Method for processing data, method for training data processing model, and computing device

Abstract

Embodiments of the present disclosure provide a method for processing data, a method for training a data processing model, and a computing device. The method for processing data comprises: receiving user behavior data of a user; and inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model, and obtaining a data processing result corresponding to the user. The data processing model is obtained by training using sequential sample features and non-sequential sample features. The sequential sample features and the non-sequential sample features are obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model, obtaining initial sequential sample features and initial non-sequential sample features, and performing feature enhancement on the initial sequential sample features and the initial non-sequential sample features. The problem of poor model training efficiency caused by sample data is prevented, meeting the requirements of practical application.

Description

Data processing method, data processing model training method and computing device

[0001] This disclosure claims priority to the Chinese patent application filed with the China Patent Office on December 29, 2023, with application number 2023118601863 and application name “Data processing method, data processing model training method and computing device”, the entire content of which is incorporated by reference into this disclosure. Technical Field

[0002] The embodiments of the present disclosure relate to the field of computer technology, and in particular to a data processing method, a data processing model training method, and a computing device. Background Art

[0003] With the continuous development of computers and artificial intelligence technology, machine learning has achieved success in many fields. As a result, many neural network models trained based on sample data have emerged.

[0004] In order to improve the performance of neural network models, existing technologies often only focus on the innovation of neural network model structure, which leads to the gradual complexity of the model, resulting in poor model training efficiency and inability to meet the needs of practical applications.

[0005] Summary of the Invention

[0006] In view of this, embodiments of the present disclosure provide a data processing method. One or more embodiments of the present disclosure also relate to another data processing method, a data processing model training method, an object recommendation model training method, a data processing device, another data processing device, a data processing model training device, an object recommendation model training device, a computing device, a computer-readable storage medium, and a computer program to address technical deficiencies in the prior art.

[0007] According to a first aspect of an embodiment of the present disclosure, there is provided a data processing method, including:

[0008] Receiving user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0009] The sequential behavior data and / or the non-sequential behavior data are input into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are used to extract features of sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features.

[0010] According to a second aspect of an embodiment of the present disclosure, there is provided a data processing apparatus, including:

[0011] A data receiving module is configured to receive user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0012] The data processing module is configured to input the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain the data processing results corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are used to extract features of sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features.

[0013] According to a third aspect of an embodiment of the present disclosure, a data processing model training method is provided, comprising:

[0014] Determining sequential samples and non-sequential samples for a data processing model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users;

[0015] Inputting the sequential samples and the non-sequential samples into the data processing model to be trained, performing feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtaining initial sequential sample features and initial non-sequential sample features;

[0016] Performing feature enhancement on the initial sequential sample feature to obtain the sequential sample feature, and performing feature enhancement on the initial non-sequential sample feature to obtain the non-sequential sample feature;

[0017] Based on the sequential sample features and the non-sequential sample features, the data processing model to be trained is trained to obtain a data processing model.

[0018] According to a fourth aspect of an embodiment of the present disclosure, a data processing model training device is provided, comprising:

[0019] a sample determination module configured to determine sequential samples and non-sequential samples for training a data processing model, wherein the sequential samples and the non-sequential samples are user behavior data of sample users;

[0020] a feature extraction module configured to input the sequential samples and the non-sequential samples into the data processing model to be trained, perform feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtain initial sequential sample features and initial non-sequential sample features;

[0021] a sample enhancement module configured to perform feature enhancement on the initial sequential sample features to obtain the sequential sample features, and to perform feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features;

[0022] The model training module is configured to train the data processing model to be trained based on the sequence sample features and the non-sequence sample features to obtain a data processing model.

[0023] According to a fifth aspect of an embodiment of the present disclosure, another data processing method is provided, which is applied in the cloud, including:

[0024] receiving user behavior data of a user sent by a terminal, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0025] Inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, the sequential sample features and the non-sequential sample features are obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features;

[0026] The data processing result is sent to the terminal.

[0027] According to a sixth aspect of an embodiment of the present disclosure, another data processing device is provided, which is applied to a cloud, including:

[0028] a data receiving module configured to receive user behavior data of a user sent by a terminal, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0029] a data processing module configured to input the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, the sequential sample features and the non-sequential sample features are obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features;

[0030] The result sending module is configured to send the data processing result to the terminal.

[0031] According to a seventh aspect of an embodiment of the present disclosure, a method for training an object recommendation model is provided, comprising:

[0032] Determining sequential samples and non-sequential samples for the object recommendation model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users for the sample objects;

[0033] Inputting the sequential samples and the non-sequential samples into the to-be-trained object recommendation model, performing feature extraction on the sequential samples and the non-sequential samples in the to-be-trained object recommendation model to obtain initial sequential sample features and initial non-sequential sample features;

[0034] Performing feature enhancement on the initial sequential sample feature to obtain the sequential sample feature, and performing feature enhancement on the initial non-sequential sample feature to obtain the non-sequential sample feature;

[0035] Based on the sequential sample features and the non-sequential sample features, the object recommendation model to be trained is trained to obtain an object recommendation model.

[0036] According to an eighth aspect of an embodiment of the present disclosure, there is provided an object recommendation model training device, comprising:

[0037] A sample determination module is configured to determine sequential samples and non-sequential samples for the object recommendation model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users for the sample objects;

[0038] a feature extraction module configured to input the sequential samples and the non-sequential samples into the to-be-trained object recommendation model, perform feature extraction on the sequential samples and the non-sequential samples in the to-be-trained object recommendation model, and obtain initial sequential sample features and initial non-sequential sample features;

[0039] a sample enhancement module configured to perform feature enhancement on the initial sequential sample features to obtain the sequential sample features, and to perform feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features;

[0040] The model training module is configured to train the object recommendation model to be trained based on the sequential sample features and the non-sequential sample features to obtain an object recommendation model.

[0041] According to a ninth aspect of an embodiment of the present disclosure, there is provided a computing device, including:

[0042] memory and processor;

[0043] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above-mentioned two data processing methods, data processing model training methods, or object recommendation model training methods are implemented.

[0044] According to the tenth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, which stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned two data processing methods, data processing model training methods, or object recommendation model training methods.

[0045] According to the eleventh aspect of an embodiment of the present disclosure, a computer program is provided, wherein, when the computer program is executed in a computer, the computer is caused to execute the steps of the above two data processing methods, data processing model training methods, or object recommendation model training methods.

[0046] One or more embodiments of the present disclosure provide a data processing method, comprising: receiving user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data; inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are used to extract features of sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features.

[0047] Specifically, the data processing model in the data processing method is obtained by training using sequential sample features and non-sequential sample features. The sequential sample features and non-sequential sample features are obtained by feature enhancement of the initial sequential sample features and initial non-sequential sample features obtained by feature extraction of sequential samples and non-sequential samples in the data processing model; thereby, the influence of sample data on model training is fully considered during the model training process, avoiding the problem of poor model training efficiency caused by sample data; and, the trained data processing model is used to process the sequential behavior data and / or non-sequential behavior data contained in the user behavior data, and obtain data processing results, thereby meeting the needs of practical applications through the data processing model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] FIG1 is an application diagram of a data processing method provided by an embodiment of the present disclosure;

[0049] FIG2 is a flow chart of a data processing method provided by one embodiment of the present disclosure;

[0050] FIG3 is a schematic diagram of a data processing model training method provided by an embodiment of the present disclosure;

[0051] FIG4 is a flow chart of a data processing model training method provided by one embodiment of the present disclosure;

[0052] FIG5 is a flowchart of constructing a co-occurrence matrix in a data processing model training method provided by one embodiment of the present disclosure;

[0053] FIG6 is a schematic diagram of data expansion in a data processing model training method provided by one embodiment of the present disclosure;

[0054] FIG7 is a schematic diagram of sequential contrastive learning model training in a data processing model training method provided by one embodiment of the present disclosure;

[0055] FIG8 is a schematic diagram of sequence feature data enhancement in a data processing model training method provided by one embodiment of the present disclosure;

[0056] FIG9 is a schematic diagram of two feature shielding methods in a data processing model training method provided by an embodiment of the present disclosure;

[0057] FIG10 is a schematic diagram of a non-sequential contrastive learning model training method in a data processing model training method according to an embodiment of the present disclosure;

[0058] FIG11 is a flowchart of a processing process of a data processing model training method provided by one embodiment of the present disclosure;

[0059] FIG12 is a flowchart of another data processing method provided by one embodiment of the present disclosure;

[0060] FIG13 is a flowchart of an object recommendation model training method provided by one embodiment of the present disclosure;

[0061] FIG14 is a structural block diagram of a computing device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0062] The following description sets forth many specific details to facilitate a full understanding of the present disclosure. However, the present disclosure can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present disclosure. Therefore, the present disclosure is not limited to the specific implementations disclosed below.

[0063] The terms used in one or more embodiments of the present disclosure are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of the present disclosure. The singular forms "a", "the", and "the" used in one or more embodiments of the present disclosure and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of the present disclosure refers to and includes any or all possible combinations of one or more associated listed items.

[0064] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of one or more embodiments of the present disclosure, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0065] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of the present disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0066] In one or more embodiments of the present disclosure, a large model refers to a deep learning model with large-scale model parameters, which typically contains hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than ten trillion model parameters. A large model can also be called a cornerstone model / foundation model (Foundation Model), which is pre-trained by using large-scale unlabeled corpus to produce a pre-trained model with more than 100 million parameters. This model can adapt to a wide range of downstream tasks, and the model has good generalization capabilities, such as a large-scale language model (LLM), a multi-modal pre-training model, etc.

[0067] When large models are used in practice, only a small number of samples are needed to fine-tune the pre-trained model and it can be applied to different tasks. Large models can be widely used in natural language processing (NLP), computer vision and other fields. Specifically, they can be applied to computer vision tasks such as visual question answering (VQA), image caption (IC), and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summary generation, and machine translation. The main application scenarios of large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0068] First, the terms involved in one or more embodiments of the present disclosure are explained.

[0069] CL4SRec: short for Contrastive Learning for Sequential Recommendation, is a contrastive learning model that only considers sequence features in recommendation scenarios.

[0070] CL4CTR: A Contrastive Learning Framework for CTR Prediction is a contrastive learning model that only considers non-sequential features in recommendation scenarios.

[0071] Transformer: is a deep learning model for natural language processing (NLP) tasks, and is now mostly used as the encoder of the model.

[0072] MaskNet: Introducing Feature-Wise Multiplication to CTR Ranking Models by Instance-Guided Mask is a CTR model that can better mine cross-features.

[0073] CTR: The CTR (Click-Through Rate) model is a machine learning model that is used to predict whether a user will click on a page or link.

[0074] EasyRec: is a machine learning framework designed for recommendation scenarios. It has implemented a variety of machine learning models for common recommendation tasks.

[0075] NCE Loss: Short for Noise-Contrastive Estimation Loss, it is a loss function used for training classification models.

[0076] BERT model: is a language representation model. BERT stands for Bidirectional Encoder Representations from Transformers. BERT aims to pre-train deep bidirectional representations by jointly conditioning on the left and right context in all layers.

[0077] Duorec: Short for Contrastive Learning for Representation Degeneration Problem in Sequential Recommendation. It is an improved contrastive learning recommendation model based on CL4SRec that only considers sequential data.

[0078] DNN: refers to deep neural network algorithm.

[0079] With the continuous development of computer and artificial intelligence technologies, machine learning has achieved success in various fields, leading to the emergence of many neural network models trained based on sample data. However, to improve the performance of neural network models, the model training process often focuses solely on innovations in the neural network model structure, resulting in increasingly complex models without considering the inherent issues with the sample data. This results in poor model training efficiency and an inability to meet the needs of practical applications. For example, in one or more embodiments provided herein, with the success of machine learning in various fields, researchers have built many modern recommendation systems based on highly expressive deep neural architectures. Recommendation systems are intelligent systems that can recommend personalized content to users. They primarily predict future needs by analyzing users' historical behavior and interests, simplifying the decision-making process for users while providing a better service experience. However, these recommendation models built based on deep networks often have high data requirements and require a large amount of data to train the models. In recommendation systems, personalized recommendations rely on user-generated data, resulting in high data acquisition costs. Moreover, most users interact (e.g., consume or click on) only a small portion of the larger set of items. Therefore, data sparsity is one of the main reasons that hinders recommendation models from fully realizing their potential. Although the diversity and sparsity of data in recommendation scenarios often limit the performance of recommendation models, and there may be missing or too few samples in actual application scenarios, resulting in poor model training efficiency, researchers often only focus on innovation in model structure, resulting in the gradual complexity of the model, while ignoring the problems with the data itself.

[0080] To address the aforementioned issues, while contrastive learning algorithms have been proven effective in multiple fields for addressing data sparsity, research on contrastive learning algorithms in the recommendation field has largely focused on temporal features, lacking consideration for non-sequential features. Furthermore, researchers' excessive focus on innovation in model structure has led to increasingly complex models and prohibitively high training costs, hindering their practical application. For example, this disclosure provides two models. The first is the CL4SRec model; however, since models like CL4SRec are contrastive learning models designed for recommendation scenarios that only consider sequential features, they only consider sequential data and have a single feature set, making them difficult to apply in real-world scenarios. The second is the CL4CTR model; however, models like CL4CTR are contrastive learning models designed for recommendation scenarios that only consider non-sequential features. They only consider non-sequential data, have high training speed and cost, and are also difficult to transfer to real-world scenarios. Consequently, the sparsity and long-tail nature of data in recommendation scenarios, along with the question of how to efficiently apply self-supervision and contrastive learning algorithms in these scenarios, have become pressing challenges.

[0081] In the present disclosure, a data processing method is provided. The present disclosure also relates to another data processing method, a data processing model training method, an object recommendation model training method, a data processing device, another data processing device, a data processing model training device, an object recommendation model training device, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0082] Referring to Figure 1 , a schematic diagram illustrating an application of a data processing method provided according to an embodiment of the present disclosure is provided, and the data processing method provided by the present disclosure is described in the context of a product recommendation scenario. User behavior data can be understood as data generated when a user browses product pages, purchases products, adds products to favorites, and so on. This user behavior data includes sequential behavior data and / or non-sequential behavior data; this sequential behavior data can be a record of user interactions with products. For example, an interaction record can be composed of three time series features: product ID, product category ID, and user behavior sequence. Non-sequential behavior data can include user attribute data such as user age and gender, as well as product attribute data such as product price and product capacity. Based on this, referring to Figure 1 , a user provides their user behavior data to a server 104 via a terminal 102. This user behavior data includes sequential behavior data and / or non-sequential behavior data. After receiving the sequential behavior data and / or non-sequential behavior data, the server 104 inputs the data into a product recommendation model for processing, thereby obtaining predicted product recommendations for the user. Subsequently, product recommendations can be made to users based on the product recommendation results, thereby meeting the needs of practical applications and providing users with a good shopping experience.

[0083] Referring to FIG. 2 , FIG. 2 shows a flow chart of a data processing method provided according to an embodiment of the present disclosure, which specifically includes the following steps.

[0084] Step 202: Receive user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data.

[0085] The user behavior data can be understood as data generated during user operations. When the data processing method provided in the embodiments of the present disclosure is applied to different scenarios, the operations and user behavior data will also vary. For example, when the data processing method is applied to a product recommendation scenario, the operations can be understood as user operations such as browsing product pages, purchasing products, and adding products to favorites. The user behavior data can be understood as data generated during the user's operations such as browsing product pages, purchasing products, and adding products to favorites. It should be noted that users can perform the aforementioned operations on products on an internet shopping platform; the user behavior data can be the user's user behavior data stored on the internet shopping platform. For another example, when the data processing method is applied to a tourist attraction recommendation scenario, the operations can be understood as user operations such as browsing tourist attraction pages, purchasing tourist attraction tickets, and adding tourist attraction tickets to favorites. The user behavior data can be understood as data generated during the user's operations such as browsing tourist attraction pages, purchasing tourist attraction tickets, and adding tourist attraction tickets to favorites. It should be noted that users can perform the aforementioned operations on tourist attraction tickets on an internet tourist attraction ticket purchasing platform; the user behavior data can be the user's user behavior data stored on the internet tourist attraction ticket purchasing platform.

[0086] The sequence-type behavior data can be understood as behavior data that is sorted into a sequence according to specific parameters. For example, the sequence-type behavior data can be time-series behavior data, and the time-series behavior data can be behavior data sorted by time. In one or more embodiments provided in the present disclosure, the sequence-type behavior data can be a sequence-type behavior feature, and the sequence-type behavior feature can be a record of interaction between a user and a product. For example, an interaction record can be composed of three time series features consisting of a product ID, a product category ID, and a user behavior sequence, where different time series features are one-to-one corresponding within the same time period.

[0087] The non-sequential behavior data can be understood as behavior data that is not sorted into a sequence according to specific parameters. For example, the non-sequential behavior data can be user attribute data such as user age and user gender, as well as product attribute data such as product price and product capacity. In one or more embodiments provided in the present disclosure, the non-sequential behavior data can be non-sequential behavior features; for example, the non-sequential behavior features can be user attribute features such as user age and user gender, as well as product attribute features such as product price and product capacity.

[0088] In one or more embodiments provided by the present disclosure, the data processing method further includes receiving user behavior data of a user, and determining whether the user behavior data contains sequential behavior data and / or non-sequential behavior data; performing feature extraction on the sequential behavior data and / or non-sequential behavior data contained in the user behavior data to obtain sequential behavior features and / or non-sequential behavior features; thereby avoiding the problem of a larger model caused by configuring a network layer for feature extraction in the data processing model, and allowing the data processing model to be compatible with more hardware devices. Subsequently, the sequential behavior features and / or non-sequential behavior features are input into the data processing model to obtain the data processing results corresponding to the user. For example, in the data processing method provided by the present disclosure, the user behavior data can be a data set; after receiving the data set provided by the user, sequential behavior data and / or non-sequential behavior data are determined from the data set; then, feature extraction is performed on the sequential behavior data and / or non-sequential behavior data to obtain data features; and then, based on the feature type of the data features, the data features are divided into sequential behavior features and non-sequential behavior features.

[0089] Step 204: Input the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user.

[0090] Wherein, the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are obtained by extracting features of sequential samples and non-sequential samples in the data processing model, obtaining initial sequential sample features and initial non-sequential sample features, and performing feature enhancement on the initial sequential sample features and the initial non-sequential sample features.

[0091] The data processing model can be understood as a neural network model capable of processing sequential and / or non-sequential behavior data. When the data processing method provided by the embodiments of the present disclosure is applied to different scenarios, the data processing model will also vary. For example, when the data processing method is applied to a product recommendation scenario, the data processing model may be a product recommendation model. This product recommendation model processes the user behavior data generated during various user operations on products and outputs corresponding product recommendation results to the user's corresponding terminal. This product recommendation result can then be used to make product recommendations for the user. For another example, when the data processing method is applied to a scenic spot recommendation scenario, the data processing model may be a scenic spot recommendation model. This scenic spot recommendation model processes the user behavior data generated during various user operations on scenic spots and outputs corresponding scenic spot recommendation results to the user. This scenic spot recommendation result can then be used to make scenic spot recommendations for the user. Based on this, the data processing model can be understood as an object recommendation model. This object recommendation model processes the sequential and / or non-sequential behavior data to obtain object recommendation results for the user.

[0092] In one or more embodiments provided herein, the data processing model includes a first data processing module and a second data processing module; wherein the first data processing module can perform feature extraction on sequential behavior data to obtain sequential behavior features of the sequential behavior data, and encode the sequential behavior features using an encoder to obtain object recommendation results determined using the sequential behavior data. wherein the second data processing module can perform feature extraction on non-sequential behavior data to obtain non-sequential behavior features of the non-sequential behavior data, and encode the non-sequential behavior features using an encoder to obtain object recommendation results determined using the non-sequential behavior data. It should be noted that when sequential behavior data or non-sequential behavior data is input into the data processing model, the data processing model can use the first data processing module or the second data processing module for processing to obtain object recommendation results for the user. When sequential behavior data and non-sequential behavior data are input into the data processing model, the data processing model can use the first data processing module and the second data processing module for processing, and the object recommendation results output by the first data processing module and the second data processing module are both used as the object recommendation results for the user.

[0093] The data processing result can be understood as the result of the data processing model processing the sequential behavior data and / or the non-sequential behavior data. When the data processing method provided in the embodiment of the present disclosure is applied to different scenarios, the data processing result is also different. For example, when the data processing method is applied to the product recommendation scenario, the data processing result is a product recommendation result. The product recommendation result can be the product information of the recommended product that needs to be recommended to the user. For example, the product information can include: product type: skin care products, product price: 100-150 yuan, product brand, product capacity, etc. The product information can be used to recommend products to the user. For another example, when the data processing method is applied to the scenic spot recommendation scenario, the data processing result is a scenic spot recommendation result. The scenic spot recommendation result can be the scenic spot information of the recommended scenic spot that needs to be recommended to the user. For example, the scenic spot information can include: scenic spot type: farmhouse, scenic spot price: 100-150 yuan, scenic spot region, scenic spot name, etc. The scenic spot information can be used to recommend scenic spots to the user.

[0094] The sequential sample can be understood as the sequential behavior data in the user behavior data of the sample user, that is, the sequential sample can be understood as the sequential behavior data used as a training sample. For example, the sequential sample can be the interaction record between the user and the product as a training sample, and the interaction record can be composed of three time series features consisting of the product ID, the product category ID, and the user behavior sequence, wherein different time series features are one-to-one corresponding within the same time period. The non-sequential sample can be understood as the sequential behavior data in the user behavior data of the sample user, that is, the non-sequential sample can be understood as the non-sequential behavior data used as a training sample. For example, user attribute data such as user age and user gender as training samples, and product attribute data such as product price and product capacity. In one or more embodiments provided in the present disclosure, the non-sequential behavior data can be non-sequential behavior features; for example, the non-sequential behavior features can be user attribute features such as user age and user gender, and product attribute features such as product price and product capacity. The non-sequential sample features can be understood as sample features extracted from non-sequential samples and enhanced; these non-sequential sample features are used to train the data processing model. The sequential sample features can be understood as sample features extracted from sequential samples and enhanced; these sequential sample features are used to train the data processing model.

[0095] Specifically, the data processing method in the embodiment of the present disclosure, after receiving the user behavior data of the user, will determine the sequential behavior data and / or non-sequential behavior data from the user behavior data; then input the sequential behavior data and / or non-sequential behavior data into the data processing model, thereby obtaining the data processing results corresponding to the user. In one or more embodiments provided in the present disclosure, inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user includes: inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model; performing feature extraction on the sequential behavior data using a first data processing module in the data processing model to obtain sequential behavior features of the sequential behavior data, and encoding the sequential behavior features using an encoder in the first data processing module to obtain a first data processing result, and / or performing feature extraction on the non-sequential behavior data using a second data processing module in the data processing model to obtain non-sequential behavior features of the non-sequential behavior data, and encoding the non-sequential behavior features using an encoder in the second data processing module to obtain a second data processing result; and determining the data processing result corresponding to the user based on the first data processing result and / or the second data processing result.

[0096] The data processing model in the data processing method provided by the present disclosure can process sequential behavior data through a first data processing module, and / or process non-sequential behavior data through a second data processing module, thereby fully considering the influence of sequential behavior data and non-sequential behavior data on the data processing model in determining the data processing results, thereby obtaining more accurate data processing results and meeting the needs of actual application scenarios.

[0097] In one or more embodiments provided by the present disclosure, after determining the data processing result corresponding to the user, the data processing result can be sent to the user.

[0098] In one or more embodiments provided by the present disclosure, to address problems encountered during model training, the data processing method provided by the present disclosure proposes a data processing model for full-feature contrastive learning, which can simultaneously consider both sequential and non-sequential data features. Furthermore, improvements are made to the model's training speed prediction and cost, thereby optimizing the model from the perspectives of training time and training cost. While improving the model's prediction effect, it also provides assistance for actual recommendation scenario predictions, meeting the needs of actual application scenarios. Specifically, during the training process for this data processing model, the data processing method provided by the present disclosure takes into account that existing related work often focuses on innovation in model structure and ignores optimizing model prediction results from a data perspective. For example, contrastive learning models in recommendation scenarios often study sequential features and non-sequential features separately, which is inconsistent with actual application scenarios. Based on this, the data processing method provided by the present disclosure provides a data set expansion method based on self-supervision to optimize the model's prediction effect from the perspective of sample data. In addition, in terms of model structure, the data processing method provided by the embodiments of the present disclosure can simultaneously provide a contrastive learning model for sequential and non-sequential features, thereby meeting the needs of actual application scenarios. Specifically, the training process of the data processing model includes: before inputting the sequential behavior data and / or the non-sequential behavior data into the data processing model and obtaining the data processing result corresponding to the user, the training process also includes:

[0099] Determining the sequential samples and the non-sequential samples for a data processing model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users;

[0100] Inputting the sequential samples and the non-sequential samples into the data processing model to be trained, performing feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtaining initial sequential sample features and initial non-sequential sample features;

[0101] Performing feature enhancement on the initial sequential sample feature to obtain the sequential sample feature, and performing feature enhancement on the initial non-sequential sample feature to obtain the non-sequential sample feature;

[0102] Based on the sequential sample features and the non-sequential sample features, the data processing model to be trained is trained to obtain the data processing model.

[0103] The data processing model to be trained can be understood as a data processing model that needs to be trained;

[0104] Specifically, the data processing method provided by the present disclosure needs to train the data processing model before inputting the sequential behavior data and / or the non-sequential behavior data into the data processing model to obtain the data processing results corresponding to the user. The training steps may be: determining the sequential samples and non-sequential samples required for model training for the data processing model to be trained; wherein the sequential samples and non-sequential samples are the user behavior data of the sample user; that is, the sequential behavior data contained in the user behavior data of the sample user is used as a sequential sample, and the non-sequential behavior data is used as a non-sequential sample.

[0105] After obtaining the sequential and non-sequential samples, the sequential and non-sequential samples are input into the data processing model to be trained. In the data processing model to be trained, feature extraction is performed on the sequential samples to obtain initial sequential sample features, and feature extraction is performed on the non-sequential samples to obtain initial non-sequential sample features.

[0106] After obtaining the initial sequential sample features and the initial non-sequential sample features, they need to be enhanced. For example, the initial sequential sample features can be enhanced by using vector masking, vector clipping, or vector reordering to obtain sequential sample features. The initial non-sequential sample features can also be enhanced by using feature masking to obtain non-sequential sample features.

[0107] After obtaining the features of the sequential and non-sequential samples, the data processing model to be trained is trained using the features until the training stop condition is met, thereby obtaining a trained data processing model. Subsequently, the sequential and / or non-sequential behavioral data are processed based on this data processing model to meet the needs of practical applications.

[0108] It should be noted that the above is the training step for training the data processing model in this embodiment. The training step belongs to the same concept as the steps in the data processing model training method in one or more embodiments of the present disclosure. For details not described in detail in the training step, please refer to the corresponding or corresponding description in one or more embodiments of the present disclosure, and no further details will be given here.

[0109] In one or more embodiments provided herein, during model training, the data processing method provided herein takes into account the significant impact of sample size on model training efficiency. To improve the performance of the trained data processing model, the data processing method provided herein expands the sample data, thereby increasing the sample size and further ensuring the performance of the trained data processing model. The specific method is as follows.

[0110] The step of enhancing the initial sequence sample features to obtain the sequence sample features includes:

[0111] Determining similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample feature, and performing sample expansion based on the user behavior objects and the similar behavior objects to obtain an expanded sequence-type sample feature;

[0112] Feature enhancement is performed on the initial sequence sample features and the expanded sequence sample features to obtain the sequence sample features of the data processing model to be trained.

[0113] The user behavior object can be understood as the behavior object targeted by the user behavior. For example, when the data processing method is applied to a product recommendation scenario, the user behavior object can be understood as the target product corresponding to the user behavior, that is, the product corresponding to the user's browsing, purchasing, and adding behaviors. For another example, when the data processing method is applied to a scenic spot recommendation scenario, the user behavior object can be understood as the target scenic spot corresponding to the user behavior, that is, the scenic spot corresponding to the user's browsing, purchasing, and adding behaviors. For another example, when the data processing method is applied to a food recommendation scenario, the user behavior object can be understood as the target food corresponding to the user behavior, that is, the food corresponding to the user's browsing, purchasing, and adding behaviors. Similar behavior objects can be understood as behavior objects that are relatively similar to the user behavior object; for example, when the user behavior object is brand A facial cleanser, the similar behavior object can be brand B facial cleanser.

[0114] Specifically, after extracting features from sequence samples in the data processing model to be trained and obtaining initial sequence sample features, sample expansion can be performed on the initial sequence sample features. First, it is necessary to determine the user behavior objects in the initial sequence sample features. There can be at least two user behavior objects, and based on the similarity results between the user behavior objects, corresponding similar behavior objects are determined for each user behavior object; then, based on the user behavior objects and similar behavior objects, a sample expansion operation is performed to obtain expanded sequence sample features. Based on this, by determining the initial sequence sample features and the expanded sequence sample features as the sample features of the data processing model to be trained, the samples of the data processing model to be trained are expanded, thereby increasing the number of training samples.

[0115] After obtaining the initial sequence type sample features and the expanded sequence type sample features, feature enhancement is performed on the initial sequence type sample features and the expanded sequence type sample features, and the enhanced initial sequence type sample features and the enhanced expanded sequence type sample features are used as sequence type sample features.

[0116] The data processing model in the data processing method provided by one or more embodiments of the present disclosure is obtained by training with sequential sample features and non-sequential sample features. The sequential sample features and non-sequential sample features are obtained by performing feature enhancement on initial sequential sample features and initial non-sequential sample features obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model. Thus, the influence of sample data on model training is fully considered during the model training process, thereby avoiding the problem of poor model training efficiency caused by sample data. Moreover, the trained data processing model is used to process the sequential behavior data and / or non-sequential behavior data contained in the user behavior data, and a data processing result is obtained, thereby meeting the needs of practical applications through the data processing model.

[0117] Referring to Figure 3, Figure 3 shows a schematic diagram of a model structure in a data processing model training method provided according to an embodiment of the present disclosure, wherein the data processing model provided by the present disclosure is a full-feature contrast learning model, which is a multi-task learning model composed of a sequential feature contrast learning model, a non-sequential feature contrast learning model and a multi-layer perceptron model. The specific training process includes extracting features from the original data provided by the user (such as sequential samples and non-sequential samples) to obtain sequential features and non-sequential features. The non-sequential features are input into the non-sequential feature contrast learning model to obtain the model processing results, and the loss value of the non-sequential feature contrast learning model is calculated based on the model processing results. The non-sequential features are input into the embedding layer in the multi-layer perceptron model, the non-sequential feature vector of the non-sequential features, the non-sequential feature vector is input into the multi-layer perceptron in the multi-layer perceptron model to obtain the processing results, and the loss value of the multi-layer perceptron model is calculated based on the processing results. Data augmentation is performed on sequential features and then fed into a sequential feature contrastive learning model for processing. The model processing results are then used to calculate the sequential feature contrastive learning model loss. The model is trained using these three loss values ​​until the training termination criteria are met. This fully accounts for the impact of sample data during model training, avoiding issues with poor model training efficiency caused by sample data.

[0118] Referring to FIG4 , FIG4 shows a flow chart of a data processing model training method provided according to an embodiment of the present disclosure, which specifically includes the following steps.

[0119] Step 402: Determine sequential samples and non-sequential samples for training a data processing model, wherein the sequential samples and the non-sequential samples are user behavior data of sample users.

[0120] In the data processing model training method provided by the present disclosure, the user behavior data of the sample user can be a data set, which includes sequential behavior data and non-sequential behavior data; after obtaining the data set provided by the sample user, the sequential behavior data and the non-sequential behavior data are determined from the data set; then, feature extraction is performed on the sequential behavior data and the non-sequential behavior data used as training samples to obtain sequential sample features and non-sequential sample features.

[0121] Step 404: Input the sequential samples and the non-sequential samples into the data processing model to be trained, perform feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtain initial sequential sample features and initial non-sequential sample features.

[0122] In one or more embodiments provided in the present disclosure, the sequential sample may be sequential behavior data used as a training sample, and the non-sequential sample may be non-sequential behavior data used as a training sample. After determining the sequential sample and the non-sequential sample, the sequential sample and the non-sequential sample will be input into the data processing model to be trained. Using the feature extraction module in the data processing model to be trained, feature extraction is performed on the sequential sample to obtain initial sequential sample features, and feature extraction is performed on the non-sequential sample to obtain initial non-sequential sample features. Among them, the feature extraction module can be understood as a module for feature extraction in the data processing model to be trained, and the feature extraction module can be one or more network layers, or the feature extraction module can be a neural network model for feature extraction in the data processing model to be trained. Subsequently, the initial sequential behavior sample features and the initial non-sequential behavior sample features are used to perform model training on the data processing model to be trained.

[0123] Step 406: performing feature enhancement on the initial sequential sample features to obtain the sequential sample features, and performing feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features.

[0124] The feature enhancement of the initial sequence sample feature to obtain the sequence sample feature includes:

[0125] Determining similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample feature, and performing sample expansion based on the user behavior objects and the similar behavior objects to obtain an expanded sequence-type sample feature;

[0126] Feature enhancement is performed on the initial sequence sample features and the expanded sequence sample features to obtain the sequence sample features of the data processing model to be trained.

[0127] The expanded sequence sample feature can be understood as sample data obtained by expanding the user behavior element in the initial sequence sample feature and the similar behavior element corresponding to the user behavior element.

[0128] Specifically, after extracting features from the sequence samples in the data processing model to be trained and obtaining the initial sequence sample features, the initial sequence sample features can be expanded. First, it is necessary to determine at least two user behavior objects in the initial sequence sample features, and determine corresponding similar behavior objects for each user behavior object through the similarity results between the at least two user behavior objects; then, based on the user behavior objects and the similar behavior objects, perform sample expansion operations to obtain expanded sequence sample features. After obtaining the initial sequence sample features and the expanded sequence sample features, feature enhancement is performed on the initial sequence sample features and the expanded sequence sample features, and the enhanced initial sequence sample features and the enhanced expanded sequence sample features are used as sequence sample features. In this way, the samples of the data processing model to be trained are expanded and the number of training samples is increased.

[0129] It should be noted that in one or more embodiments provided herein, feature enhancement may be performed on both the initial sequence-based sample features and the expanded sequence-based sample features, wherein the steps for feature enhancement for both the initial sequence-based sample features and the expanded sequence-based sample features are identical. For the steps for feature enhancement for the expanded sequence-based sample features, please refer to the section on feature enhancement for the initial sequence-based sample features, and thus, redundant description is avoided.

[0130] In one or more embodiments provided by the present disclosure, due to problems such as data diversity and sparsity in some application scenarios, sample data may be missing or too little in actual application scenarios, which often limits the performance of the data processing model. Therefore, in response to the situation where there are too few data samples, the data processing model training method in one or more embodiments of the present disclosure proposes a self-supervised data expansion scheme, which is a scheme for expanding the original data set based on a self-supervised expansion method. It can expand the original data set from the perspective of data to optimize the prediction effect of the model. Specifically, in the process of expanding the sample data, it is necessary to determine the similar behavior objects corresponding to the user behavior objects in the initial sequence type sample features, and use the user behavior objects and the similar behavior objects to expand the samples to obtain expanded sequence type sample features. The specific method is as follows. The determination of similar behaviors corresponding to user behaviors in the initial sequence type includes steps one to three:

[0131] Step 1: Determine at least two user behavior objects in the initial sequence sample features.

[0132] It should be noted that each initial sequence sample feature contains at least two user behavior objects.

[0133] Step 2: Determine similarity results between a target behavior object and other behavior objects among the at least two user behavior objects, wherein the target behavior object is any one of the at least two user behavior objects, and the other behavior objects are other user behavior objects among the at least two user behavior objects except the target behavior object.

[0134] Specifically, determining similar results between the target behavior object and other behavior objects in the at least two user behavior objects includes:

[0135] Constructing a user behavior object matrix based on the at least two user behavior objects;

[0136] Obtaining a similarity result matrix corresponding to the at least two user behavior objects by performing similarity calculation on the user behavior object matrix;

[0137] The similarity result matrix is ​​used to determine similarity results between the target behavior object and other behavior objects in the at least two user behavior objects.

[0138] Among them, the user behavior object matrix can be understood as a matrix composed of user behavior object matrices; taking the user behavior object as the product corresponding to the user behavior as an example, in the process of calculating the similarity between products, it is necessary to first create a co-occurrence matrix of products. It should be noted that there are at least two sample users in the present disclosure, and each sample user has a corresponding initial sequential sample feature and non-sequential sample feature. The co-occurrence matrix is ​​composed of products in the initial sequential sample features of multiple sample users. Then, for each sample user, the products in the list of products interacted by each sample user are added 1 in the co-occurrence matrix, thereby adjusting the co-occurrence matrix to obtain an adjusted co-occurrence matrix, which can be understood as a user behavior object matrix. For example, see Figure 5, which shows a flowchart of constructing a co-occurrence matrix in a data processing model training method provided according to an embodiment of the present disclosure. In the product list in Figure 5 , A, B, and D can be understood as different products, i.e., the user behavior objects. In the co-occurrence matrix in Figure 5 , A, B, C, D, and E can be understood as different products. Based on this, when calculating product similarity, we first determine the products included in each initial sequential sample feature and build a product list based on these products (that is, based on each user's initial sequential sample feature, we build a list of each user's favorite products). Secondly, for each user, we add 1 to the co-occurrence matrix for each pair of products in their product list. Specifically, taking product category 1 as an example, which contains three products, A, B, and D. We match these three products pairwise, adding a value of 1 to the intersection of products A and B in co-occurrence matrix 1; adding a value of 1 to the intersection of products A and D; and adding a value of 1 to the intersection of products B and D. This completes the operation of adding 1 to the co-occurrence matrix for each pair of products in product list 1, and determines the corresponding co-occurrence matrix 1 for product list 1. In the subsequent similarity calculation process, these co-occurrence matrices can be added together to obtain a target matrix containing all user behavior objects. Finally, the target matrix is ​​normalized to obtain the cosine similarity matrix between products.

[0139] The similarity result matrix can be understood as a matrix used to represent the similarity results between the target behavior object and the other behavior objects in at least two user behavior objects; for example, the similarity result matrix can be a cosine similarity matrix. Taking the user behavior objects as the products corresponding to the user behavior as an example, when calculating the similarity between products, after obtaining the adjusted co-occurrence matrix, the co-occurrence matrices are added together and the added matrix is ​​normalized to obtain the cosine similarity matrix between the products.

[0140] The similarity result can be understood as data representing the degree of similarity between two user behavior objects; for example, the similarity result can be a similarity, which can be any value in the range [0, 1]. When the similarity is closer to 1, the two user behavior objects are more similar, and when the similarity is closer to 0, the two user behavior objects are less similar.

[0141] Specifically, the data processing model training method provided by the present disclosure can construct a user behavior object matrix based on at least two user behavior objects; then, by performing similarity calculation on the user behavior object matrix, a similarity result matrix corresponding to at least two user behavior objects is obtained; for example, performing similarity calculation on the user behavior object matrix can be understood as adding the user behavior object matrices and normalizing the matrix obtained by addition to obtain a cosine similarity matrix. For another example, performing similarity calculation on the user behavior object matrix can also be understood as adding the user behavior object matrices to obtain a target matrix, and calculating the target matrix for Euclidean distance calculation to obtain a Euclidean distance matrix (i.e., a similarity result matrix) containing the Euclidean distance calculation result. The Euclidean distance value of the calculation result represents the similarity between the two user behavior objects; the larger the Euclidean distance value, the smaller the similarity, and the smaller the Euclidean distance value, the greater the similarity. Determine the similarity results between the target behavior object and other behavior objects in at least two user behavior objects from the similarity result matrix. Therefore, based on the similarity results, similar behavior objects that are relatively similar to each user behavior object can be accurately determined, which facilitates the subsequent sample expansion based on relatively similar similar behavior objects, while increasing the number of sequential sample features and improving the performance of sequential sample features.

[0142] The data processing model training method provided by this disclosure is illustrated using its application in a product recommendation scenario as an example. For situations where sequential feature data may have a limited number of samples, a replacement data augmentation method based on collaborative filtering is proposed to augment the original data. For ease of explanation, assume that a dataset contains three sequential features: product ID, product category ID, and user behavior sequence. These three sequential features constitute an initial sequential sample feature. First, a correspondence table between user IDs and product IDs is created. Next, a relationship matrix between different products is constructed. Using the cosine similarity calculation method, the similarity between products is calculated to obtain a similarity result matrix, thereby determining similarity results. Specifically, the user behavior object can be the product corresponding to the user behavior. Based on this, in the process of calculating the similarity between products, a product co-occurrence matrix is ​​first created. Then, for each user, the co-occurrence matrix is ​​normalized by adding 1 to the co-occurrence matrix for each product in the product list they interacted with. The product list refers to the correspondence list between users and products.

[0143] It should be noted that in one or more embodiments provided by the present disclosure, it is also necessary to construct a relationship table between sample users and user behavior objects. For example, a relationship table between sample users and commodities is created. Based on the interaction information between users and commodities in the original data set (i.e., the initial sequence sample features), a corresponding list of user IDs and commodity IDs is created, and the corresponding list contains all commodity IDs that each user has interacted with. Among them, the interaction between sample users and commodities in one or more embodiments provided by the present disclosure refers to the cross-operation behavior of sample users on commodities, such as browsing commodities, purchasing commodities, and collecting commodities.

[0144] Step 3: Determine a corresponding similar behavior object for the target behavior object based on the similarity results.

[0145] In one or more embodiments provided in the present disclosure, the similarity result is similarity;

[0146] The determining a corresponding similar behavior object for the target behavior object based on the similarity result includes:

[0147] Determining a target similarity greater than or equal to a preset similarity threshold from the similarities between the target behavior object and the other behavior objects;

[0148] Determine other behavior objects corresponding to the target similarity as at least two candidate similar behavior objects of the target behavior object;

[0149] Sort the at least two candidate similar behavior objects based on the target similarity to obtain a candidate similar behavior object sequence;

[0150] A preset first number of the candidate similar behavior objects are selected from the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0151] The preset similarity threshold can be set according to the actual application scenario, and this disclosure does not impose any specific restrictions on this. For example, when the similarity is any value in the interval [0, 1], the preset similarity threshold can be a value of 0.7. Candidate similar behavior objects can be understood as user behavior objects that are relatively similar to the target behavior object among other behavior objects. The preset first number can be set according to the actual application scenario, and this disclosure does not impose any specific restrictions on this. For example, the preset first number can be 10 or 5.

[0152] The process of selecting a preset first number of candidate similar behavior objects from the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object can be understood as selecting the top 10 or top 5 candidate similar behavior objects in the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0153] It should be noted that, when ranking at least two candidate similar behavior objects based on target similarity, it can be understood that the candidate similar behavior object with greater target similarity is ranked at the front, and the candidate similar behavior object with smaller target similarity is ranked at the back.

[0154] Continuing with the above example, where the target behavior object is the target product and the other behavior objects are other products, after calculating the similarity between the products using cosine similarity to obtain a similarity matrix, the similarity between the target product and the other products can be determined. After obtaining the similarity, it is necessary to compare this similarity with the preset similarity threshold of 0.7 to determine a target similarity greater than or equal to the preset similarity threshold of 0.7. This target similarity indicates that the target product is relatively similar to the other products. Based on the other products corresponding to the target similarity, a set of similar products to the target product is constructed. Finally, the target product similarity is used to sort the top five other products as the set of similar products to the target product, thereby accurately determining the similar behavior objects for the target behavior object. The target product is any product included in the product ID sequence feature. In other words, a similar product set is determined for each product included in the product ID sequence feature. Other products are products other than the target product in the product ID sequence feature.

[0155] In one or more embodiments provided by the present disclosure, if no similar product set is obtained after collaborative filtering calculation, the category to which the target product belongs is found, and the products ranked in the top 20% of frequency in the category are selected as the similar items set of the target product, so as to facilitate subsequent sample expansion based on similar behavior objects. Specifically, the similarity result is the similarity degree;

[0156] Accordingly, determining a corresponding similar behavior object for the target behavior object based on the similarity result includes:

[0157] The determining a corresponding similar behavior object for the target behavior object based on the similarity result includes:

[0158] When the similarity between the target behavior object and the other behavior objects is less than a preset similarity threshold, determining the object type information of the target behavior object;

[0159] Determine, from the at least two user behavior objects, at least two candidate similar behavior objects corresponding to the object type information, and determine the object quantity of each candidate similar behavior object;

[0160] Sort the at least two candidate similar behavior objects based on the number of objects to obtain a candidate similar behavior object sequence;

[0161] A preset second number of the candidate similar behavior objects are selected from the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0162] The object type information can be understood as information representing the type of the user behavior object. For example, if the user behavior object is the product corresponding to the user behavior, the object type information can be the product type. For example, the product type can be facial cleanser type, jacket type, marker type, etc.

[0163] The preset second number can be set according to actual application scenarios, and the present disclosure does not specifically limit this. For example, the preset second number can be 20% or the first 30%.

[0164] From the candidate similar behavior object sequence, a preset second number of candidate similar behavior objects are selected from top to bottom as similar behavior objects corresponding to the target behavior object. This can be understood as selecting the top 20% of the candidate similar behavior objects in the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0165] Continuing with the above example, after calculating the similarity between products using the cosine similarity calculation method to obtain a similarity matrix, the similarity between the target product and other products can be determined. After obtaining the similarity, it is necessary to compare the similarity with the preset similarity threshold of 0.7. When the similarities are all less than the preset similarity threshold of 0.7, determine the product type of the target product, and determine the same type of products corresponding to the product type from other products. Use the same type of products as candidate similar products for the target product. Then determine the number of times each product in the candidate similar products appears, that is, the number of products for each candidate similar product in the candidate similar products. Use the number of products to sort the candidate similar products, and use the top 20% of the candidate similar products as the similar product set for the target product; thereby accurately determining similar behavior objects for the target behavior object.

[0166] In one or more embodiments provided by the present disclosure, the target behavior object may be the user behavior object with the earliest interaction time in the initial sequence-type sample feature; that is, the target behavior object may be the user behavior object at the first position in the sequence in the initial sequence-type sample feature. The interaction time of the target behavior object is earlier than the interaction time of other user behavior objects in at least two user behavior objects. For example, the target behavior object is a target product, which is the product with the earliest time point in the time series-type feature. That is to say, the target product ranked first in the time series-type feature. Therefore, for the earliest product ID in the time series-type feature, a set of similar products similar to it is determined.

[0167] In one or more embodiments provided herein, when a dataset has a small number of samples, a dataset expansion method based on self-supervision is proposed to expand the original dataset, thereby improving the model's prediction performance from the data itself. During the sample data expansion process, after determining the similar behavior objects corresponding to the user behavior object, the user behavior object and the similar behavior objects need to be sampled. The specific steps are as follows: the number of initial sequence-type sample features is at least two;

[0168] The sample expansion based on the user behavior object and the similar behavior object to obtain the expanded sequence sample features includes:

[0169] Determining user behavior object information of each initial sequence-type sample feature in at least two initial sequence-type sample features;

[0170] Based on the user behavior object information, determining a sequence sample feature to be expanded from the at least two initial sequence sample features;

[0171] Based on the similar behavior objects, the user behavior objects in the sequence type sample features to be expanded are processed to obtain expanded sequence type sample features.

[0172] The user behavior object information may include the number of user behavior objects included in the initial sequence-type sample feature, or the object sizes of multiple user behavior objects included in the initial sequence-type sample feature. Subsequently, based on the number of objects or the object sizes, a sequence-type sample feature to be expanded may be determined from at least two initial sequence-type sample features.

[0173] Specifically, the number of user behavior objects in each of at least two initial sequence type sample features is determined; based on the comparison of the number of user behavior objects with a preset object number threshold, the initial sequence type sample feature in which the number of user behavior objects is greater than the preset object number threshold among the at least two initial sequence type sample features is used to determine the sequence type sample feature to be expanded; based on similar behavior objects, the user behavior objects in the sequence type sample feature to be expanded are processed to obtain the expanded sequence type sample feature; thereby, the original data set is expanded, and the prediction effect of the model is improved from the data itself level.

[0174] Continuing with the above example, where user behavior object information refers to the number of objects of user behavior objects, based on this, first, it is necessary to determine the number of product IDs contained in each serial behavior feature, and determine the product ID time series feature with a number of product IDs greater than 10 as the serial behavior feature to be expanded. It should be noted that the number of product IDs also represents the length of the product ID time series feature. If the number of product IDs is 10, the length of the product ID time series feature is also 10. Secondly, from the product ID time series features with a length greater than 10, determine the product ID (target product) with the earliest appearance time; and determine the set of similar products corresponding to the product ID. Based on this set of similar products, the product ID time series feature is expanded to obtain an expanded serial sample feature.

[0175] In one or more embodiments provided in the present disclosure, processing the user behavior objects in the sequence-type sample features to be expanded based on the similar behavior objects to obtain the expanded sequence-type sample features includes:

[0176] Determining a behavior object to be expanded and object type information to be expanded corresponding to the behavior object to be expanded from at least two user behavior objects included in the sequence sample feature to be expanded;

[0177] Selecting a target similar behavior object from similar behavior objects corresponding to the behavior object to be expanded, and determining target object type information of the target similar behavior object;

[0178] The target similar behavior object is used to replace the behavior object to be expanded in the sequence type sample feature to be expanded, and the target object type information is used to replace the object type information to be expanded corresponding to the behavior object to be expanded in the sequence type sample feature to be expanded, so as to obtain the expanded sequence type sample feature.

[0179] The behavior object to be expanded can be understood as the user behavior object to be expanded, the object type information to be expanded can be understood as the object type information of the behavior object to be expanded, and the target object type information can be understood as the object type information of the target similar behavior object.

[0180] Specifically, a behavior object to be expanded is determined from at least two user behavior objects included in the sequence-type sample feature to be expanded. The behavior object to be expanded can be understood as the user behavior object located at the first sequence position in the sequence-type sample feature to be expanded, or the behavior object to be expanded can be understood as the user behavior object generated earliest in the sequence-type sample feature to be expanded.

[0181] After determining the behavior object to be expanded, the type information of the object to be expanded corresponding to the behavior object to be expanded is determined; then, a target similar behavior object is selected from the similar behavior objects corresponding to the behavior object to be expanded, and the target object type information of the target similar behavior object is determined; the target similar behavior object can be any similar behavior object among the similar behavior objects. Alternatively, the target similar behavior object can be a similar behavior object among the similar behavior objects that has a greater degree of similarity with the behavior object to be expanded.

[0182] By replacing the behavior object to be expanded in the sequence sample feature to be expanded with the target similar behavior object, and by replacing the object type information to be expanded corresponding to the behavior object to be expanded in the sequence sample feature to be expanded with the target object type information, the expanded sequence sample feature is obtained; thereby expanding the original data set and improving the prediction effect of the model from the data level itself.

[0183] Continuing with the above example, FIG6 is a schematic diagram of data expansion in a data processing model training method provided by an embodiment of the present disclosure; referring to FIG6, the data 1 contained in the original data set can be understood as the initial sequence sample feature, and the original data set can contain multiple data. The initial sequence sample feature includes a product ID feature (i.e., a product ID time series feature), a category ID feature (i.e., a category ID time series feature), and a user behavior sequence feature. Based on this, after determining a set of similar products, a similar product ID is randomly selected from the set of similar products, and the target product in the product ID time series feature is replaced with the similar product ID. Then, the category ID time series feature and the user behavior sequence feature corresponding to the product ID time series feature are determined. For the product category ID time series feature, the category ID of the target product is replaced with the category ID corresponding to the similar product; the user behavior sequence feature remains unchanged, thereby constructing a new sequence sample feature (i.e., data 2 in the data set after data expansion in FIG6) based on the original user behavior record.

[0184] Step 406: performing feature enhancement on the initial sequential sample features to obtain the sequential sample features, and performing feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features.

[0185] In one or more embodiments provided herein, the data processing model training method of the present disclosure proposes a full-feature contrast learning model that simultaneously considers both sequential and non-sequential data features, innovates the data enhancement methods of existing models, and optimizes existing models from the perspectives of training time and training cost, surpassing the prediction effects of existing models while also meeting the needs of practical application scenarios. Specifically, the feature enhancement of the initial sequential sample features to obtain the sequential sample features includes:

[0186] Inputting the initial sequence sample features into the first data processing module in the data processing model to be trained;

[0187] Processing the initial sequence sample features using a feature processing unit in the first data processing module to obtain sequence sample features to be enhanced;

[0188] A target data enhancement rule is determined from at least two data enhancement rules, and the target data enhancement rule is used to perform data enhancement on the sequence sample feature to be enhanced to obtain the sequence sample feature.

[0189] The first data processing module can be understood as a module in the data processing model to be trained that processes features of sequential samples. The first data processing module can be a data processing model, for example, a sequential contrastive learning model. Alternatively, the first data processing module can be one or more network layers that process features of sequential samples.

[0190] The feature processing unit can be understood as an input layer or embedding layer integrated in the first data processing module, which is used to process the sequence sample features. The sequence sample features to be enhanced can be understood as the features obtained by inputting the sequence sample features to be enhanced into the feature processing unit.

[0191] The data enhancement rule can be set according to the actual application scenario, and the present disclosure does not impose specific restrictions on this. For example, the data enhancement rule can be a way of performing vector enhancement (i.e., feature enhancement). For example, the vector enhancement operation includes a vector masking operation, a vector clipping operation, and a vector reordering operation. Among them, the vector masking operation sets a learnable tensor, which can be used to perform a masking operation on the feature vector (i.e., the initial sequence sample feature, the expanded sequence sample feature); the vector clipping operation can perform a random clipping operation on the feature vector; and the vector reordering operation can perform a reordering operation on the feature vector. In one or more embodiments provided in the present disclosure, the data enhancement operation can be implemented by a data enhancement module in a data processing model. The data enhancement module can be understood as one or more network layers in a data processing model; or, the data enhancement module can be understood as a sub-model in a data processing model for performing data enhancement. Based on the data processing module, the three vector enhancement schemes provided in the present disclosure perform data enhancement after the original data passes through the embedding layer (embedding layer), thereby maintaining the intrinsic correlation between users and products in historical behaviors and improving the performance of the enhanced sample data.

[0192] The target data enhancement rule can be understood as a data enhancement rule used to enhance the features of the sequence sample to be enhanced among the at least two data enhancement rules. The target data enhancement rule can be set according to the actual application scenario. The target data enhancement rule can be one or at least two.

[0193] Continuing with the above example, FIG7 is a schematic diagram of a sequence contrast learning model training method in a data processing model training method provided by an embodiment of the present disclosure. Referring to FIG7 , the records in FIG7 can be understood as initial sequence sample features. For the initial sequence sample features, the sequence contrast learning model splices the sequence features to obtain spliced ​​sequence features; the spliced ​​sequence features are input into the input layer of the EasyRec (i.e., sequence contrast learning model) integration, and the corresponding feature vector representation (i.e., sequence sample features to be enhanced) is obtained through the input layer of the EasyRec integration; finally, a vector enhancement operation is performed on the feature vector to obtain a positive sample pair in contrast learning (i.e., the above-mentioned sequence sample features), thereby innovating the data enhancement method of the existing model and optimizing the existing model from the perspective of training time and training cost, while surpassing the prediction effect of the existing model, it also meets the needs of actual application scenarios.

[0194] Based on Figure 7, it can be seen that all the sequence sample feature pairs obtained after feature enhancement are subsequently input into the encoder in the sequence contrastive learning model. According to the results output by the encoder, the loss function is used to calculate and obtain the loss value of the sequence contrastive learning function.

[0195] It should be noted that the encoder of the sequential contrastive learning model in Figure 7 shares parameters with the encoders in other models (the encoder in the multi-layer perceptron model), which can be understood as obtaining a vector representation of the sequence features. Different task encoders are different, and the encoder is mainly used to obtain a vector representation of the features. The shared parameters in Figure 7 are for the contrastive learning task on the right and the CTR task on the left (optional). The same encoder is required to generate the vector representation of the sequence features (for example, both use transfomer). The shared parameters are to control the vector representations generated by the sequence features on both sides to be the same.

[0196] In one or more embodiments provided herein, a full-feature contrastive learning model is proposed that optimizes the data augmentation methods of existing models, further optimizing existing models in terms of training time and training cost, thereby surpassing the prediction performance of existing models. Specifically, determining a target data augmentation rule from at least two data augmentation rules, and using the target data augmentation rule to perform data augmentation on the sequence sample features to be enhanced to obtain the sequence sample features includes:

[0197] Selecting two target data enhancement rules from a preset first data enhancement rule, a preset second data enhancement rule, and a preset third data enhancement rule;

[0198] Performing data enhancement on the sequence-type sample feature to be enhanced using the two target data enhancement rules to obtain a first enhanced sequence-type sample feature and a second enhanced sequence-type sample feature;

[0199] The first enhanced sequence-type sample feature and the second enhanced sequence-type sample feature are used as the sequence-type sample feature.

[0200] Among them, the preset first data enhancement rule, the preset second data enhancement rule, and the preset third data enhancement rule can be set according to the actual application scenario, and the present disclosure does not impose specific restrictions on this. For example, the first data enhancement rule can be a vector mask operation, the preset second data enhancement rule can be a vector clipping operation, and the preset third data enhancement rule can be a vector reordering operation. Based on this, selecting two data enhancement rules can be understood as randomly selecting two data enhancement rules from the preset first data enhancement rule, the preset second data enhancement rule, and the preset third data enhancement rule. Moreover, randomly selecting two data enhancement rules can be selecting two identical data enhancement rules. For example, randomly selecting the preset first data enhancement rule to perform two data enhancements on the target sequence behavior vector twice.

[0201] Continuing with the above example, FIG8 is a schematic diagram of sequence feature data enhancement in a data processing model training method provided by an embodiment of the present disclosure. Based on FIG8 , three data enhancement schemes, namely, vector masking, vector clipping, and vector reordering, provided by the data processing model training method of the present disclosure are introduced in detail.

[0202] Among them, the vector mask can be understood as item mask, which is a technology based on the BERT model applied in natural language processing. In the BERT model, it is called zero-masking, which is used to extract key information from the text and perform tasks such as text classification and named entity recognition. This product masking technology can solve the problems of missing contextual information, ambiguity and context in the text. The first data processing module (such as the CL4SRec model) in the data processing model training method provided by the present invention randomly masks a certain proportion of items in the historical sequence of each user (for example, replacing them with 0), but this will destroy the intrinsic attributes of the product to a certain extent.

[0203] Based on this, in order to avoid the problem that the product shielding operation will destroy the intrinsic properties of the product, the data processing model training method provided in this disclosure is improved. The specific improvement steps are:

[0204] First, for each user historical behavior sequence s u , which contains I=(i1,i2,…,i n )n items (i represents the feature set at a certain moment in the user's history, with a size of m. In this dataset, i represents the three features of product ID, product category ID, and behavior type). u After the embedding layer, we get V=(υ1,υ2,…,υ n ), the dimension of υ is n*E, E is the embedding size, and its calculation formula is shown in the following formula (1):

[0205] Secondly, randomly block a certain proportion of items aug mask =(index1,index2,...,index mask ), where L mask =p mask *n,L mask refers to the length of the sequence after masking; p mask refers to the ratio of masking the sequence length; aug mask Indicates the high-dimensional vector index selected to be masked in the sequence. Index refers to the high-dimensional vector index, which will be replaced by a learnable vector [υector mIn summary, the calculation formulas of the product shielding algorithm are shown in formula (2) and formula (3):

[0206] in, Refers to the user's historical behavior sequence s after vector masking u .

[0207] The types of products a user browses are likely to remain stable over time. For example, if a user wants to buy a facial cleanser, they will likely click on multiple brands or different merchants to decide which one to buy. Therefore, it can be seen that the user's historical interactions with products are likely to be similar over time. Therefore, through this product masking algorithm, two different views of the same user's historical sequence obtained through this algorithm can still retain the user's purchase intention. At the same time, the learnable variables allow the model to learn and discover more intrinsic connections between users and products.

[0208] Among them, vector cropping can be understood as item cropping. Item cropping is a technology widely used in various image recognition and analysis tasks in computer vision. Its purpose is to crop out unnecessary areas in the image, thereby focusing on important areas and improving the efficiency and accuracy of image processing. The first data processing module (such as the CL4SRec model) in the data processing model training method provided in the present disclosure randomly crops the original data of each user; and in the data processing model training method provided in the present disclosure, specifically, the original data is first passed through the embedding layer and then randomly cropped to complete data enhancement. The specific execution steps are:

[0209] First, for each user historical behavior sequence s u After the embedding layer, we get V=(υ1,υ2,...υ n ), which have the same meanings as the parameters described in the product mask operation and are not described in detail here.

[0210] Secondly, randomly cut a certain proportion of items aug crop =(index1,index2,...,index crop ), where L crop =p crop *n,L crop refers to the length of the sequence after pruning, p crop refers to the ratio of the length of the retained sequence; aug crop Indicates the index of the high-dimensional vector to be pruned in the sequence. cropThe index corresponding to the item contained in can be deleted to obtain the final data enhanced view. In summary, the calculation formula of the product clipping algorithm is shown in the following formula (4).

[0211] in, Refers to the user's historical behavior sequence s after vector clipping u .

[0212] The data augmentation achieved by cropping in this project can be demonstrated in two ways. First, it provides a partial view of a user's historical sequence. Without conducting a comprehensive analysis of the user, it enhances the user representation model by learning their general preferences. Second, in contrastive learning algorithms, if the views obtained after two cropping steps do not intersect, this can be treated as a prediction task for a future point in time, allowing the model to predict future changes in user purchasing preferences.

[0213] Vector reordering can be understood as item reordering. Sequential recommendation tasks often predict the data for the next time point based on the passage of time, assuming that items interacting with users are sequentially related. However, due to complex external environmental factors, the order in which users interact with items can be flexible. Therefore, item reordering is used for data augmentation.

[0214] Based on this, for the initial sequence sample features, after inputting them into the sequence contrast learning model. First, the sequence contrast learning model splices the initial sequence sample features through the embedding layer to obtain the spliced ​​initial sequence sample features; the spliced ​​sequence features are input into the input layer of the model integration to obtain the corresponding feature vector (i.e., the sequence sample features to be enhanced); finally, the feature vector is subjected to two random data enhancement operations. The data enhancement operation can be the above-mentioned vector mask, vector clipping, or vector reordering. Both data enhancement operations are performed by randomly selecting data enhancement from the above three data enhancement methods. After data enhancement, two enhanced sequence sample features corresponding to the two data enhancement operations are obtained (i.e., positive sample pairs in contrast learning). Thereby, the data enhancement method of the existing model is optimized, and the existing model is further optimized from the perspectives of training time and training cost, surpassing the prediction effect of the existing model.

[0215] In one or more embodiments provided herein, the data processing model training method of the present disclosure proposes a full-feature contrast learning model that simultaneously considers both sequential data features and non-sequential data features, and innovates the data enhancement method of the existing model. By enhancing the features of non-sequential data features, the existing model is optimized from the perspective of training time and training cost, improving the model prediction effect while also meeting the needs of actual application scenarios. The specific method is as follows. The feature enhancement of the initial non-sequential sample features to obtain the non-sequential sample features includes:

[0216] Inputting the initial non-sequential sample features into a second data processing module in the data processing model to be trained;

[0217] Performing feature masking processing on the initial non-sequential sample features to obtain enhanced non-sequential sample features;

[0218] The enhanced non-sequential sample features are processed by a feature processing unit in the second data processing module to obtain the non-sequential sample features.

[0219] The second data processing module can be understood as a module in the data processing model to be trained that processes non-sequential sample features. The second data processing module can be one or more data processing models, for example, a sequential contrastive learning model and / or a multi-layer perceptron model. Alternatively, the second data processing module can be one or more network layers that process non-sequential sample features.

[0220] In the data processing model training method provided by the present disclosure, feature masking can be used to enhance the features of the initial non-sequential sample features. In one or more embodiments provided by the present disclosure, in the data processing model training method provided by the present disclosure, two data enhancement methods for the initial non-sequential sample features will be introduced. The data enhancement method can be a feature shielding (or called a feature mask, vector mask) method. Since non-sequential sample features generally do not depend on the time sequence, product reordering is no longer applicable, and product clipping is similar to feature shielding. Therefore, data shielding is improved here. Two feature shielding methods are provided in the data processing model training method provided by the present disclosure, and any one of them is subsequently determined as a data enhancement method. The specific two feature shielding methods are shown in Figure 9, which is a schematic diagram of two feature shielding methods in a data processing model training method provided by an embodiment of the present disclosure; specifically, the feature shielding method 1 on the left side of Figure 9 is the first feature shielding method.

[0221] The first feature shielding method treats each initial non-sequential sample feature as a whole and directly shields it. The shielding method is to directly set the feature to be shielded to a 0 vector. Specifically, assuming that the current non-sequential feature is F = (f1, f2, ..., f n ) There are n features in total. First, after the data input layer (embedding layer), V = (υ1, υ2, ..., υ n ), the dimension of υ is n*E, where E is the embedding layer size. Secondly, a certain proportion of items are randomly masked aug mask =(index1,index2,...,index mask ), where L mask =p mask *n,aug mask Indicates the index to be masked in the sequence. The vector will be replaced by a zero vector of size 1*E. In summary, the calculation formula of the first feature masking method is shown in the following formula (5) and formula (6):

[0222] in, It refers to the non-sequential features after feature masking using the first feature masking method.

[0223] Feature shielding method 2 on the right side of Figure 9 is the second feature shielding method. This second feature shielding method is similar to the model-level data enhancement method mentioned in the Duorec model. It mainly randomly shields the feature vector representation after the data input layer. The blocks framed by the dotted box in Figure 9 are the shielded information. The specific method is to multiply the original feature matrix with a Bernoulli random variable matrix to achieve random shielding of some feature information with a probability of p. In summary, the calculation formulas of the second feature shielding method are shown in the following formulas (7) and (8): B Bernoulli(p) formula (8).

[0224] in, It refers to the non-sequential features after feature masking using the second feature masking method. B Bernoulli(p) refers to the Bernoulli random variable matrix.

[0225] The feature processing unit can be understood as the input layer or embedding layer integrated in the second data processing module, which is used to process the enhanced non-sequential sample features. The encoding unit can be understood as the encoder in the non-sequential data processing module, and the encoding unit can adopt MaskNet.

[0226] It should be noted that the Transfomer encoder usually has a good effect when processing time series related features, but its training time is long and it is not very applicable when the number of features is large. Therefore, the data processing model training method provided by the present disclosure adopts MaskNet as the feature representation module (i.e., encoder) of the non-sequential data processing module (i.e., non-sequential contrastive learning model). Compared with Transfomer, MaskNet has a better performance in encoding large-scale features. Specifically, MaskNet proposes an instance-guided mask scheme, which uses element-wise product in the feature embedding layer and feedforward layer in the DNN. The instance-guided masking method can dynamically integrate global context information into the feature embedding layer and feedforward layer, and can highlight important features when the data set contains a large number of types and quantities. This shows that the model has good performance in actual application scenarios. Therefore, the data processing model training method provided by the present disclosure selects it as the encoder to constitute the feature representation module.

[0227] Continuing with the above example, FIG10 is a schematic diagram of a non-sequential contrastive learning model training method in a data processing model training method provided by an embodiment of the present disclosure. Referring to FIG10 , for the non-sequential features in the data set (i.e., the initial non-sequential sample features), they are input into the non-sequential contrastive learning model. First, a data enhancement operation is performed on the non-sequential features, and the data enhancement operation selects a feature masking method, that is, a number of non-sequential features are randomly masked with the features as a whole. Secondly, the enhanced non-sequential features are subjected to the input layer of the EasyRec integration to obtain the corresponding data-enhanced feature vector representation (i.e., the non-sequential sample features).

[0228] As shown in Figure 10, after obtaining the features of non-sequential samples, they are input into the encoder (MaskNet can be used as the encoder because it can highlight important features and has a faster training speed). Finally, based on the results output by the encoder, the loss value is calculated using the loss function. L2_loss can be used as the loss function.

[0229] It should be noted that the other models included in Figure 10 can be understood as the first data processing module and the data prediction module in one or more embodiments provided in the present disclosure; for the explanation of other models in Figure 10, please refer to the description of the first data processing module and the data prediction module in one or more embodiments provided in the present disclosure.

[0230] Step 408: Based on the sequential sample features and the non-sequential sample features, the data processing model to be trained is trained to obtain a data processing model.

[0231] In one or more embodiments provided by the present disclosure, in the data processing model training method provided by the present disclosure, the application-related work of contrastive learning in recommendation scenarios usually considers sequential sample features and non-sequential sample features separately, while the full-feature contrastive learning model provided by the present disclosure considers both types of features at the same time, improves and optimizes the data enhancement method of the CL4SRec model, and provides the performance of the full-feature contrastive learning model as part of the sequential contrastive learning model.

[0232] In one or more embodiments provided in the present disclosure, the training of the data processing model to be trained based on the sequential sample features and the non-sequential sample features to obtain the data processing model includes:

[0233] Processing the sequence sample features using a first data processing module in the data processing model to be trained to obtain a first sample processing result;

[0234] Processing the non-sequential sample features using a second data processing module in the data processing model to be trained to obtain a second sample processing result;

[0235] The first data processing module is trained based on the first sample processing result, and the second data processing module is trained based on the second sample processing result, until a model training stop condition is reached, thereby obtaining the data processing model.

[0236] The first sample processing result can be understood as the loss value of training the first data processing module, and the second sample processing result can be understood as the loss value of training the second data processing module.

[0237] Specifically, in the data processing model training method provided by the present disclosure, after the initial sequential sample features and the initial non-sequential sample features are enhanced to obtain sequential sample features and non-sequential sample features, the sequential sample features can be input into the encoding unit in the first data processing module for encoding processing, thereby obtaining a first sample processing result determined based on the sequential sample data. For the non-sequential sample features, they can be input into the encoding unit in the second data processing module for encoding processing, thereby obtaining a second sample processing result determined based on the non-sequential sample data. Thereafter, the parameters of the first data processing module are adjusted based on the first sample processing result, and the parameters of the second data processing module are adjusted based on the second sample processing result, until the model training stop condition is reached, and the trained data processing model is obtained. Therefore, the influence of sample data on model training is fully considered in the process of model training, the problem of poor model training efficiency caused by sample data is avoided, and the needs of practical applications are met.

[0238] Continuing with the above example, after obtaining the enhanced sequential sample features, they are input into the encoder of the sequential contrastive learning model. All enhanced sequential sample features are input into the encoder of the sequential contrastive learning model. Based on the encoder output, the loss function is used to calculate the sequential contrastive learning function loss value. After obtaining the enhanced non-sequential sample features, they are input into the encoder of the non-sequential contrastive learning model. Based on the encoder output, the loss function is used to calculate the non-sequential contrastive learning function loss value. After obtaining the sequential contrastive learning function loss value and the non-sequential contrastive learning function loss value, the parameters of the sequential contrastive learning model are adjusted based on the sequential contrastive learning function loss value; the parameters of the non-sequential contrastive learning model are adjusted based on the non-sequential contrastive learning function loss value until the model training stop condition is met, resulting in a trained full-feature contrastive learning model.

[0239] In one or more embodiments provided by the present disclosure, in the data processing model training method provided by the present disclosure, the application-related work of contrastive learning in recommendation scenarios usually considers sequential features and non-sequential features separately. The full-feature contrastive learning model provided by the present disclosure takes both types of features into consideration at the same time; by optimizing the CL4CTR model in terms of training speed and training cost, the encoder is replaced with MaskNet so that it can be implemented in real application scenarios and serve as part of the non-sequential contrastive learning model. Finally, the sequential contrastive learning model and the non-sequential contrastive learning model are merged as a full-feature contrastive learning model to provide assistance for the prediction of actual recommendation application scenarios. Specifically, the second data processing module includes a non-sequential data processing module and a data prediction module;

[0240] The step of processing the non-sequential sample features using the second data processing module in the data processing model to be trained to obtain a second sample processing result includes:

[0241] Using the encoding unit in the non-sequential data processing module to encode the non-sequential sample features to obtain the non-sequential data processing result;

[0242] Processing the non-sequential sample features using the data prediction module to obtain a data prediction result;

[0243] The non-sequential data processing result and the data prediction result are used as the second sample processing result.

[0244] The non-sequential data processing module can be understood as a non-sequential contrastive learning model within the data processing model. The data prediction module can be understood as a multi-layer perceptron model within the data processing model. The data processing model can be a full-feature contrastive learning model. The non-sequential data processing result can be understood as the loss value obtained by training the non-sequential data processing module, for example, the loss value of a non-sequential contrastive learning function. The data prediction result can be understood as the loss value obtained by training the data prediction module, for example, the loss value of a multi-layer perceptron model.

[0245] The encoding unit can be understood as an encoder in a non-sequential data processing module, and the encoding unit can adopt a transfomer.

[0246] Continuing with the above example, after using the non-sequential contrastive learning model to perform feature enhancement and obtain the enhanced non-sequential sample features, they will be input into the encoder of the non-sequential contrastive learning model (MaskNet can be selected as the encoder because it can highlight important features while having a faster training speed). Then, based on the results output by the encoder, the loss function is used to calculate the loss value of the non-sequential contrastive learning function, where L2_loss can be selected as the loss function. In addition, the non-sequential sample features are input into the multi-layer perceptron model, and the loss value of the multi-layer perceptron model is calculated based on the results output by the multi-layer perceptron model. The loss value of the non-sequential contrastive learning function and the loss value of the multi-layer perceptron model are used as the second sample processing results for model training. Subsequently, the parameters of the non-sequential contrastive learning model can be adjusted based on the loss value of the non-sequential contrastive learning function; the parameters of the multi-layer perceptron model can be adjusted based on the loss value of the multi-layer perceptron model until the model training stop condition is reached.

[0247] In the embodiment provided by the present disclosure, the non-sequential sample features are input into the non-sequential data processing module, data enhancement is performed on the initial non-sequential sample features, and the obtained enhanced non-sequential sample features are input into the feature processing unit in the non-sequential data processing module to obtain non-sequential sample features; finally, the non-sequential sample features are input into the encoding unit in the non-sequential data processing module for encoding processing to obtain non-sequential data processing results, and the non-sequential sample features are processed by the data prediction module to obtain data prediction results. By using the non-sequential data processing results and the data prediction results as the second sample processing results, subsequent model training based on the second sample processing results is facilitated, and the model performance indicators of the current application scenario are further optimized to meet the needs of actual applications.

[0248] In one embodiment provided by the present disclosure, the processing of the non-sequential sample features by the data prediction module to obtain a data prediction result includes:

[0249] Inputting the non-sequential sample features into the data prediction module, and processing the non-sequential sample features using a feature processing unit in the data prediction module to obtain processed non-sequential sample features;

[0250] The processed non-sequential sample features are input into the encoding unit in the data prediction module for encoding processing to obtain the data prediction result.

[0251] The feature processing unit can be understood as the input layer or embedding layer integrated in the data prediction module. The encoding unit can be understood as the encoder in the data prediction module, and the encoding unit can adopt a multi-layer perceptron (MLP).

[0252] Continuing with the above example, the non-sequential sample features are input into the embedding layer of the multi-layer perceptron model to obtain the corresponding feature vector representation (i.e., the processed non-sequential sample features). This feature vector representation is then input into the multi-layer perceptron for binary classification prediction (which can be click-through rate prediction). The multi-layer perceptron loss value is calculated based on the prediction results to predict the click-through rate, which facilitates the subsequent improvement of the model performance based on the multi-layer perceptron loss value.

[0253] In one or more embodiments provided by the present disclosure, the loss function module of this model is mainly composed of loss_ctr (i.e., data prediction result) calculated by the basic model (i.e., multi-layer perceptron model), cl_loss calculated by the sequence contrastive learning model seq_modelThe first sample processing result is composed of the cl_lossnone_seq_model (i.e., the non-sequential data processing result) calculated by the non-sequential contrastive learning model. loss_ctr is a binary classification task. The cl_lossnone_seq_model calculated by the contrastive model in this model differs from the calculation method of the contrastive learning function loss in the sequential contrastive learning model. In practical application scenarios, the number of non-sequential features is large, for example, thousands or more. Converting them to a low-dimensional vector representation results in a large vector matrix. Using the calculation method used in the sequential contrastive learning model would result in excessive model computation and resource utilization. Compared to the calculation method of the contrastive learning function loss in the sequential contrastive learning model, the process of calculating cl_lossnone_seq_model for the contrastive model is simplified (feature alignment loss is included in cl_lossnone_seq_model).

[0254] Continuing with the above example, after obtaining the three loss values ​​(i.e., model prediction results), we can use these predictions to observe the model's performance in the application scenario. The model is trained based on these loss values ​​until the training stopping criteria are met. This training stopping criteria can be set based on the actual application scenario, for example, when the loss values ​​reach convergence.

[0255] The data processing model training method provided by the disclosed embodiments, after determining the sequential samples and non-sequential samples of the data processing model to be trained, performs feature extraction on the sequential samples and non-sequential samples in the data processing model to be trained, and performs feature enhancement on the initial sequential sample features and initial non-sequential sample features obtained by feature extraction to obtain sequential sample features and non-sequential sample features for model training; then, the data processing model to be trained is trained based on the sequential sample features and non-sequential sample features to obtain the data processing model. Thus, the influence of sample data on model training is fully considered during the model training process, avoiding the problem of poor model training efficiency caused by sample data, and meeting the needs of practical applications.

[0256] The following further illustrates the data processing model training method provided by the present disclosure using its application in a product recommendation scenario, with reference to FIG11 . FIG11 illustrates a flowchart of the processing process of a data processing model training method provided by one embodiment of the present disclosure, specifically including the following steps.

[0257] Step 1102: Receive a data set provided by a customer, and perform feature extraction on the data in the data set to obtain data features.

[0258] Step 1104: According to the data feature type, the data features are divided into sequential features (ie, sequential sample features in the above embodiment) and non-sequential features (ie, non-sequential sample features in the above embodiment).

[0259] Sequential features refer to the interaction records between users and products. For example, an interaction record can be composed of three time series features: product ID, product category ID, and user behavior sequence. Different time series features have a one-to-one correspondence within the same time period.

[0260] Non-sequential features include user-side data such as user age and gender, and product-side features such as product price and product capacity. In other words, all data features other than sequential features are non-sequential features.

[0261] Step 1106: Expand the original sequence features based on the replacement data enhancement method of collaborative filtering.

[0262] The specific steps are:

[0263] 1. Create a user-item relationship table. Based on the user-item interaction information in the original dataset, create a table that maps user IDs to item IDs. This table contains the IDs of all items that each user has interacted with.

[0264] 2. Calculate the similarity between items. First, create an item co-occurrence matrix. Next, for each user, add 1 to the co-occurrence matrix for each item in their interacted item list. Finally, normalize the co-occurrence matrix to obtain the cosine similarity matrix between items.

[0265] Step 1108: Determine the similarity between the target product and other products, and use the top five products ranked by similarity as a set of similar products to the target product.

[0266] The target product is the product with the oldest time point in the time series. In other words, for the oldest product ID in the time series, a set of similar products is determined.

[0267] It should be noted that if no similar product set is obtained after collaborative filtering calculation, the category to which the target product belongs is found, and the top 20% of products with the highest frequency of occurrence in the category are selected as the similar item set of the target product.

[0268] Step 1110: Construct a new sequence feature data.

[0269] The specific steps are:

[0270] 1. Determine the time series features of product IDs whose length is greater than 10.

[0271] 2. From the product ID time series features with a length greater than 10, determine the product ID (target product) with the longest appearance time; and determine the set of similar products corresponding to this product ID.

[0272] 3. Randomly select a product ID from the set of similar products and replace the target product.

[0273] 4. Determine the category ID and behavior sequence time features corresponding to the product ID time series feature. For the product category ID time series feature, replace the target product's category ID with the category ID corresponding to a similar product; the behavior sequence time series feature remains unchanged, thus constructing a new data record based on the original user behavior record.

[0274] Step 1112: For the sequence features, input them into the sequence contrastive learning model to obtain the loss value.

[0275] The specific steps are:

[0276] 1. Concatenate a product ID, a product category ID, and a user behavior sequence to obtain the concatenated sequence features.

[0277] 2. Input the concatenated sequence features into the embedding layer of the model to obtain the corresponding feature vector representation.

[0278] 3. Perform two random vector enhancement operations on the feature vector representation to obtain a positive sample pair in contrastive learning, which is two enhanced sequence features.

[0279] Among them, vector enhancement operations include vector masking, vector cropping, and vector reordering. The two random operations can be a data enhancement operation method.

[0280] 4. After obtaining the two enhanced sequence features, the two enhanced sequence features are input into the corresponding encoders respectively, and the loss value is calculated according to the output of the encoder.

[0281] Step 1114: For non-sequential features, input them into a non-sequential contrastive learning model to obtain a loss value.

[0282] The specific steps are:

[0283] 1. Perform data enhancement operations on non-sequential features to obtain enhanced non-sequential features.

[0284] Among them, the data enhancement method selects the feature mask mode.

[0285] 2. Input the enhanced non-sequential features into the embedding layer of the model to obtain the enhanced feature vector representation of the corresponding data.

[0286] 3. Input the feature vector representation after data enhancement into the encoder and calculate the loss value based on the output of the encoder.

[0287] Step 1116: Input the non-sequential features into the embedding layer to obtain the corresponding feature vector representation, and input the feature vector representation into the multi-layer perceptron to perform binary classification prediction; calculate the loss value based on the prediction result.

[0288] Step 1118: After completing the process of steps 1112 to 1116 above, the model can be trained according to the model prediction results (the above three loss values) until the model training stopping condition is reached.

[0289] Based on this, the data processing model training method provided in this disclosure takes into account the situations in which missing or insufficient data samples may occur in real-world application scenarios. Based on the collaborative filtering algorithm, a self-supervised dataset expansion method is proposed to expand the original dataset. By expanding the original dataset from the perspective of sample data, the model's prediction effect is optimized. Compared with related work focusing on model optimization, the training cost and speed required by this method are significantly superior. Furthermore, the use of self-supervision and contrastive learning can effectively address the sparsity and long-tail problems of recommendation scenario data. In combination with actual application needs, a contrastive learning model is built that considers both sequential and non-sequential features to optimize the current model performance indicators.

[0290] Referring to FIG. 12 , FIG. 12 shows a flow chart of another data processing method provided according to an embodiment of the present disclosure. The data processing method is applied in the cloud and specifically includes the following steps.

[0291] Step 1202: receiving user behavior data of a user sent by a terminal, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0292] Step 1204: Inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are used to extract features of sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features;

[0293] Step 1206: Send the data processing result to the terminal.

[0294] In one or more embodiments provided by the present disclosure, a data processing model in another data processing method is obtained by training using sequential sample features and non-sequential sample features. The sequential sample features and non-sequential sample features are obtained by performing feature enhancement on initial sequential sample features and initial non-sequential sample features obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model. Thus, the influence of sample data on model training is fully considered during model training, thereby avoiding the problem of poor model training efficiency caused by sample data. Moreover, the trained data processing model is used to process the sequential behavior data and / or non-sequential behavior data contained in the user behavior data sent by the terminal to obtain a data processing result, and the data processing result is sent to the terminal, thereby meeting the needs of practical applications through the data processing model.

[0295] The above is a schematic scheme of another data processing method of this embodiment. It should be noted that the technical scheme of the other data processing method and the technical scheme of the above-mentioned data processing method are based on the same concept. For details not described in detail in the technical scheme of the other data processing method, please refer to the description of the technical scheme of the above-mentioned data processing method.

[0296] Referring to FIG13 , FIG13 shows a flowchart of an object recommendation model training method provided according to an embodiment of the present disclosure, which specifically includes the following steps.

[0297] Step 1302: Determine sequential samples and non-sequential samples for the object recommendation model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users for the sample objects;

[0298] Step 1304: Input the sequential samples and the non-sequential samples into the to-be-trained object recommendation model, perform feature extraction on the sequential samples and the non-sequential samples in the to-be-trained object recommendation model to obtain initial sequential sample features and initial non-sequential sample features;

[0299] Step 1306: performing feature enhancement on the initial sequential sample feature to obtain the sequential sample feature, and performing feature enhancement on the initial non-sequential sample feature to obtain the non-sequential sample feature;

[0300] Step 1308: Based on the sequential sample features and the non-sequential sample features, the object recommendation model to be trained is trained to obtain an object recommendation model.

[0301] The sample object can be understood as a behavior object serving as a sample, and the behavior object can be understood as the behavior object targeted by the sample user's user behavior. For example, when the object recommendation method is applied to a product recommendation scenario, the sample object can be understood as a sample product. When the object recommendation method is applied to a scenic spot recommendation scenario, the sample object can be understood as a sample scenic spot. When the object recommendation method is applied to a food recommendation scenario, the sample object can be understood as a sample food. The object recommendation model can be understood as the data processing model in the above-mentioned data processing model training method.

[0302] The object recommendation model training method provided by the disclosed embodiments, after determining the sequential samples and non-sequential samples of the data processing model to be trained, performs feature extraction on the sequential and non-sequential samples in the object recommendation model to be trained, and performs feature enhancement on the initial sequential and non-sequential sample features obtained by feature extraction to obtain sequential and non-sequential sample features for model training; then, the object recommendation model to be trained is trained based on the sequential and non-sequential sample features to obtain an object recommendation model. Thus, the influence of sample data on model training is fully considered during the model training process, avoiding the problem of poor model training efficiency caused by sample data, and meeting the needs of practical applications.

[0303] The above is a schematic scheme of an object recommendation model training method of this embodiment. It should be noted that the technical scheme of this object recommendation model training method and the technical scheme of the aforementioned data processing model training method are based on the same concept. For details not described in detail in the technical scheme of the object recommendation model training method, please refer to the description of the technical scheme of the aforementioned data processing model training method.

[0304] Corresponding to the above method embodiment, the present disclosure further provides an embodiment of a data processing device, the device comprising:

[0305] A data receiving module is configured to receive user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0306] The data processing module is configured to input the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain the data processing results corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are used to extract features of sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features.

[0307] Optionally, the data processing device further includes a model training module configured to:

[0308] Determining the sequential samples and the non-sequential samples for a data processing model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users;

[0309] Inputting the sequential samples and the non-sequential samples into the data processing model to be trained, performing feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtaining initial sequential sample features and initial non-sequential sample features;

[0310] Performing feature enhancement on the initial sequential sample feature to obtain the sequential sample feature, and performing feature enhancement on the initial non-sequential sample feature to obtain the non-sequential sample feature;

[0311] Based on the sequential sample features and the non-sequential sample features, the data processing model to be trained is trained to obtain the data processing model.

[0312] Optionally, the model training module is further configured to:

[0313] Determining similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample feature, and performing sample expansion based on the user behavior objects and the similar behavior objects to obtain an expanded sequence-type sample feature;

[0314] Feature enhancement is performed on the initial sequence sample features and the expanded sequence sample features to obtain the sequence sample features of the data processing model to be trained.

[0315] A data processing model in a data processing device provided by an embodiment of the present disclosure is obtained by training with sequential sample features and non-sequential sample features. The sequential sample features and non-sequential sample features are obtained by performing feature enhancement on initial sequential sample features and initial non-sequential sample features obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model. Thus, the influence of sample data on model training is fully considered during the model training process, thereby avoiding the problem of poor model training efficiency caused by sample data. Moreover, the trained data processing model is used to process the sequential behavior data and / or non-sequential behavior data contained in the user behavior data, and a data processing result is obtained, thereby meeting the needs of practical applications through the data processing model.

[0316] The above is a schematic diagram of a data processing device according to this embodiment. It should be noted that the technical solution of the data processing device and the technical solution of the above-mentioned data processing method are based on the same concept. For details not described in detail in the technical solution of the data processing device, please refer to the description of the technical solution of the above-mentioned data processing method.

[0317] Corresponding to the above method embodiment, the present disclosure also provides an embodiment of a data processing model training device, the device comprising:

[0318] a sample determination module configured to determine sequential samples and non-sequential samples for training a data processing model, wherein the sequential samples and the non-sequential samples are user behavior data of sample users;

[0319] a feature extraction module configured to input the sequential samples and the non-sequential samples into the data processing model to be trained, perform feature extraction on the sequential samples and the non-sequential samples in the data processing model to be trained, and obtain initial sequential sample features and initial non-sequential sample features;

[0320] a sample enhancement module configured to perform feature enhancement on the initial sequential sample features to obtain the sequential sample features, and to perform feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features;

[0321] The model training module is configured to train the data processing model to be trained based on the sequence sample features and the non-sequence sample features to obtain a data processing model.

[0322] Optionally, the sample enhancement module is further configured to:

[0323] Determining similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample feature, and performing sample expansion based on the user behavior objects and the similar behavior objects to obtain an expanded sequence-type sample feature;

[0324] Feature enhancement is performed on the initial sequence sample features and the expanded sequence sample features to obtain the sequence sample features of the data processing model to be trained.

[0325] Optionally, the sample enhancement module is further configured to:

[0326] Determining at least two user behavior objects in the initial sequence sample features;

[0327] Determining similarity results between a target behavior object and other behavior objects among the at least two user behavior objects, wherein the target behavior object is any one of the at least two user behavior objects, and the other behavior objects are other user behavior objects among the at least two user behavior objects except the target behavior object;

[0328] A corresponding similar behavior object is determined for the target behavior object based on the similarity result.

[0329] Optionally, the sample enhancement module is further configured to:

[0330] Constructing a user behavior object matrix based on the at least two user behavior objects;

[0331] Obtaining a similarity result matrix corresponding to the at least two user behavior objects by performing similarity calculation on the user behavior object matrix;

[0332] The similarity result matrix is ​​used to determine similarity results between the target behavior object and other behavior objects in the at least two user behavior objects.

[0333] Optionally, the similarity result is similarity;

[0334] The sample enhancement module is further configured to:

[0335] Determining a target similarity greater than or equal to a preset similarity threshold from the similarities between the target behavior object and the other behavior objects;

[0336] Determine other behavior objects corresponding to the target similarity as at least two candidate similar behavior objects of the target behavior object;

[0337] Sort the at least two candidate similar behavior objects based on the target similarity to obtain a candidate similar behavior object sequence;

[0338] A preset first number of the candidate similar behavior objects are selected from the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0339] Optionally, the similarity result is similarity;

[0340] The sample enhancement module is further configured to:

[0341] When the similarity between the target behavior object and the other behavior objects is less than a preset similarity threshold, determining the object type information of the target behavior object;

[0342] Determine, from the at least two user behavior objects, at least two candidate similar behavior objects corresponding to the object type information, and determine the object quantity of each candidate similar behavior object;

[0343] Sort the at least two candidate similar behavior objects based on the number of objects to obtain a candidate similar behavior object sequence;

[0344] A preset second number of the candidate similar behavior objects are selected from the candidate similar behavior object sequence from top to bottom as similar behavior objects corresponding to the target behavior object.

[0345] Optionally, the number of the initial sequence-type sample features is at least two;

[0346] The sample enhancement module is further configured to:

[0347] Determining user behavior object information of each initial sequence-type sample feature in at least two initial sequence-type sample features;

[0348] Based on the user behavior object information, determining a sequence sample feature to be expanded from the at least two initial sequence sample features;

[0349] Based on the similar behavior objects, the user behavior objects in the sequence type sample features to be expanded are processed to obtain expanded sequence type sample features.

[0350] Optionally, the sample enhancement module is further configured to:

[0351] Determining a behavior object to be expanded and object type information to be expanded corresponding to the behavior object to be expanded from at least two user behavior objects included in the sequence sample feature to be expanded;

[0352] Selecting a target similar behavior object from similar behavior objects corresponding to the behavior object to be expanded, and determining target object type information of the target similar behavior object;

[0353] The target similar behavior object is used to replace the behavior object to be expanded in the sequence type sample feature to be expanded, and the target object type information is used to replace the object type information to be expanded corresponding to the behavior object to be expanded in the sequence type sample feature to be expanded, so as to obtain the expanded sequence type sample feature.

[0354] Optionally, the sample enhancement module is further configured to:

[0355] Inputting the initial sequence sample features into the first data processing module in the data processing model to be trained;

[0356] Processing the initial sequence sample features using a feature processing unit in the first data processing module to obtain sequence sample features to be enhanced;

[0357] A target data enhancement rule is determined from at least two data enhancement rules, and the target data enhancement rule is used to perform data enhancement on the sequence sample feature to be enhanced to obtain the sequence sample feature.

[0358] Optionally, the sample enhancement module is further configured to:

[0359] Selecting two target data enhancement rules from a preset first data enhancement rule, a preset second data enhancement rule, and a preset third data enhancement rule;

[0360] Performing data enhancement on the sequence-type sample feature to be enhanced using the two target data enhancement rules to obtain a first enhanced sequence-type sample feature and a second enhanced sequence-type sample feature;

[0361] The first enhanced sequence-type sample feature and the second enhanced sequence-type sample feature are used as the sequence-type sample feature.

[0362] Optionally, the sample enhancement module is further configured to:

[0363] Inputting the initial non-sequential sample features into a second data processing module in the data processing model to be trained;

[0364] Performing feature masking processing on the initial non-sequential sample features to obtain enhanced non-sequential sample features;

[0365] The enhanced non-sequential sample features are processed by a feature processing unit in the second data processing module to obtain the non-sequential sample features.

[0366] Optionally, the model training module is further configured to:

[0367] Processing the sequence sample features using a first data processing module in the data processing model to be trained to obtain a first sample processing result;

[0368] Processing the non-sequential sample features using a second data processing module in the data processing model to be trained to obtain a second sample processing result;

[0369] The first data processing module is trained based on the first sample processing result, and the second data processing module is trained based on the second sample processing result, until a model training stop condition is reached, thereby obtaining the data processing model.

[0370] Optionally, the second data processing module includes a non-sequential data processing module and a data prediction module;

[0371] The model training module is further configured to:

[0372] Using the encoding unit in the non-sequential data processing module to encode the non-sequential sample features to obtain the non-sequential data processing result;

[0373] Processing the non-sequential sample features using the data prediction module to obtain a data prediction result;

[0374] The non-sequential data processing result and the data prediction result are used as the second sample processing result.

[0375] Optionally, the model training module is further configured to:

[0376] Inputting the non-sequential sample features into the data prediction module, and processing the non-sequential sample features using a feature processing unit in the data prediction module to obtain processed non-sequential sample features;

[0377] The processed non-sequential sample features are input into the encoding unit in the data prediction module for encoding processing to obtain the data prediction result.

[0378] The data processing model training device provided by the disclosed embodiments, after determining the sequential samples and non-sequential samples of the data processing model to be trained, performs feature extraction on the sequential samples and non-sequential samples in the data processing model to be trained, and performs feature enhancement on the initial sequential sample features and initial non-sequential sample features obtained by feature extraction to obtain sequential sample features and non-sequential sample features for model training; then, the data processing model to be trained is trained based on the sequential sample features and non-sequential sample features to obtain a data processing model. Thus, the influence of sample data on model training is fully considered during the model training process, avoiding the problem of poor model training efficiency caused by sample data, and meeting the needs of practical applications.

[0379] The above is a schematic diagram of a data processing model training device according to this embodiment. It should be noted that the technical solution of this data processing model training device and the technical solution of the aforementioned data processing model training method are based on the same concept. For details not described in detail in the technical solution of the data processing model training device, please refer to the description of the technical solution of the aforementioned data processing model training method.

[0380] Corresponding to the above method embodiment, the present disclosure also provides another data processing device embodiment, which is applied to the cloud and includes:

[0381] a data receiving module configured to receive user behavior data of a user sent by a terminal, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data;

[0382] a data processing module configured to input the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training sequential sample features and non-sequential sample features, the sequential sample features and the non-sequential sample features are obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and feature enhancement is performed on the initial sequential sample features and the initial non-sequential sample features;

[0383] The result sending module is configured to send the data processing result to the terminal.

[0384] In one or more embodiments provided by the present disclosure, a data processing model in another data processing device is obtained by training with sequential sample features and non-sequential sample features. The sequential sample features and non-sequential sample features are obtained by performing feature enhancement on initial sequential sample features and initial non-sequential sample features obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model. Thus, the influence of sample data on model training is fully considered during model training, thereby avoiding the problem of poor model training efficiency caused by sample data. Moreover, the trained data processing model is used to process the sequential behavior data and / or non-sequential behavior data contained in the user behavior data sent by the terminal to obtain a data processing result, and the data processing result is sent to the terminal, thereby meeting the needs of practical applications through the data processing model.

[0385] The above is a schematic diagram of another data processing device according to this embodiment. It should be noted that the technical solution of this another data processing device and the technical solution of the aforementioned another data processing method are based on the same concept. For details not described in detail in the technical solution of the another data processing device, please refer to the description of the technical solution of the aforementioned another data processing method.

[0386] Corresponding to the above method embodiment, the present disclosure also provides an embodiment of an object recommendation model training device, the device comprising:

[0387] A sample determination module is configured to determine sequential samples and non-sequential samples for the object recommendation model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users for the sample objects;

[0388] a feature extraction module configured to input the sequential samples and the non-sequential samples into the to-be-trained object recommendation model, perform feature extraction on the sequential samples and the non-sequential samples in the to-be-trained object recommendation model, and obtain initial sequential sample features and initial non-sequential sample features;

[0389] a sample enhancement module configured to perform feature enhancement on the initial sequential sample features to obtain the sequential sample features, and to perform feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features;

[0390] The model training module is configured to train the object recommendation model to be trained based on the sequential sample features and the non-sequential sample features to obtain an object recommendation model.

[0391] The object recommendation model training device provided by the disclosed embodiments, after determining sequential and non-sequential samples for a data processing model to be trained, performs feature extraction on the sequential and non-sequential samples in the object recommendation model to be trained. The device then enhances the initial sequential and non-sequential sample features obtained from the feature extraction to obtain sequential and non-sequential sample features for model training. The object recommendation model to be trained is then trained based on these sequential and non-sequential sample features to obtain an object recommendation model. This fully considers the impact of sample data on model training during model training, avoids the problem of poor model training efficiency caused by sample data, and meets the needs of practical applications.

[0392] The above is a schematic diagram of an object recommendation model training device according to this embodiment. It should be noted that the technical solution of the object recommendation model training device and the technical solution of the object recommendation model training method described above are based on the same concept. For details not described in detail in the technical solution of the object recommendation model training device, please refer to the description of the technical solution of the object recommendation model training method described above.

[0393] Figure 14 shows a block diagram of a computing device 1400 according to one embodiment of the present disclosure. Components of the computing device 1400 include, but are not limited to, a memory 1410 and a processor 1420. The processor 1420 is connected to the memory 1410 via a bus 1430, and a database 1450 is used to store data.

[0394] The computing device 1400 also includes an access device 1440 that enables the computing device 1400 to communicate via one or more networks 1460. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1440 may include one or more of any type of network interface (e.g., a network interface controller (NIC)) whether wired or wireless, such as an IEEE 802.12 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, or a near field communication (NFC) interface.

[0395] In one embodiment of the present disclosure, the aforementioned components of computing device 1400 and other components not shown in FIG14 may also be connected to each other, for example, via a bus. It should be understood that the block diagram of the computing device structure shown in FIG14 is for illustrative purposes only and does not limit the scope of the present disclosure. Those skilled in the art may add or replace other components as needed.

[0396] Computing device 1400 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, personal digital assistant, laptop computer, notebook computer, netbook computer, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or personal computer (PC). Computing device 1400 may also be a mobile or stationary server.

[0397] Among them, the processor 1420 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned two data processing methods, data processing model training methods, or object recommendation model training methods.

[0398] The above is a schematic diagram of a computing device according to this embodiment. It should be noted that the technical solution of this computing device is based on the same concept as the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solutions of the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods.

[0399] An embodiment of the present disclosure also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned two data processing methods, data processing model training methods, or object recommendation model training methods.

[0400] The above is a schematic diagram of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium is based on the same concept as the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solutions of the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods.

[0401] An embodiment of the present disclosure further provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above two data processing methods, data processing model training methods, or object recommendation model training methods.

[0402] The above is an illustrative solution of a computer program according to this embodiment. It should be noted that the technical solution of this computer program is based on the same concept as the technical solutions of the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solutions of the two aforementioned data processing methods, data processing model training methods, and object recommendation model training methods.

[0403] The foregoing description describes specific embodiments of the present disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0404] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0405] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present disclosure are not limited by the order of the actions described, because according to the embodiments of the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of the present disclosure.

[0406] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0407] The preferred embodiments of the present disclosure disclosed above are only used to help illustrate the present disclosure. The optional embodiments do not describe all details in detail, nor do they limit the invention to only the specific embodiments described. Obviously, many modifications and variations can be made based on the content of the embodiments of the present disclosure. The present disclosure selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present disclosure, so that those skilled in the art can better understand and utilize the present disclosure. The present disclosure is limited only by the claims and their full scope and equivalents.

Claims

1. A data processing method, comprising: Receiving user behavior data of a user, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data; Inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is obtained by training with sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are obtained by extracting features from sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and performing feature enhancement on the initial sequential sample features and the initial non-sequential sample features.

2. The data processing method according to claim 1, before inputting the sequential behavior data and / or the non-sequential behavior data into the data processing model to obtain the data processing result corresponding to the user, further comprising: Determining the sequential samples and the non-sequential samples of the data processing model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users; Inputting the sequential samples and the non-sequential samples into the data processing model to be trained, and extracting features from the sequential samples and the non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features; Performing feature enhancement on the initial sequential sample features to obtain the sequential sample features, and performing feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features; Training the data processing model to be trained based on the sequential sample features and the non-sequential sample features to obtain the data processing model.

3. The data processing method according to claim 2, the performing feature enhancement on the initial sequential sample features to obtain the sequential sample features, comprising: Determining similar behavior objects corresponding to the user behavior objects in the initial sequential sample features, and performing sample expansion based on the user behavior objects and the similar behavior objects to obtain expanded sequential sample features; Performing feature enhancement on the initial sequential sample features and the expanded sequential sample features to obtain the sequential sample features of the data processing model to be trained.

4. A data processing model training method, comprising: Determining sequential samples and non-sequential samples of a data processing model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of sample users; Inputting the sequential samples and the non-sequential samples into the data processing model to be trained, and extracting features from the sequential samples and the non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features; Perform feature enhancement on the initial sequence-type sample features to obtain the sequence-type sample features, and perform feature enhancement on the initial non-sequence-type sample features to obtain the non-sequence-type sample features; Based on the sequence-type sample features and the non-sequence-type sample features, train the data processing model to be trained to obtain a data processing model.

5. The data processing model training method according to claim 4, wherein the performing feature enhancement on the initial sequence-type sample features to obtain the sequence-type sample features includes: Determine similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample features, and based on the User behavior objects and the similar behavior objects, perform sample augmentation to obtain augmented sequence-type sample features; Perform feature enhancement on the initial sequence-type sample features and the augmented sequence-type sample features to obtain the sequence-type sample features of the data processing model to be trained.

6. The data processing model training method according to claim 5, wherein the determining similar behavior objects corresponding to the user behavior objects in the initial sequence-type sample features includes: Determine at least two user behavior objects in the initial sequence-type sample features; Determine the similarity results between the target behavior object and other behavior objects among the at least two user behavior objects, where the target behavior object is any one of the at least two user behavior objects, and the other behavior objects are the other user behavior objects among the at least two user behavior objects except the target behavior object; Based on the similarity results, determine corresponding similar behavior objects for the target behavior object.

7. The data processing model training method according to claim 6, wherein the determining the similarity results between the target behavior object and other behavior objects among the at least two user behavior objects includes: Construct a user behavior object matrix based on the at least two user behavior objects; Through similarity calculation on the user behavior object matrix, obtain a similarity result matrix corresponding to the at least two user behavior objects; Through the similarity result matrix, determine the similarity results between the target behavior object and other behavior objects among the at least two user behavior objects.

8. The data processing model training method according to claim 6, wherein the similarity result is a similarity degree; The determining corresponding similar behavior objects for the target behavior object based on the similarity results includes: Among the similarity degrees between the target behavior object and the other behavior objects, determine a target similarity degree greater than or equal to a preset similarity threshold; Determine the other behavior objects corresponding to the target similarity degree as at least two candidate similar behavior objects of the target behavior object; Based on the target similarity degree, sort the at least two candidate similar behavior objects to obtain a candidate similar behavior object sequence; From the candidate similar behavior object sequence, select a preset first number of the candidate similar behavior objects from top to bottom as the similar behavior objects corresponding to the target behavior object.

9. The data processing model training method according to claim 6, wherein the similarity result is a similarity degree; The determining, based on the similarity result, of a corresponding similar behavior object for the target behavior object includes: determining object type information of the target behavior object when the similarity degree between the target behavior object and the other behavior objects is less than a preset similarity threshold; determining at least two candidate similar behavior objects corresponding to the object type information from the at least two user behavior objects, and determining the object quantity of each candidate similar behavior object; sorting the at least two candidate similar behavior objects based on the object quantity to obtain a candidate similar behavior object sequence; selecting, from the candidate similar behavior object sequence, a preset second quantity of the candidate similar behavior objects from top to bottom as the similar behavior objects corresponding to the target behavior object.

10. The data processing model training method according to any one of claims 6-9, wherein the quantity of the initial sequential sample features is at least two; The obtaining of augmented sequential sample features based on the user behavior object and the similar behavior object includes: determining user behavior object information of each of the at least two initial sequential sample features; determining a to-be-augmented sequential sample feature from the at least two initial sequential sample features based on the user behavior object information; processing the user behavior object in the to-be-augmented sequential sample feature based on the similar behavior object to obtain augmented sequential sample features.

11. The data processing model training method according to claim 10, wherein the processing of the user behavior object in the to-be-augmented sequential sample feature based on the similar behavior object to obtain augmented sequential sample features includes: determining a to-be-augmented behavior object and to-be-augmented object type information corresponding to the to-be-augmented behavior object from among the at least two user behavior objects included in the to-be-augmented sequential sample feature; selecting a target similar behavior object from the similar behavior objects corresponding to the to-be-augmented behavior object, and determining target object type information of the target similar behavior object; replacing the to-be-augmented behavior object in the to-be-augmented sequential sample feature with the target similar behavior object, and replacing the to-be-augmented object type information corresponding to the to-be-augmented behavior object in the to-be-augmented sequential sample feature with the target object type information to obtain the augmented sequential sample features.

12. The data processing model training method according to claim 4, wherein the obtaining of the sequential sample features by enhancing the features of the initial sequential sample features includes: inputting the initial sequential sample features into a first data processing module in the data processing model to be trained; processing the initial sequential sample features by a feature processing unit in the first data processing module to obtain to-be-enhanced sequential sample features; Determine a target data augmentation rule from at least two data augmentation rules, and use the target data augmentation rule to perform data augmentation on the sequence-type sample features to be augmented, to obtain the sequence-type sample features.

13. The data processing model training method according to claim 12, wherein the determining a target data augmentation rule from at least two data augmentation rules, and using the target data augmentation rule to perform data augmentation on the sequence-type sample features to be augmented, to obtain the sequence-type sample features, comprises: Select two target data augmentation rules from a preset first data augmentation rule, a preset second data augmentation rule, and a preset third data augmentation rule; Use the two target data augmentation rules to perform data augmentation on the sequence-type sample features to be augmented, to obtain a first augmented sequence-type sample feature and a second augmented sequence-type sample feature; Use the first augmented sequence-type sample feature and the second augmented sequence-type sample feature as the sequence-type sample features.

14. The data processing model training method according to claim 4, wherein the performing feature augmentation on the initial non-sequence-type sample features to obtain the non-sequence-type sample features comprises: Input the initial non-sequence-type sample features into a second data processing module in the data processing model to be trained; Perform feature masking processing on the initial non-sequence-type sample features to obtain augmented non-sequence-type sample features; Use a feature processing unit in the second data processing module to process the augmented non-sequence-type sample features to obtain the non-sequence-type sample features.

15. The data processing model training method according to claim 4, wherein the training the data processing model to be trained based on the sequence-type sample features and the non-sequence-type sample features to obtain the data processing model comprises: Use a first data processing module in the data processing model to be trained to process the sequence-type sample features to obtain a first sample processing result; Use a second data processing module in the data processing model to be trained to process the non-sequence-type sample features to obtain a second sample processing result; Train the first data processing module based on the first sample processing result, and train the second data processing module based on the second sample processing result until a model training stop condition is reached, to obtain the data processing model.

16. The data processing model training method according to claim 15, wherein the second data processing module comprises a non-sequence-type data processing module and a data prediction module; The using a second data processing module in the data processing model to be trained to process the non-sequence-type sample features to obtain a second sample processing result comprises: Use an encoding unit in the non-sequence-type data processing module to perform encoding processing on the non-sequence-type sample features to obtain the non-sequence-type data processing result; Use the data prediction module to process the non-sequence-type sample features to obtain a data prediction result; Use the non-sequence-type data processing result and the data prediction result as the second sample processing result.

17. The data processing model training method according to claim 16, wherein the step of using the data prediction module to process the non-sequential sample features to obtain a data prediction result includes: Inputting the non-sequential sample features into the data prediction module, and using the feature processing unit in the data prediction module to process the non-sequential sample features to obtain processed non-sequential sample features; Inputting the processed non-sequential sample features into the encoding unit in the data prediction module for encoding processing to obtain the data prediction result.

18. A data processing method applied to the cloud, including: Receiving user behavior data of a user sent by a terminal, wherein the user behavior data includes sequential behavior data and / or non-sequential behavior data; Inputting the sequential behavior data and / or the non-sequential behavior data into a data processing model to obtain a data processing result corresponding to the user, wherein the data processing model is trained by using sequential sample features and non-sequential sample features, and the sequential sample features and the non-sequential sample features are obtained by performing feature extraction on sequential samples and non-sequential samples in the data processing model to obtain initial sequential sample features and initial non-sequential sample features, and then performing feature enhancement on the initial sequential sample features and the initial non-sequential sample features; Sending the data processing result to the terminal.

19. An object recommendation model training method, including: Determining sequential samples and non-sequential samples of an object recommendation model to be trained, wherein the sequential samples and the non-sequential samples are user behavior data of a sample user for a sample object; Inputting the sequential samples and the non-sequential samples into the object recommendation model to be trained, and performing feature extraction on the sequential samples and the non-sequential samples in the object recommendation model to be trained to obtain initial sequential sample features and initial non-sequential sample features; Performing feature enhancement on the initial sequential sample features to obtain the sequential sample features, and performing feature enhancement on the initial non-sequential sample features to obtain the non-sequential sample features; Training the object recommendation model to be trained based on the sequential sample features and the non-sequential sample features to obtain an object recommendation model.

20. A computing device, including: A memory and a processor; The memory is used for storing computer-executable instructions, and the processor is used for executing the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the data processing method according to any one of claims 1 to 3, the data processing model training method according to any one of claims 4 to 17, the data processing method according to any one of claims 18, or the object recommendation model training method according to any one of claims 19 are implemented.

21. A computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the data processing method according to any one of claims 1 to 3, the data processing model training method according to any one of claims 4 to 17, the data processing method according to any one of claims 18, or the object recommendation model training method according to any one of claims 19 are implemented.

22. A computer program, when the computer program is executed in a computer, causes the computer to execute the steps of the data processing method according to any one of claims 1 to 3, the data processing model training method according to any one of claims 4 to 17, the data processing method according to any one of claims 18, or the object recommendation model training method according to any one of claims 19.

Images