CN-122022876-A - Self-adaptive retail consumption behavior prediction method and system

CN122022876ACN 122022876 ACN122022876 ACN 122022876ACN-122022876-A

Abstract

The invention discloses a self-adaptive retail consumption behavior prediction method and a self-adaptive retail consumption behavior prediction system, which are applied to the field of retail. The method comprises the steps of S1, deploying edge nodes and multi-mode sensors on a retail site, extracting behavior characteristics by an edge lightweight neural network, carrying out anonymization encryption uploading, destroying original data, S2, dynamically adjusting real-time and historical data weight according to a scene, fusing the characteristics by a multi-head self-attention network, S3, constructing a three-level HMM (hidden Markov model) for microscopic behavior, mesoscopic intention and macroscopic decision, carrying out hierarchical reasoning and prediction, and S4, carrying out federal learning update model, and carrying out incremental aggregation deployment of encryption parameters only. The system comprises edge perception and privacy protection, multi-source fusion and self-adaptive weighting, hierarchical HMM prediction, federal learning and self-adaptive modules. The invention has the advantages of taking privacy and instantaneity into consideration, improving prediction accuracy and scene suitability, and supporting retail accurate marketing.

Inventors

HU XIANDE
LI BAILIANG
HUANG HAIPENG
XIA XINGLONG

Assignees

苏州易泰勒电子科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251212

Claims (8)

1. An adaptive retail consumer behavior prediction method, comprising the steps of: Step S1, deploying edge nodes on a retail site, collecting consumer behavior data through a multi-mode sensor, extracting behavior characteristics in real time by adopting a lightweight neural network at an edge end, carrying out anonymization and encryption processing on the characteristics, uploading the characteristics, and destroying original data at the edge end; S2, receiving the characteristics processed in the step S1, introducing a self-adaptive weight dynamic adjustment algorithm, dynamically adjusting the weight proportion of real-time characteristic data and historical data based on the identified current retail scene, and carrying out weighted fusion on the time behavior characteristics by adopting an attention mechanism; S3, constructing a microscopic behavior, a middle-view intention and a macroscopic decision three-level hidden Markov model, wherein the fusion characteristic output in the step S2 is used as an observation sequence of the microscopic behavior hidden Markov model, the output hidden state is used as an observation sequence of the middle-view intention hidden Markov model, and the output hidden state is used as an observation sequence of the macroscopic decision hidden Markov model, so that layering reasoning and prediction from specific behaviors to abstract decisions are realized; And S4, updating the hierarchical hidden Markov model by adopting a federal learning framework, locally training the model by each edge node, uploading the encrypted model parameter increment to a central server to aggregate to generate a global model, and transmitting the global model to each edge node through a safety channel to realize self-adaptive deployment.
2. The method according to claim 1, wherein in step S1: the multi-mode sensor comprises an infrared array sensor and a depth camera; The behavior characteristics comprise residence time, taking frequency and article comparison behavior; The anonymizing process includes generating a random unique identifier for the feature.
3. The method according to claim 1, wherein in step S2: the self-adaptive weight dynamic adjustment algorithm extracts local characteristics through a convolutional neural network, identifies the current scene type and dynamically adjusts the weight according to the scene type; the attention mechanism is a multi-headed self-attention network.
4. The method according to claim 1, wherein in the step S3: The hidden state quantity of the microscopic behavior, the mesoscopic intention and the macroscopic decision hidden Markov model is sequentially increased; The micro-behavior hidden Markov model adopts a BN-LSTM network to extract features, and the mesoscopic intention and macro-decision hidden Markov model adopts a GRU network.
5. The method of claim 4, wherein the microscopic behavior hidden markov model has a hidden state number of 32, the mesoscopic behavior hidden markov model has a hidden state number of 64, and the macroscopic decision hidden markov model has a hidden state number of 128.
6. The method according to claim 1, wherein in the step S3, the state transition probabilities of the respective hidden markov models are calculated by means of a conditional random field model.
7. The method of claim 1, wherein in step S4, the model update process further comprises constructing an instant rewards function by smoothing error metrics, weight change penalty terms, and uncertainty metrics to evaluate predictive effects and guide model optimization.
8. An adaptive retail consumer behavior prediction system for implementing the method of any one of claims 1-7, comprising: the edge perception and privacy protection module is configured to collect data through a multi-mode sensor at a retail site, and extract, anonymize and encrypt behavior characteristics at an edge end; The multisource fusion and self-adaptive weighting module is configured to receive the characteristics, dynamically adjust data weights based on the current scene, and conduct characteristic fusion by adopting an attention mechanism; The hierarchical hidden Markov model prediction module is configured to perform hierarchical reasoning and prediction on the fused features by using a three-level hidden Markov model; and the federal learning and self-adapting module is configured to update and deploy the hierarchical hidden Markov model through a federal learning framework.

Description

Self-adaptive retail consumption behavior prediction method and system Technical Field The invention relates to the technical field of intersection of artificial intelligence and the Internet of things, and in particular relates to a self-adaptive retail consumption behavior prediction method and system based on multi-mode data fusion, hierarchical modeling and edge calculation, which can realize accurate and real-time prediction from microscopic behaviors to macroscopic decisions of consumers and ensure data privacy and scene suitability. Background With the rapid development of smart retail, accurate understanding and prediction of consumer behavior has become a key to enhancing retail industry performance. However, existing consumer behavior analysis techniques still have significant limitations: First, the data source is single and the analysis level is shallow. The traditional method depends on historical transaction records or single visual monitoring data, and is difficult to capture microscopic behavior sequences (such as picking up commodities, repeatedly comparing, watching for a long time and the like) of consumers in the shopping process, so that instantaneous shopping intention cannot be deeply inferred, and the prediction accuracy is low. Secondly, privacy security and processing efficiency are difficult to be compatible. Existing solutions generally need to directly transmit original data such as videos, images and the like containing a large amount of personal biological information to a cloud server for processing. The method has the advantages that not only is the huge privacy leakage risk caused, but also a heavy burden is caused on network bandwidth and cloud computing resources, and the requirement of off-line retail scenes on real-time response is difficult to meet. Again, the model is stiff and lacks scene adaptation. Most of prediction models adopt fixed parameters, and cannot be adaptively adjusted according to the obvious difference of consumer behavior patterns under different retail scenes (such as fresh supermarkets and electronic product shops), so that the model generalization capability is weak, and the performance is obviously reduced when the model is deployed across scenes. Finally, there is a lack of hierarchical modeling of the decision process. The prior art generally attempts to map the original behavior directly to the final decision in a single, flat model, ignoring the progressive logic of "microscopic behavior- & gt mesoscopic intent- & gt macroscopic decision" in the consumer decision process, resulting in a rough prediction result and poor interpretability. The Hidden Markov Model (HMM) is a probability-based time sequence modeling tool, and is characterized in that unobservable hidden states (such as shopping intention) are inferred through observable sequences (such as consumer behavior data), state transition probabilities are utilized to describe state evolution, and the hidden Markov model is suitable for time sequence prediction. However, in the prior art, HMMs are mostly used for single-level simple behavior mapping, a multi-level progressive inference architecture is not constructed, and a multi-mode data and scene self-adaptive mechanism are not combined, so that prediction precision and generalization capability of the HMMs in complex retail scenes are limited. Therefore, there is an urgent need in the art for a hierarchical consumption behavior prediction scheme that can achieve high efficiency, accuracy, and adapt to different retail environments while protecting privacy. Disclosure of Invention The primary object of the present invention is to overcome the above-mentioned drawbacks of the prior art, and to provide a method and a system for adaptive retail consumer behavior prediction, which can achieve privacy protection, real-time processing and accurate prediction. The core aim is to realize full chain, hierarchical understanding and prediction of consumers from microscopic behaviors, mesoscopic intention to macroscopic decision, and improve the self-adaptive capability of the system in different retail scenes. In order to achieve the above purpose, the present invention adopts the following technical scheme: an adaptive retail consumption behavior prediction method, comprising the steps of: Step S1, deploying edge nodes on a retail site, collecting consumer behavior data through a multi-mode sensor, extracting behavior characteristics in real time by adopting a lightweight neural network at an edge end, carrying out anonymization and encryption processing on the characteristics, uploading the characteristics, and destroying original data at the edge end; S2, receiving the characteristics processed in the step S1, introducing a self-adaptive weight dynamic adjustment algorithm, dynamically adjusting the weight proportion of real-time characteristic data and historical data based on the identified current retail scene, and carrying out weighted fusi