CN-122021967-A - Intelligent recommendation algorithm based on machine learning

CN122021967ACN 122021967 ACN122021967 ACN 122021967ACN-122021967-A

Abstract

The invention discloses an intelligent recommendation algorithm based on machine learning, and relates to the technical field of recommendation systems. The method 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, wherein a mixed model comprising a multi-head self-attention mechanism and a long-period interest fusion module is constructed by combining GBDT with PCA to extract characteristics, adam algorithm training is adopted, early stop strategy optimization is combined, greedy algorithm optimization diversity is adopted when a recommendation list is generated, and the model updating period is dynamically adjusted according to user activity. The method and the device improve recommendation accuracy, instantaneity and diversity, can accurately match personalized requirements of users, and avoid information cocoons.

Inventors

CHEN FEI
XU HAITAO
SHEN DAWEI
ZHAO JIANHUA

Assignees

北京阿尔法风控科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. An intelligent recommendation algorithm based on machine learning is characterized by comprising the following steps: S1, multi-source data acquisition and preprocessing, namely acquiring multi-source user related data and article related data, cleaning, standardizing and extracting features of original data to form a standardized feature data set; S2, constructing a machine learning recommendation model, namely constructing the machine learning recommendation model for representing the association of the user and the article preference based on the extracted features; S3, model training and optimization, namely after the standardized feature data set is divided, training a machine learning recommendation model, and combining verification and test of optimization model parameters to obtain an optimal recommendation model; And S4, intelligent recommendation generation, namely acquiring target user data, preprocessing the target user data, inputting an optimal recommendation model, obtaining a preference evaluation result of the user on candidate articles, and generating and outputting a personalized recommendation list according to the preference evaluation result.
2. The intelligent recommendation algorithm based on machine learning according to claim 1, wherein in step S1, the user basic information comprises user age, gender, occupation, region and registration duration, the user behavior data comprises user click, collection, purchase, scoring and browsing duration of an article, the article attribute data comprises article category, price, brand, specification and release time, and the interaction context data comprises interaction time, terminal type and network environment.
3. The intelligent recommendation algorithm based on machine learning of claim 1, wherein in step S1, feature extraction is performed by combining Principal Component Analysis (PCA) with a gradient lifting tree (GBDT), features strongly related to user preferences are screened by GBDT, and feature redundancy is reduced by PCA dimension reduction processing.
4. The intelligent recommendation algorithm based on machine learning according to claim 1 is characterized in that the machine learning recommendation model is a hybrid model and comprises a feature coding layer, a depth interaction layer and a recommendation prediction layer, wherein the depth interaction layer introduces a multi-head self-attention mechanism, captures user-object interaction features with different dimensions through a plurality of attention heads, and performs splicing and fusion to strengthen feature interaction characterization capability.
5. The intelligent recommendation algorithm based on machine learning according to claim 1, wherein in step S3, the standardized feature data set is divided according to the proportion of a training set to a verification set to a test set=7:2:1, a cross entropy loss function is adopted for training to construct a training target, an adaptive momentum optimization algorithm (Adam) is adopted for iterative training, and an early-stop strategy is adopted to avoid overfitting, wherein the early-stop strategy specifically comprises stopping training and saving current model parameters as optimal parameters when a model loss function value on the verification set is not reduced continuously for a preset round.
6. The intelligent recommendation algorithm based on machine learning according to claim 1, further comprising the step S5 of dynamically updating the model, wherein feedback data of a recommendation list of a user is collected in real time, the feedback data comprise clicking, purchasing, skipping and scoring of recommended articles by the user, the feedback data are included in the standardized characteristic data set of the step S1, the training process of the step S3 is repeated periodically to perform incremental training on the optimal recommendation model, and dynamic optimization of the model is achieved.
7. The intelligent recommendation algorithm based on machine learning of claim 6, wherein in step S5, the period of periodic repeated training is dynamically adjustable according to user activity, the training period is shortened when the user activity is higher than a preset threshold, and the training period is prolonged when the user activity is lower than the preset threshold.
8. The intelligent recommendation algorithm based on machine learning of claim 1 is characterized in that in step S4, after a personalized recommendation list is generated, diversity optimization is further required to be carried out on the recommendation list, and a greedy algorithm is adopted to remove articles with similarity higher than a preset threshold value in the recommendation list, so that category diversity and attribute diversity of the recommended articles are ensured.
9. The intelligent recommendation algorithm based on machine learning of claim 1, wherein the machine learning recommendation model further introduces a long-term interest and short-term interest fusion module of a user, long-term interest features of the user are extracted through a long-term short-term memory network (LSTM), short-term interest features of the user are extracted through a time sequence convolution network (TCN), and comprehensive interest features of the user are obtained through attention weighted fusion.
10. The intelligent recommendation algorithm based on machine learning according to claim 1, wherein in step S1, the data normalization process adopts a min-max normalization method to map data to a [0,1] interval, and the calculation formula is: wherein x is the original data, x' is the normalized data, As a minimum value of the original data, Is the maximum value of the original data.

Description

Intelligent recommendation algorithm based on machine learning Technical Field The invention relates to the technical field of computers, in particular to an intelligent recommendation algorithm based on machine learning. Background With the rapid development of internet technology, the problem of information overload is increasingly prominent, and a recommendation system is widely applied to various internet products such as an electronic commerce platform, a short video platform, a content information platform and the like as a core technology for solving the problem. The recommendation system has the core aims of accurately pushing personalized content for users from mass objects according to interest preferences of the users, improving user experience and improving user viscosity and conversion efficiency of a platform. The conventional recommendation algorithm mainly comprises a traditional collaborative filtering algorithm, a recommendation algorithm based on content, a machine learning recommendation algorithm of a single model and the like, and the conventional recommendation algorithm has the defects that firstly, data utilization is insufficient, most algorithms only depend on single type user or article data, multisource heterogeneous data cannot be fused, so that recommendation characteristic dimension is insufficient, secondly, a characteristic processing mode is rough, characteristic redundancy is high, core characteristics which are strongly related to user preference cannot be effectively screened out, recommendation accuracy is affected, thirdly, model interaction characterization capability is limited, a traditional model is difficult to capture a multi-dimensional complex interaction relation between a user and an article, so that recommendation accuracy is insufficient, fourthly, the model lacks dynamic adaptability, parameters are fixed after training is completed, dynamic change of interest of the user cannot be followed timely, recommendation effect is gradually attenuated after long-term use, fifthly, recommendation diversity is insufficient, a problem of 'information cocoons' is easy to occur, interest of the user is limited, long-term interest and interest of the user are not sufficiently fused, real requirements of the user are difficult to comprehensively describe, and recommendation deviation is caused. Disclosure of Invention (One) solving the technical problems Aiming at the defects of the prior art, the invention provides an intelligent recommendation algorithm based on machine learning, which is used for solving the problems. (II) technical scheme In order to achieve the purpose, the invention provides the following technical scheme that the intelligent recommendation algorithm based on machine learning comprises the following steps: S1, multi-source data acquisition and preprocessing The method comprises the steps of collecting multisource user-related data and article-related data, wherein the user-related data comprises user basic information and user behavior data, the article-related data comprises article attribute data, collecting interaction context data in the interaction process of a user and an article, cleaning, standardizing and extracting features of the collected original data, and finally forming a standardized feature data set. The user basic information comprises user age, gender, occupation, region and registration time, user behavior data comprises clicking, collecting, purchasing, grading and browsing time of the user on the object, object attribute data comprises object category, price, brand, specification and release time, and interaction context data comprises interaction time, terminal type and network environment. In the data cleaning process, sample data with the missing value ratio exceeding a preset threshold (such as 30%) are removed, abnormal values are identified and deleted by a box diagram method, and duplicate removal processing is carried out on the repeated data. The data standardization process adopts a min-max standardization method to map data to a [0,1] interval, and a calculation formula is as follows: wherein x is the original data, x' is the normalized data, As a minimum value of the original data,Is the maximum value of the original data. The feature extraction adopts a mode of combining Principal Component Analysis (PCA) with gradient lifting tree (GBDT), the importance evaluation is carried out on the standardized features through a GBDT model, the features which are strongly related to the user preference (such as the features 80% before the feature importance scoring are screened), the dimension reduction treatment is carried out on the screened features through PCA, and the principal components with the accumulated variance contribution rate reaching a preset threshold (such as 90%) are reserved so as to reduce the feature redundancy and improve the training efficiency of the follow-up model. S2, constructing a machine learn