Data processing method and system

Abstract

Embodiments of the present disclosure relate to data processing methods and systems. A data processing method is provided, comprising: dividing a plurality of content representations into at least one batch of content representations for at least one round of feature clustering iteration computation; performing the at least one round of feature clustering iteration computation, wherein in each round of iteration: dividing the batch of content representations for the round of iteration into a plurality of groups of content representations, the plurality of groups of content representations for being respectively distributed to a corresponding plurality of processing units; distributing a plurality of cluster centers to the plurality of processing units and distributing the plurality of groups of content representations to the corresponding plurality of processing units, each processing unit computing a similarity measure between each content representation in the group of content representations distributed thereto and each cluster center in the plurality of cluster centers; updating the plurality of cluster centers based on the computed similarity measure of each content representation. The present disclosure improves feature clustering and increases implicit image diversity.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to data processing, and more particularly to data processing methods, systems and media for implicit profiling mining. BACKGROUND

[0002] With the development of the Internet and mobile Internet, personalized services have become an important strategy for user needs and enterprise competition. Personalized recommendation, advertising, precision marketing and other applications based on user profiling have become a popular technology in the field of Internet and mobile Internet. In order to better achieve personalized services, it is necessary to analyze the user profile, that is, by analyzing the user's behavior, interest, preference and other information, the user's profile is obtained, so as to provide more personalized services for the user.

[0003] At present, the implicit profiling mining method based on content representation and feature clustering has become a popular user profiling analysis method. This method analyzes the user's behavior, interest, preference and other information to obtain the user's profile, thereby providing more personalized services for the user. SUMMARY

[0004] Embodiments of the present disclosure relate to data processing methods, systems and media for implicit profiling mining.

[0005] According to one aspect of the present disclosure, a data processing method is provided, comprising: dividing a plurality of content representations into at least one batch of content representations for at least one round of feature clustering iteration calculation; performing the at least one round of feature clustering iteration calculation, wherein in each iteration: dividing the batch of content representations for the iteration into a plurality of groups of content representations, the plurality of groups of content representations being for respectively distributing to a corresponding plurality of processing units; distributing a plurality of cluster centers to the plurality of processing units and distributing the plurality of groups of content representations to the corresponding plurality of processing units, each processing unit calculating a similarity measure between each content representation in the group of content representations distributed to it and each cluster center in the plurality of cluster centers; determining a cluster center to which each content representation belongs based on the calculated similarity measure of each content representation; and updating the plurality of cluster centers by obtaining an updated cluster center for each cluster center by performing a pooling average on the content representations belonging to the cluster center, thereby obtaining an updated plurality of cluster centers.

[0006] In some embodiments, the data processing method can further comprise obtaining the calculated similarity measure of the content representation from the plurality of processing units.

[0007] In some embodiments, each processing unit computing the similarity measure between each of the set of content representations distributed thereto and each of the plurality of cluster centers further comprises dividing the set of content representations into a plurality of subsets of content representations, and computing the similarity measure in batches for the plurality of subsets of content representations, which comprises computing, in turn for each subset of content representations, the similarity measure between each of the subset of content representations and each of the plurality of cluster centers.

[0008] In some embodiments, the data processing method further comprises, after the end of the iteration of feature clustering computation, indexing the plurality of cluster centers using a retrieval framework, obtaining an ID for each cluster center, for each of a first plurality of content representations in the plurality of content representations, retrieving, using the retrieval framework and based on the similarity measure of the content representation and the cluster centers, the IDs of a first number of cluster centers that are most similar to the content representation as a set of training data corresponding to the content representation, thereby obtaining a first plurality of sets of training data, and training a first sequential representation model using the obtained first plurality of sets of training data, obtaining a trained first sequential representation model.

[0009] In some embodiments, the data processing method further comprises, for each cluster center in the plurality of cluster centers, obtaining a set of similar cluster centers, which comprises inputting the ID of the cluster center into the trained first sequential representation model and obtaining an output cluster center ID, and storing the pair of input cluster center ID and output cluster center ID as a set of similar cluster centers.

[0010] In some embodiments, the data processing method further comprises, for each of a plurality of content representations associated with each of a plurality of users in the plurality of users, obtaining a second plurality of sets of training data by computing, using the retrieval framework and based on the similarity measure of the content representation and the cluster centers, a respective plurality of cluster centers that are most similar to the plurality of content representations, obtaining, according to the number of content representations of the user aggregated at each of the plurality of cluster centers, the IDs of a second number of plurality of cluster centers that are ranked highest as a set of training data, and training a second sequential representation model using the obtained second plurality of sets of training data, obtaining a trained second sequential representation model.

[0011] In some embodiments, the data processing method further comprises, for each of the plurality of content representations, obtaining a set of associated cluster centers, which comprises, for each of the plurality of cluster centers, obtaining a set of associated cluster centers, which comprises inputting the ID of the cluster center into the trained second sequential representation model and obtaining an output cluster center ID, and storing the pair of input cluster center ID and output cluster center ID as a set of associated cluster centers.

[0012] In some embodiments, the data processing method can further include: retrieving, by using the retrieval framework, a first number of cluster centers that are most similar to the content representation of the real-time content based on a similarity measure of the content representation of the real-time content to the cluster centers; filtering at least one similar cluster center from the first number of cluster centers based on a plurality of similar cluster center groups of the plurality of content representations, to obtain a second number of cluster centers; expanding the second number of cluster centers based on a plurality of related cluster center groups of the plurality of content representations, to obtain a third number of cluster centers; and establishing the implicit profile based on the third number of cluster centers.

[0013] According to another aspect of the present disclosure, there is provided a data processing system comprising a plurality of graphics processing units, wherein a master graphics processing unit of the plurality of graphics processing units is configured to perform operations of the method as described above.

[0014] According to another aspect of the present disclosure, there is provided a computer readable storage medium having stored thereon computer readable program instructions which, when executed by a processor, perform operations of the method as described above. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and other objects, features and advantages of the present disclosure will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings, in which like reference characters designate like elements in the several views.

[0016] Figure 1 is a block diagram of an example system architecture 100 for implicit profile mining according to an embodiment of the present disclosure.

[0017] Figure 2 is a flowchart of an example data processing method according to an embodiment of the present disclosure.

[0018] Figure 3 shows a flowchart of a method for training a first sequential representation model and obtaining similar cluster center groups according to an embodiment of the present disclosure.

[0019] Figure 4 shows a flowchart of a method for training a second sequential representation model and obtaining related cluster center groups according to an embodiment of the present disclosure.

[0020] Figure 5 shows a flowchart of a method for establishing an implicit profile according to an embodiment of the present disclosure.

[0021] Figure 6 is a schematic diagram showing a general hardware environment of a device that can implement an embodiment according to the present disclosure. DETAILED DESCRIPTION

[0022] The following description is presented to enable any person skilled in the art to practice the embodiments and is provided in the context of particular applications and its requirements. Various modifications to the embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of the embodiments. Thus, the embodiments are not limited to the described embodiments but rather the intent is to convey the widest scope of the principles and features disclosed herein.

[0023] The present disclosure relates to data processing methods, systems and media for improved implicit profile mining. In one aspect, the present disclosure utilizes a plurality of processing units (e.g., a plurality of graphic processing units) to perform feature aggregation iterative computation, and the total content representations to be processed are batched according to the capability of the plurality of processing units. Each batch of content representations is assigned to the plurality of processing units for processing. A master processing unit is responsible for generating and updating cluster centers, and distributing the cluster centers and content representations to be processed by each processing unit before each round of feature aggregation iterative computation. In each round of iteration, each processing unit performs similarity measurement computation on its own content representation with respect to the cluster centers distributed by the master processing unit. The master processing unit updates the cluster centers based on the computed similarity measurements of all content representations.

[0024] By distributing the feature aggregation iterative computation on the plurality of processing units and performing in parallel, the present disclosure embodiments can achieve a large number of cluster centers that cannot be achieved on a single processing unit, allowing for accurate and diverse analysis of user profiles, on the one hand. On the other hand, the present disclosure embodiments achieve an improvement in computing speed, enabling feature aggregation for the same number of content representations to be achieved in a shorter time.

[0025] According to the present disclosure embodiments, in each iteration computation, the content representations to be processed by a processing unit can be further divided into a plurality of subgroups, and the similarity measurements of the content representations in each subgroup with respect to the cluster centers are computed in turn, and then the computation results are merged. This can further reduce the amount of computation performed on a single processing unit each time, avoiding the situation of running out of memory of the processing unit.

[0026] According to the present disclosure embodiments, a retrieval framework is used to store the computed cluster centers and establish an index for all cluster centers. Using the most similar cluster centers of each content representation of the plurality of content representations as training data, a first sequence model can be trained, so that the trained sequence representation model can accurately predict similar cluster centers for input cluster centers. The trained first sequence representation model can predict the most similar cluster center for each content representation given by the retrieval framework, thereby constructing a similar cluster center group. The similar cluster center group can be used to filter similar cluster centers for content representations obtained in real time. This can improve the accuracy of implicit profile analysis.

[0027] According to an embodiment of the present disclosure, the second sequence model can be trained using the most concentrated multiple cluster centers of each user's content representation as training data, so that the trained second sequence model can accurately predict the associated but dissimilar cluster centers for each input cluster center. The trained first sequence representation model can predict the associated cluster center of each content representation given by the retrieval framework for the most similar cluster center, thereby constructing an associated cluster center group. The associated cluster center group can be used to expand the associated cluster center for the content representation obtained in real time. This can improve the diversity of implicit portrait analysis.

[0028] The embodiment of the present disclosure can efficiently and accurately mine the interests and preferences of users, improve the diversity of implicit portraits, improve the accuracy and efficiency of the recommendation system, and allow more personalized and accurate services to be provided to users. At the same time, the embodiment of the present disclosure can also effectively protect user privacy and ensure that the personal information of users is not leaked.

[0029] Figure 1 is a block diagram of an example system architecture 100 for implicit portrait mining according to an embodiment of the present disclosure.

[0030] As shown in Figure 1 The system architecture 100 includes a content representation unit 101, a cluster center calculation unit 103, a retrieval unit 105, and a sequence representation unit 107. These units can be implemented via one or more central processing units and / or one or more graphics processing units.

[0031] The content representation unit 101, for example, uses a single / multi-modal pre-trained representation model to represent the text, images, graphics, and videos involved in user interactions (creation, strong interaction, weak interaction, etc.), obtains vectors corresponding to the content, referred to as content representations. One content representation is, for example, a 768-dimensional vector. The content representation unit 101 can use existing user consumption content to obtain a large amount of content representations. The process of obtaining content representations is the process of extracting content features.

[0032] In some embodiments, a text encoder BERT (Bidirectional Encoder Representation from Transformers) based on a CLIP (Contrastive Language-Image Pre-training) model can be used to extract text features, and a ViT (Vision Transformer) can be used to extract image features.

[0033] Taking text as an example, the following steps can be used for feature extraction:

[0034] (1) Convert the original data (e.g., a text) into the input format required by BERT, i.e., convert the text into tokens (tags), and then add special tokens (e.g., CLSC, SEP, etc.) as the beginning and end as input data

[0035] (2) Input the input data into the BERT model to obtain the output vector of each token.

[0036] (3) Average pool all output vectors to obtain a vector of, for example, 768 dimensions as the feature representation of the text, i.e., content representation.

[0037] In some embodiments, the content representation unit 101 can process the content representation, for example, truncate the full-precision content representation (e.g., 32 bits) to half-precision content representation (e.g., 16 bits) to reduce the occupation of video memory while minimizing the loss of precision.

[0038] In other embodiments, the content representation unit can preprocess the original data, for example, data cleaning, data normalization, data sampling, etc. Data cleaning can remove useless data and noise data to ensure the quality and accuracy of the data. Data normalization can scale the value range of the data to [0, 1] to facilitate subsequent feature extraction. Data sampling can sample large-scale data to facilitate feature extraction and clustering within an acceptable time.

[0039] The cluster center calculation unit 101 can use an unsupervised clustering algorithm (e.g., K-means, DSCAN) to cluster the multiple content representations obtained by the content representation unit 101 to obtain one or more cluster centers corresponding to the multiple content representations.

[0040] Taking text content as an example, feature clustering can use a K-means algorithm. For example, feature clustering can include the following steps:

[0041] (1) Take all content representations (i.e., feature vectors) to be clustered as input data and set the number of clusters K.

[0042] (2) Randomly initialize K cluster centers (i.e., cluster centers).

[0043] (3) Calculate the distance between each content representation and each cluster center, and assign the content representation to the cluster where the nearest cluster center is located.

[0044] (4) Recalculate the center of each cluster.

[0045] (5) Repeat steps (3) and (4) until the cluster centers no longer change or a maximum number of iterations is reached. The cluster centers obtained at this time are taken as the cluster centers corresponding to all content features.

[0046] In the prior art, all content features are clustered by a central processing unit or a single graphics processing unit to obtain cluster centers corresponding to all content features. However, because the central processing unit or the single graphics processing unit has limited computing resources, clustering features by the central processing unit or the single graphics processing unit can only produce a small number of cluster centers, which cannot provide fine-grained feature description, and on the other hand, the convergence speed of feature clustering is slow and time-consuming. For example, a single graphics processing unit may only allow 20 cluster centers to be aggregated when performing feature clustering on 10 million content features, and it takes several days.

[0047] According to embodiments of the present disclosure, feature clustering can be implemented via multiple processing units (such as central processing units, graphics processing units, etc.). The processing units implementing the cluster center computing unit can belong to the same computing node or different computing nodes.

[0048] In some embodiments, feature clustering can be implemented via multiple graphics processing units. In other words, the cluster center computing unit 101 can be implemented by multiple graphics processing units. The following describes an example in which the cluster center computing unit 101 is implemented by multiple graphics processing units. However, those skilled in the art can understand that the implementation of the cluster center computing unit 101 is not limited thereto.

[0049] Unlike the prior art in which a single central processing unit or a single graphics processing unit is used to perform feature clustering on all content features, in embodiments of the present disclosure, all content features are divided into multiple batches of content features, and each batch of content features is used for a round of feature clustering iteration. Each batch of content features is divided into multiple groups of content features for distribution to multiple graphics processing units.

[0050] The multiple graphics processing units include, for example, a master graphics processing unit and slave graphics processing units. The master graphics processing unit can perform calculations like the slave graphics processing units, and can also have control functions. For example, the master graphics processing unit can generate cluster centers, update cluster centers, divide content features, distribute content features, etc.

[0051] For the first round of feature clustering iteration computation, the master GPU can divide the first batch of content representations for the first round of feature clustering iteration computation into a first plurality of groups of content representations, the number of groups being the same as the number of GPUs. The master GPU can randomly generate K (e.g., 100,000) cluster centers, i.e., initial cluster centers. The master GPU sends all the initial cluster centers to each GPU and a group of content representations corresponding to the GPU. Each GPU computes similarity measures (e.g., cosine similarity measures) between each content representation in the group and each of the initial cluster centers based on the received initial cluster centers and the received group of content representations, and sends the computed similarity measures to the master GPU. After the master GPU receives the computed similarity measures from all the GPUs, the master GPU updates the initial cluster centers based on all the similarity measures to obtain updated cluster centers.

[0052] Next, a second round of feature clustering iteration computation is entered. In the second round of iteration, the master GPU can divide the second batch of content representations into a second plurality of groups of content representations. The master GPU can send the updated cluster centers to each GPU and a group of content representations corresponding to the GPU. Similarly, each GPU computes similarity measures (e.g., cosine similarity measures) between each content representation in the group and each of the updated cluster centers based on the received updated cluster centers and the received group of content representations, and sends the computed similarity measures to the master GPU. After the master GPU receives the computed similarity measures from all the GPUs, the master GPU updates the cluster centers again based on all the similarity measures to obtain further updated cluster centers.

[0053] Multiple rounds of feature clustering iteration computation can be performed until the cluster centers no longer change or a maximum threshold number of iterations is reached. The final obtained cluster centers can be output at this time. These cluster centers represent, for example, points of interest discovered based on the plurality of content representations.

[0054] According to embodiments of the present disclosure, since the iteration processing is performed on multiple GPUs based on the batching of content representations, on one hand, fast feature convergence is possible, and on the other hand, more cluster centers can be achieved, thereby efficiently allowing the diversity of user profiling analysis.

[0055] In some embodiments, each processing unit, when computing the similarity measure between each content representation in a group of content representations assigned to it and each cluster center in the plurality of cluster centers, can further divide the group of content representations into a plurality of subgroups of content representations, and then compute the similarity measure in batches for the plurality of subgroups of content representations. For example, the similarity measure between each content representation in each subgroup of content representations and each cluster center in the plurality of cluster centers is computed sequentially for each subgroup of content representations. The results of the computation for each subgroup of content representations are then combined together. This can further reduce the requirement on the computing power of each graphics processing unit and avoid exhausting the memory of the graphics processing unit when facing the same computing demand.

[0056] According to embodiments of the present disclosure, by utilizing multiple graphics processing units to perform the feature clustering iterative computation in batches, 10 million content representations can be clustered into, for example, 100 thousand clusters in less than one day. This can greatly improve the efficiency and at the same time allow fine-grained user profiling analysis.

[0057] The retrieval unit 105 can include a retrieval framework such as HNSW (Hierarchical Navigable Small World), Faiss, Milvus, Proxima, Vearch, Jina, etc. The retrieval unit 105 can receive the cluster centers computed by the cluster center computation unit 103, store all the cluster centers and index all the cluster centers using the retrieval framework. Each cluster center has a cluster center ID (identifier). Using the retrieval framework, one or more cluster centers with the highest similarity to a content representation of content interacted by a user in real time can be efficiently retrieved. The retrieval is, for example, based on the retrieval framework and a cosine similarity measure of the content representation and the cluster centers.

[0058] The sequence representation unit 107 can utilize one or more sequence representation models (e.g., word2vector, deepwalk, etc.) to determine similar cluster centers and / or related cluster centers for assisting cluster center filtering and cluster center expansion when performing user profiling analysis in real time.

[0059] As mentioned above, the cluster centers calculated by the cluster center calculation unit 103 can have respective cluster center IDs. In some embodiments, for a given content representation, using the retrieval model, based on a similarity measure between the content representation and the cluster centers, the top, for example, 10 cluster centers can be determined, and thus the IDs of these 10 cluster centers. The set of these 10 cluster center IDs (e.g., also preserving the ranking of the similarity measure) can form a set of training data. In the same way, a large number of sets of training data can be obtained from the large number of content representations obtained by the content representation unit 101. These sets of training data can be used, for example, to train a word2vector model. Specifically, taking the set of cluster center IDs as an example, a number of the cluster center IDs can be masked as input data for training the word2vector model. The model outputs predicted cluster center IDs, and the model parameters are adjusted according to whether the predicted cluster center IDs are consistent with the actual cluster center IDs. By using a plurality of sets of cluster center IDs corresponding to a plurality of content representations to train the word2vector model, a trained word2vector model is obtained.

[0060] At this time, the ID of each cluster center calculated by the cluster center calculation unit 103 is input into the trained word2vector model, and the predicted similar cluster center ID is obtained, and (input cluster center ID, output cluster center ID) is stored as a set of similar cluster centers.

[0061] In some embodiments, each of the plurality of content representations calculated by the content representation unit 101 is associated with a respective user. A user can be associated with a plurality of content representations. The plurality of content representations associated with a user can be used to obtain training data for training, for example, a deepwalk model. In some embodiments, using the retrieval framework and based on a similarity measure between each of the content representations associated with the user and the cluster centers, a plurality of cluster centers corresponding to the plurality of content representations of the user can be determined. That is, the cluster centers to which the plurality of content representations of the user are aggregated are determined. The number of content representations aggregated at each cluster center can be different for different cluster centers. The number of content representations of the user aggregated at each cluster center can be calculated, ranked, and the top, for example, 20 cluster centers are determined, and thus the IDs of these 20 cluster centers are obtained. The set of these 20 IDs (e.g., preserving the ranking) can form a set of training data.

[0062] When training the deepwalk model using such training data, similar to training the word2vector, a set of cluster center IDs is input into the deepwalk model, with a number of cluster center IDs being masked. The parameters of the deepwalk model are adjusted according to the output of the deepwalk model. By using a plurality of sets of cluster center IDs to train the deepwalk model, a trained deepwalk model is obtained.

[0063] At this time, the ID of each cluster center calculated by the cluster center calculation unit 103 is input into the trained deepwalk model, the predicted associated cluster center ID is obtained, and (input cluster center ID, output cluster center ID) is stored as a set of associated cluster centers.

[0064] Figure 2 is a flowchart of an example data processing method 200 according to an embodiment of the present disclosure.

[0065] As shown in Figure 1 , the method 200 can include an operation 201 of obtaining a plurality of sets of content representations.

[0066] The content representation can be a vector representing the content. For example, one content representation is a 728-dimensional 32-bit vector.

[0067] Obtaining a plurality of sets of content representations can include one or more pre-processing of raw data before the content is represented. For example, data cleaning can be performed on the raw data to remove useless data and noise data, to ensure the quality and accuracy of the data. The data can be normalized to scale the value range of the data to [0, 1] to facilitate subsequent feature extraction. Large-scale data can be sampled to limit the number of contents ultimately represented to facilitate feature extraction and clustering within an acceptable time.

[0068] The method 200 can further include an operation 202 of dividing the plurality of content representations into at least one batch of content representations for at least one round of feature clustering iteration calculation.

[0069] The size of each batch of content representations can be determined according to the total computing processing capacity of the plurality of graphics processing units participating in the calculation, for example.

[0070] The method 200 can further include an operation 203 of performing the at least one round of feature clustering iteration calculation, in each iteration: dividing a batch of content representations for the iteration into a plurality of sets of content representations for distribution to a corresponding plurality of processing units respectively; distributing a plurality of cluster centers to the plurality of processing units and distributing the plurality of sets of content representations to the corresponding plurality of processing units, each processing unit calculating a similarity measure between each content representation in a set of content representations distributed to it and each cluster center in the plurality of cluster centers; determining a cluster center to which each content representation belongs based on the calculated similarity measure of each content representation; updating the plurality of cluster centers by, for each cluster center, performing a pooling average on the content representations belonging to the cluster center to obtain an updated cluster center, thereby obtaining an updated plurality of cluster centers.

[0071] The method 200 can also include operation 204, at which the feature clustering iteration computation is ended when the plurality of cluster centers is no longer changed or reaches a threshold number of iteration times, and the final plurality of cluster centers is output.

[0072] The processing unit can be a central processing unit and / or a graphics processing unit. In some embodiments, the method is performed by a plurality of graphics processing units including a master graphics processing unit and slave graphics processing units, and the plurality of rounds of feature clustering iteration computation can be implemented as follows.

[0073] (1) The master graphics processing unit divides the first batch of content representations for the first round of iteration computation into a first plurality of groups of content representations corresponding to the plurality of graphics processing units.

[0074] (2) The master graphics processing unit initializes K cluster centers.

[0075] (3) The master graphics processing unit sends the K initial cluster centers and a corresponding group of content representations to each graphics processing unit. (It can be understood that in the case where the master graphics processing unit also participates in the computation, “each graphics processing unit” includes the master graphics processing unit, and the sending can be understood as the master graphics processing unit internally distributing.)

[0076] (4) Upon receiving the K initial cluster centers and its own group of content representations, each graphics processing unit calculates a similarity measure (e.g., a cosine similarity measure) between each content representation in the group of content representations and each initial cluster center. Taking K as 100,000 and a group of content representations containing 10,000 content representations as an example, each graphics processing unit calculates 1 billion similarity measures. For each content representation, there are 100,000 similarity measures corresponding to the distance between the content representation and the 100,000 cluster centers, wherein the smallest similarity measure indicates that the content representation should be attributed to the cluster center associated with the similarity measure. The master graphics processing unit aggregates the similarity measures of each content representation calculated by each graphics processing unit (including the master graphics processing unit in the case where the master graphics processing unit participates in the computation).

[0077] (5) Based on the similarity measures of each content representation calculated, the cluster center to which each content representation is attributed is determined. The master graphics processing unit can collect the similarity measures of each content representation calculated by each graphics processing unit. For each content representation, the cluster center to which the content representation is attributed is determined according to the smallest similarity measure in the calculated similarity measures.

[0078] (6) The master graphics processing unit updates the plurality of cluster centers by obtaining an updated cluster center for each cluster center by performing a pooling average on the content representations attributed to the cluster center, thereby obtaining an updated plurality of cluster centers.

[0079] (7) The master GPU divides the second batch of content representations for the second round of iteration computation into a second plurality of groups of content representations corresponding to the plurality of GPUs;

[0080] (8) The master GPU sends the K updated cluster centers and the corresponding group of content representations to each GPU in the first round;

[0081] (9) Upon receiving the K updated cluster centers and its own group of content representations, the master GPU computes a similarity measure (e.g., cosine similarity measure) between each content representation in the group and each updated cluster center.

[0082] (10) The master GPU determines the cluster center to which each content representation belongs based on the computed similarity measure of each content representation.

[0083] (11) The master GPU updates the plurality of cluster centers by obtaining updated cluster centers by performing a pooling average on the content representations belonging to each cluster center for each cluster center, thereby obtaining an updated plurality of cluster centers.

[0084] Subsequent iterations are performed in a similar manner until the plurality of cluster centers no longer changes or a threshold number of iterations is reached, at which point the feature clustering iteration computation ends, and the final plurality of cluster centers is output.

[0085] Those skilled in the art can understand that the number of processing units, the order of operation steps, etc. can be modified as needed without deviating from the teachings of the present disclosure.

[0086] Figure 3 A flowchart of a method 300 for training a first-order representation model and obtaining a similar group of cluster centers according to an embodiment of the present disclosure is shown.

[0087] As shown in Figure 3 , the method 300 can include operation 301, in which a plurality of cluster centers is indexed using a retrieval framework to obtain an ID of each cluster center.

[0088] The retrieval framework is, for example, a retrieval framework such as HNSW, Faiss, Milvus, Proxima, Vearch, Jina, etc. as described above. The plurality of cluster centers here is, for example, the final cluster centers aggregated according to the method of Figure 2 . The retrieval framework stores and indexes the plurality of cluster centers. The retrieval framework is capable of returning similarity search results.

[0089] The method 300 can include operation 303, in which, for each content representation in a first plurality of content representations in the plurality of content representations, a first number of cluster centers most similar to the content representation are retrieved as a set of training data corresponding to the content representation, using the retrieval framework and based on a similarity measure of the content representation to the cluster centers, thereby obtaining a first plurality of sets of training data.

[0090] The plurality of content representations herein, for example, are Figure 2 All content representations used for feature clustering. The first plurality of content representations can be all the plurality of content representations, or a part of the plurality of content representations. As long as the first plurality of content representations are sufficient to train a good first sequential feature model.

[0091] For each content representation in the first plurality of content representations, a predetermined number (e.g., 10) of cluster centers most similar to the content representation can be retrieved, using the retrieval framework and based on a similarity measure of the content representation to the cluster centers. That is, for each content representation, a set of 10 most similar cluster center IDs can be obtained. In the set, the similarity can be ranked. Such a set of cluster center IDs can be taken as a set of training data. By obtaining a set of cluster center IDs for each content representation in the first plurality of content representations, a first plurality of sets of training data can be obtained.

[0092] The method 300 can include operation 305, in which the first sequential feature model is trained using the obtained first plurality of sets of training data, thereby obtaining a trained first sequential feature model.

[0093] The first sequential feature model is, for example, a word2vector model. Those skilled in the art can understand that other similar models can be used.

[0094] A set of training data, i.e., a set of cluster center IDs, is input into the word2vector model with a part of the IDs masked, the word2vector model outputs predicted IDs, and then the parameters of the word2vector model are adjusted according to the output.

[0095] The method 300 can include operation 307, in which, for each cluster center in the plurality of cluster centers, a similar cluster center set is obtained, including: inputting the cluster center ID into the trained first sequential feature model and obtaining an output cluster center ID; and storing the pair of input cluster center ID and output cluster center ID as a similar cluster center set.

[0096] After the word2vector model is trained, for each cluster center calculated by the method 200, find a similar cluster center using the trained word2vector model. Specifically, input the ID of each cluster center obtained by the method 200 into the trained word2vector model, and the trained word2vector model will output the predicted ID of the cluster center most similar to the cluster center. The input ID and the output ID are stored as a similar cluster center group. Finally, the same number of similar cluster center groups as the cluster centers obtained by the method 200 will be obtained. Figure 2 Figure 2 The similar cluster center group means that the contents represented by the two cluster center IDs in the group are very similar, and when performing user portrait analysis or pushing data to users, only one of the cluster centers can be considered. Therefore, the similar cluster center group can be used for redundant filtering of cluster centers.

[0097] The similar cluster center group means that the contents represented by the two cluster center IDs in the group are very similar, and when performing user portrait analysis or pushing data to users, only one of the cluster centers can be considered. Therefore, the similar cluster center group can be used for redundant filtering of cluster centers.

[0098] Figure 4 A flowchart of a method 400 for training a second sequence representation model and obtaining a related cluster center group according to an embodiment of the present disclosure is shown.

[0099] As shown in Figure 4 , the method 400 can include operation 401, in which a plurality of cluster centers are indexed using a retrieval framework to obtain an ID of each cluster center.

[0100] Operation 401 is similar to operation 301 in Figure 3 , and will not be described here.

[0101] The method 400 can include operation 403, in which for a plurality of content representations associated with each user in a plurality of users, the following operations are performed to obtain a second plurality of training data: using a retrieval framework and based on a similarity measure of content representations and cluster centers, calculating a corresponding plurality of cluster centers most similar to the plurality of content representations, according to the number of content representations of the user aggregated at each cluster center in the plurality of cluster centers, sorting from high to low, obtaining the ID of the second number of the plurality of cluster centers with the highest ranking as a group of training data.

[0102] The plurality of content representations here are, for example, the content representations of the user in the plurality of users. Figure 2 ​The plurality of content representations are used to generate training data. Specifically, the plurality of content representations associated with a user are extracted, for which the retrieval framework is utilized and based on the similarity measure of each content representation to a cluster center, the ID of the most similar cluster center corresponding to each content representation can be obtained, thereby obtaining a plurality of IDs of the most similar cluster centers corresponding to the plurality of content representations respectively. Then, the specific number of content features aggregated at each cluster center in the plurality of content features of the user is calculated. The plurality of cluster centers are sorted in descending order according to the number of content features of the user aggregated at each cluster center, and the IDs of the top-ranked predetermined number (e.g., 20) of cluster centers are used to form a set of IDs as a training data. In the set, the order of high and low ranking can be preserved.

[0103] For a plurality of users, a plurality of sets of such ID sets can be obtained accordingly as a plurality of sets of training data.

[0104] The method 400 can include operation 405, in which a second sequence representation model is trained using the obtained second plurality of sets of training data, and a trained second sequence representation model is obtained.

[0105] The second sequence representation model can be, for example, a deepwalk model. Those skilled in the art can understand that other similar models can be used.

[0106] A set of training data, i.e., a set of IDs of a cluster center, is input into the deepwalk model with a portion of the IDs masked as input data, and the deepwalk model outputs predicted IDs, and then the parameters of the deepwalk model are adjusted according to the output.

[0107] The method 400 can include operation 407, in which, for each cluster center in the plurality of cluster centers, an associated cluster center group is obtained, including: inputting the ID of the cluster center into the trained second sequence representation model and obtaining the output cluster center ID; and storing the pair of input cluster center ID and output cluster center ID as an associated cluster center group.

[0108] After the deepwalk model is trained, the trained deepwalk model is used to find the associated cluster center group for each cluster center calculated by the method of Figure 2 Specifically, the ID of each cluster center is input into the trained deepwalk model, and the trained deepwalk model outputs the predicted ID of the cluster center associated with the cluster center but not similar to the cluster center. The input ID and the output ID are stored as an associated cluster center group. Finally, the same number of associated cluster center groups as the number of cluster centers can be obtained.

[0109] An associated cluster center group means that the content of two cluster center IDs in the group is associated but not similar (e.g. one cluster center is a cat and one cluster center is a dog, which is associated but not similar), when conducting user portrait analysis or pushing data to a user, the first cluster center ID can be based on the first cluster center ID, while also considering the second cluster center ID, i.e. the associated cluster center ID of the first cluster center ID. The associated ID can help to achieve accurate and rich data pushing to the user.

[0110] Figure 5 A flow chart of a method 400 for establishing a user portrait according to an embodiment of the present disclosure is shown.

[0111] As shown in Figure 5 , the method 500 can include operation 501, in which a retrieval framework is utilized to retrieve a first number of cluster centers with the highest similarity to the content representation of real-time content based on a similarity measure of the content representation of the real-time content to the cluster centers.

[0112] The retrieval framework can be a retrieval framework that stores the cluster centers obtained by the method shown in Figure 2 .

[0113] Real-time content is generated, for example, when a user conducts content interaction. The content interaction can include, for example, creating content (e.g. publishing text, pictures, videos, picture-text posts, etc.), interacting with content (e.g. browsing, liking, commenting, etc.), and the like. The real-time content can be represented as a vector, i.e. a content representation.

[0114] When receiving the content representation of real-time content, a retrieval framework can be utilized to retrieve a predetermined number (e.g. 10) of cluster centers with the highest similarity to the content representation based on a similarity measure (e.g. cosine similarity) of the real-time content to the cluster centers, more specifically, the IDs of these cluster centers.

[0115] The method 500 can further include operation 502, in which at least one similar cluster center is filtered from the first number of cluster centers based on a plurality of similar cluster center groups, resulting in a second number of cluster centers.

[0116] The plurality of similar cluster center groups are, for example, similar cluster center groups obtained by the method shown in Figure 3 . In some embodiments, the first number of cluster centers can retain the order of similarity from high to low. That is, the first number of cluster centers is a sequence of cluster centers with similarity from high to low.

[0117] Filtering the first number of cluster centers for at least one similar cluster center can include matching the first number of cluster centers with each cluster center group in the plurality of similar cluster center groups, and if there are two cluster centers matching a cluster center group, only one of the two cluster centers is kept. The second number of cluster centers after filtering can also be a sequence of cluster centers with different similarity levels.

[0118] The cluster center filtering can avoid causing unnecessary redundancy when building a user profile or pushing data to a user.

[0119] The method 500 can further include operation 503, in which, based on the plurality of associated cluster center groups, the second number of cluster centers is expanded to obtain a third number of cluster centers.

[0120] The plurality of associated cluster center groups are, for example, obtained by the method of Figure 4 The expansion of the second number of cluster centers can include sorting the cluster centers, finding the associated cluster center for each cluster center, and inserting the associated cluster center into the sequence of cluster centers until the number of obtained cluster centers reaches a predetermined third number (e.g., 50). In some embodiments, the expansion can include multiple expansions, i.e., after the first round of expansion, if the third number has not been reached, the expanded cluster centers can be expanded again until the number of obtained cluster centers reaches the predetermined third number.

[0121] Those skilled in the art can understand that the first number, the second number, and the third number in the method can be set as needed.

[0122] The method 500 can further include operation 504, in which, based on the third number of cluster centers, an implicit profile is built.

[0123] In this step, the implicit profile of the user is analyzed / built based on the final cluster centers obtained through the above filtering and expansion, so as to achieve accurate and rich data pushing.

[0124] Figure 6 is a schematic diagram showing a general hardware environment of a device that can implement embodiments according to the present disclosure.

[0125] Reference is now made to Figure 6 , which shows a schematic diagram of an example of a computing node 600. The computing node 600 is merely one example of a suitable computing node and is not intended to suggest any limitation as to the scope of use or functionality of embodiments described herein. Regardless, the computing node 600 is capable of being implemented and / or performing any of the functionality set forth hereinabove.

[0126] In computing node 600, there is a computer system / server 6012, which is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well- known computing systems, environments, and / or configurations that can be suitable for use with computer system / server 6012 include, but are not limited to, personal computers, server computers, thin clients, thick clients, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed cloud computing environments that include any of the above systems or devices, and the like.

[0127] Computer system / server 6012 can be described in the general context of computer system-executable instructions, such as program modules, being executed by a computer system. Generally, program modules can include routines, programs, objects, components, logic, data structures, and so on that perform particular tasks or implement particular abstract data types. Computer system / server 6012 can be practiced in distributed cloud computing environments with remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules can be located in both local and remote computer system storage media including memory storage devices.

[0128] As Figure 6 shown in FIG. 11, computer system / server 6012 in computing node 600 is shown in the form of a general-purpose computing device. The components of computer system / server 6012 can include, but are not limited to, one or more processors or processing units 6016, a system memory 6028, and a bus 6018 that couples various system components including system memory 6028 to processor 6016.

[0129] The one or more processor(s) or processing unit(s) 6016 can include central processing units and / or graphics processing units for implementing the methods and systems described in this disclosure.

[0130] Bus 6018 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor or local bus using any of a variety of bus architectures, and the like. As an example and not by way of limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, Peripheral Component Interconnect (PCI) bus, Peripheral Component Interconnect Express (PCIe), and Advanced Microcontroller Bus Architecture (AMBA) OMI.

[0131] Computer system / server 6012 typically includes a variety of computer system readable media. These media can be any available media that is accessible by computer system / server 6012 and includes both volatile and non-volatile media, removable and non-removable media.

[0132] The system memory 6028 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 6030 and / or cache memory 6032. Computer system / server 6012 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, a storage system 6034 can be provided for reading from and writing to non-removable, non-volatile magnetic media (not shown and typically called a "hard drive"). Although not specifically shown, a magnetic disk drive can also be used for reading from and writing to a removable, non-volatile magnetic disk (e.g., a "floppy disk"), and an optical disk drive can be used for reading from or writing to a removable, non-volatile optical disk (such as a CD-ROM, DVD-ROM or other optical media). In these instances, each can be connected to bus 6018 by one or more data media interfaces. As will be further depicted and described below, memory 6028 can include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the disclosure.

[0133] By way of example, and not limitation, as depicted and described below, program / utility 6040 having a set (at least one) of program modules 6042 can be stored in memory 6028. Each of the operating system, one or more application programs, other program modules, and program data or some combination thereof, can include implementation of a network environment. Program modules 6042 generally carry out the functions and / or methodologies described in embodiments as described herein.

[0134] The computer system / server 6012 can also communicate with one or more external devices 6014 such as a keyboard, a pointing device, a display 6024, etc.; one or more devices that enable a user to interact with the computer system / server 6012; and / or one or more devices that enable the computer system / server 6012 to communicate with one or more other computing devices. Such communication can occur via Input / Output (I / O) interfaces 22. Still yet, the computer system / server 6012 can communicate with one or more networks such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet) via network adapter 20. As depicted, network adapter 20 communicates with the other components of the computer system / server 6012 via bus 6018. It should be appreciated that although not shown, other hardware and / or software components could be used in conjunction with the computer system / server 6012. Examples include, but are not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0135] The present disclosure can be embodied as a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present disclosure.

[0136] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0137] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0138] Computer readable program instructions for carrying out operations of the present disclosure can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state setting data, or any combination of one or more of the above in any combination, written in any combination of one or more programming languages, including object oriented programming languages such as Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to customize the electronic circuitry, in order to perform aspects of the present disclosure.

[0139] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present disclosure. It will 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 readable program instructions.

[0140] These computer readable program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions can also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0141] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0142] The flow diagrams and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow diagrams and block diagrams can represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical functions (s). In some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations of blocks in the block diagrams and / or flow diagrams, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or combinations of special purpose hardware and computer instructions.

[0143] Those skilled in the art will further appreciate that the various operations illustrated in the figures can not be performed in the order shown. In some embodiments, the operations can be performed in a different order, or some operations can be performed concurrently. Those skilled in the art will also appreciate that some operations can be performed by special purpose hardware-based systems, which can be utilized in place of, or in conjunction with, computer software instructions.

[0144] The description of various embodiments of the present disclosure has been presented for purposes of illustration, but is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application, or technical improvement over technology found in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A data processing method for implicit image mining, comprising: dividing a plurality of content representations into at least one batch of content representations for at least one round of feature clustering iteration calculation, wherein the content representations comprise feature vectors corresponding to contents obtained by representing texts, images and videos involved in user interactions; performing the at least one round of feature clustering iteration calculation, wherein in each iteration round: dividing the batch of content representations for the iteration round into a plurality of groups of content representations for respectively distributing to a corresponding plurality of processing units; distributing a plurality of cluster centers to the plurality of processing units and distributing the plurality of groups of content representations to the corresponding plurality of processing units, each processing unit calculating a similarity measure between each content representation in the group of content representations distributed thereto and each cluster center in the plurality of cluster centers; determining a cluster center to which each content representation belongs based on the calculated similarity measure of each content representation; updating the plurality of cluster centers by obtaining an updated cluster center for each cluster center by performing a pooling average on the content representations belonging to the cluster center, thereby obtaining an updated plurality of cluster centers. 2.The data processing method of claim 1, further comprising: obtaining the calculated similarity measures of the content representations from the plurality of processing units.

3. The data processing method of claim 1, wherein, The calculating, by each processing unit, of the similarity measure between each content representation in the group of content representations distributed thereto and each cluster center in the plurality of cluster centers further comprises: dividing the group of content representations into a plurality of subgroups of content representations, calculating the similarity measures in batches for the plurality of subgroups of content representations, comprising: sequentially calculating, for each subgroup of content representations, the similarity measure between each content representation in the subgroup of content representations and each cluster center in the plurality of cluster centers. 4.The data processing method of claim 3, further comprising: after the feature clustering iteration calculation is completed, indexing the plurality of cluster centers using a retrieval framework to obtain an ID of each cluster center; for each content representation in a first plurality of content representations in the plurality of content representations, retrieving, using the retrieval framework and based on the similarity measures of the content representations and the cluster centers, a first number of IDs of the most similar cluster centers to the content representation as a training data set corresponding to the content representation, thereby obtaining a first plurality of training data sets; training a first sequential representation model using the obtained first plurality of training data sets to obtain a trained first sequential representation model. 5.The data processing method of claim 4, further comprising: for each cluster center in the plurality of cluster centers, obtaining a similar cluster center group, comprising: inputting the ID of the cluster center into the trained first sequential representation model and obtaining an output cluster center ID; and storing the pair of input cluster center ID and output cluster center ID as a similar cluster center group. 6.The data processing method of claim 5, further comprising: for a plurality of content representations associated with each user in a plurality of users in the plurality of content representations, performing the following operations to obtain a second plurality of training data sets: using a search framework and based on similarity measures of the content representations to the cluster centers, computing a respective plurality of cluster centers that are most similar to the plurality of content representations, obtaining, as a second set of training data, a second number of cluster centers ordered from high to low according to the number of content representations of the user aggregated at each cluster center in the plurality of cluster centers; training a second sequential representation model using the obtained second set of training data to obtain a trained second sequential representation model.

7. The data processing method of claim 6, further comprising: for each cluster center in the plurality of cluster centers, obtaining an associated cluster center group, comprising: inputting the cluster center ID into the trained second sequential representation model and obtaining an output cluster center ID; and storing the pair of input cluster center ID and output cluster center ID as an associated cluster center group.

8. The data processing method of claim 7, further comprising: using a search framework, based on similarity measures of content representations of real-time content to cluster centers, searching for a first number of cluster centers that are most similar to the content representations of the real-time content; based on the plurality of similar cluster center groups of the plurality of content representations, filtering at least one similar cluster center from the first number of cluster centers to obtain a second number of cluster centers; based on the plurality of associated cluster center groups of the plurality of content representations, expanding the second number of cluster centers to obtain a third number of cluster centers; and establishing an implicit profile based on the third number of cluster centers. a master graphics processing unit in the plurality of graphics processing units is configured to perform operations of the method of any of claims 1-8.

9. A data processing system comprising a plurality of graphics processing units, wherein, 10. A computer-readable storage medium having stored thereon computer-readable program instructions, which instructions, when executed by a processor, perform operations of the method of any of claims 1-8. ​

Images