CN-121981793-A - Dynamic commodity recommendation system based on multi-mode deep learning and user real-time behaviors

CN121981793ACN 121981793 ACN121981793 ACN 121981793ACN-121981793-A

Abstract

The invention discloses a dynamic commodity recommendation system based on multi-mode deep learning and user real-time behaviors, which comprises a multi-mode commodity feature extraction module, a user real-time behavior sensing module, a dynamic recommendation decision module and a recommendation result output module, wherein the multi-mode commodity feature extraction module is used for carrying out deep fusion on commodity text, images and video multi-mode data by adopting an improved transducer model to extract commodity feature vectors, the user real-time behavior sensing module is used for capturing user behavior data in real time through a streaming computing frame to construct user interest vectors, the dynamic recommendation decision module is used for screening candidate commodities according to the commodity feature vectors and the user interest vectors, and utilizing the improved DeepFM model to generate a recommendation commodity list by combining scene factors, and the recommendation result output module is used for displaying recommendation results according to user terminal types. The method has the advantages of high recommendation accuracy, high response speed and wide scene suitability, and effectively solves the problems of one-sided characteristic, response lag and poor dynamic suitability of the conventional commodity recommendation system.

Inventors

CHEN XIAOYAN
GONG WEIQING
Lv Xuanmin
TAO TING
LI CAO

Assignees

江苏省常州技师学院

Dates

Publication Date: 20260505
Application Date: 20251224

Claims (10)

1. A dynamic commodity recommendation system based on multi-modal deep learning and user real-time behavior, comprising: The multi-mode commodity feature extraction module is used for carrying out deep fusion on multi-mode data of commodity texts, images and videos by adopting an improved transducer model to extract commodity feature vectors; the user real-time behavior perception module is used for capturing user behavior data in real time through the streaming computing framework and constructing user interest vectors; The dynamic recommendation decision module is used for screening candidate commodities according to commodity feature vectors and user interest vectors, and generating a recommendation commodity list by utilizing the improved DeepFM model in combination with scene factors; and the recommendation result output module is used for displaying the recommendation result according to the type of the user terminal.
2. The dynamic commodity recommendation system based on multi-modal deep learning and user real-time behavior according to claim 1, wherein said multi-modal commodity feature extraction module comprises: The commodity information acquisition unit is used for acquiring multi-mode commodity information data comprising commodity texts, images and videos; The data preprocessing unit is used for respectively preprocessing the acquired commodity information data according to the data type; And the multi-mode feature extraction unit is used for extracting commodity features from the preprocessed commodity information data by utilizing the improved transducer model, and obtaining commodity feature vectors by weighting and fusing.
3. The dynamic commodity recommendation system based on multi-modal deep learning and user real-time behavior according to claim 2, wherein said improved transducer model is specifically: The method comprises the steps of taking preprocessed commodity information data as model input data, dividing the input data into a full sequence and a plurality of subsequences divided according to different time, inputting the full sequence into Encoder of an improved transducer model, then carrying out MultiheadAttention, skipConnection and FeedForward operation for a plurality of times to obtain a high-dimensional matrix, sequentially inputting the multiple subsequences generated at different time into a Decoder of the improved transducer model, and calculating and extracting commodity characteristics at Encoder-DecoderAttention of the Decoder.
4. The dynamic commodity recommending system based on multi-mode deep learning and user real-time behavior according to claim 1, wherein said data preprocessing unit respectively preprocesses the collected commodity information data according to the data type comprises: Text data, namely, word segmentation is carried out on commodity titles and details, after stop words are filtered, text is converted into 128-dimensional Word vectors through Word2Vec, then local semantic features are extracted through TextCNN, and 256-dimensional text feature vectors are output; extracting features of commodity image by ResNet-50 model, removing the last full-connection layer, retaining 4096-dimensional convolution features, and globally averaging and pooling to compress to 256-dimensional image feature vector; extracting video key frames (1 frame per second), extracting each frame of characteristics by adopting a ResNet-50 model which is the same as that of the image data, and obtaining 256-dimensional video characteristic vectors through time sequence average pooling.
5. The dynamic commodity recommending system based on multi-modal deep learning and user real-time behavior according to claim 1, wherein said user real-time behavior sensing module comprises: the data acquisition unit is used for acquiring real-time operation data of a user on the transaction platform; the user behavior analysis unit is used for analyzing the real-time operation data to form a user real-time behavior sequence; the behavior weight distribution unit is used for distributing weights to the user behaviors according to the user real-time behavior sequence; the user interest vector construction unit is used for capturing the dependency relationship of the user real-time behavior sequence by adopting a time sequence attention mechanism and generating the user interest vector.
6. The dynamic commodity recommending system based on multi-modal deep learning and user real-time behavior according to claim 1, wherein said dynamic recommending decision module comprises: the light matching unit is used for mapping commodity feature vectors and user interest vectors to a hash bucket by adopting a local sensitive hash algorithm, screening a plurality of candidate commodities, introducing a commodity similarity threshold value to filter low-correlation candidate commodities, and obtaining a candidate commodity list; And the depth ordering unit is used for constructing an improved DeepFM model, inputting candidate commodity feature vectors, user interest vectors and scene features, outputting commodity recommendation scores, and arranging the commodity recommendation scores according to the descending order of the scores to obtain a recommended commodity list.
7. The dynamic commodity recommending system based on multi-mode deep learning and real-time behavior of a user according to claim 6, wherein said dynamic recommending decision module further comprises a scene adapting unit for adjusting the ranking weight of said deep ranking unit according to a real-time scene based on a preset scene adapting rule base.
8. The dynamic commodity recommending system based on multi-mode deep learning and real-time behavior of a user according to claim 1, wherein said recommending result output module is further used for monitoring behavior data of the user on the recommended commodity, and updating the recommending list when the user triggers a new behavior.
9. The dynamic commodity recommending system based on multi-mode deep learning and user real-time behavior according to claim 1, wherein the recommending result outputting module introduces commodity diversity constraint when displaying recommending results, and the commodity diversity constraint is not more than 40% in the list of recommending commodities.
10. The dynamic commodity recommendation system according to any one of claims 1-9, wherein said system further comprises: And the feedback optimization module is used for acquiring feedback data of the recommended result from the user and iterating and continuously optimizing parameters of the improved DeepFM model based on the feedback data.

Description

Dynamic commodity recommendation system based on multi-mode deep learning and user real-time behaviors Technical Field The invention relates to the technical field of big data, in particular to a dynamic commodity recommendation system based on multi-mode deep learning and real-time behaviors of users. Background With the rapid development of information technology, the popularity of the internet and mobile devices has increased exponentially in the amount of data generated by users. The data not only contains basic information of the user, but also contains multi-dimensional information such as browsing history, purchase records, social media interaction, geographic positions and the like of the user. These data provide a rich resource for the merchandise recommendation system, but at the same time present unprecedented challenges. Currently, there are many commodity recommendation systems on the market, mostly based on traditional machine learning algorithms or simple statistical models. These systems improve the accuracy of the recommendation to some extent, but still have the problems of difficulty in capturing complex and variable interest preferences of users, difficulty in processing large-scale and high-dimensional data, low recommendation efficiency, and difficulty in adapting to dynamically changing market environments and user behaviors. Some prior art techniques employ deep learning algorithms for recommendation, such as recommendation models based on Recurrent Neural Networks (RNNs) or Convolutional Neural Networks (CNNs). However, these models remain challenging to recommend, and major problems include: The multi-mode feature fusion is insufficient, wherein commodity data comprises multi-mode information such as texts (commodity titles, details), images, videos and the like, the existing system is independently processed in multiple pairs of single modes, and a deep fusion mechanism is lacked, so that feature representation is on one side, and commodity core value cannot be comprehensively reflected; The user behavior response is lagged, that is, most systems rely on historical behavior data to construct a static recommendation model, the real-time behavior capture such as current browsing, clicking and purchasing of lamps is not timely carried out on users, the recommendation result has long update period, and the instant requirement change of the users cannot be adapted; The dynamic adaptability is poor, the recommendation strategy is fixed, and the dynamic adjustment cannot be performed according to the scenes such as platform flow fluctuation, commodity inventory change, sales promotion activities and the like, so that the sales promotion commodity is not fully exposed or the sales stagnation commodity is excessively recommended. Therefore, although the prior art tries to introduce a deep learning model to improve the feature extraction capability, the prior art has the defects of insufficient multi-mode commodity feature fusion, corresponding lag of user behaviors, poor suitability of a recommendation strategy and the like. Therefore, a dynamic commodity recommendation system with both accuracy and real-time performance is needed to solve the above-mentioned core pain. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide a dynamic commodity recommendation system based on multi-mode deep learning and real-time behaviors of users, which can realize deep mining of commodity multi-dimensional characteristics, accurate capturing of real-time demands of users and intelligent iteration of recommendation strategies and can solve the problems of one-sided characteristics, response lag and poor dynamic adaptability of the existing system. In order to achieve the above purpose, the invention adopts the following technical scheme: a dynamic commodity recommendation system based on multi-mode deep learning and real-time behavior of users is characterized by comprising the following components: The multi-mode commodity feature extraction module is used for carrying out deep fusion on multi-mode data of commodity texts, images and videos by adopting an improved transducer model to extract commodity feature vectors; the user real-time behavior perception module is used for capturing user behavior data in real time through the streaming computing framework and constructing user interest vectors; The dynamic recommendation decision module is used for screening candidate commodities according to commodity feature vectors and user interest vectors, and generating a recommendation commodity list by utilizing the improved DeepFM model in combination with scene factors; and the recommendation result output module is used for displaying the recommendation result according to the type of the user terminal. Further, the multi-mode commodity feature extraction module includes: The commodity information acquisition unit is used for acquiring multi-mode commodity information d