51 results about "Early stopping" patented technology

The invention discloses a multi-degree-of-freedom mechanical arm control method based on quantum particle swarm optimization of a migration strategy. The method comprises the steps that system modeling is conducted; mPC to-be-adjusted parameter definition and constraint processing are carried out; adopting, adopting, and introducing penalty terms to construct a composite fitness function; updating particle positions by adopting a basic or enhanced quantum updating mode; triggering conditions are judged in a group diversity measurement mode, and when the conditions are met, a migration strategy and dynamic subgroup division are executed; the global optimal solution and the fitness value are loaded into a real-time MPC controller, and online adjustment is carried out; a prediction equation is constructed, tracking errors and energy consumption optimization are converted into a standard quadratic programming problem, and the solving precision is dynamically adjusted in combination with Cholesky pre-decomposition, a structured sparse solver and a warm-start and early stop strategy; and a closed-loop adaptive control system is constructed. According to the method, the precision, robustness and real-time performance of trajectory tracking control of the mechanical arm can be effectively improved.

The invention discloses a work order semantics and load curve linked metering anomaly positioning method, which relates to the technical field of electric power metering, and is characterized in that an OBIS event is used as a strong anchor, work order time is used as a soft anchor, and acquisition time is used as a sampling anchor to construct a unified time axis, and an optimal window, an operation mask, and machine account confidence and time alignment mapping are output. Decomposing the load curve into trend, period and disturbance residual errors in the optimal window, and decoding OBIS events and work order elements in combination with the morphological dictionary and neighborhood consistency to form three-domain evidence vectors; and fusing the reconstructed residual error, the density rare degree and the single-class interval to generate a candidate list, and solidifying an evidence chain by mask shunting and early stop noise reduction. According to the method, events and semantics are taken as priori, forms are taken as likelihood, a virtual repair operator is introduced to perform anti-fact verification, non-technical loss can be diffused in a transformer area map to cut off an overhaul reconstruction edge, an evidence card write-back closed loop is generated, and recall, cause determination and positioning are improved.

The invention provides a communication radiation source individual identification method based on time domain and time sequence coding constellation diagram feature fusion. The method is used for realizing stable identification of different transmitters under non-ideal conditions of carrier frequency offset, noise interference and the like. The method comprises the following steps: firstly, carrying out normalization and fixed-length interception on a received I / Q signal, and calculating an oversampling multiple according to a symbol rate and a time resolution; then, a time domain and time sequence coding constellation diagram double-branch feature extraction network is constructed, a time domain structure feature is extracted from a time domain branch by adopting a one-dimensional modified ResNet18, and a constellation diagram with time information is generated from a time sequence coding constellation diagram branch through a time sequence coding constellation module and the feature that the phase changes along with time is extracted; and carrying out joint modeling on the two types of features through a trainable weighted fusion mechanism. In the training stage, the performance of the model is improved by adopting cross entropy loss, an Adam optimizer, learning rate scheduling and an early stop strategy. In the reasoning process, input signals are classified by using the model weight with the highest verification accuracy, and a communication radiation source individual identification result is output. Experimental results show that the method can effectively enhance the frequency offset robustness and generalization ability of the model, realizes high-precision identification of the individual communication radiation source, and is suitable for the fields of security access of the Internet of Things, spectrum monitoring, electronic countermeasures and the like.

The invention discloses a hydropower station reservoir runoff intelligent forecasting method based on a large language model, and belongs to the technical field of hydrological forecasting, and the method comprises the following steps: S1, data collection and preprocessing; s2, constructing a BERT-Hydrology deep learning model for the runoff forecast of the reservoir of the hydropower station; s3, designing a weighted mixed loss function, and introducing an optimizer, learning rate scheduling and an early stop mechanism to improve the convergence speed and numerical stability of model training; and S4, establishing an evaluation system, performing rolling forecasting, and outputting an evaluation index and a significance test result. According to the intelligent hydropower station reservoir runoff forecasting method based on the large language model, the accuracy and the stability of runoff forecasting are remarkably improved by constructing the runoff forecasting model, excavating the hidden rule in the hydrological sequence and establishing the multi-step forecasting mapping, and the forecasting accuracy and the forecasting stability are improved. And technical support is provided for intelligent scheduling of a hydropower station reservoir and efficient management of water resources.

This invention discloses a deep learning-based method for predicting the phenological synchronization of specialized butterflies and their host plants, belonging to the interdisciplinary field of biodiversity conservation, ecological prediction, and deep learning technology. The method includes: collecting environmental climate variables, butterfly occurrence data, and phenological time-series data of their host plants; preprocessing the data to create training and validation sets; constructing a dual-label system with the day of year (DOY) of butterfly occurrence peaks (e.g., emergence peak, egg-laying peak) as the primary label and the normalized value of synchronization deviation as the secondary label; building a deep learning model with ecological weights, including an embedding layer, an LSTM layer, and a fusion layer; inputting a vector concatenated from daily or weekly climate time-series features and ecological dynamic features; and outputting the primary and secondary labels through a dual-output layer; designing a hybrid loss function that integrates the main regression loss, ecological synchronization constraint loss, and small-sample regularization loss; training the model using the Adam optimizer and implementing an early stopping mechanism; and validating and evaluating the model using regression accuracy and ecological synchronization accuracy indicators. The method includes modules for data acquisition, preprocessing, label construction, model training, prediction evaluation, and storage. This invention improves the accuracy of phenological synchronous prediction by embedding ecological dynamic feature weights and dual-label constraints, adapts to small sample scenarios, and provides technical support for the monitoring and protection management of rare butterfly species (such as the Golden Birdwing and the Chinese Tiger Swallowtail).

The invention provides a multi-agent combined model automatic training method and a related product. According to one specific mode of the method, a decision-making agent corresponding to a target business scene is utilized to determine a lower-round hyper-parameter optimization execution agent for controlling a lower-round hyper-parameter optimization scheme of a target machine learning model; and based on the generated parameter adjustment record, a next-round hyper-parameter optimization execution agent is utilized to execute corresponding control operation so as to stop training of the target machine learning model, or replace a hyper-parameter search space of the target machine learning model, or determine next-round hyper-parameter configuration of the target machine learning model. And training the target machine learning model according to the next-round hyper-parameter configuration. Therefore, a traditional Bayesian hyper-parameter optimization method is intelligently transformed; the model training frequency is reduced; the model construction efficiency is improved; the model performance is improved; and self-adaptive model training is realized by automatically judging early stopping and continuing training.

The invention discloses a data processing method and device, electronic equipment and a storage medium, and relates to the technical field of computers. A specific embodiment of the method comprises the following steps: receiving an experiment trigger instruction, and obtaining corresponding historical sample data to determine experiment sample data and contrast sample data; querying an early stop index type, determining an early stop index based on the early stop index type, and calculating a corresponding early stop threshold value; executing an experimental operation program on the experimental sample data to obtain an experimental result, and calculating a first conversion quantity and a second conversion quantity of the early stop index in the experimental result and the contrast sample data according to the early stop threshold value to determine a change parameter of the early stop index; and in response to the change parameter meeting the early stop condition, stopping the experimental operation program. According to the implementation mode, the problem that time and resource waste is easily caused by a row mode that the experiment result is analyzed only after the execution of the preset sample size is finished or the experiment is executed for the preset time period in the experiment can be solved.

The invention discloses a synonymous mutation harmfulness prediction method based on a gene language model, and the method comprises the steps: (1) data obtaining and screening: employing a data set existing in a laboratory, and guaranteeing the data reliability and task pertinence; (2) carrying out feature annotation: respectively carrying out embedded feature extraction on the reference sequence and the mutation sequence by utilizing four representative gene language models GPN-MSA, Hyena DNA, SpliceBERT and Caduceus; (3) feature processing and fusion: performing further extraction on GPN-MSA and SpliceBERT embedding by adopting dynamic convolution and bidirectional LSTM (Long Short Term Memory); extracting the embedding of the Hyena DNA by adopting TextCNN (Convolutional Neural Network); linear dimensionality reduction is carried out on Caduceus embedding, and then splicing and dimensionality reduction are carried out on the four features; (4) model training: training in a five-fold cross validation mode, and preventing over-fitting through an early stop mechanism to obtain an optimal weight; and (5) model prediction: performing harmfulness prediction by taking 0.5 as a threshold value. According to the method, the complementary information of the multi-gene language model is fused, the prediction precision and robustness of the synonymous mutation harmfulness are remarkably improved, and a new technical support is provided for research on the synonymous mutation harmfulness.

The application discloses a kind of fusion attention mechanism and the industrial bearing vibration time series signal fault prediction method and system of LSTM.This method includes: collecting bearing vibration signal and carrying out filtering, noise reduction and normalization preprocessing;BiLSTM and coordinate attention mechanism combined deep learning model is constructed to extract bidirectional time series features and enhance key fault features;Multi-objective composite loss function and Adam optimizer are used for model training, and early stopping mechanism is introduced to prevent overfitting;Real-time vibration signal is used to identify fault type and degree evaluation using the trained model, and multi-scale spectral kurtosis features and nonlinear dynamics parameters are used for quantitative analysis;Finally, output fault diagnosis report, and trigger multi-level early warning based on adaptive threshold value.The application can realize high-precision, high-reliability bearing fault prediction and health state evaluation, and is suitable for industrial equipment intelligent operation and maintenance.

The invention belongs to the field of machine learning, and discloses a reasoning acceleration method and system for a diffusion large language model, which are used for distinguishing a calculation activity area and a stability area of the diffusion large language model; in the subsequent diffusion step, only the increment output by the active area is calculated, and the increment is fused with the calculation result of the stable area to obtain an intermediate result; calculating a token-level generation entropy based on the intermediate calculation result, thereby identifying a decoded token set to achieve an early stop; and finally, dynamically adjusting hardware calculation scheduling. According to the method, the temporal locality of a calculation unit between continuous diffusion steps is utilized, repeated calculation of complete forward transmission is avoided through a cache stable region, and redundant calculation and memory bandwidth waste are reduced; the token level generation entropy value dynamically monitors the convergence state of the token, the reasoning process is submitted to be terminated, and excessive denoising and invalid calculation caused by fixed-step reasoning are avoided; dynamic scheduling adapts to a sparse calculation mode and irregular memory access required by the model, and the problem of mismatching of hardware and software is relieved.

The invention discloses a cross-regional edge node congestion collaborative early warning method based on federated learning. The method comprises the steps of data set construction, federated learning framework initialization, local model training, global model aggregation, congestion detection and congestion early warning. The invention relates to the technical field of cross-regional edge network intelligent monitoring, in particular to a cross-regional edge node congestion collaborative early warning method based on federated learning, which comprises the following steps of: initializing a sequence model at an edge node, constructing local loss combined with binary cross entropy and time sequence smooth punishment, and stably training by Adam and early stop; through weighted average aggregation disturbance parameters based on data volume and loss, performance degradation after fusion is prevented by adopting momentum fusion and a rollback threshold value; and meanwhile, an attention mechanism is embedded to realize joint prediction of probability, severity and duration, cross-regional propagation parameters of the graph neural network are utilized, loss is verified, and invalid sharing is rejected, so that prediction accuracy and timeliness are improved, and node congestion is effectively relieved.

The application relates to a SO4 based on deep cross network (DCN V2) feature enhancement and integrated learning 2‑ The concentration driving factor analysis method comprises the following steps: collecting time series data of atmospheric pollutant concentration and meteorological parameters in a target area, and performing pretreatment and quality control; a DCN V2 model integrating a linear cross network, an exponential cross network and a deep neural network is constructed to realize nonlinear high-order cross feature enhancement of original features; a two-layer Stacking integrated learning model taking XGBoost, LightGBM and CatBoost as heterogeneous base models and linear regression as a meta model is constructed, and global optimization is performed through Bayesian optimization; and the SHAP method is applied to quantify the marginal contribution and direction of each feature, and identify the key driving factors of SO4 2‑ concentration. The application integrates optimization strategies such as GPU acceleration, early stopping mechanism, automatic task detection, parallel training and backup training, realizes the automation, high efficiency and reproducibility of the analysis process, and guarantees the physical interpretability of the driving factor analysis, thereby providing scientific and technological support for atmospheric pollution prevention and decision-making.

The invention belongs to the field of artificial intelligence safety, and discloses a large model safety protection method and system and electronic equipment. In order to solve the problem that safety and efficiency cannot be considered at the same time through existing one-step filtering or static rules, sensitive word hard filtering, injection detection and content safety double-model preliminary screening are provided, the risk difference of model output results and the comprehensive risk score of uncertainty indexes are calculated, and a preset dynamic determination threshold value is combined; and distributing the user request to low-risk, medium-risk and high-risk three-path processing. Wherein the low risk uses a lightweight model for rapid verification, the high risk is directly rejected, the medium risk executes a pre-formulated adaptive streaming generation control strategy, and an adaptive sliding window is adopted for monitoring generation and triggering early stop and self-repair. According to the scheme, the average delay and calculation cost can be remarkably reduced while the cue word injection and harmful content recognition accuracy is met, and therefore the safety and practicability of a large model are improved.

The application discloses a kind of big model fine-tuning systems and methods based on multimodal gradient coordination, the method contains: the gradient similarity cosθ is calculated, and its with two-stage alternating optimization process determines update direction;With differentiable structure search and Gumbel-Softmax automatically determine LoRA insertion layer position;With Top-K gate and dynamic sparsity rate pruning cross-modal prompt unit;Based on convergence curvature κ analysis and adjust early stopping threshold, prevent invalid training;With cosθ, Kt, κ closed-loop feedback, unified correction gradient, structure, sparse and termination condition, realize ≤5 round global convergence verification.The application solves the problems of gradient conflict, structure fixation and early stopping failure in the prior art, basic DGC-LoRA, 1000 samples 5 rounds of convergence, 0.1% parameters reach 93.1% accuracy, invalid training is reduced by 40%, A100 throughput increases by 88.3%, and extreme conflict is self-healed within 3 rounds.

The application discloses a database training method and device for regional energy consumption prediction and path optimization, an optimization method and system. The STIRPAT model is trained through the standardized multivariate time series data set; the standardized multivariate time series data set is divided to obtain the batched LSTM training data, the batched early stop verification data and the hyperparameter verification set; the LSTM time series prediction model is trained through the batched LSTM training data, the batched early stop verification data and the hyperparameter verification set to obtain the trained LSTM time series prediction model; the optimal fusion weight is solved according to the trained STIRPAT model and the trained LSTM time series prediction model in combination with the hyperparameter verification set, and the STIRPAT-LSTM fusion prediction scheme is constructed. The Pareto optimal solution set finally generated by the application has both the causal logic support and the accurate time series prediction basis, and the optimality and the feasibility are taken into account, thereby providing reliable technical support for regional energy planning.

Certain aspects of the present disclosure provide techniques and apparatus for machine learning. In an example method, a set of tokens having a probability distribution is generated using a secondary generative machine learning model associated with a primary generative machine learning model. An entropy of the set of tokens is computed based on the probability distribution, and one or more stopping criteria for the secondary generative machine learning model are determined. A next token is generated using the primary generative machine learning model after exiting from the secondary generative machine learning model based on the first entropy and the one or more stopping criteria.

According to various implementations, generally disclosed herein is a hybrid and hierarchical neural architecture search (NAS) approach. The approach includes performing a search space partitioning scheme to divide the search space into sub-search spaces. The approach further includes performing a first type of NAS, such as a Multi-trial NAS, to cover a search across the sub-search spaces. The approach also includes performing a second type of NAS, such as a One-Shot NAS, to cover each sub-search space. The approach further includes automatically stopping the second type of NAS based on one or more early stopping criteria.

The application discloses a kind of based on early stop judgment and multi-head decoding question and answer processing method, it is related to big model question and answer processing technical field, method includes: obtaining the question data that user inputs on man-machine interactive platform, the question data is input into preset big model and is handled, obtains the answer data corresponding to the question data;The preset big model includes hidden layer, self-att layer and FeedFroward layer, training is carried out by setting early stop judgment mechanism and multi-head decoding mechanism.This scheme is through early stop judgment mechanism and multi-head decoding mechanism, optimization decoding and reasoning stage, reduce the amount of calculation, improve training efficiency, by the above-mentioned mode can improve the training efficiency of preset big model, simultaneously can be caused by reducing the amount of calculation and the training precision of preset big model is higher, obtains the answer data that question data of user is closer.

The invention discloses a thinking chain reasoning early stop control method and system based on entropy-probe double signals, and the method comprises the steps: extracting the internal hidden state and external output probability distribution of a model in real time with a logic step as a unit in the process of generating a thinking chain through autoregression of a large-scale language model; utilizing a lightweight probe to map the hidden state into a reasoning convergence confidence coefficient, and calculating a step entropy as an uncertainty index; carrying out joint evaluation on the double-characteristic signals; carrying out truncation reasoning immediately through a fast channel when the signals are extremely strong, and carrying out multi-step consistency verification through a stable window mechanism when the signals tend to converge but fluctuate; and once an early stop condition is triggered, immediately terminating generation of a subsequent redundant thinking chain and outputting a final answer. According to the method, redundant steps can be effectively recognized and cut off, reasoning delay and calculation cost are remarkably reduced on the premise that base model parameters are not changed, the system robustness is improved while the accuracy is kept, and the error risk caused by excessive thinking is relieved.

The application discloses a database logical topology construction method based on a large language model, which comprises the following steps: obtaining target database Schema information, using a large language model to complete the semantic metadata of a missing service to generate enhanced metadata; based on the enhanced metadata, performing multi-stage reasoning arrangement to construct a global logical topology network, using an adaptive partition strategy and an early stop mechanism to solve the context limitation problem of the large language model; performing information entropy determination on key data fields to identify feature types, using an orthogonal boundary synthesis method to construct a virtual sample set covering global feature values to obtain an enhanced topology network; establishing a Schema change monitoring mechanism, triggering an incremental adsorption algorithm to integrate change elements, and maintaining the validity of the correlation relationship through a timeliness confidence decay model to obtain a dynamic global logical topology. The application realizes automatic correlation mining and topology dynamic maintenance of a foreign key-free database, and improves the SQL generation accuracy of an intelligent query scene.

The invention discloses a gate identity recognition acceleration method based on deep learning, and aims to solve the problem of high end-to-end time delay caused by frame combination waiting and queuing congestion due to asynchronous arrival of three modes of visible light, infrared and depth. Joint optimization is carried out on joint arrival-quality distribution prediction, single-mode feature extraction, delay perception double-channel attention fusion, micro early stop and similarity retrieval, and adaptive decision is carried out based on waiting cost and target time delay quantile. The technical effects that the average time delay and the tail time delay are remarkably reduced while the identity recognition accuracy and the living body judgment reliability are guaranteed, and the advanced retrieval rate and the passing throughput are improved are achieved.

The invention provides a noise tag self-adaptive processing method suitable for a flash memory channel. The noise tag self-adaptive processing method comprises the following steps: reading flash memory voltage data and an error tag; training by stages, and synchronously adopting an early stop strategy to avoid overfitting; in the reasoning stage, a log-likelihood ratio and a correct label are generated, and model performance is evaluated; the threshold quantization log-likelihood ratio is optimized through a dynamic programming algorithm, and the coding error rate performance is further improved. The method can effectively adapt to interference such as programming noise and data retention noise in a flash memory channel, improves the storage reliability of the flash memory, and is suitable for solid-state driving of multi-level storage units such as a single-level storage unit, a multi-level storage unit and a three-level storage unit.

According to the adversarial sample generation-based data-category-free federal learning method provided by the invention, the adversarial network is generated to synthesize the data, and the limitation that original data cannot be accessed in federal learning is ingeniously bypassed. And actively'erasing 'the memory of the target category in the model through the joint effect of classification loss and confrontation loss. Meanwhile, mechanisms such as parameter regularization, feature matching and gradient penalty effectively constrain the forgetting process, and it is ensured that non-target performance is reserved to the maximum extent. The dynamic scheduling and early stop mechanism further optimizes the training efficiency and the final effect.

The application discloses a container throughput prediction method based on stacked ensemble learning, and relates to the technical field of intelligent ports. In the system operation, multi-source data is collected from a port operation system, an economic statistics platform and a shipping database, a comprehensive feature system containing throughput, freight rate, transportation time, policy variable, seasonal characteristics and macroeconomic indicators is constructed, and serialization and standardization processing are performed, and an improved model is constructed. The improved CNN-LSTM model introduces a deep separable convolution, a bidirectional LSTM and an improved attention mechanism to enhance the representation ability of key time steps, adopts Bayesian optimization to automatically search for hyperparameters, combines an early stopping strategy to control overfitting, simultaneously realizes multi-model integration based on inverse error weighting and meta-learning linear regression model, adaptively adjusts the rolling prediction window size according to the data coefficient of variation, generates future multi-period throughput prediction results through multi-step rolling prediction, and performs denormalization output.

The present application relates to the technical field of space structure design and parameter optimization, and particularly relates to a net shell structure node parameter automatic optimization method based on a neural network model, comprising: collecting "input-output" data pairs of a single-layer cylindrical net shell constructed by an aluminum alloy plate type node, and input parameters comprising design parameters and initial node parameters of the single-layer cylindrical net shell. The present application first obtains node optimization data under different parameters by a genetic algorithm combined with finite element software and programming tools, and supplements calculation examples to construct a training data set covering the range of common engineering design parameters, then trains a three-layer fully connected neural network model containing two layers of hidden layers, and cooperates with an early stop strategy and a learning rate decay strategy to ensure the model accuracy. When applied, the model can automatically output the optimized node parameters after inputting the design parameters of the target single-layer cylindrical net shell, without the repeated iteration and manual intervention of traditional methods, greatly shortening the optimization time consumption, and efficiently meeting the batch optimization demand under multiple working conditions.

This invention discloses an adaptive early-stopping generation method and system for language models based on dynamic confidence and domain reinforcement, applied to the inference process of large-scale language models based on autoregressive generation. During model generation, this method performs real-time confidence assessment on the generated text and constructs a multi-factor coupled dynamic confidence threshold function by combining task domain features, generated content structure, and quality constraints. When the generated text meets the pre-set confidence conditions and passes the final quality check, the generation process is terminated early, thereby reducing redundant generation. This invention does not require modification of the original model parameters or structure, is applicable to various inference service frameworks, and can effectively reduce generation latency and computational resource consumption while ensuring generation accuracy and completeness, thus improving the overall efficiency of large-scale model inference systems.

The invention discloses an emotion feature auxiliary personality recognition method and system based on multi-task learning. According to the method, a dual-branch network comprising an emotion branch and a personality branch is constructed by loading and preprocessing multi-modal features, emotion tags and personality tags of an MDPE data set; the emotion branch adopts a Transform encoder to model sequence features and outputs an emotion prediction result; the personality branch adopts a multi-layer perceptron to extract personality embedded representation; further, constructing an emotion feature auxiliary module, performing attention weighted pooling on an emotion prediction result to obtain an emotion abstract representation, splicing the emotion abstract representation with a personality embedded representation, and then realizing information fusion through cross attention and a selectable gating mechanism; and finally, obtaining a personality prediction result through residual connection and layer normalization. A joint loss function is designed to coordinate double-task joint training, the model training effect is guaranteed through an optimization algorithm, a learning rate attenuation strategy and an early stop mechanism, and information sharing and collaborative optimization of emotion recognition and personality recognition are achieved.

This invention discloses a multi-resolution end-to-end and online parameter update method for depth perception. It includes: (1) constructing a multi-view vision acquisition system to obtain left and right view images and depth information and completing calibration and alignment; (2) constructing a progressive multi-resolution end-to-end disparity inference network to achieve staged disparity prediction and support dynamic early stopping output; (3) converting depth information into disparity form to construct sparse supervision signals; (4) constructing an asynchronous online adaptive mechanism decoupled from the inference process and the model update process to complete model parameter optimization; (5) introducing a supervised quality assessment strategy to control update triggering and improve online learning stability; and (6) converting the disparity results into a depth map as the final output. This invention, through decoupling inference and learning, progressive computation, and a quality-gated update mechanism, achieves continuous adaptive optimization of the model while ensuring real-time performance, significantly improving perception accuracy and robustness in complex environments.

The invention belongs to the cross technical field of high-performance concrete design and machine learning, and discloses a concrete-force displacement curve prediction method based on LSTM, and the prediction method specifically comprises five modules: a data set construction module, a data preprocessing module, an LSTM model construction module, a model training module, and a curve prediction and evaluation module. Cooperative application of structural composition, parameter configuration and model training becomes key elements for breaking through a traditional prediction method. The structure is composed of an LSTM layer, a batch normalization layer, a full connection layer 1, a dropout layer, a full connection layer 2, a dropout layer and a full connection layer 3 in sequence. Through reasonable parameter configuration, interlayer transmission among all neural network layers is achieved, training stability is optimized through a loss function module, an optimizer module, a scheduling strategy module, an early stop mechanism module and a training parameter module in model training, an accurate force-displacement curve is finally output, and precision is verified through evaluation index scores such as MSE, RMSE, MAE and R.

The invention discloses an intelligent recommendation algorithm based on machine learning, and relates to the technical field of recommendation systems. The algorithm comprises the steps of multi-source data acquisition and preprocessing, machine learning recommendation model construction, model training and optimization, intelligent recommendation generation and model dynamic updating. Features are extracted through combination of GBDT and PCA, a hybrid model containing a multi-head self-attention mechanism and a long and short term interest fusion module is constructed, an Adam algorithm is adopted for training and is combined with early stop strategy optimization, diversity is optimized through a greedy algorithm when a recommendation list is generated, and a model updating period is dynamically adjusted according to user activeness. According to the method, recommendation accuracy, real-time performance and diversity are improved, personalized demands of users can be accurately matched, and information of cocoon rooms is avoided.