in poor embedding representation learning performance. Summary of the Invention

[0004] This application provides an information processing method, apparatus, electronic device, storage medium, and program product that can effectively improve the learning effect of embedded representation learning.

[0005] This application provides an information processing method, including:

[0006] Obtain multiple feature information of the attribute information of the object to be processed;

[0007] The feature information quantity is determined, and the multiple feature information is grouped according to the feature information quantity to obtain multiple feature groups;

[0008] The multiple feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures, the lower hierarchical structure includes the feature groups in the upper hierarchical structure.

[0009] Embedding representation learning is performed based on the hierarchical structure to obtain the embedded representation of the attribute information of the object to be processed, so as to recommend content to the object to be processed based on the embedded representation.

[0010] This application also provides an information processing apparatus, including:

[0011] The acquisition unit is used to acquire multiple feature information of the attribute information of the object to be processed.

[0012] A grouping unit is used to determine the feature information quantity of the feature information, and to group the multiple feature information according to the feature information quantity to obtain multiple feature groups;

[0013] The merging unit is used to perform multiple cumulative merging of the multiple feature groups to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures of the multiple hierarchical structures, the lower hierarchical structure includes the feature groups in the upper hierarchical structure.

[0014] The learning unit is used to perform embedding representation learning based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed, so as to recommend content to the object to be processed based on the embedding representation.

[0015] In some embodiments, the merging unit includes:

[0016] A subunit is defined to determine the group information content of each of the feature groups;

[0017] The sorting subunit is used to sort the feature groups of the multiple feature groups in descending order of the group information content to obtain multiple sorted feature groups.

[0018] The merging subunit is used to perform multiple cumulative merging operations on the sorted feature groups, starting from the first feature group, to obtain the multiple hierarchical structures.

[0019] In some embodiments, merging subunits is further used for:

[0020] The first feature group among the sorted feature groups is determined as the first hierarchical structure;

[0021] The first feature group is merged with the next feature group of the first feature group in the sorted multiple feature groups to obtain the next hierarchical structure of the first hierarchical structure;

[0022] The next hierarchical structure is taken as the new first feature group, and the process of merging the first feature group with the next feature group of the first feature group in the sorted multiple feature groups is repeated until the last feature group in the sorted multiple feature groups is merged to obtain the multiple hierarchical structures.

[0023] In some embodiments, the learning unit includes:

[0024] The learning sub-unit is used to learn the embedding representation for each of the hierarchical structures respectively, so as to obtain the sub-embedding representation corresponding to each hierarchical structure;

[0025] A sub-unit is generated to generate an embedded representation of the attribute information of the object to be processed based on the sub-embedded representation.

[0026] In some embodiments, the learning subunit is specifically used for:

[0027] For each of the hierarchical structures, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure;

[0028] Embedding representation learning is performed based on the filtered hierarchical structure to obtain the sub-embedded representation corresponding to the hierarchical structure.

[0029] In some embodiments, the learning subunit is further used for:

[0030] Obtain the frequency threshold corresponding to the hierarchical structure, wherein the frequency threshold is related to the hierarchy of the hierarchical structure;

[0031] The feature information in the hierarchical structure that has a frequency less than or equal to the frequency threshold is filtered to obtain the filtered hierarchical structure.

[0032] In some embodiments, the grouping unit is specifically used for:

[0033] Obtain the number of groupings and the total number of features of the multiple feature information;

[0034] Based on the total number of features and the number of groups, the number of group features allocated to the feature group is calculated and determined;

[0035] The multiple feature information are sorted in descending order of the amount of feature information to obtain the sorted feature information.

[0036] Based on the number of group features, the sorted multiple feature information is grouped to obtain multiple feature groups, each feature group including feature information corresponding to the number of group features.

[0037] This application also provides an electronic device, including a memory storing multiple instructions; the processor loads instructions from the memory to execute steps in any of the information processing methods provided in this application.

[0038] This application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the information processing methods provided in this application.

[0039] This application also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of any information processing method provided in this application.

[0040] This application embodiment first obtains multiple feature information of the attribute information of the object to be processed; then determines the feature information quantity of the feature information, and groups the multiple feature information according to the feature information quantity to obtain multiple feature groups; then, the multiple feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. In the multiple hierarchical structures, the lower level structure includes the feature groups in the upper level structure; finally, embedding representation learning is performed based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed, so as to perform content recommendation on the object to be processed based on the embedding representation. In other words, by grouping multiple feature information by feature information quantity to obtain multiple feature groups, and then accumulating and merging multiple feature groups multiple times, multiple hierarchical structures arranged in a hierarchical manner can be obtained. Since the feature information quantity contained in the multiple hierarchical structures is different, using multiple hierarchical structures for embedding representation learning can ensure that different hierarchical structures can correspond to different embedding dimensions, realize dynamic adjustment of embedding dimensions, thereby making more effective use of computing resources and avoiding unnecessary overhead. Furthermore, since in multiple hierarchical structures, the lower level includes the feature groups of the upper level, by simply ensuring that the hierarchical structure closer to the upper level contains features with a large amount of information, these features can also appear in the lower level. This allows high-frequency features to receive greater information representation when using multiple hierarchical structures for embedding representation learning, while low-frequency features receive less information representation, thus avoiding underfitting and overfitting. This improves the learning effect of embedding learning and, consequently, increases the accuracy of content recommendation when using the learned embedding representations. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1a This is a schematic diagram of a scenario for the information processing method provided in an embodiment of this application;

[0043] Figure 1b This is a flowchart illustrating the information processing method provided in an embodiment of this application;

[0044] Figure 2aThis is a flowchart illustrating the information processing method provided in this application embodiment when applied to the server.

[0045] Figure 2b This is a schematic diagram showing the distribution of multiple hierarchical structures provided in the embodiments of this application;

[0046] Figure 2c This is a schematic diagram of the embedded representation learning process provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of the structure of an information processing device provided in an embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] This application provides an information processing method, apparatus, electronic device, storage medium, and program product.

[0051] Specifically, the information processing device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0052] In some embodiments, the information processing device may also be integrated into multiple electronic devices, such as multiple servers, with the information processing method of this application being implemented by the multiple servers.

[0053] In some embodiments, the server may also be implemented as a terminal.

[0054] For example, refer to Figure 1a The information processing device can be integrated into an electronic device, which can specifically be... Figure 1aOn the server side, the electronic device can acquire multiple feature information of the attribute information of the object to be processed; determine the feature information quantity of the feature information, and group the multiple feature information according to the feature information quantity to obtain multiple feature groups; perform multiple accumulation and merging of the multiple feature groups to obtain multiple hierarchical structures arranged in a hierarchical manner, in which the lower level structure includes the feature groups in the upper level structure in two adjacent hierarchical structures; perform embedding representation learning based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed, and perform content recommendation based on the embedding representation of the object to be processed.

[0055] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0056] The following explains the technical terms used in this embodiment:

[0057] AUC, short for Area Under Curve, is the area under the ROC curve and is an important metric for evaluating the performance of binary classification models. AUC values ​​range from 0 to 1; a higher AUC value indicates a stronger discriminative ability of the model.

[0058] A / B testing (also known as split testing or bucket testing) is a common experimental method used to evaluate performance differences between two or more versions (variants).

[0059] GMV, short for Gross Merchandise Volume, is an important business metric used to evaluate the performance differences between different versions (often referred to as variant A and variant B).

[0060] It is understood that, in the specific embodiments of this application, data such as the attribute information of the object to be processed are involved. When the following embodiments of this application are applied to specific products or technologies, permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0061] In this embodiment, a smart contract-based information processing method involving blockchain technology is provided. This method can be applied to the aforementioned blockchain system, such as... Figure 1b As shown, the specific process of this information processing method can be as follows:

[0062] 101. Obtain multiple feature information of the attribute information of the object to be processed.

[0063] In this context, the object to be processed refers to the object that requires information processing. In this embodiment, the object to be processed can refer to an entity or target that needs to receive recommended content in the recommendation system. For example, the object to be processed can include, but is not limited to, entities such as users, items, tags, and locations.

[0064] The attribute information can refer to information describing various aspects of the object to be processed. Optionally, this attribute information can include different types of attribute information such as basic attributes, social attributes, and geographical attributes. For example, taking a user as the object to be processed, the attribute information of the object to be processed can include: user unique identifier (user ID), geographical location, etc. As another example, taking an item as an example, the attribute information of the object to be processed can include: item unique identifier (item ID), category, price, brand, listing time, etc. Taking a tag as an example, the attribute information of the object to be processed can include: tag name, tag category, associated objects, etc. Optionally, this attribute information can include text information, image information, audio information, etc., without limitation.

[0065] Feature information (hereinafter also referred to as features) can refer to quantitative indicators or variables extracted from attribute information and used for model training and prediction. Feature information can be represented by feature values. For example, attribute information can be encoded using feature encoding methods (such as One-Hot encoding) to obtain corresponding feature values, and these feature values ​​are used as feature information.

[0066] In some implementations, raw data of the object to be processed can be collected. This raw data can be data related to the object obtained from various sources. For example, the sources of this raw data include, but are not limited to, databases, log files, and application programming interfaces (APIs). Then, attribute information of the object to be processed is extracted from this raw data, and this attribute information is encoded using a specified feature encoding method to obtain feature information. The raw data can be unprocessed or untransformed relevant data of the object to be processed.

[0067] 102. Determine the feature information quantity of the feature information, and group multiple feature information according to the feature information quantity to obtain multiple feature groups.

[0068] The information content of feature information refers to the amount of effective information that features can provide about the target variable or model performance.

[0069] Understandably, the higher the information content of the features, the better the model can typically utilize those features for prediction and classification. Increasing the information content of statistical features helps in selecting the most effective features from a variety of options for model training, thereby improving model performance.

[0070] In some implementations, when calculating the amount of feature information, entropy, importance, singular values, and the size of the embedding layer searched by Neural Architecture Search (NAS) can be used to represent the amount of feature information.

[0071] Entropy is used to measure the uncertainty of random variables. Features with lower entropy provide more information.

[0072] Feature importance is an indicator provided by many machine learning models (such as random forests and gradient boosting trees) to evaluate the impact of features on model performance. The higher the feature importance, the greater the information content of the feature.

[0073] Singular values ​​are one of the results of Singular Value Decomposition (SVD) of a matrix, used to represent the importance of each feature in the matrix. Larger singular values ​​indicate that the feature contains more information.

[0074] The size of the embedding layer reflects the model's ability and complexity in capturing feature information; a larger embedding layer can usually capture more feature information.

[0075] In some implementations, in step 102, grouping multiple feature information according to the amount of feature information to obtain multiple feature groups may include:

[0076] S1. Obtain the number of groups and the total number of features for multiple feature information.

[0077] The number of groups refers to the number of feature information that needs to be divided into feature groups. This number of groups can be a hyperparameter, and can be customized according to actual needs.

[0078] S2. Calculate and determine the number of group features to be allocated to the feature groups based on the total number of features and the number of groups.

[0079] Among them, the number of group features refers to the number of feature information that needs to be allocated in a single feature group.

[0080] In some implementations, the quotient of the total number of features and the number of groups can be used to determine the number of group features per feature group. For example, if the total number of features is 500 and the number of groups is 5, then the number of group features per feature group can be determined to be 100. Optionally, if the result of dividing the total number of features by the number of groups includes a remainder, then the feature information corresponding to the remainder can be added to any feature group. For example, if the total number of features is 500 and the number of groups is 3 (such as feature group 1, feature group 2, and feature group 3), then each feature group has 166 features to be assigned, leaving 2 features unassigned. These 2 features can then be assigned to any one of feature group 1, feature group 2, or feature group 3, thus obtaining the number of group features per feature group.

[0081] S3. Sort the multiple feature information in descending order of feature information quantity to obtain the sorted feature information.

[0082] For example, consider multiple feature information items including feature information 1, feature information 2, feature information 3, and feature information 4. The feature information quantity of feature information 1 is 'a', the feature information quantity of feature information 2 is 'b', the feature information quantity of feature information 3 is 'c', and the feature information quantity of feature information 4 is 'd'. If b > a > d > c, then the sorted feature information items are feature information 2, feature information 1, feature information 4, and feature information 3. Here, a, b, c, and d are positive integers.

[0083] S4. Based on the number of group features, the sorted feature information is grouped to obtain multiple feature groups. Each feature group includes feature information corresponding to the number of group features.

[0084] For example, if the sorted feature information includes 300 features, and the number of group features is 100, then the first to the 100th features can be designated as the first feature group, the 101st to the 200th features as the second feature group, and the 201st to the 300th features as the third feature group. This ensures that features with a large amount of information can be assigned to the same feature group.

[0085] In other implementations, the number of target group features corresponding to each feature group can be obtained. This number of target group features is a hyperparameter that can be customized according to actual needs. Then, based on this number of target group features, the sorted multiple feature information is grouped to obtain multiple feature groups. Each feature group includes feature information corresponding to the number of group features. Optionally, the number of target group features corresponding to each feature group can be the same or different, and this is not limited here. For example, suppose we need to set the number of target features for three feature groups (e.g., feature group 1, feature group 2, and feature group 3). The pre-set number of target group features for feature group 1 can be M, for feature group 2 it can be N, and for feature group 3 it can be K, where M, N, and K are positive integers. As one implementation, M, N, and K can satisfy the condition M = N = K. As another implementation, M, N, and K can satisfy the condition M > N > K. As yet another implementation, M, N, and K can satisfy a preset proportional relationship.

[0086] 103. Multiple feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures, the lower level structure includes the feature groups in the upper level structure.

[0087] In this embodiment, cumulative merging refers to merging multiple feature groups multiple times, with each merge based on the result of the previous merge. For example, multiple feature groups include feature group 1, feature group 2, feature group 3, and feature group 4. Merging feature group 1 and feature group 2 yields the first merge result. Then, merging the first merge result with feature group 3 yields the second merge result. Finally, merging the second merge result with feature group 4 yields the fourth merge result. This process of cumulative merging multiple feature groups is repeated.

[0088] In some implementations, in step 103, multiple feature groups are cumulatively merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. Specific implementations may include:

[0089] A1. Determine the group information content of each feature group.

[0090] The group information content of a feature group can refer to the sum of the feature information content of all the feature information contained in the feature group. For example, if feature group A includes feature information 1, feature information 2, feature information 3, and feature information 4, and the feature information content of feature information 1 is a, the feature information content of feature information 2 is b, and the feature information content of feature information 3 is c, then the group information content of feature group A is (a+b+c).

[0091] A2. Sort the feature groups of multiple feature groups in descending order of information content to obtain sorted feature groups.

[0092] In some implementations, since the multiple feature groups are allocated according to the amount of feature information in step S4, the order of the multiple feature groups obtained in step S4 can be directly used to obtain the sorted multiple feature groups. For example, the sorted multiple feature groups can be a first feature group, a second feature group, and a third feature group.

[0093] A3. Starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures.

[0094] In some implementations, in step A3, starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures. Specific implementations may include:

[0095] A31. Determine the first feature group among the sorted feature groups as the first hierarchical structure.

[0096] The hierarchical structure can be used to represent feature information at different levels. In this embodiment, the hierarchical structure can be a hierarchical structure in Multi-Level Dynamic Embedding (MLDE) to represent feature information at different granularities. Each hierarchical structure can capture abstract features at different levels, thus forming a hierarchical embedding representation. This hierarchical structure can effectively handle complex feature relationships and can be effectively applied to fields such as recommendation systems, natural language processing, and image recognition.

[0097] Using the example above, for instance, the sorted feature groups can be the first feature group, the second feature group, and the third feature group. Then, the first feature group can be determined as the first hierarchical structure (hereinafter referred to as the first hierarchical structure).

[0098] A32. Merge the first feature group with the next feature group of the first feature group in the sorted feature groups to obtain the next hierarchical structure of the first hierarchical structure.

[0099] Using the example above, the first hierarchical structure and the second feature group can be merged to obtain the next hierarchical structure of the first hierarchical structure (hereinafter referred to as the second hierarchical structure). The second hierarchical structure includes all feature information in the first feature group and all feature information in the second feature group.

[0100] A33. Take the next hierarchical structure as the new first feature group, return to execute the step of merging the first feature group with the next feature group of the first feature group in the sorted multiple feature groups to obtain the next hierarchical structure of the first hierarchical structure, until it is merged into the last feature group in the sorted multiple feature groups to obtain multiple hierarchical structures.

[0101] Following the example above, the second-level structure and the third feature group can be merged to obtain the next level structure of the second-level structure (hereinafter referred to as the third-level structure). The third-level structure includes all feature information from the first feature group, all feature information from the second feature group, and all feature information from the third feature group. Since the third feature group is the last feature group among multiple feature groups, the multiple level structures obtained in this example include the aforementioned first-level structure, second-level structure, and third-level structure. The second-level structure is located below the first-level structure, and the third-level structure is located below the second-level structure. As shown in Table 1, in any two adjacent level structures, the lower-level level structure includes all feature groups from the upper-level level structure and has one more feature group than the upper-level level structure.

[0102] Table 1

[0103] Hierarchical structure Included feature groups First-level structure First feature group Second-level structure First feature group, second feature group Third-level structure First feature group, second feature group, third feature group

[0104] In this implementation, the next hierarchical structure of the first hierarchical structure is obtained by merging the first feature group with the next feature group of the first feature group in the sorted multiple feature groups. Then, the next hierarchical structure is used as the new first feature group, and the process of merging the first feature group with the next feature group of the first feature group in the sorted multiple feature groups is repeated until the last feature group in the sorted multiple feature groups is merged, resulting in multiple hierarchical structures. This increases the frequency of feature groups with higher information content in the multiple hierarchical structures and reduces the frequency of feature groups with lower information content. This facilitates subsequent embedding representation learning using multiple hierarchical structures, allowing high-frequency features to receive greater information and low-frequency features to receive less information, thus avoiding underfitting and overfitting problems.

[0105] In other embodiments, in step A3, starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures. Specific implementations may include:

[0106] The first feature group among the sorted feature groups is determined as the first hierarchical structure.

[0107] The feature groups that are ranked after the first feature group in the sorted feature groups are taken as candidate feature groups. The first feature group and the candidate feature groups are merged to obtain the next level structure of the first level structure, and the next level structure is taken as the candidate level structure.

[0108] The candidate hierarchical structure is used as the new first feature group. The process is repeated to select the feature groups that are ranked after the first feature group from the sorted feature groups as candidate feature groups. The first feature group and the candidate feature groups are then merged to obtain the next hierarchical structure of the first hierarchical structure. The next hierarchical structure is then used as the candidate hierarchical structure. This process is repeated until the last feature group in the sorted feature groups is merged to obtain multiple hierarchical structures.

[0109] For example, if multiple feature groups include a first feature group, a second feature group, a third feature group, a fourth feature group, and a fifth feature group, and the number of feature groups to be merged each time is 2, then the first feature group can be used as the first hierarchical structure. Then, the second and third feature groups can be merged with the first hierarchical structure to obtain the second hierarchical structure. Then, the fourth and fifth feature groups can be merged with the second hierarchical structure to obtain the third hierarchical structure, thereby obtaining multiple hierarchical structures.

[0110] The number of feature groups to be merged in each merge can be the same or different, depending on the actual needs, and is not limited here. As an example, as shown in Table 2, the number of feature groups to be merged in each merge can increase with the number of merges. The following is an example of multiple feature groups including the first feature group, the second feature group, the third feature group, the fourth feature group, the fifth feature group, and the sixth feature group. The first feature group can be used as the first level structure. Then, based on the fact that the number of feature groups in the first merge can be 1, the first level structure can be merged with the second feature group to obtain the second level structure. Then, based on the fact that the number of feature groups in the second merge can be 2, the second level structure, the third feature group, and the fourth feature group can be merged to obtain the third level structure. Finally, based on the fact that the number of feature groups in the third merge can be 3, the third level structure, the fourth feature group, the fifth feature group, and the sixth feature group can be merged to obtain the fourth level structure.

[0111] Table 2

[0112]

[0113]

[0114] In this embodiment, by increasing the number of feature groups to be merged each time, the representation of low-frequency features can be further reduced, thereby improving the effect of embedding representation learning using multiple hierarchical structures.

[0115] 104. Based on hierarchical structure, embedding representation learning is performed to obtain the embedding representation of the attribute information of the object to be processed, and content recommendation is performed on the object to be processed based on the embedding representation.

[0116] Embedding representation is a technique that transforms discrete, high-dimensional sparse features into low-dimensional dense vectors. In natural language processing, recommender systems, and image recognition, embedding representation is widely used to capture semantic relationships and similarities between features. The key to embedding representation is mapping the original features to a continuous vector space, such that similar features are close together in this space. Since transforming the relevant features of the object to be processed (such as users, items, etc.) into low-dimensional dense vectors (embedded representations) can better capture the relationships between features, thereby improving the accuracy of recommendations, in this embodiment, the recommendation model can be trained based on embedding representation, and then the trained recommendation model can be used to effectively recommend content to the object to be processed.

[0117] In some implementations, a multi-level dynamic embedding method can be used to learn embedding representations based on the multiple hierarchical structures obtained above. For example, after the multiple feature information is divided into multiple hierarchical structures, an embedding layer is constructed for each hierarchical structure to perform embedding learning. Then, the weights of the embedding layers are dynamically adjusted according to the feedback during the training process of embedding learning. Finally, the outputs of the multiple embedding layers are combined to form a comprehensive feature representation.

[0118] In some implementations, the specific implementation of embedding representation learning based on the hierarchical structure in step 104 to obtain the embedding representation of the attribute information of the object to be processed may include:

[0119] B1. Perform embedding representation learning on each hierarchical structure to obtain the sub-embedded representation corresponding to each hierarchical structure.

[0120] For example, an embedding layer can be defined for each hierarchical structure. This embedding layer maps discrete feature indices to low-dimensional dense vectors. Then, the embedding layer is trained. Specifically, the parameters of the embedding layer can be optimized by training a model. During training, the model adjusts the weights of the embedding layer based on backpropagation of the loss function, so that the embedding vectors can better capture the relationships between features. After the embedding layer training is complete, each hierarchical structure's features will have a corresponding low-dimensional dense vector; this vector is the sub-embedding representation of that hierarchical structure. Here, the aforementioned model can refer to a machine learning model used to train and optimize the embedding layer parameters.

[0121] In some implementations, in step B1, embedding representation learning is performed on each hierarchical structure to obtain the sub-embedded representation corresponding to each hierarchical structure. Specific implementations may include:

[0122] B11. For each hierarchical structure, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure.

[0123] Noisy features refer to features that do not contribute to model performance or may even negatively impact it. These noisy features may be caused by data acquisition errors, outliers, or other reasons. For example, noisy features may be features with low frequency of occurrence, thus retaining more frequently occurring and effective features. For instance, if a user's features appear infrequently, it indicates that the user may be a new or inactive user, and their feature information may be insufficient. Similarly, if an item's features appear infrequently, it indicates that the item may be an unpopular item, and insufficient feature information may affect model performance. Optionally, the frequency of occurrence can be the number of times the feature information appears within a specified range. Optionally, the specified range can be a time range or a data range, without limitation.

[0124] In some implementations, in step B11, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure. Specific implementations may include:

[0125] Obtain the frequency threshold corresponding to the hierarchical structure. The frequency threshold is related to the hierarchy of the hierarchical structure.

[0126] The feature information that appears in the hierarchical structure at a frequency less than or equal to the frequency threshold is filtered to obtain the filtered hierarchical structure.

[0127] For example, multiple hierarchical structures can be arranged from top to bottom in order of increasing feature information content, and lower-level hierarchical structures can be associated with smaller frequency thresholds. For instance, the multiple hierarchical structures include hierarchical structure 1, hierarchical structure 2, and hierarchical structure 3 arranged from top to bottom. The frequency threshold associated with hierarchical structure 1 is k1, the frequency threshold associated with hierarchical structure 2 is k2, and the frequency threshold associated with hierarchical structure 3 is k3, where k1 > k2 > k3.

[0128] In this embodiment, since the lower the level of the hierarchical structure, the greater the amount of feature information it contains, and the greater the possibility of more noise feature information, setting a smaller frequency threshold can more effectively filter out the noise feature information in the current hierarchical structure.

[0129] B12. Based on the filtered hierarchical structure, perform embedding representation learning to obtain the sub-embedded representations corresponding to the hierarchical structure.

[0130] In this embodiment, by filtering the noisy feature information in each hierarchical structure, a filtered hierarchical structure is obtained. Then, based on the filtered hierarchical structure, embedding representation learning is performed to obtain the sub-embedding representation corresponding to the hierarchical structure. This reduces the recommendation model's dependence on outliers and uncommon features when training the recommendation model using sub-embedding representation, making the recommendation model more stable and reliable. It also reduces the input dimension of the recommendation model, lowers computational complexity, and speeds up training and inference.

[0131] B2. Based on the sub-embedded representation, generate an embedded representation of the attribute information of the object to be processed.

[0132] In some implementations, combining the sub-embedded representations of various hierarchical structures to generate a comprehensive embedded representation is a key step in multi-level dynamic embedding methods. By combining the various sub-embedded representations, a comprehensive and rich embedded representation of the attribute information of the object to be processed can be generated for subsequent tasks, such as content recommendation, classification, or regression. In this embodiment, this embedded representation can be used to train a recommendation model. The trained recommendation model can output recommended content that the object to be processed is interested in, or recommended content that is related to or similar to the object to be processed, based on the embedded representation of the attribute information of the input object to be processed. For example, for users, it can recommend advertisements, text content, video content, etc. that they are interested in; or for items, it can recommend related or similar products.

[0133] As can be seen, in this embodiment, multiple feature information of the attribute information of the object to be processed is first obtained; then the feature information quantity is determined, and the multiple feature information is grouped according to the feature information quantity to obtain multiple feature groups; then, the multiple feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. In the multiple hierarchical structures, the lower level structure includes the feature groups in the upper level structure; finally, embedding representation learning is performed based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed, so as to perform content recommendation on the object to be processed based on the embedding representation. In other words, by grouping multiple feature information by feature information quantity to obtain multiple feature groups, and then accumulating and merging multiple feature groups multiple times, multiple hierarchical structures arranged in a hierarchical manner can be obtained. Since the feature information quantity contained in the multiple hierarchical structures is different, using multiple hierarchical structures for embedding representation learning can ensure that different hierarchical structures can correspond to different embedding dimensions, realize dynamic adjustment of embedding dimensions, and thus make more effective use of computing resources and avoid unnecessary overhead. Furthermore, since in multiple hierarchical structures, the lower level includes the feature groups of the upper level, by simply ensuring that the hierarchical structure closer to the upper level contains features with high information content, these features will also appear in the lower level. Conversely, features with low information content, being assigned to later-ranked feature groups, can only be assigned to lower-level hierarchical structures. This increases the embedding dimension of features with high information content and decreases the embedding dimension of features with low information content. Without increasing model complexity, this enhances the model's expressive power, allowing high-frequency features to receive greater information and low-frequency features to receive less information, thus avoiding underfitting and overfitting. Consequently, when using the learned embedding representations for content recommendation, this improves recommendation accuracy.

[0134] The method described in the above embodiments will be further described in detail below.

[0135] In this embodiment, an information processing method will be applied to... Figure 1a Taking the server-side as an example, the method of this application embodiment will be described in detail.

[0136] like Figure 2a As shown, the specific process of an information processing method is as follows:

[0137] 201. The server obtains multiple feature information of the attribute information of the object to be processed.

[0138] In some implementations, the server can receive attribute information of the object to be processed sent by the terminal, thereby obtaining the attribute information of the object to be processed.

[0139] 202. The server determines the amount of feature information and groups multiple feature information according to the amount of feature information to obtain multiple feature groups.

[0140] For example, if the feature information of multiple feature information is [1, 2, 4, 7, 9], and it is divided into four groups according to the equal distance of 3, then [1-3) is one group, [3-6), [6, 9), [9, 12) is another group, and thus it can be divided into four groups: [1, 2], [4], [7], [9].

[0141] In some implementations, in step 202, grouping multiple feature information according to the amount of feature information to obtain multiple feature groups may include:

[0142] Obtain the number of groupings and the total number of features from multiple feature information;

[0143] The number of group features allocated to a feature group is calculated and determined based on the total number of features and the number of groups.

[0144] Sort the multiple feature information in descending order of the amount of feature information to obtain the sorted feature information;

[0145] Based on the number of group features, the sorted feature information is grouped to obtain multiple feature groups, each of which includes feature information corresponding to the number of group features.

[0146] For example, the server can statistically analyze the feature information content of the feature information. This feature information content can be represented using entropy, importance, singular values, or even the embedding layer size obtained from NAS search. It can be understood that different methods of representing information content have varying degrees of difficulty and stability. Therefore, an appropriate feature information content representation can be adopted according to actual needs. For instance, in advertising recommendation scenarios, feature information is relatively sparse, meaning it appears infrequently. In this case, statistical entropy can be used to obtain the feature information content, making the representation more effective. Conversely, in content recommendation scenarios, feature information is relatively dense, meaning it appears frequently. In this case, statistical frequency can be used to obtain the feature information content, making the representation more effective.

[0147] 203. The server performs multiple cumulative merging operations on multiple feature groups to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures, the lower level structure includes the feature groups in the upper level structure.

[0148] In some implementations, in step 203, multiple feature groups are cumulatively merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. Specific implementations may include:

[0149] Determine the group information content of each feature group;

[0150] Sort the feature groups of multiple feature groups in descending order of group information content to obtain sorted feature groups.

[0151] Starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures.

[0152] In some implementations, the specific implementation of the step "starting from the first feature group among the sorted feature groups, performing multiple cumulative merging operations on the sorted feature groups to obtain multiple hierarchical structures" may include:

[0153] The first feature group among the sorted feature groups is determined as the first hierarchical structure;

[0154] The first feature group is merged with the next feature group of the first feature group in the sorted feature groups to obtain the next hierarchical structure of the first hierarchical structure;

[0155] The next hierarchical structure is taken as the new first feature group. The process is repeated to merge the first feature group with the next feature group of the first feature group in the sorted multiple feature groups to obtain the next hierarchical structure of the first hierarchical structure. This process continues until the last feature group in the sorted multiple feature groups is merged to obtain multiple hierarchical structures.

[0156] 204. The server performs embedding representation learning on each hierarchical structure to obtain the sub-embedded representation corresponding to each hierarchical structure.

[0157] In some implementations, in step 204, embedding representation learning is performed on each hierarchical structure to obtain the sub-embedded representation corresponding to each hierarchical structure. Specific implementations may include:

[0158] For each hierarchical structure, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure.

[0159] Embedding representation learning is performed based on the filtered hierarchical structure to obtain the sub-embedded representations corresponding to the hierarchical structure.

[0160] In some implementations, the specific implementation of the step "filtering the noise feature information in the hierarchical structure to obtain the filtered hierarchical structure" may include:

[0161] Obtain the frequency threshold corresponding to the hierarchical structure. The frequency threshold is related to the hierarchy of the hierarchical structure.

[0162] The feature information that appears in the hierarchical structure at a frequency less than or equal to the frequency threshold is filtered to obtain the filtered hierarchical structure.

[0163] 205. The server generates an embedded representation of the attribute information of the object to be processed based on the sub-embedded representation, and then recommends content based on the embedded representation of the object to be processed.

[0164] As an example, the information processing method in steps 201 to 205 above can be applied to the actual recommendation system scenario. For example, this information processing method can be used to perform multi-level dynamic embedding of various feature information such as user attributes, item attributes, and user interests, and then use the obtained embedding representation to train the recommendation model so as to perform content recommendation through the trained recommendation model.

[0165] Optionally, the scenarios for this recommendation system may include, but are not limited to: content recommendation, advertising recommendation, social media platforms, personalized recommendation service platforms, etc.

[0166] As a more specific example, a specific implementation of the above-described multi-layered dynamic embedding may include the following steps:

[0167] Step 1: Obtain the attribute information and other data of the object to be processed, and perform statistical analysis on the feature information in the data.

[0168] For example, each feature in the data can consist of one or more feature values ​​v from the vocabulary V. If it is a single-value feature (such as user ID, item ID, etc.), there is only one feature value (such as v1). If it is a multi-value feature (such as a list of purchased items), there are multiple feature values ​​(such as v2, v3, etc.).

[0169] Step 2: Group the feature information according to the amount of feature information. For example, if there are 500 feature information items, these 500 feature information items need to be sorted in descending order of feature information amount, and then the sorted results are grouped to obtain multiple feature groups. For example, each time a specified number of feature information items are traversed, they can be divided into a feature group. Specifically, they can be divided into R groups, where R is a hyperparameter. For example, R can be set to 3 to 5 groups.

[0170] Step 3: Sort multiple feature groups and stack them based on the sorted feature groups to obtain multiple hierarchical structures.

[0171] like Figure 2bAs shown, the feature information can be divided into four feature groups: feature group 1, feature group 2, feature group 3, and feature group 4. The amount of information in feature group 1 is greater than that in feature group 2, which is greater than that in feature group 3, and which is greater than that in feature group 4. Therefore, the sorted feature groups are: feature group 1, feature group 2, feature group 3, and feature group 4. The feature information from these four feature groups can then be stacked to obtain hierarchical structures 1, 2, 3, and 4. Hierarchical structure 1 is composed of the feature information from feature group 1; hierarchical structure 2 is formed by stacking the feature information from feature group 1 and feature group 2; hierarchical structure 3 is formed by stacking the feature information from feature group 1, feature group 2, and feature group 3; and hierarchical structure 4 is formed by stacking the feature information from feature group 1, feature group 2, feature group 3, and feature group 4.

[0172] Step four: Learn sub-embedded representations by utilizing the feature information contained in each of the multiple hierarchical structures, and generate embedded representations of the attribute information of the object to be processed based on each sub-embedded representation.

[0173] Following the example above, each of the four hierarchical structures can correspond to a sub-embedding representation. This sub-embedding representation can be a set of vectors after each feature information in the corresponding hierarchical structure is converted into a vector. Through this multi-level embedding expression, different features can have different embedding dimensions. Features with a large amount of information have a high embedding dimension, while features with a small amount of information have a low embedding dimension. This allows high-frequency features to have a larger amount of information expressed, while low-frequency features have a lower expression, thus avoiding underfitting and overfitting problems.

[0174] Optionally, a corresponding feature filtering threshold can be set for each level of the hierarchical structure. This feature filtering threshold can be used to filter out noisy feature information in the corresponding hierarchical structure. For example, the feature filtering thresholds for hierarchical structures 1 to 4 can be set sequentially from large to small.

[0175] Understandably, since each sub-embedded representation has the same dimension, this stacked hierarchical structure allows features with high information content to have the largest dimension in the entire embedding space, while features with low information content have the smallest dimension. For example, if the vector dimension of the sub-embedded representation corresponding to the feature information of each feature group is d, after layering, feature group 1 can appear in 4 levels, so the vector dimension of the feature information in feature group 1 becomes 4d, while feature group 4 only appears in 1 level, so the vector dimension of the feature information in feature group 4 remains d. Furthermore, the feature information content of the feature information in feature group 4 is inherently less than that of the feature information in feature group 1. Therefore, during the embedding representation learning process, the feature information in feature group 1, which has high feature information content, will receive greater representation, while the feature information in feature group 4, which has low feature information content, will have reduced representation.

[0176] For example, in practical applications, such as Figure 2c As shown, a specific implementation of embedding representation learning using MLDE can include three steps: constructing an embedding table, embedding lookup, and pooling. A specific example is as follows:

[0177] An embedding table can be pre-built, and the attribute information of the object to be processed can be obtained. This attribute information can be a labeled dataset D = {(x1,y1),(x2,y2),...,(x...}. |D| ,y |D| )}, where x |D| For the Dth original data in the dataset, y |D| For the corresponding original data x |D| The tags are defined as follows: For each x, x contains multiple sets of original features, such as product information, user click interests, user conversion interests, and user basic attributes. Each original feature can be a single-valued feature, such as user ID, or a multi-valued feature, such as the product categories the user has clicked in the past 60 days.

[0178] Then, for each original feature v, its corresponding feature value (v1, v2, ..., v) can be assigned. t By mapping between embeddinglookup and the embedding table, each feature value is mapped to the embedding vector space, resulting in a feature value-corresponding vector (e1, e2, ..., e...). t ), where e i ∈1×d, where d is the dimension. This can be understood as the embedding table being a matrix. The feature values ​​can be converted into embedded representations using the embedding function g(x; E) and the embedding table. Since the feature information is grouped in this embodiment, multiple feature groups (hereinafter referred to as field groups) are obtained. Therefore, the embedded representation can be represented as vectors obtained by the embedding function within multiple field groups.

[0179] Next, a pooling operation is performed on the vectors obtained within each field group. Optionally, the pooling operation can include sum pooling or average pooling. The resulting vectors after pooling are fed into the subsequent neural network f(e; θ) for supervised learning using a loss function L. This process is defined as the following optimization problem:

[0180] argmin L(E,θ),X(x,y)∈D,l(f(g(x;E);θ),y)

[0181] Where L(E,θ) is the objective function of the above optimization problem; L is the loss function; E is the embedding table; θ is the model parameters, such as the weights and biases of the neural network; D is the training dataset; g(x;E) is the embedding layer, which maps the input feature x to the embedding vector; f(g(x;E);θ) is the forward propagation process of the model, which converts the embedding vector g(x;E) into the predicted value through the model f; l(f(g(x;E);θ),y) is the specific expression of the loss function.

[0182] In this embodiment, MLDE optimizes the process as follows: MLDE sorts the feature information according to its magnitude and groups the feature information. Based on the grouping results, it further hierarchically divides the data into multiple hierarchical structures. Each hierarchical structure learns a different embedding representation g for its features. k (x;E k Using this representation, the optimization problem can be redefined as follows:

[0183] argmin L(E,θ),X(x,y)∈D,l(f(g1(x;E1),g2(x;E2),g k (x;E k );θ),y)

[0184] The optimization method described above allows for the concatenation of multiple sub-embedded representations, which are then fed into a neural network for supervised learning to obtain the final embedded representation. This embedded representation can then be used as input to a recommendation model for training, enabling the model to make accurate content recommendations after training.

[0185] As an example, Table 3 shows the test results of various embedding learning methods used in the information processing method and related technologies of this embodiment to test the learning effect.

[0186] Table 3

[0187]

[0188]

[0189] As shown in Table 3, in actual testing, the information processing method provided in this embodiment reduced model complexity (e.g., compared to other learning methods, the input vector dimension of this method is reduced to (1.7 x 4 x N)) while also achieving the best AUC performance. In addition to offline metrics, improvements were also achieved in metrics such as GMV in A / B experiments. This not only improved the performance of the main objective but also allowed some sparse objectives to be learned better, with lower prediction bias. Thus, a better embedding representation learning effect was obtained.

[0190] In summary, the information processing method provided in this embodiment addresses the issue of increased storage and computational overhead in related embedding learning methods. From an interpretability perspective, it offers a grouping representation based on feature information content. This increases the embedding dimension of features with high information content and decreases the embedding dimension of features with low information content, thereby enhancing the model's expressive power without increasing model complexity and alleviating overfitting and underfitting issues on different features. Furthermore, this embodiment not only groups the features but also performs hierarchical processing based on the grouping results. This grouping and hierarchical structure enables dynamic embedding, thus optimizing the multi-level dynamic embedding method and improving the learning effect of embedding representation learning.

[0191] To better implement the above methods, this application also provides an information processing device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers.

[0192] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the information processing device specifically integrated into information processing as an example.

[0193] For example, such as Figure 3 As shown, the information processing device may include an acquisition unit 301, a grouping unit 302, a merging unit 303, and a learning unit 304, as follows:

[0194] The acquisition unit 301 is used to acquire multiple feature information of the attribute information of the object to be processed;

[0195] Grouping unit 302 is used to determine the feature information quantity of feature information, and to group multiple feature information according to the feature information quantity to obtain multiple feature groups;

[0196] The merging unit 303 is used to perform multiple cumulative merging of multiple feature groups to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures of the multiple hierarchical structures, the lower hierarchical structure includes the feature groups in the upper hierarchical structure.

[0197] Learning unit 304 is used to learn embedding representations based on hierarchical structure to obtain embedding representations of the attribute information of the object to be processed, so as to recommend content based on the embedding representations of the object to be processed.

[0198] In some embodiments, the merging unit 303 includes:

[0199] Determine sub-units to determine the group information content of each feature group;

[0200] The sorting subunit is used to sort the feature groups of multiple feature groups in descending order of group information content, resulting in sorted feature groups.

[0201] The merging sub-unit is used to start from the first feature group in the sorted feature groups and perform multiple cumulative merging operations on the sorted feature groups to obtain multiple hierarchical structures.

[0202] In some embodiments, merging subunits is further used for:

[0203] The first feature group among the sorted feature groups is determined as the first hierarchical structure;

[0204] The first feature group is merged with the next feature group of the first feature group in the sorted feature groups to obtain the next hierarchical structure of the first hierarchical structure;

[0205] The next hierarchical structure is taken as the new first feature group. The process is repeated to merge the first feature group with the next feature group of the first feature group in the sorted multiple feature groups to obtain the next hierarchical structure of the first hierarchical structure. This process continues until the last feature group in the sorted multiple feature groups is merged to obtain multiple hierarchical structures.

[0206] In some embodiments, the learning unit 304 includes:

[0207] Learning sub-units are used to learn the embedding representation for each hierarchical structure, so as to obtain the sub-embedding representation corresponding to each hierarchical structure;

[0208] The generation sub-unit is used to generate an embedded representation of the attribute information of the object to be processed based on the sub-embedded representation.

[0209] In some embodiments, the learning subunit is specifically used for:

[0210] For each hierarchical structure, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure.

[0211] Embedding representation learning is performed based on the filtered hierarchical structure to obtain the sub-embedded representations corresponding to the hierarchical structure.

[0212] In some embodiments, the learning subunit is further used for:

[0213] Obtain the frequency threshold corresponding to the hierarchical structure. The frequency threshold is related to the hierarchy of the hierarchical structure.

[0214] The feature information that appears in the hierarchical structure at a frequency less than or equal to the frequency threshold is filtered to obtain the filtered hierarchical structure.

[0215] In some embodiments, the grouping unit 302 is specifically used for:

[0216] Obtain the number of groupings and the total number of features from multiple feature information;

[0217] The number of group features allocated to a feature group is calculated and determined based on the total number of features and the number of groups.

[0218] Sort the multiple feature information in descending order of the amount of feature information to obtain the sorted feature information;

[0219] Based on the number of group features, the sorted feature information is grouped to obtain multiple feature groups, each of which includes feature information corresponding to the number of group features.

[0220] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0221] This application also provides an electronic device, which can be a terminal, a server, or other similar device. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0222] In some embodiments, the information processing device may also be integrated into multiple electronic devices, such as multiple servers, with the information processing method of this application being implemented by the multiple servers.

[0223] In this embodiment, the electronic device of this embodiment will be used as an example for detailed description, such as... Figure 4 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0224] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, an input module 404, and a communication module 405. Those skilled in the art will understand that... Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0225] The processor 401 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 402, and by calling data stored in the memory 402, it performs various functions and processes data, thereby performing overall detection of the electronic device. In some embodiments, the processor 401 may include one or more processing cores; in some embodiments, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 401.

[0226] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0227] The electronic device also includes a power supply 403 that supplies power to the various components. In some embodiments, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0228] The electronic device may also include an input module 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0229] The electronic device may also include a communication module 405. In some embodiments, the communication module 405 may include a wireless module, through which the electronic device can perform short-range wireless transmission, thereby providing users with wireless broadband internet access. For example, the communication module 405 can be used to help users send and receive emails, browse web pages, and access streaming media.

[0230] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions.

[0231] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0232] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0233] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the information processing methods provided in embodiments of this application.

[0234] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0235] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the above embodiments.

[0236] Since the instructions stored in the storage medium can execute the steps of any of the information processing methods provided in the embodiments of this application, the beneficial effects that any of the information processing methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0237] The foregoing has provided a detailed description of an information processing method, apparatus, electronic device, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An information processing method, characterized in that, include: Obtain multiple feature information of the attribute information of the object to be processed; The feature information quantity is determined, and the multiple feature information is grouped according to the feature information quantity to obtain multiple feature groups; The multiple feature groups are accumulated and merged multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures, the lower hierarchical structure includes the feature groups in the upper hierarchical structure. Embedding representation learning is performed based on the hierarchical structure to obtain the embedded representation of the attribute information of the object to be processed, so as to recommend content to the object to be processed based on the embedded representation.

2. The method according to claim 1, characterized in that, The process of accumulating and merging the multiple feature groups multiple times to obtain multiple hierarchical structures arranged in a hierarchical manner includes: Determine the group information content of each of the feature groups; The feature groups of the multiple feature groups are sorted in descending order of the amount of information in each group to obtain multiple sorted feature groups. Starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain the multiple hierarchical structures.

3. The method according to claim 2, characterized in that, Starting from the first feature group among the sorted feature groups, the sorted feature groups are accumulated and merged multiple times to obtain the multiple hierarchical structures, including: The first feature group among the sorted feature groups is determined as the first hierarchical structure; The first feature group is merged with the next feature group of the first feature group in the sorted plurality of feature groups to obtain the next hierarchical structure of the first hierarchical structure; The next hierarchical structure is taken as the new first feature group, and the process of merging the first feature group with the next feature group of the first feature group in the sorted multiple feature groups is repeated until the last feature group in the sorted multiple feature groups is merged to obtain the multiple hierarchical structures.

4. The method according to claim 1, characterized in that, The embedding representation learning based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed includes: Embedding representation learning is performed on each of the hierarchical structures to obtain the sub-embedding representation corresponding to each hierarchical structure; Based on the sub-embedded representation, an embedded representation of the attribute information of the object to be processed is generated.

5. The method according to claim 4, characterized in that, The step of learning the embedding representation for each of the hierarchical structures to obtain the sub-embedding representation corresponding to each hierarchical structure includes: For each of the hierarchical structures, the noise feature information in the hierarchical structure is filtered to obtain the filtered hierarchical structure; Embedding representation learning is performed based on the filtered hierarchical structure to obtain the sub-embedded representation corresponding to the hierarchical structure.

6. The method according to claim 5, characterized in that, The step of filtering the noise feature information in the hierarchical structure to obtain the filtered hierarchical structure includes: Obtain the frequency threshold corresponding to the hierarchical structure, wherein the frequency threshold is related to the hierarchy of the hierarchical structure; The feature information in the hierarchical structure that has a frequency less than or equal to the frequency threshold is filtered to obtain the filtered hierarchical structure.

7. The method according to any one of claims 1 to 6, characterized in that, The step of grouping the multiple feature information according to the amount of feature information to obtain multiple feature groups includes: Obtain the number of groupings and the total number of features of the multiple feature information; Based on the total number of features and the number of groups, the number of group features allocated to the feature group is calculated and determined; The multiple feature information is sorted in descending order of the amount of feature information to obtain the sorted feature information. Based on the number of group features, the sorted multiple feature information is grouped to obtain multiple feature groups, each feature group including feature information corresponding to the number of group features.

8. An information processing device, characterized in that, include: The acquisition unit is used to acquire multiple feature information of the attribute information of the object to be processed. A grouping unit is used to determine the feature information quantity of the feature information, and to group the multiple feature information according to the feature information quantity to obtain multiple feature groups; The merging unit is used to perform multiple cumulative merging of the multiple feature groups to obtain multiple hierarchical structures arranged in a hierarchical manner. In two adjacent hierarchical structures of the multiple hierarchical structures, the lower hierarchical structure includes the feature groups in the upper hierarchical structure. The learning unit is used to perform embedding representation learning based on the hierarchical structure to obtain the embedding representation of the attribute information of the object to be processed, so as to recommend content to the object to be processed based on the embedding representation.

9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to perform the steps of the information processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the information processing method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the steps of the information processing method according to any one of claims 1 to 7.

Images