Task processing method, object recommendation method, model training method, and computing device

Abstract

Embodiments of the present disclosure provide a task processing method, an object recommendation method, a model training method, and a computing device. The task processing method comprises: acquiring structural data of a target task, wherein the structural data comprises multi-type task data conforming to a specific format; extracting data features of the multi-type task data; on the basis of at least two interaction operations, performing interaction feature analysis on the data features of the multi-type task data respectively, to obtain interaction features corresponding to the various interaction operations; on the basis of the interaction features corresponding to the various interaction operations, determining a target interaction feature corresponding to the target task; and executing the target task on the basis of the target interaction feature to obtain a task processing result of the target task. By means of adaption to multi-type task data in a specific format, a target interaction feature which is an arithmetic interaction feature is extracted, serving as a high-accuracy data feature, and a target task is executed on the basis of the target interaction feature, improving the accuracy of task processing.

Description

[0001] This disclosure claims priority to Chinese patent application number 202410125040.2, filed with the China Patent Office on January 29, 2024, entitled "Task Processing, Object Recommendation, Model Training Method, and Computing Device," the entire contents of which are incorporated herein by reference. Technical Field: Embodiments of the present disclosure relate to the field of data processing technology, and more particularly to a task processing method, an object recommendation method, and a computing device. Background: With the advancement of data processing technology, task processing performance has been improved by executing corresponding target tasks based on the data features of structured data in different fields and scenarios. Currently, deep learning algorithms (e.g., Transformer models) or alternative algorithms (e.g., XGBoost tree models) are used to perform feature mining on multiple types of task data in structured data to obtain data features for each type of task data. Based on these data features, tasks in different fields and scenarios are executed to achieve task processing. However, structured data includes multiple types of task data conforming to specific formats. Deep learning algorithms or alternative algorithms cannot adaptively mine the arithmetic interaction features of structured data, resulting in insufficient accuracy of data features and, in turn, insufficient accuracy in task processing. Therefore, a highly accurate task processing method is urgently needed. In view of this, embodiments of the present disclosure provide a task processing method. One or more embodiments of the present disclosure also relate to an object recommendation method, a model training method, a task processing apparatus, an object recommendation apparatus, a model training apparatus, a computing device, a computer-readable storage medium, and a computer program product to address technical deficiencies in the prior art. One aspect of an embodiment of the present disclosure provides a task processing method, comprising: obtaining structured data of a target task, wherein the structured data includes multiple types of task data conforming to a specific format; extracting data features of the multiple types of task data; performing interaction feature analysis on the data features of the multiple types of task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations; determining a target interaction feature corresponding to the target task based on the interaction features corresponding to the various interaction operations; and executing the target task based on the target interaction feature to obtain a task processing result for the target task.In one embodiment of the present disclosure, structured data for a target task is obtained, where the structured data includes multiple types of task data in a specific format. Data features of the multiple types of task data are extracted. Interaction feature analysis is performed on the data features of the multiple types of task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations. A target interaction feature corresponding to the target task is determined based on the interaction features corresponding to the various interaction operations. The target task is executed based on the target interaction feature to obtain a task processing result for the target task. The data features of the multiple types of task data in the specific format are extracted. Interaction feature analysis is performed on the data features of the multiple types of task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations. Based on the interaction features corresponding to the various interaction operations, a target interaction feature is extracted based on the interaction features corresponding to the various interaction operations. This arithmetic interaction feature, adapted to the multiple types of data in the specific format, is a highly accurate data feature. Executing the target task based on the target interaction feature improves task processing accuracy. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a flowchart of a task processing method provided by one embodiment of the present disclosure; Figure 2 is a structural diagram of a task processing model in a task processing method provided by one embodiment of the present disclosure; Figure 3 is a structural diagram of an arithmetic unit in a task processing method provided by one embodiment of the present disclosure; Figure 4 is a schematic diagram of attention calculation in a task processing method provided by one embodiment of the present disclosure; Figure 5 is a flowchart of an object recommendation method provided by one embodiment of the present disclosure; Figure 6 is a flowchart of a model training method provided by one embodiment of the present disclosure; Figure 7 is a schematic diagram of the structure of a task processing device provided by one embodiment of the present disclosure; Figure 8 is a schematic diagram of the structure of an object recommendation device provided by one embodiment of the present disclosure; Figure 9 is a schematic diagram of the structure of a model training device provided by one embodiment of the present disclosure; and Figure 10 is a block diagram of the structure of a computing device provided by one embodiment of the present disclosure. DETAILED DESCRIPTION The following description sets forth numerous specific details to facilitate a thorough 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 may make similar generalizations without departing from the scope of the present disclosure. Therefore, the present disclosure is not limited to the specific implementations disclosed below. The terminology used in one or more embodiments of the present disclosure is for the purpose of describing specific embodiments only and is not intended to limit the present disclosure. As used in one or more embodiments of the present disclosure and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, 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 encompasses any and all possible combinations of one or more of the associated listed items.It should be understood that while terms such as "first" and "second" 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 used solely to distinguish information of the same type from one another. For example, "first" could also be referred to as "second," and similarly, "second" could also be referred to as "first," without departing from the scope of one or more embodiments of the present disclosure. Depending on the context, the term "if" as used herein could be interpreted as "at the time of," "at the time of," or "in response to a determination." Furthermore, 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, and displayed data, etc.) involved in one or more embodiments of the present disclosure are all authorized by the user or fully authorized by all parties. The collection, use, and processing of such data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding entry points are provided for users to choose to authorize or deny such data. In one or more embodiments of the present disclosure, a "large model" refers to a deep learning model with large-scale model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than ten trillion model parameters. Large models, also known as foundation models, are pre-trained on large-scale unlabeled corpora to produce pre-trained models with over 100 million parameters. These models are adaptable to a wide range of downstream tasks and have good generalization capabilities. Examples include large language models (LLMs) and multimodal pre-trained models.

[0002] (multi-modal pre-training model). In practical applications, large models only require a small number of samples to fine-tune the pre-trained model and can be applied to different tasks. Large models can be widely used in fields such as natural language processing (NLP) and computer vision. Specifically, they can be applied to computer vision tasks such as visual question answering (VQA), image captioning (IC), and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. Major application scenarios for large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. First, the terms used in one or more embodiments of this disclosure are explained. Structured data: Data with a specific format, typically presented in a table format in a database, where rows and columns clearly define data attributes. Rows represent a type, and columns represent data, including numerical data and categorical data. Deep Learning: A branch of machine learning, it is an algorithm that uses artificial neural networks as its architecture to learn data representations. Alternative Algorithms (Iterative Ensemble Learning Methods): A machine learning strategy for iteratively building powerful predictive models by iteratively training and combining a series of "weak" learners to form a "strong" learner. In each iteration, the new weak learners focus on information or errors that were not fully captured by the previous model, specifically those data points that performed poorly in the previous iteration. Convolutional Neural Networks (CNN) Model: A multi-layer deep learning model with forward and backpropagation, featuring convolutional filters that process feature data. Recurrent Neural Networks (RNN) Model: A recursive deep learning model that processes vector representations recursively, with intermediate layers connected in a chain-like fashion. Long Short Term Memory (LSTM) model: A deep learning model that has the ability to memorize long-term and short-term information and has a convolution filter for processing feature data.Deep Self-Attention Model (Transformer Model): A deep learning architecture based on the attention mechanism for processing sequential data, such as natural language. Bidirectional Encoder Representations from Transformers (BERT Model): A specialized Transformer model trained using a bidirectional Transformer encoder and large-scale unlabeled text data.

[0003] XGBoost (extreme Gradient Boosting): A machine learning algorithm that belongs to the gradient boosting decision tree

[0004] An optimized implementation of Gradient Boosting Decision Trees (GBDT). Attention: A technique that enables a model to selectively focus on important parts of an input sequence based on contextual information and adjust the weights of the output representation accordingly. When processing long sequences, it emphasizes important segments of the input sequence by calculating an attention weight (a weighted coefficient distribution), rather than treating all input units (tokens) equally. The attention mechanism can be divided into two steps: 1. Calculate the attention weight QKL using the query information Q and the key information K. 2. Calculate the weighted average of the input sequence based on the attention weight QKT. The specific calculation formula of the attention mechanism is shown in Formula 1:

[0005] Attention = softmax ({Shou - V) Formula 1 Where Attention is the attention feature, Q is the query information, K is the key information, V is the value information, The scaling parameter, softmax(), is the normalization of softmax. When the query and key information are based on the same input, the attention mechanism is a self-attention mechanism. When the query and key information are based on different inputs, the attention mechanism is a cross-attention mechanism.

[0006] CPC (Cost Per Click): Get corresponding income for each click, CPC=consumption / click volume.

[0007] CPA (Cost Per Action): Each action receives corresponding revenue, CPA=CPC / CVR.

[0008] CVR (Conversion Rate): Conversion rate = conversion volume / click volume.

[0009] CTR (Click-Through-Rate): Click-Through rate = number of clicks / number of impressions.

[0010] ROI (Return on Investment) refers to the value returned from investment, measuring the input-output ratio of advertising. Return on investment = order amount / consumption = average amount per order / CPA

[0011] The Top-K screening strategy is widely used in fields such as information retrieval, recommendation systems, and ranking optimization. Its core concept is to identify the K most important, highest-scoring elements (or items) within a given dataset or list of scores. Currently, deep learning models or alternative algorithmic models are used to learn the data features (Xi, X2, X3…) of various types of task data within structured data. These features are then weighted to produce arithmetic interaction features (X = aXi + [3X2 + YX3…]). Generalized additive models enhanced by deep learning algorithms are also used.

[0012] Generalized Additive Models (GAMs) enhance the feature representation capabilities of data features for various types of task data (Xj, X2\ X3'...), but essentially only enhance the feature representation of data features for a single type of task data. Weighted processing based on the data features of each type of task data is still required to obtain arithmetic interaction features (X' = aXi' + 6X2' + YX3'...). o Combining deep learning models with tree models combines the advantages of deep learning models' deep feature extraction capabilities with the advantages of tree models' feature sparsity when processing structured data. However, essentially, they still enhance the data features of data for a single type of task data (X", \2", X3"...), and still require weighted processing based on the data features of each type of task data to obtain arithmetic interaction features (X" = aXi" + 6X2" + YX3"...). o However, multi-type task data that conforms to a specific format not only requires learning the data features of single-type task data, It is also necessary to learn the interactive computational features of the data features of multiple types of task data. For example, if the target task is a motion analysis task, the structured data for the target task includes multiple types of task data that conform to a feature format, including speed and time. Motion analysis requires not only analyzing the data features of a single type of task data, such as average speed, speed probability distribution, average motion time, and speed extremes, but also mining the arithmetic interactive features of the data features of multiple types of task data, such as acceleration and displacement distance, to improve the accuracy of motion analysis. How to mine the arithmetic interactive features of the data features of multiple types of task data to improve the accuracy of task processing is an urgent problem that needs to be solved. To address this problem, the present disclosure provides a task processing method. This method extracts data features (X1, X2, X3, etc.) of multiple types of task data that conform to a specific format. Based on at least two interactive operations, feature analysis is performed to obtain interactive features corresponding to each interactive operation (X1, X3, X236, X3922). Based on the interactive features corresponding to each interactive operation, the method extracts the target interactive features (X = 1, X2, X3922) adapted to the multi-type data in the specific format. The arithmetic interaction feature F(X", X236, X3,9,22>) is a highly accurate data feature. This feature improves task processing accuracy by integrating the target interaction feature into the task. This disclosure also relates to an object recommendation method, a task processing apparatus, an object recommendation device, a computing device, a computer-readable storage medium, and a computer program product, each of which is described in detail in the following embodiments.Referring to Figure 1, a flowchart of a task processing method provided by one embodiment of the present disclosure is shown, comprising the following specific steps: Step 102: Obtain structured data for a target task, wherein the structured data includes multiple types of task data conforming to a specific format. This embodiment of the present disclosure is applicable to applications, websites, or mini-programs with task processing functions, including but not limited to project management tools, data analysis websites, search engines, advertising recommendation tools, product recommendation tools, e-commerce platforms, and health management mini-programs. A target task is a data processing task to be performed based on structured data. For example, in a project management tool, the target task is to generate a project schedule for a new project. In another example, in a search engine, the target task is to return relevant result pages based on a user's query keywords. In another example, in an advertising recommendation tool, the target task is to accurately push advertising content that matches the interests and preferences of a specific user. In another example, in a data analysis website, the target task is to analyze target channels and resource allocation plans for advertising delivery. The structured data of a target task is structured data used to execute the target task and serves as input data for executing the target task. The structured data of a target task is formatted data that can be presented in a table format to enable structured data access. For example, in a project management tool, the structured data of a target task includes structured data such as project content, personnel arrangements, and schedules. Another example is a search engine, where the structured data of a target task includes structured data such as query keyword tables, page keywords, and page links. Another example is an advertising recommendation tool, where the structured data of a target task includes structured data such as user data tables, user profile data, and advertising information tables. Another example is a data analysis website, where the structured data of a target task includes structured data such as a summary table of advertising distribution channels, the amount of advertising distribution channel resources, and the number of user views of the target application. A specific format is the format in which multiple types of task data are formed within structured data. Specifically, it is a pre-determined standard format for structured data that ensures correct data analysis and processing. It should be noted that the specific format is the basis for the feature analysis based on at least two interactive operations in the subsequent step 106. Only the specific format can clearly specify which type of task data is to be processed. For example, multi-type task data includes velocity and time. Only with a specific format (velocity in column 1, time in column 2) can the data in columns 1 and 2 be divided to obtain acceleration data, and the data in columns 1 and 2 be divided to obtain displacement data. Multi-type task data is a combination of different types of task data used to execute a target task and serves as input data for executing the target task.Different types of task data form specific formats, such as rows with types and columns with data, or rows with types and columns with data. Each type of task data has a corresponding attribute type and attribute data within that attribute type. Attribute data can be categorized as numerical data or categorical data. For example, in a project management tool, multiple types of task data include: a project basic information table, which contains task data such as project name, start date, and end date; a personnel information table, which contains task data such as employee name, position, and skill tags; and a task list, which contains task data such as task description, responsible person, deadline, and priority. Another example is a search engine, which contains web page records in an index, which contain data such as web page links, page content, keywords, and keyword weights. For example, in an ad recommendation tool, multiple types of task data include: a user profile table containing task data such as age, gender, region, browsing history, and purchase history; an ad creative library containing task data such as ad titles, images, links, and ad categories. For another example, on a data analysis website, multiple types of task data include: an ad distribution channel statistics table containing task data such as ad distribution channel, ad distribution channel click-through rate, and ad distribution channel conversion rate; and target application user page views and user behavior data. Obtaining structured data for a target task can involve obtaining structured data uploaded by users for the target task, automatically capturing or generating the required structured data based on data analysis rules, or synchronized or periodically pulled from different data sources via an application programming interface (API), without limitation. Optionally, after obtaining the structured data for the target task, the process further includes preprocessing the structured data. Data preprocessing includes, but is not limited to, data padding, data standardization, data conversion, data merging, and abnormal data removal. For example, on a data analysis website for applications, accurate user recommendations are made by analyzing application usage data and user data of potential users.The app developer uploads app usage data to a data analysis website and obtains user data of potential users from the website. The structured data required for generating user recommendations includes a data table in a specific format (data table format: M types, 1000 data items per type). This data table includes 10 types: consumption, clicks, conversions, impressions, and order amounts, as well as data for the past quarter for these 10 types. The developer then obtains structured data for the target task, including multi-type task data in a specific format. This lays the foundation for subsequent data feature extraction and provides specific format support for subsequent interaction feature analysis based on at least two interactive operations. Step 104: Extract data features for the multi-type task data. The data features of the multi-type task data are quantized feature sequences (feature vectors) of the multi-type task data. Extracting these features lays the foundation for subsequent feature analysis of the corresponding interactive operations between the multi-type task data. Data features are low-dimensional feature vectors used for feature preprocessing of interactive operations or deep learning. Data features can include embedding features, hash coding features, and LDA features (Linear Discriminant Analysis Features). Extracting data features from multi-type task data can be achieved through table lookup, for example, querying an embedding table to extract data features from the multi-type task data. Alternatively, high-dimensional, discrete multi-type task data can be converted into a low-dimensional, continuous feature sequence representation through linear transformation, without limitation. For example, a d-dimensional linear transformation is performed on 10 types of numerical data in a data table. Then, an embedding table is queried for the 10 types of category data in the data table to obtain a 10xd-dimensional matrix. The embedded feature sequence Xo for the 10 types of data extracts data features from the multi-type task data, providing a quantitative feature foundation for subsequent feature analysis. Step 106: Based on at least two interactive operations, interactive feature analysis is performed on the data features of the multi-type task data to obtain interactive features corresponding to each interactive operation. Interactive operations are operations that perform mathematical operations, logical reasoning, or other complex calculations on the relationships between different data items. Interactive operations are used to reveal the deep features hidden behind the original data and achieve more accurate feature analysis.Interactive operations include first-order interactive operations such as addition, subtraction, negation, extreme value, mean, sum, OR, and XOR, as well as higher-order interactive operations such as multiplication, division, competition, exponential operation, and logarithm. Interactive operations corresponding to multi-type task data include first-order interactive operations between single types, multi-order interactive operations between single types, first-order interactive operations between multiple types, and higher-order interactive operations between multiple types. The interactive features corresponding to various interactive operations are deep feature sequences determined by the interactive operations between different data items. These are new deep features generated based on the data characteristics of multi-task data. Interactive features uncover data relationships or patterns not directly apparent in the original task data. For example, in a financial risk control model, a user's credit score and recent login frequency are weighted and summed to generate a new feature that comprehensively evaluates the user's account activity and credit level. This new feature can help determine the risk level of a loan application. For example, in the product recommendation scenario of an e-commerce platform, a user's purchase history and the currently browsed product category can be analyzed for similarity using cosine similarity in interaction feature analysis, thereby generating interaction features that indicate user interest matching. Another example is the dot product of vehicle speed and driving direction in the intelligent transportation field, which can generate an interaction feature representing the intensity of vehicle travel in a specific direction, helping to optimize route planning or predict traffic flow. Interaction feature analysis is an operation that deeply explores the hidden data interaction relationships and patterns between multi-type task data. By combining, transforming, and abstracting raw data features, deeper interaction data features are extracted, resulting in new data features that more accurately reflect the structure between the data. Based on at least two interaction operations, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to each interaction operation. Specifically, based on the data features of the multi-type task data, feature attention calculation is performed on the corresponding interaction operations between the multi-type task data to obtain interaction features corresponding to each interaction operation. Optionally, feature attention calculations are performed on the corresponding interaction operations between multiple types of task data using computational channels corresponding to at least two interaction operations, obtaining interaction features corresponding to each interaction operation. Feature attention calculation is a feature analysis based on the attention mechanism. For the specific calculation formula, see Formula 1 and related content. For example, the computational channels corresponding to the two interaction operations are: a first-order attention stream and a high-order attention stream. Ten prompt features P are allocated to each attention stream, each corresponding to a d-dimensional vector, resulting in a 10 x d-dimensional matrix.Based on the embedded feature sequence X of 10 types of data, key information K = XWK and value information V = XW are generated. V Where WK is the pre-trained parameter matrix (d x d dimensions) for key information, and WV is the pre-trained parameter matrix (d x d dimensions) for value information. Attention calculations are performed on the corresponding first-order and higher-order interaction operations between the 10 types of data to obtain interaction features C)A (10 x d dimensions) and C)M (10 x d dimensions) corresponding to the two interaction operations. Based on at least two interaction operations, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to each interaction operation. This fully explores the data relationships and patterns between the data features of the multi-type task data, laying the foundation for the subsequent extraction of target interaction features corresponding to at least two interaction operations for multi-type task data adapted to a specific format. Step 108: Based on the interaction features corresponding to the various interaction operations, the target interaction features corresponding to the target task are determined. The target interaction feature corresponding to the target task is a deep feature sequence determined based on the interaction features corresponding to various interactive operations. It can be a deep data feature that better represents the structured data of the target task from the interaction features corresponding to various interactive operations, or it can be a deep data feature that better represents the structured data of the target task obtained by feature fusion of the interaction features corresponding to various interactive operations. The target interaction feature is an arithmetic interaction feature adapted to multi-type task data in a specific format. For example, in advertising delivery analysis, interaction features include CPC feature Featurecpc, CPA feature FeaturecpA, CVR feature Feature CV R> CTR Feature CTR and ROI feature RO j, the target interaction feature is the feature obtained by weighting CPC feature, CPA feature, CVR feature, CTR feature and ROI feature: F (Featurecpc, Feature CPA , Feature CV R, Feature CTR , Feature ROj), this feature represents a more comprehensive and accurate feature of advertising delivery. In an optional embodiment of the present disclosure, step 108 includes the following specific steps: determining a target interaction feature corresponding to the target task from the interaction features corresponding to various interaction operations. For example, from the interaction features C)A (10×d dimensions) and C)M (10×d dimensions) corresponding to two interaction operations, interaction feature ()M is determined as the target interaction feature O (10×d dimensions). From the interaction features corresponding to various interaction operations, the target interaction feature corresponding to the target task is determined. Determining the interaction feature that better represents the structured data of the target task improves the accuracy of the target interaction feature and the accuracy of subsequent task processing. In an optional embodiment of the present disclosure, step 108 includes the following specific steps: performing feature fusion on the interaction features corresponding to various interaction operations to obtain the target interaction feature corresponding to the target task. Feature fusion includes concatenation, normalization, weighting, full connection, residual connection, feedforward, concatenation, and superposition (addition). For example, the interaction features C)A (10 x d dimensions) and 0M (10 x d dimensions) corresponding to the two interaction operations are concatenated to obtain a 20 x d matrix. The full connection is then transposed and integrated to obtain the target interaction feature 0 corresponding to the target task: FC(VConcat(OA, (}M)T)T, where FC() represents full connection processing, Concat () represents concatenation processing, and O represents the target interaction feature (10 x d dimensions). The interaction features corresponding to various interaction operations are fused to obtain the target interaction feature corresponding to the target task. This obtains interaction features that comprehensively represent the structured data of the target task, improving the accuracy of the target interaction feature and the accuracy of subsequent task processing. Based on the interaction features corresponding to various interaction operations, the target interaction feature corresponding to the target task is determined. The target interaction feature, an arithmetic interaction feature adapted to a specific format and extracted from multi-type data, is a highly accurate data feature, laying the foundation for deep features for subsequent target task execution. Step 110: Execute the target task based on the target interaction feature to obtain the task processing result of the target task. The task processing result of the target task is the specific output or decision result obtained by executing the target task based on the target interaction feature. For example, in a project management tool, the task processing result of the target task is the project schedule for the new project. In another example, in a search engine, the task processing result of the target task is the returned relevant result page. For example, in an advertising recommendation tool, the target task's task processing result is the pushed advertising content. For another example, on a data analysis website, the target task's task processing result is the target channel and delivery resource allocation plan. Steps 104-108 can be considered a form of feature engineering, and step 110 is a downstream task performed based on the feature engineering results. It should be noted that to optimize the accuracy of the extracted target interaction features, steps 104-108 can be iteratively performed to execute the target task based on the target interaction features and obtain the target task's task processing result. Specifically, a task processing model is used to execute the target task based on the target interaction features and obtain the target task processing result. The task processing model is a deep learning model with task processing capabilities, including but not limited to CNN models, RNN models, LSTM models, Transformer models, BERT models, and large language models. For example, the large language model is used to perform precise user recommendations based on the target interaction features O (10 x d dimensions) to obtain precise users for the application app. The developers of the application app complete the application delivery to precise users, achieving precise application delivery.In an embodiment of the present disclosure, data features of multi-type task data conforming to a specific format are extracted. Based on at least two interactive operations, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to each interactive operation. Based on the interaction features corresponding to each interactive operation, a target interaction feature is extracted, which is an arithmetic interaction feature adapted to the multi-type data in the specific format. This is a highly accurate data feature. Executing the target task based on the target interaction feature improves task processing accuracy. In an optional embodiment of the present disclosure, step 106 includes the following specific steps: determining query information corresponding to each interactive operation based on prompt features corresponding to the at least two interactive operations, and determining key information and value information for each type of task data based on the data features of the multi-type task data. The prompt feature is used to prompt feature analysis of the corresponding interactive operations between the multi-type task data; and feature calculation is performed on the query information corresponding to each interactive operation and the key information and value information of each type of task data to obtain interaction features corresponding to each interactive operation. The prompt feature is a feature vector of a prompt function pre-trained in each interaction feature analysis, and is used to prompt feature analysis of the corresponding interactive operations between the multi-type task data. Pre-trained prompt features capture the data interactions between multi-type task data in a specific format. This, combined with the specific format, ensures correct data analysis and processing. For example, the prompt feature P is a 10 x d-dimensional matrix, where each d-dimensional vector represents guidance information for a type of task data, helping the model understand multi-type task data or the data relationships between them. Optionally, the feature size of the prompt feature is N, which is a sum of the pre-set number Np and the feature dimension d of the multi-type task data features, i.e., N. p xd matrix. The introduction and number of hint features address the time and memory inefficiency of the attention mechanism when faced with a large number of features. This reduces the model complexity from quadratic to linear in the number of input features, enabling better scaling for large-scale data and significantly improving its ability to process large-scale data. Furthermore, hint features ignore unimportant correlations, preventing overfitting and improving robustness to data noise. Query information is a feature vector used as query information in the attention mechanism. It is determined based on hint features and is used to find more relevant key information. Key information is a feature vector used as key information in the attention mechanism. It is generated by transforming model parameters from data features of multiple types of task data. The key information is compared with the query information to determine the data relevance between different data items. For example, key information K = XW K, where X is the embedded feature sequence of the 10 types of data, and WK is the pre-trained parameter matrix of the key information. The key information is used to match the query information Q and measure the data relevance between the 10 types of data. The value information is the feature vector used as the value information in the attention mechanism. It is generated by transforming the data features of the multi-type task data through the model parameters. The value information does not directly participate in the data relevance calculation. Instead, when a key information highly matches the query information, the corresponding value information will be selected and aggregated into the final result. For example, the value information V = XW V , where X is the embedded feature sequence of 10 types of data, WV is the pre-trained parameter matrix for value information, and value information V contains the "high-value" data features that actually need to be passed to subsequent stages. It should be noted that the attention mechanism in the disclosed embodiment, by constructing query information, key information, and value information, and performing feature attention calculations on at least two interactive operations, can mine the complex data interaction relationships between multiple types of task data, thereby generating interactive features that accurately reflect the corresponding interactive operations. Feature calculations are performed on the query information corresponding to each interactive operation, and the key information and value information of each type of task data, to obtain interactive features corresponding to each interactive operation. Specifically, a Top-K screening strategy is used to perform feature calculations on the query information corresponding to each interactive feature, and the key information and value information of each type of task data, to obtain interactive features corresponding to each interactive operation. For example, the two interactive operations correspond to the following computing channels: the first-order attention flow and the high-order attention flow. Based on the 10 prompt features P corresponding to the first-order attention flow, the query information corresponding to each interactive feature is determined, and based on the embedded feature sequence X of the 10 types of data, the key information K = XW is generated. K Sum value information V = xw v Based on the 10 prompt features P corresponding to the high-order attention flow, the query information corresponding to each interaction feature is determined, and based on the embedded feature sequence X of the 10 types of data, the key information K = log (X) WK and the value information V = log (X) WV are generated. Using the Top-K screening strategy, the corresponding first-order attention calculation is performed on the query information P corresponding to each interaction feature, the key information K and the value information V of the 10 types of data. A = softmax ( (topK ((PK AT) / Vd) V) and high-order attention calculation ()M = esoftmax ( (topK ((PK, T) / VH) V) ), to obtain interaction features (iQ xd dimensions) and ()M (10 xd dimensions) corresponding to the two interaction operations. In the embodiment of the present disclosure, the complex data interaction relationship between multiple types of task data is mined through prompt features, thereby generating interaction features that can accurately reflect the corresponding various interaction operations, reducing complexity, preventing overfitting, and improving the efficiency of task processing. In an optional embodiment of the present disclosure, feature calculation is performed on the query information corresponding to each interaction operation, the key information and value information of each type of task data, to obtain interaction features corresponding to the various interaction operations, including the following specific steps: calculating initial weight information based on the query information corresponding to the first interaction operation and the key information of each type of task data, where the first interaction operation is any interaction operation; determining a target weight that reaches a preset threshold from the initial weight information, and determining target weight information based on the target weight; and performing weighted processing on the value information of each type of task data based on the target weight information to obtain the interaction feature corresponding to the first interaction operation. Initial weights are temporary weights generated during the calculation of any interaction operation in the attention mechanism based on the query and key information. These weights are typically represented in matrix form, typically with the same number of rows as the task data and a fixed number of columns. Each element represents the data relevance between the query and the corresponding key information. For example, for 10 task data types, the initial weights are a 10x10 matrix, where each element (i, j) represents the data relevance between the query vector Qi and the key vector Kj for the jth task data type. The preset threshold is a pre-set value or quantity used to filter out more important weights from the initial weights. In practical applications, this threshold is used to filter out key information that is not sufficiently relevant to the query, thereby focusing on the elements that have a greater impact on the results. For example, if the threshold is set to 0.7, only query information with a key information relevance greater than or equal to 0.7 will be included in the target weight. The target weight is the data relevance that exceeds the preset threshold, filtered out from the initial weights. The target weight is a subset extracted from the initial weight information based on a preset threshold. It only contains those target weights that pass the threshold screening. It is usually in the form of a matrix. This matrix also has the same number of rows as the type task data, but the number of columns may be reduced to less than the number of original key information because not all key information meets the threshold condition.For example, in a high-order interactive operation scenario, after threshold filtering of the initial weight information, the target weights may include only the high-weight values ​​for a few types of task data most relevant to the query information, or may include the high-weight values ​​for a few types of task data most relevant to the query information. Other weights are set to large negative numbers and subsequently normalized to zero. For example, for the computational channels corresponding to any interactive operation: the first-order attention stream and the higher-order attention stream, the query information corresponding to each interactive feature is determined based on the 10 prompt features P corresponding to the first-order attention stream. Key information K = XW is generated based on the embedded feature sequence X of the 10 types of data. K Sum value information V = xw v , based on the 10 prompt features P corresponding to the high-order attention flow, determine the query information corresponding to each interaction feature, and based on the embedded feature sequence X of the 10 types of data, generate key information K = log (X) WK and value information V = log (X) W v o Calculate the initial weight PK based on the query information P corresponding to the interaction feature and the key information K of each type of task data A T, from the initial weights, determine the weights that reach the preset K target weights, the remaining weights are set to large negative numbers, and determine the target weights (PK A T) ', based on the normalized target weight softmax (PK A T) ', weighted processing is performed on the value information of each type of task data to obtain the interaction features corresponding to the two interaction operations 0A = softmax ( (topK ((PK T )7Vd) V) and C)M = e softmax ( (topK C(PK T)7Vd) v). In the disclosed embodiment, the Top-K strategy is used to more effectively aggregate and mine the complex data interaction relationships between multi-type task data, strengthen highly correlated interaction features, reduce complexity, prevent overfitting, and improve task processing efficiency. In an optional embodiment of the disclosed embodiment, the at least two interaction operations include first-order interaction operations and higher-order interaction operations. Accordingly, step 106 includes the following specific steps: performing interaction feature analysis on the data features of the multi-type task data based on the first-order interaction operations to obtain interaction features corresponding to the first-order interaction operations; and performing interaction feature analysis on the data features of the multi-type task data based on the higher-order interaction operations to obtain interaction features corresponding to the higher-order interaction operations. First-order interaction operations are operations that perform first-order mathematical operations, logical reasoning, or other complex calculations on different data in the multi-type task data, including but not limited to addition, subtraction, negation, extreme value, mean, sum, or, and exclusive-or operations. High-order interactive operations are high-order mathematical operations, logical reasoning or other complex calculations between different data in multi-type task data. This operation mainly focuses on the data relationship between multiple feature dimensions, including but not limited to This includes, but is not limited to, exponential processing, multiplication, geometric mean, cumulative distribution processing, and probability density processing. After order reduction, the interaction feature becomes a sparser feature. When performing high-order interaction operations, to simplify the processing, reduce computational complexity, or improve analysis efficiency, order reduction is used to convert the data features of the task data to a lower-dimensional interaction dimension. The reduced-order interaction features generated in this process are the reduced-order interaction features corresponding to the high-order interaction operations. Based on the data features of multi-type task data, feature analysis is performed on the high-order interaction operations between the multi-type task data to obtain the reduced-order interaction features corresponding to the high-order interaction operations. Specifically, based on the data features of the multi-type task data, feature attention calculation is performed on the high-order interaction operations between the multi-type task data to obtain the reduced-order interaction features corresponding to the high-order interaction operations. Optionally, when the order reduction process is logarithmic processing, a ReLU function is applied and a small positive number epsilon is added to each element in the data features of the multi-type task data to avoid taking the logarithm of zero elements, which would result in an infinite order reduction result. For example, in the high-order attention flow, the embedded feature sequence X of 10 types of data is logarithmically processed log (RELU (X) + epsilon), and based on the embedded feature sequence log (RELU (X) + epsilon) of 10 types of data, key information K = log (RELU (X) + epsilon) and value information V = log

[0013] (RELU (X) + epsilon) W v , where WK is the pre-trained parameter matrix (dXd dimensions) for key information and WV is the pre-trained parameter matrix (dXd dimensions) for value information. The corresponding high-order attention calculations are performed on the corresponding high-order interaction operations between the 10 types of data: softmax ((topK ((PK T )yVd) V), obtain the reduced-order interaction features corresponding to the high-order interaction operation, perform exponential processing on the reduced-order interaction features corresponding to the high-order interaction operation, and obtain the interaction features corresponding to the high-order interaction operation C)M = e softmax ( (topK ((PK T )' / Vd) V) oIn an embodiment of the present disclosure, by performing rank-up and rank-down processing before and after feature analysis, the efficiency and accuracy of feature analysis are improved. In an optional embodiment of the present disclosure, the multi-type task data includes numerical task data and category task data. Accordingly, step 104 includes the following specific steps: performing data feature conversion on the numerical task data to obtain data features of the numerical task data; and querying a preset category embedding table based on the category task data to obtain data features of the category task data. Multi-type task data includes numerical task data and category task data, requiring differentiated feature extraction. Numerical task data is task data of a numerical data type, which is either a continuous variable or a discrete variable, and is quantitative data with a clear numerical meaning. For example, this includes the purchase amount and browsing time of users on e-commerce platforms. Another example is the frequency of user clicks on web pages in search engines. In financial risk control, this includes customer credit scores and repayment histories. For example, in advertising analysis, CPC (cost per click), CPM (cost per thousand impressions), and CTR (click-through rate) are all typical numerical task data. Category task data is task data of the category data type and is non-quantitative, lacking a clear numerical meaning. For example, in product recommendation scenarios on e-commerce platforms, product categories (such as "electronics," "apparel," and "home furnishings") serve as category task data. Data feature transformations include, but are not limited to, linear and nonlinear transformations. Linear transformations include, but are not limited to, feature mapping and normalization. Through data feature transformation, data of varying scales or distributions can be addressed, allowing for standardization of individual data features. The category embedding table is a pre-defined embedding lookup table for category task data. By querying the embedding table, category task data can be converted into quantitative features, laying the foundation for subsequent feature analysis. For example, a d-dimensional linear transformation is performed on the 10 types of numerical data in the data table. By querying the embedding table for the 10 types of category data in the data table, a 10xd-dimensional matrix is ​​obtained, which is the embedded feature sequence X of the 10 types of data. In the embodiment of the present disclosure, feature extraction is performed for different task data, laying the foundation for subsequent feature analysis. In an optional embodiment of the present disclosure, step 106 includes the following specific steps: inputting data features into the encoding layer of the task processing model, and using the arithmetic unit of the encoding layer to perform interactive feature analysis on the data features of the multiple types of task data based on at least two interactive operations, to obtain interactive features corresponding to the various interactive operations.The encoding layer is a model layer used for feature analysis and processing, implementing deep feature mining. For example, methods such as addition, multiplication, and attention mechanisms are used to analyze data relationships between multiple types of task data and discover deep-level interactive features. The arithmetic unit is a model unit used for feature analysis, used to mine deep features, namely, interactive features corresponding to various interactive operations. The arithmetic unit includes first-order arithmetic units and / or high-order arithmetic units. It should be noted that, when the task processing model adopts the Transformer model architecture, the arithmetic unit replaces the Multi-Head Attention (MHA) unit in the Transformer model architecture. The specific method has been described in detail in step 106 above and will not be repeated here. For example, the embedded feature sequence X is input into the arithmetic unit of the model. The computational channels corresponding to the two interactive operations in the arithmetic unit are the first-order attention stream and the high-order attention stream. Each attention stream is assigned 10 prompt features P, each corresponding to a d-dimensional vector, resulting in a 10 x d-dimensional matrix. Based on the embedded feature sequence X of 10 types of data, key information K = XW is generated. K Sum value information V = XW VWhere WK is the pre-trained parameter matrix (d x d dimensions) for key information, and WV is the pre-trained parameter matrix (d x d dimensions) for value information. Corresponding attention calculations are performed on the corresponding first-order and high-order interaction operations between the 10 types of data, obtaining interaction features (d x d dimensions) and (M) (10 x d dimensions) corresponding to the two interaction operations. In the disclosed embodiment, the arithmetic unit performs efficient and accurate feature analysis, thereby achieving efficient and accurate task processing. Optionally, step 104 includes the following specific steps: inputting multi-type task data into the embedding layer of the task processing model and extracting data features from the multi-type task data. The task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit. Correspondingly, step 108 includes the following specific steps: inputting interaction features corresponding to various interactive operations into the feature processing unit, and determining target interaction features corresponding to the target task based on the interaction features corresponding to the various interactive operations. Correspondingly, step 110 includes the following specific steps: inputting the target interaction features into the decoding layer, executing the target task based on the target interaction features, and obtaining a task processing result for the target task. The task processing model is a deep learning model with task processing capabilities, including but not limited to CNN models, RNN models, LSTM models, Transformer models, BERT models, and large language models. The task processing model can process multi-type task data. Through steps such as feature extraction, interactive operations, and feature fusion, it generates highly accurate target interaction features. These features are then used to execute target tasks, such as project plan generation, search result ranking, advertising recommendation, or data analysis. Accordingly, the task processing model includes a project plan generation model, a search result ranking model, an advertising recommendation model, or a data analysis model. The embedding layer is a model layer used for low-dimensional feature extraction (feature embedding). For example, in natural language processing, the word embedding layer in a text processing model maps words to vectors in a low-dimensional space. When processing multi-type task data, the embedding layer converts category data (such as product categories and user tags) into numerical vectors, facilitating interactive operations and feature combination in subsequent model layers. The feature processing unit is a model unit used for feature processing. It performs further feature processing on the interaction features corresponding to various interactive operations, including but not limited to feature selection, order reduction, fusion, residual connections, and feedforward processing, to determine the target interaction features that best reflect the arithmetic interaction characteristics of the structured data for the target task.It should be noted that when the task processing model adopts a Transformer model architecture, the feature processing unit can retain residual connections and feedforward layers to enhance gradient flow and feature expression capabilities. The decoding layer is the model layer used to execute the target task. It decodes the processed target interaction features into the task processing results of the target task, which can be classification, regression, generation, or other output forms. For example, in a precise user recommendation scenario, the decoding layer generates a specific recommendation list based on the target interaction features. In another example, in a project management tool, the decoding layer generates a new project plan based on the target interaction features. In the disclosed embodiment, the specific methods of each step have been detailed in steps 104 to 110 above and will not be repeated here. For example, the 10 types of numerical data from the data table are input into the embedding layer of the modified Transformer model, where a d-dimensional linear transformation is performed. By querying the embedding table, a 10xd-dimensional matrix is ​​obtained for the 10 types of category data in the data table. The embedded feature sequence X for the 10 types of data is input into the arithmetic unit of the model. The two interactive operations in the arithmetic unit correspond to computational channels: the first-order attention stream and the high-order attention stream. Each attention stream is assigned 10 prompt features P, each corresponding to a d-dimensional vector. The prompt features are a 10xd-dimensional matrix. Based on the embedded feature sequence X for the 10 types of data, key information K = XW is generated. K Sum value information V = XW V , where WK is the parameter matrix (d × d dimensions) pre-trained for key information, and WV is the parameter matrix (d × d dimensions) pre-trained for value information. The corresponding attention calculations are performed on the corresponding first-order and high-order interaction operations between the 10 types of data, and the interaction features 0A (10 × d dimensions) and 0M corresponding to the two interaction operations are obtained.

[0014] (10 x d dimensions). The interaction features C)A and C)M corresponding to the two interaction operations are input into the feature processing unit of the model. The interaction features OA (10 x d dimensions) and OM (10 x d dimensions) corresponding to the two interaction operations are concatenated to obtain a 20 x d dimension matrix. The full connection is processed, and the transposition operation and integration are performed to obtain the target interaction feature corresponding to the target task 0 = FC(VConcat(O A , 0 M ) T ) T oThe target interaction unit is input into the model's decoding layer. Based on the target interaction features O (10 x d dimensions), precise user recommendations are performed to obtain precise users for the application app. In this disclosed embodiment, the task processing model, comprising an embedding layer, an arithmetic unit, a feature processing unit, and a decoding layer, enables efficient and accurate feature extraction, feature analysis, feature processing, and task execution. This deep learning model achieves efficient and accurate task processing. In an optional embodiment of the present disclosure, the arithmetic unit includes a first-order arithmetic unit and a high-order arithmetic unit. Accordingly, data features are input into the encoding layer of the task processing model. The arithmetic unit of the encoding layer performs interaction feature analysis on the data features of multiple types of task data based on at least two interaction operations to obtain interaction features corresponding to each interaction operation. The arithmetic unit includes the following specific steps: inputting the data features into the first-order arithmetic unit, performing interaction feature analysis on the data features of the multiple types of task data based on the first-order interaction operation to obtain interaction features corresponding to the first-order interaction operation; inputting the data features into the high-order arithmetic unit, performing interaction feature analysis on the data features of the multiple types of task data based on the high-order interaction operation to obtain interaction features corresponding to the high-order interaction operation. The arithmetic unit includes two parallel branches: the first-order arithmetic unit and the high-order arithmetic unit, each of which performs feature analysis corresponding to the two interaction operations to obtain interaction features corresponding to the two interaction operations. The first-order arithmetic unit is a branch of the arithmetic unit that performs first-order interactive operations. It defines how multiple types of task data are combined through specific mathematical or logical operations. This unit uses these first-order interactive operations to analyze relative changes, trends, or correlations between these multiple types of task data items. For example, in a project management tool, to analyze the impact of the time schedule of each phase of a project on the overall plan, the first-order arithmetic unit might use methods such as addition (accumulated time), subtraction (time difference), or ratio operations to extract deep data features from these multiple types of task data. The higher-order arithmetic unit is a branch of the arithmetic unit that performs higher-order interactive operations. It defines how multiple types of task data are combined through specific mathematical or logical operations. This unit uses these higher-order interactive operations to analyze data relationships and models between these multiple types of task data. For example, in an advertising recommendation system, the higher-order arithmetic unit is used to simultaneously consider multiple types of task data, such as user personal information, browsing history, and product attributes, and to identify implicit correlations between these multiple types of task data through methods such as tensor multiplication and attention mechanisms. In the embodiment of the present disclosure, the specific manner of each step has been described in detail in the above corresponding embodiment and will not be repeated here. Exemplarily, the data feature is input into the first-order arithmetic unit, and the data feature is input into the high-order arithmetic unit.Each of the 10 hint features P assigned to the first-order and higher-order arithmetic units corresponds to a d-dimensional vector, resulting in a 10 x d-dimensional matrix. In the first-order arithmetic unit, based on the embedded feature sequence X of the 10 data types, key information K = XWK and value information V = XW are generated. V , where WK is the pre-trained parameter matrix (d × d dimensions) for key information and WV is the pre-trained parameter matrix (d × d dimensions) for value information. In the high-order arithmetic unit, based on the embedded feature sequence X of 10 types of data, key information K = log (X) WK and value information V = log (X) W are generated. vWhere WK is the parameter matrix (d x d dimensions) pre-trained for key information, and WV is the parameter matrix (d x d dimensions) pre-trained for value information. In the first-order arithmetic unit, first-order attention calculations are performed on the corresponding first-order interaction operations between the 10 types of data, obtaining interaction features 0A (10 x d dimensions) corresponding to the first-order interaction operations. In the high-order arithmetic unit, high-order attention calculations are performed on the corresponding high-order interaction operations between the 10 types of data, obtaining interaction features 0M (10 x d dimensions) corresponding to the high-order interaction operations. In the disclosed embodiments, parallel feature analysis based on a dual-branch arithmetic unit enables efficient learning of feature arithmetic interactions. This effectively and accurately explores data relationships and patterns between data features of multi-type task data in first-order and high-order interaction operations. This lays the foundation for the subsequent extraction of target interaction features for multi-type task data adapted to specific formats, laying the foundation for interaction features corresponding to first-order and high-order interaction operations. In an optional embodiment of the present disclosure, before inputting multi-type task data into the embedding layer of the task processing model and extracting data features of the multi-type task data, the following specific steps are further included: obtaining a sample set, wherein the sample set includes sample structured data of the sample task and a label task processing result of the sample task, and the sample structured data includes multi-type sample task data conforming to a specific format; inputting the multi-type sample task data into the embedding layer of the task processing model and extracting data features of the multi-type sample task data, wherein the task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; inputting the data features into the arithmetic unit, and performing interaction feature analysis on the data features of the multi-type sample task data based on at least two interaction operations to obtain interaction features corresponding to each interaction operation; inputting the interaction features corresponding to each interaction operation into the feature processing unit, and determining a sample interaction feature corresponding to the sample task based on the interaction features corresponding to each interaction operation; inputting the sample interaction features into the decoding layer, executing the sample task based on the sample interaction features, and obtaining a predicted task processing result for the sample task; and training the task processing model based on the predicted task processing result and the label task processing result. A sample set is a collection of sample data used to train a task processing model. It can be constructed from historical data in historical task processing scenarios, obtained from open-source databases, or manually generated, without limitation. A sample task is a pending data processing task in the sample set used to train the task processing model. For example, in a project management tool, a sample task might be a task instance for generating a new project plan based on structured data from historical projects.For example, in a search engine, a sample task might be a simulated scenario that returns relevant web pages based on a given query. Sample structured data is structured data used to execute the sample task and serves as training data for training the task processing model. Sample task structured data is formatted data that can be presented in a table format to enable structured data access. Multi-type sample task data is multiple types of task data used to execute the sample task and serves as training data for training the task processing model. Multi-type task data forms a specific format. It should be noted that the specific format used in model training is consistent with the specific format used in subsequent task processing. That is, the trained task processing model is specific to multi-type task data that conforms to the specific format. The labeling task processing result is the standard output or decision result obtained from executing the sample task, which is the ideal output or decision result. The prediction task processing result is the output or decision result obtained from executing the sample task, which is a biased output or decision result. This result is compared with the actual labeling task processing result to evaluate the accuracy of the task processing model in its current training state and to update the model parameters accordingly, completing the model training. The task processing model is trained based on the prediction and labeling task processing results. Specifically, a loss value is calculated based on the prediction and labeling task processing results, and the task processing model is trained based on the loss value. The specific training method can be implemented using a gradient update method. For example, by collecting historical data, a sample set is constructed: it contains a large number of user behavior records and their corresponding real-world feedback data (e.g., labels such as clicks and purchases). Each sample consists of a set of structured data, including: Sample task: simulates the ad recommendation process for a specific user. Sample structured data: includes multiple data features, such as user personal information (age, gender, location), browsing history (product category, brand preference, browsing time), shopping cart information (added item ID and quantity), and product attributes (price, rating, sales volume). Labeling task processing results: The label of each sample represents the user's final response to the recommendation, such as actual feedback such as whether the ad was clicked or whether a purchase was made. During model training, these multi-type sample task data are input into the embedding layer of the modified Transformer model. Each raw feature undergoes embedding encoding and is converted into a corresponding embedded feature sequence. In the first-order arithmetic unit, for multi-type user browsing history data, the model calculates the time interval between adjacent browsing events or analyzes the user's behavior patterns of consecutively browsing products in the same category, extracting interaction features corresponding to the first-order interaction operations that reflect the changing trends of user interests.In the high-order arithmetic unit, the model combines multiple types of features to perform complex calculations. In the feature processing unit, the interaction features obtained in the first-order and high-order arithmetic units are integrated to generate sample interaction features. These sample interaction features are then fed into the decoding layer. The decoder generates a personalized advertising recommendation strategy for the target user based on these features and predicts the user's likely reaction to the recommendation list—that is, predicts the task processing result. By comparing the predicted task processing result with the true labeled task processing result, a loss value is calculated. Based on this loss value, the model parameters are updated using the gradient update method using backpropagation, thereby gradually improving the modified Transformer model's ability to make effective recommendations for unknown users in practical applications. In the disclosed embodiments, supervised training of the task processing model improves model performance, laying the foundation for subsequent high-accuracy task processing. In an optional embodiment of the present disclosure, data features are input into an arithmetic unit, and interaction feature analysis is performed on the data features of multiple types of sample task data based on at least two interaction operations to obtain interaction features corresponding to each interaction operation. The method includes the following specific steps: inputting the data features into the arithmetic unit, determining query information corresponding to each interaction operation based on prompt features corresponding to the at least two interaction operations, and determining key information and value information for each type of sample task data based on the data features of the multiple types of sample task data. The prompt features are used to prompt feature analysis of the corresponding interaction operations between the multiple types of sample task data to obtain query information, key information, and value information; performing feature calculation on the query information corresponding to each interaction operation and the key information and value information of each type of task data to obtain interaction features corresponding to each interaction operation; and correspondingly, training a task processing model based on the predicted task processing results and the labeled task processing results. The method includes the following specific steps: adjusting parameters and prompt features of the task processing model based on the predicted task processing results and the labeled task processing results. During model training, the prompt features are randomly generated trainable feature vectors that help the model learn arithmetic interaction rules for multiple types of task data that conform to a specific format. This reduces the model complexity from quadratic to linear in the number of input features, significantly improving its ability to handle large datasets. Furthermore, using hint features helps the model ignore unimportant data correlations, preventing overfitting and improving the model's robustness to data noise.Furthermore, the number of hint features is related to the number of task data types: For structured data with a limited number of types (e.g., no more than 500), consider setting the number of hint features, N_p, equal to the number of types, N. For datasets with a large number of features, consider initially setting the number of hint features, N_p, to 256 or 512 (this is a heuristic range and can be adjusted based on actual conditions). In each arithmetic unit of subsequent iterative processing, consider halving the number of hint features, N_p. This progressively decreasing setting helps reduce computational complexity layer by layer while allowing the model to focus on more abstract and global interactive features at deeper levels. The specific steps for feature attention calculation in the disclosed embodiments have been described in detail in the corresponding embodiments above and will not be repeated here. For example, based on the loss value, the model parameters and hint features are updated through a backpropagation gradient update method, thereby gradually improving the ability of the modified Transformer model to make effective recommendations for unknown users in practical applications. In an embodiment of the present disclosure, by training a cue feature used in attention calculation, the interactive relationships between multi-type task data conforming to a specific format are embedded in the task processing model. Furthermore, the model's time and space complexity increases linearly with the number of features, enabling better scalability to large-scale data. This addresses the inefficiency of the attention mechanism when the number of features is large, as well as the challenge of current data-dependent architectures that struggle to capture interactive features of data. Figure 2 shows the structure of a task processing model in a task processing method provided by an embodiment of the present disclosure. As shown in Figure 2, the embedding layer includes a category embedding table and a linear layer. By inputting category task data into the category embedding table and numerical task data into the linear layer, feature extraction is performed to obtain data features for the multi-type task data.

[0015] The L sequentially connected encoding layers include an arithmetic unit, an overlay layer, a feedforward layer, and an overlay layer, with a residual connection structure. Feature analysis and feature processing are performed in the encoding layer to obtain the target interaction features. Figure 3 shows the structure of the arithmetic unit in a task processing method provided by one embodiment of the present disclosure. As shown in Figure 3, the arithmetic unit includes a first-order arithmetic unit, a high-order arithmetic unit, and a concatenation and fully connected subunit. The first-order arithmetic unit sequentially connects the attention subunit and weighting subunit of the first-order interaction operation in series, while the high-order arithmetic unit sequentially connects the activation and reduction subunits, the attention subunit, weighting subunit, and promotion subunit of the high-order interaction operation in series. The concatenation and fully connected subunits perform concatenation and fully connected processing on the interaction features corresponding to the first-order interaction operation and the interaction features corresponding to the high-order interaction operation. Taking the attention subunit for first-order interactive operations as an example, key information and value information are generated based on input data features. Query information is generated based on pre-trained boosted features. An initial weight matrix is ​​obtained through matrix multiplication based on the query information and key information. A Top-K strategy is used to filter the weights to obtain a target weight matrix. Based on the target weight matrix, the value information is weighted and normalized to obtain the interaction features corresponding to the first-order interactive operation. The attention subunit for higher-order interactive operations has the same structure and processing flow, but differs in model parameters and prompt features. Figure 4 shows a schematic diagram of attention calculation in a task processing method provided by one embodiment of the present disclosure. As shown in Figure 4: Initial weights are determined based on query information and key information. A Top-K strategy is used to determine target weights greater than a preset threshold, obtaining a target weight matrix composed of the target weights. Based on the target weights, the value information is weighted and normalized to obtain the interaction matrix corresponding to the interactive operation. Referring to FIG. 5 , FIG. 5 shows a flowchart of an object recommendation method provided by an embodiment of the present disclosure, which is applied to a cloud-side device. The method includes the following specific steps: Step 502: Obtaining structured data for a target recommendation scenario, wherein the structured data includes multi-type object data conforming to a specific format. Step 504: Extracting data features of the multi-type object data. Step 506: Based on at least two interactive operations, performing interaction feature analysis on the data features of the multi-type object data to obtain interaction features corresponding to each interaction operation. Step 508: Determining target interaction features corresponding to the target recommendation scenario based on the interaction features corresponding to the various interactive operations. Step 510: Determining information to be recommended for the target recommendation scenario based on the target interaction features. Step 512: Sending the information to be recommended to the terminal device.The disclosed embodiments are applicable to applications, websites, or mini-programs with object recommendation functions. Data processing for these applications, websites, or mini-programs is implemented on cloud-side devices. The cloud-side device is a virtual network cloud-side device that provides the object recommendation function. A task processing model for the object recommendation function is deployed on the cloud-side device. The end-side device is a physical terminal device that provides the object recommendation function. The end-side and cloud-side devices are connected via a network channel for data transmission. The computing power and storage capacity of the cloud-side device are higher than those of the end-side device. The disclosed embodiments and the embodiment described in FIG1 above are based on the same inventive concept. The specific methods of steps 502 to 510 are similar to those of steps 102 to 110 above and will not be repeated here. Taking the product recommendation scenario of an e-commerce platform as an example, steps 502-512 are explained below: Step 1: Obtain structured data for the target recommendation scenario. E-commerce systems retrieve multi-type object data in a predefined, specific format from a database. This data primarily includes user behavior records (such as browsing history, purchase history, search keywords, and ratings) and product attribute information (such as product ID, category, brand, price, sales volume, and ratings). Step 2: Extracting data features from this multi-type object data. The platform first preprocesses the user behavior and product attribute data, converting the raw data into low-dimensional, quantitative feature vectors, such as user interest vectors and product feature vectors, using embedding or linear transformation techniques. Step 3: Based on at least two interactive operations, feature analysis is performed on the data features of the multi-type object data. Using methods such as matrix factorization, deep neural networks, or attention mechanisms, the interactive relationships between users and products are accurately modeled and analyzed, generating multiple types of interactive features, such as collaborative filtering features and content-based similarity features. Step 4: Determining target interactive features for the target recommendation scenario: Based on the requirements of the recommendation task, the e-commerce platform directly selects or fuses target interactive features from multiple interactive features to accurately reflect the user's potential purchase intention. Step 5: Based on the target interaction features, a list of recommended items for the target recommendation scenario is clearly constructed. The system leverages these features to accurately predict the user's preference for each candidate item through a ranking algorithm or regression model, and accordingly develops a personalized list of recommended items. Step 6: This personalized list of recommended items is accurately delivered to the user's terminal device, including a mobile app, website, or mini-program. The user will see the system's carefully tailored product recommendations on their interface, ensuring precise product recommendation service.In the disclosed embodiments, data features of multi-type object data conforming to a specific format are extracted. Based on at least two interactive operations, interaction feature analysis is performed on the data features of the multi-type object data to obtain interaction features corresponding to the various interactive operations. Based on the interaction features corresponding to the various interactive operations, a target interaction feature is extracted for the multi-type object data in the specific format. This arithmetic interaction feature is a highly accurate data feature. Based on the target interaction feature, the recommended information in the target recommendation scenario is determined, thereby improving the accuracy of object recommendations. Furthermore, this feature is implemented on cloud-side devices with high computing and storage performance, improving the accuracy and efficiency of object recommendations and enhancing the user experience. In an optional embodiment of the present disclosure, after step 502, the following specific steps are further included: if the multi-type object data in the structured data does not conform to a specific format, identifying a target format for the multi-type object data in the structured data; constructing a sample set based on the target format, wherein the sample set includes sample structured data in a target recommendation scenario and label recommendation information for the target recommendation scenario, and the sample structured data includes multi-type sample object data that conforms to the target format; inputting the multi-type sample object data into an embedding layer of an object recommendation model to extract data features of the multi-type sample object data, wherein the object recommendation model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; inputting the data features into the arithmetic unit, performing interaction feature analysis on the data features of the multi-type sample object data based on at least two interaction operations, and obtaining interaction features corresponding to various interaction operations; inputting the interaction features corresponding to various interaction operations into the feature processing unit, and determining sample interaction features corresponding to the target recommendation scenario based on the interaction features corresponding to various interaction operations; inputting the sample interaction features into the decoding layer, and determining predicted recommendation information for the target recommendation scenario based on the sample interaction features; and based on the predicted recommendation information and the label recommendation information, The object recommendation model is trained. The embodiments of this disclosure and the training embodiments of the task processing model described above have been described in detail and will not be repeated here. By collecting data from actual application scenarios and dynamically constructing sample sets to train the model, model performance is improved, the accuracy of subsequent object recommendations is improved, and the user experience is enhanced. Referring to FIG6 , FIG6 shows a flowchart of a model training method provided by an embodiment of the present disclosure, applied to a cloud-side device, and comprising the following specific steps: Step 602: Obtain a sample set, wherein the sample set includes sample structured data of a sample task and the label task processing results of the sample task, and the sample structured data includes multiple types of sample task data conforming to a specific format.Step 604: Input the multi-type sample task data into the embedding layer of the task processing model to extract data features of the multi-type sample task data. The task processing model includes an embedding layer, an encoding layer, and a decoding layer. The encoding layer includes an arithmetic unit and a feature processing unit. Step 606: Input the data features into the arithmetic unit. Based on at least two interactive operations, the data features of the multi-type sample task data are analyzed to obtain interaction features corresponding to each interactive operation. Step 608: Input the interaction features corresponding to each interactive operation into the feature processing unit. Based on the interaction features corresponding to each interactive operation, the sample interaction feature corresponding to the sample task is determined. Step 610: Input the sample interaction features into the decoding layer. The sample task is executed based on the sample interaction features to obtain a predicted task processing result for the sample task. Step 612: Train the task processing model based on the predicted task processing result and the labeling task processing result. Step 614: Send the parameters of the trained task processing model to the end-side device. The cloud-side device is a network cloud-side device that provides model training capabilities and is a virtual device. The end-side device is a physical terminal device that provides task processing functionality. The end-side device and cloud-side device are connected via a network channel for data transmission. The cloud-side device has higher computing power and storage capacity than the end-side device. This disclosed embodiment and the embodiment shown in Figure 1 above are based on the same inventive concept. The detailed description of steps 102-110 in the embodiment shown in Figure 1 above is omitted here. This disclosed embodiment improves model performance through supervised training of the task processing model, laying the foundation for subsequent high-accuracy task processing. Furthermore, implementation on cloud-side devices with high computing and storage performance improves model training efficiency and effectiveness. In an optional embodiment of the present disclosure, step 606 includes the following specific steps: inputting data features into an arithmetic unit; determining query information corresponding to each interactive operation based on prompt features corresponding to at least two interactive operations; and determining key information and value information for each type of sample task data based on the data features of the multiple types of sample task data. The prompt features are used to prompt feature analysis of the corresponding interactive operations between the multiple types of sample task data; performing feature calculation on the corresponding query information, key information, and value information for each type of task data to obtain interactive features corresponding to each interactive operation. Correspondingly, step 612 includes the following specific steps: adjusting parameters and prompt features of the task processing model based on the prediction task processing results and the labeling task processing results. This disclosed embodiment and the embodiment of FIG1 above share the same inventive concept. The disclosed embodiment has been described in detail in the model training embodiment of FIG1 above and will not be repeated here.In the embodiments of the present disclosure, by training a cue feature used in attention calculation, the interactive relationships between multiple types of task data conforming to a specific format are embedded in the task processing model. Furthermore, the model's time and space complexity increases linearly with the number of features, enabling better scalability to large-scale data. This addresses the inefficiency of the attention mechanism when the number of features is large, as well as the challenge of current data-dependent architectures that struggle to capture interactive features of the data. Corresponding to the aforementioned method embodiments, the present disclosure also provides an embodiment of a task processing device. Figure 7 shows a schematic structural diagram of a task processing device provided by one embodiment of the present disclosure. As shown in FIG7 , the apparatus includes: a first acquisition module 702 configured to acquire structured data of a target task, wherein the structured data includes multi-type task data conforming to a specific format; a first extraction module 704 configured to extract data features of the multi-type task data; a first analysis module 706 configured to perform interaction feature analysis on the data features of the multi-type task data based on at least two interaction operations, respectively, to obtain interaction features corresponding to various interaction operations; a first determination module 708 configured to determine a target interaction feature corresponding to the target task based on the interaction features corresponding to the various interaction operations; and an execution module 710 configured to execute the target task based on the target interaction feature, to obtain a task processing result of the target task. Optionally, the first analysis module 706 is further configured to: determine query information corresponding to each interactive operation based on prompt features corresponding to at least two interactive operations, and determine key information and value information for each type of task data based on data features of multiple types of task data, wherein the prompt features are used to prompt feature analysis of the corresponding interactive operations between the multiple types of task data; perform feature calculation on the query information corresponding to each interactive operation, and the key information and value information of each type of task data, to obtain interactive features corresponding to each interactive operation. Optionally, the first analysis module 706 is further configured to: calculate initial weight information based on the query information corresponding to the first interactive operation and the key information of each type of task data, wherein the first interactive operation is any interactive operation; determine a target weight that reaches a preset threshold from the initial weight information, and determine target weight information based on the target weight; and perform weighted processing on the value information of each type of task data based on the target weight information to obtain interactive features corresponding to the first interactive operation.Optionally, the at least two interaction operations include first-order interaction operations and higher-order interaction operations. Accordingly, the first analysis module 706 is further configured to: perform interaction feature analysis on the data features of the multi-type task data based on the first-order interaction operations to obtain interaction features corresponding to the first-order interaction operations; and perform interaction feature analysis on the data features of the multi-type task data based on the higher-order interaction operations to obtain interaction features corresponding to the higher-order interaction operations. Optionally, the first analysis module 706 is further configured to: perform order reduction processing on the data features of the multi-type task data to obtain reduced-order data features of the multi-type task data; perform interaction feature analysis on the reduced-order data features of the multi-type task data based on the higher-order interaction operations to obtain reduced-order interaction features corresponding to the higher-order interaction operations; and perform order enhancement processing on the reduced-order interaction features corresponding to the higher-order interaction operations to obtain interaction features corresponding to the higher-order interaction operations. Optionally, the first determination module 708 is further configured to: perform feature fusion on the interaction features corresponding to the various interaction operations to obtain target interaction features corresponding to the target task. Optionally, the multi-type task data includes numerical task data and category task data; accordingly, the first extraction module 704 is further configured to: perform linear transformation on the numerical task data to obtain data features of the numerical task data; and query a preset category embedding table based on the category task data to obtain data features of the category task data. Optionally, the first analysis module 706 is further configured to: perform data feature transformation on the numerical task data to obtain data features of the numerical task data; and query a preset category embedding table based on the category task data to obtain data features of the category task data. Optionally, the arithmetic unit includes a first-order arithmetic unit and a high-order arithmetic unit; correspondingly, the first analysis module 706 is further configured to: input the data features into the first-order arithmetic unit, perform interaction feature analysis on the data features of the multi-type task data based on the first-order interaction operation, and obtain interaction features corresponding to the first-order interaction operation; input the data features into the high-order arithmetic unit, perform interaction feature analysis on the data features of the multi-type task data based on the high-order interaction operation, and obtain interaction features corresponding to the high-order interaction operation.Optionally, the device further includes: a first training module configured to obtain a sample set, wherein the sample set includes sample structured data of a sample task and a label task processing result of the sample task, and the sample structured data includes multiple types of sample task data that conform to a specific format; inputting the multiple types of sample task data into an embedding layer of a task processing model to extract data features of the multiple types of sample task data, wherein the task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; inputting the data features into the arithmetic unit, and performing interaction feature analysis on the data features of the multiple types of sample task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations; inputting the interaction features corresponding to various interaction operations into the feature processing unit, and determining the sample interaction features corresponding to the sample task based on the interaction features corresponding to the various interaction operations; inputting the sample interaction features into the decoding layer, executing the sample task based on the sample interaction features, and obtaining a prediction task processing result of the sample task; and training the task processing model based on the prediction task processing result and the label task processing result. Optionally, the first training module is further configured to: input data features into an arithmetic unit; determine query information corresponding to each interactive operation based on prompt features corresponding to at least two interactive operations; and determine key information and value information for each type of sample task data based on the data features of the multi-type sample task data, wherein the prompt features are used to prompt feature analysis of the query information, key information, and value information for the corresponding interactive operations between the multi-type sample task data; perform feature calculation on the corresponding query information, key information, and value information for each type of task data to obtain interaction features corresponding to each interactive operation; and adjust parameters and prompt features of the task processing model based on the prediction task processing results and the labeling task processing results. In the disclosed embodiment, data features of multi-type task data conforming to a specific format are extracted; interaction feature analysis is performed on the data features of the multi-type task data based on at least two interactive operations to obtain interaction features corresponding to each interactive operation; and based on the interaction features corresponding to each interactive operation, a target interaction feature is extracted, i.e., an arithmetic interaction feature adapted to the multi-type data in the specific format. This is a highly accurate data feature. Executing the target task based on the target interaction feature improves the accuracy of task processing. The above is a schematic diagram of a task processing device according to this embodiment. It should be noted that the technical solution of the task processing device and the technical solution of the task processing method described above share the same concept. For details not described in detail in the technical solution of the task processing device, please refer to the description of the technical solution of the task processing method described above.Corresponding to the above-mentioned method embodiments, the present disclosure also provides an object recommendation device embodiment. FIG8 shows a schematic structural diagram of an object recommendation device according to one embodiment of the present disclosure. As shown in FIG8 , the device is applied to a cloud-side device and includes: a second acquisition module 802 configured to acquire structured data for a target recommendation scenario, wherein the structured data includes multi-type object data conforming to a specific format; a second extraction module 804 configured to extract data features of the multi-type object data; a second analysis module 806 configured to perform interaction feature analysis on the data features of the multi-type object data based on at least two interaction operations, respectively, to obtain interaction features corresponding to each interaction operation; a second determination module 808 configured to determine a target interaction feature corresponding to the target recommendation scenario based on the interaction features corresponding to the various interaction operations; a recommendation module 810 configured to determine information to be recommended for the target recommendation scenario based on the target interaction feature; and a sending module 812 configured to send the information to be recommended to a terminal device. Optionally, the apparatus further includes: a second training module configured to, if the multi-type object data in the structured data does not conform to a specific format, identify a target format of the multi-type object data in the structured data; construct a sample set based on the target format, wherein the sample set includes sample structured data in a target recommendation scenario and label recommendation information of the target recommendation scenario, and the sample structured data includes multi-type sample object data that conforms to the target format; input the multi-type sample object data into an embedding layer of an object recommendation model to extract data features of the multi-type sample object data, wherein the object recommendation model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; input the data features into the arithmetic unit, and perform interaction feature analysis on the data features of the multi-type sample object data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations; input the interaction features corresponding to various interaction operations into the feature processing unit, and determine sample interaction features corresponding to the target recommendation scenario based on the interaction features corresponding to various interaction operations; input the sample interaction features into the decoding layer, and determine predicted recommendation information in the target recommendation scenario based on the sample interaction features; based on the predicted recommendation information and the label recommendation information, Train the object recommendation model.In an embodiment of the present disclosure, data features of multi-type object data conforming to a specific format are extracted. Based on at least two interactive operations, interaction feature analysis is performed on the data features of the multi-type object data to obtain interaction features corresponding to each interactive operation. Based on the interaction features corresponding to each interactive operation, a target interaction feature is extracted for the multi-type object data in the specific format. This arithmetic interaction feature is a highly accurate data feature. Based on the target interaction feature, the recommended information in the target recommendation scenario is determined, improving the accuracy of object recommendations. Furthermore, implementation on cloud-side devices with high computing and storage performance improves the accuracy and efficiency of object recommendations, enhancing the user experience. The above is a schematic diagram of an object recommendation device according to this embodiment. It should be noted that the technical solution of this object recommendation device and the technical solution of the object recommendation method described above share the same concept. For details not described in detail in the technical solution of the object recommendation device, please refer to the description of the technical solution of the object recommendation method. Corresponding to the above method embodiment, the present disclosure also provides an embodiment of a model training device. Figure 9 shows a schematic structural diagram of a model training device according to one embodiment of the present disclosure. As shown in FIG9 , the apparatus is applied to a cloud-side device and includes: a third acquisition module 902 configured to acquire a sample set, wherein the sample set includes sample structured data of a sample task and a label task processing result of the sample task, and the sample structured data includes multiple types of sample task data in a specific format; a third extraction module 904 configured to input the multiple types of sample task data into an embedding layer of a task processing model to extract data features of the multiple types of sample task data, wherein the task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; a third analysis module 906 configured to input the data features into the arithmetic unit, perform interaction feature analysis on the data features of the multiple types of sample task data based on at least two interaction operations, and obtain interaction features corresponding to the various interaction operations; a third determination module 908 configured to input the interaction features corresponding to the various interaction operations into the feature processing unit, and determine the sample interaction features corresponding to the sample task based on the interaction features corresponding to the various interaction operations; a decoding module 910 configured to input the sample interaction features into the decoding layer, and execute the sample task based on the sample interaction features. Obtain the predicted task processing result of the sample task; a training module 912, configured to train the task processing model based on the predicted task processing result and the label task processing result; a parameter sending module 914, configured to send the parameters of the trained task processing model to the end-side device.Optionally, the third analysis module 906 is further configured to: input data features into an arithmetic unit; determine query information corresponding to each interactive operation based on prompt features corresponding to at least two interactive operations; and determine key information and value information for each type of sample task data based on the data features of the multi-type sample task data, wherein the prompt features are used to prompt feature analysis of the corresponding interactive operations between the multi-type sample task data; perform feature calculation on the corresponding query information, key information, and value information of each type of task data to obtain interaction features corresponding to each interactive operation; and correspondingly, the training module 912 is further configured to: adjust parameters and prompt features of the task processing model based on the prediction task processing results and the labeling task processing results. In this disclosed embodiment, supervised training of the task processing model improves model performance, laying a foundation for subsequent high-accuracy task processing. Furthermore, implementation on cloud-side devices with high computing and storage performance improves model training efficiency and effectiveness. The above is a schematic scheme of a model training device in this embodiment. It should be noted that the technical solution of the model training device and the technical solution of the aforementioned model training method share the same concept. For details not described in detail in the technical solution of the model training device, please refer to the description of the technical solution of the aforementioned model training method. Figure 10 shows a block diagram of a computing device provided by one embodiment of the present disclosure. The components of computing device 1000 include, but are not limited to, a memory 1010 and a processor 1020. Processor 1020 is connected to memory 1010 via a bus 1030, and database 1050 is used to store data. Computing device 1000 also includes an access device 1040, which enables computing device 1000 to communicate via one or more networks 1060. Examples of these 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 1040 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 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. In one embodiment of the present disclosure, the aforementioned components of the computing device 1000 and other components not shown in FIG. 10 may also be connected to each other, for example, via a bus. It should be understood that the computing device structure block diagram shown in FIG. 10 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. The computing device 1000 can 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). The computing device 1000 can also be a mobile or stationary server. The processor 1020 is configured to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the aforementioned task processing method, object recommendation method, or model training method. The above is a schematic diagram of a computing device of this embodiment. It should be noted that the technical solution of this computing device is based on the same concept as the technical solutions of the aforementioned task processing method, object recommendation method, and model training method. 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 aforementioned task processing method, object recommendation method, or model training method. An embodiment of the present disclosure further provides a computer-readable storage medium storing a computer program / instruction, which implements the steps of the above-mentioned task processing method, object recommendation method or model training method when executed by a processor.The above is an illustrative embodiment 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 technical solutions of the aforementioned task processing method, object recommendation method, and model training method. 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 aforementioned task processing method, object recommendation method, or model training method. An embodiment of the present disclosure also provides a computer program product, including a computer program / instructions. When executed by a processor, the computer program / instructions implement the steps of the aforementioned task processing method, object recommendation method, or model training method. The above is an illustrative embodiment of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product is based on the same concept as the technical solutions of the aforementioned task processing method, object recommendation method, and model training method. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solutions of the aforementioned task processing method, object recommendation method, or model training method. The above 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 or sequential order shown to achieve the desired results. In certain embodiments, multi-tasking and parallel processing are also possible or may be advantageous. The computer instructions include computer program code, which may be in source code form, object code form, executable files, 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 drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. It should be noted that the content of the computer-readable medium may be appropriately increased or decreased based on 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. It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all expressed as a series of action combinations. However, those skilled in the art should be aware that the embodiments of the present disclosure are not limited to 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 this specification are preferred embodiments, and the actions and modules described are not necessarily required for the embodiments of the present disclosure. In the above embodiments, the description of each embodiment has its own focus. For portions not described in detail in a particular embodiment, reference should be made to the relevant descriptions of other embodiments. The preferred embodiments of the present disclosure disclosed above are merely provided to illustrate the present disclosure. The optional embodiments do not describe all details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and variations are possible based on the content of the embodiments of the present disclosure. These embodiments are selected and described in detail in this disclosure to better explain the principles and practical applications of the embodiments of the present disclosure, thereby enabling those skilled in the art to better understand and utilize the present disclosure. The present disclosure is limited only by the claims and their full scope and equivalents.

Claims

Claims 1. A task processing method, comprising: The method includes obtaining structured data of a target task, wherein the structured data includes multiple types of task data conforming to a specific format; extracting data features of the multiple types of task data; performing interaction feature analysis on the data features of the multiple types of task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations; determining a target interaction feature corresponding to the target task based on the interaction features corresponding to the various interaction operations; and executing the target task based on the target interaction feature to obtain a task processing result for the target task.

2. The method according to claim 1, wherein the performing interaction feature analysis on the data features of the multi-type task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations comprises: Based on prompt features corresponding to at least two interactive operations, query information corresponding to each interactive operation is determined, and based on data features of the multiple types of task data, key information and value information of each type of task data are determined, wherein the prompt features are used to prompt feature analysis of corresponding interactive operations between the multiple types of task data; and feature calculation is performed on the query information corresponding to each interactive operation and the key information and value information of each type of task data to obtain interaction features corresponding to the various interactive operations.

3. The method according to claim 2, wherein performing feature calculation on the query information corresponding to each interactive operation and the key information and value information of each type of task data to obtain the interaction features corresponding to each interactive operation comprises: calculating initial weight information based on the query information corresponding to the first interactive operation and the key information of each type of task data, wherein: The first interactive operation is any interactive operation; Determine a target weight that reaches a preset threshold from the initial weight information, and determine target weight information based on the target weight; and perform weighted processing on the value information of each type of task data based on the target weight information to obtain an interaction feature corresponding to the first interaction operation.

4. The method according to any one of claims 1 to 3, wherein the at least two interactive operations include a first-order interactive operation and a high-order interactive operation; and performing interactive feature analysis on the data features of the multi-type task data based on the at least two interactive operations to obtain interactive features corresponding to the various interactive operations comprises: Based on the first-order interactive operation, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to the first-order interactive operation; based on the high-order interactive operation, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to the high-order interactive operation.

5. The method according to claim 4, wherein the performing interaction feature analysis on the data features of the multi-type task data based on the high-order interaction operation to obtain interaction features corresponding to the high-order interaction operation comprises: performing order reduction processing on the data features of the multi-type task data to obtain reduced-order data features of the multi-type task data; Based on the high-order interactive operation, interactive feature analysis is performed on the reduced-order data features of the multi-type task data to obtain reduced-order interactive features corresponding to the high-order interactive operation; and up-order processing is performed on the reduced-order interactive features corresponding to the high-order interactive operation to obtain interactive features corresponding to the high-order interactive operation.

6. The method according to any one of claims 1 to 5, wherein the interactive features corresponding to the various interactive operations are: Determining target interaction features corresponding to the target task includes: The interaction features corresponding to the various interaction operations are subjected to feature fusion to obtain the target interaction features corresponding to the target task.

7. The method according to any one of claims 1 to 6, wherein the multi-type task data includes numerical task data and category task data; and extracting data features of the multi-type task data includes: performing data feature conversion on the numerical task data to obtain data features of the numerical task data; For the category task data, based on the category task data, a preset category embedding table is queried to obtain data features of the category task data.

8. The method according to claim 1, wherein the performing interaction feature analysis on the data features of the multi-type task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations comprises: The data features are input into a coding layer of a task processing model. An arithmetic unit of the coding layer is used to perform interaction feature analysis on the data features of the multi-type task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations.

9. The method according to claim 8, wherein the arithmetic unit includes a first-order arithmetic unit and a high-order arithmetic unit; and inputting the data features into the encoding layer of the task processing model, and using the arithmetic unit of the encoding layer to perform interaction feature analysis on the data features of the multi-type task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations, comprising: The data features are input into the first-order arithmetic unit, and based on the first-order interactive operation, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to the first-order interactive operation; the data features are input into the high-order arithmetic unit, and based on the high-order interactive operation, interaction feature analysis is performed on the data features of the multi-type task data to obtain interaction features corresponding to the high-order interactive operation.

10. The method according to any one of claims 1 to 9, further comprising, before extracting data features of the multi-type task data: A sample set is obtained, wherein the sample set includes sample structured data of a sample task and a label task processing result of the sample task, wherein the sample structured data includes multiple types of sample task data conforming to a specific format; the multiple types of sample task data are input into an embedding layer of a task processing model to extract data features of the multiple types of sample task data, wherein the task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; the data features are input into the arithmetic unit, and interaction feature analysis is performed on the data features of the multiple types of sample task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations; the interaction features corresponding to the various interaction operations are input into the feature processing unit, and sample interaction features corresponding to the sample task are determined based on the interaction features corresponding to the various interaction operations; the sample interaction features are input into the decoding layer, the sample task is executed based on the sample interaction features, and a prediction task processing result of the sample task is obtained; and the task processing model is trained based on the prediction task processing result and the label task processing result.

11. The method according to claim 10, wherein inputting the data features into the arithmetic unit and performing interaction feature analysis on the data features of the multi-type sample task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations comprises: The data features are input into the arithmetic unit, and based on the prompt features corresponding to at least two interactive operations, query information corresponding to each interactive operation is determined, and based on the data features of the multi-type sample task data, key information and value information of each type of sample task data are determined, wherein the prompt features are used to prompt the multi-type sample task data. performing feature analysis on query information, key information, and value information based on corresponding interactive operations between the two tasks; performing feature calculation on the query information corresponding to the interactive operations, and the key information and value information of the various types of task data to obtain interactive features corresponding to the various interactive operations; training the task processing model based on the predicted task processing results and the labeled task processing results, including: adjusting the parameters of the task processing model and the prompt features based on the predicted task processing results and the labeled task processing results.

12. An object recommendation method, applied to a cloud-side device, comprising: Obtaining structured data in a target recommendation scenario, wherein the structured data includes multi-type object data conforming to a specific format; extracting data features of the multi-type object data; performing interaction feature analysis on the data features of the multi-type object data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations; determining a target interaction feature corresponding to the target recommendation scenario based on the interaction features corresponding to the various interaction operations; determining information to be recommended in the target recommendation scenario based on the target interaction feature; and sending the information to be recommended to a terminal device.

13. The method according to claim 12, further comprising, after obtaining structured data in a target recommendation scenario: If the multi-type object data in the structured data does not conform to the specific format, identifying a target format of the multi-type object data in the structured data; A sample set is constructed based on the target format, wherein the sample set includes sample structured data in the target recommendation scenario and label recommendation information for the target recommendation scenario, and the sample structured data includes multiple types of sample object data that conform to the target format; the multiple types of sample object data are input into an embedding layer of an object recommendation model to extract data features of the multiple types of sample object data, wherein the object recommendation model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; the data features are input into the arithmetic unit, and interaction feature analysis is performed on the data features of the multiple types of sample object data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations; the interaction features corresponding to the various interaction operations are input into the feature processing unit, and sample interaction features corresponding to the target recommendation scenario are determined based on the interaction features corresponding to the various interaction operations; the sample interaction features are input into the decoding layer, and predicted recommendation information for the target recommendation scenario is determined based on the sample interaction features; and the object recommendation model is trained based on the predicted recommendation information and the label recommendation information.

14. A model training method, applied to a cloud-side device, comprising: Acquiring a sample set, wherein the sample set includes sample structured data of a sample task and a label task processing result of the sample task, wherein the sample structured data includes multiple types of sample task data conforming to a specific format; inputting the multiple types of sample task data into an embedding layer of a task processing model to extract data features of the multiple types of sample task data, wherein the task processing model includes an embedding layer, an encoding layer, and a decoding layer, and the encoding layer includes an arithmetic unit and a feature processing unit; inputting the data features into the arithmetic unit, and performing interaction feature analysis on the data features of the multiple types of sample task data based on at least two interaction operations to obtain interaction features corresponding to the various interaction operations; inputting the interaction features corresponding to the various interaction operations into the feature processing unit, and determining a sample interaction feature corresponding to the sample task based on the interaction features corresponding to the various interaction operations; Inputting the sample interaction features into the decoding layer, executing the sample task based on the sample interaction features to obtain a prediction task processing result of the sample task; training the task processing model based on the prediction task processing result and the label task processing result; and sending the parameters of the trained task processing model to the end-side device.

15. The method according to claim 14, wherein inputting the data features into the arithmetic unit and performing interaction feature analysis on the data features of the multi-type sample task data based on at least two interaction operations to obtain interaction features corresponding to various interaction operations comprises: The data features are input into the arithmetic unit, and based on the prompt features corresponding to at least two interactive operations, the query information corresponding to each interactive operation is determined, and based on the data features of the multiple types of sample task data, the key information and value information of each type of sample task data are determined, wherein the prompt features are used to prompt feature analysis of the corresponding interactive operations between the multiple types of sample task data; feature calculation is performed on the corresponding query information, the key information and the value information of the various types of task data to obtain the interactive features corresponding to the various interactive operations; the task processing model is trained based on the predicted task processing results and the labeled task processing results, including: adjusting the parameters of the task processing model and the prompt features based on the predicted task processing results and the labeled task processing results.

16. A computing device, comprising: memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer program / instructions are executed by the processor, the steps of the method according to any one of claims 1 to 15 are implemented.

17. A computer-readable storage medium storing a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 15.

18. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 15. 27

Images