CN-116467639-B - New equipment cold start method for smart home portrait construction

CN116467639BCN 116467639 BCN116467639 BCN 116467639BCN-116467639-B

Abstract

The invention relates to the technical field of deep learning, in particular to a new equipment cold start method used in smart home portrait construction, which comprises the steps of dividing a historical record data set of household equipment to obtain an input data set, combining a label classification model and a meta learning model, then performing meta training on a training task set to obtain an optimal initialization parameter set, generating a test task set from the input data set, performing meta testing based on the optimal initialization parameter set to obtain a label classification result, and performing cold start based on the label classification result. The method solves the problem that the usage records of the new equipment cannot be correctly labeled and classified under a small sample scene when the user portrait is constructed.

Inventors

WANG JIAHAO
YAN HANG
FANG JIANPING
SUN ZUTONG
LIN JIANXUAN
LIAO XIAOSHUN

Assignees

电子科技大学长三角研究院(湖州)
厦门智小金智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20230403

Claims (4)

1. A new device cold start method for smart home portrait construction, comprising the steps of: s1, dividing a historical record data set of household equipment to obtain an input data set; S2, combining a label classification model and a meta learning model, and then performing meta training on a training task set to obtain an optimal initialization parameter set, wherein the label classification model is a TSformer model, and the meta learning model is a MAML model; s3, generating a test task set from the input data set, and performing meta-test based on the optimal initialization parameter set to obtain a label classification result; s4, performing cold start based on the label classification result; Combining the label classification model and the meta learning model, and then performing meta training on the training task set to obtain an optimal initialization parameter set, wherein the method comprises the following steps: s21, combining the Tsformer model and the MAML model to obtain a meta-learning model; S22, splitting any Task sample of the training Task set to obtain a first support set and a first query set; S23, carrying out one-step or multi-step back propagation updating on the original initialization parameters in the meta-learning model through the first support set to obtain parameters; S24, the parameters are sent to the first query set for testing in the meta-learning model, and after loss is calculated, the original initialization parameters are updated to obtain optimal initialization parameters; S25, circulating the steps S22 to S24, and training the preset number of Task samples to obtain an optimal initialization parameter set.
2. A new equipment cold start method for smart home portrayal construction according to claim 1, The dividing the historical record data set of the household equipment to obtain an input data set comprises the following steps: And dividing the historical record data set of the equipment by taking the Task as a unit to obtain an input data set.
3. A new equipment cold start method for smart home portrayal construction according to claim 1, The TSformer model comprises a time sequence feature processing module and an extraction channel feature processing module; The time sequence feature processing module comprises a position coding module, a Multi-Head Attention module and two time sequence feature processing modules The system comprises a module, a fully-connected neural network module, a merging module, a GRU (gate-controlled loop) neural network unit, a fully-connected network and a softmax classifier.
4. A new equipment cold start method for smart home portrayal construction according to claim 1, Generating a test task set from the input data set, performing meta-test based on the optimal initialization parameter set to obtain a label classification result, including: S31, generating a test task set from the input data set; s32, splitting the test task set to obtain a second support set and a second query set; S33, performing one or several iterations on the optimal initialization parameter set on the second support set to obtain a label classification result; And S34, evaluating the label classification result by using the second query set to obtain an evaluation result.

Description

New equipment cold start method for smart home portrait construction Technical Field The invention relates to the technical field of deep learning, in particular to a new equipment cold start method used in smart home portrait construction. Background Along with the development of social and economic levels, the technological content of household products is continuously improved, and more intelligent devices enter into thousands of families, so that the household life of people becomes more comfortable and convenient. Based on this background, how to analyze users in smart home scenarios, understand users, is becoming a significant and indispensable task. With the rapid development of deep learning theory, the construction of User portraits (User portraits) is attracting researchers' interest by analyzing User behavior habits in home scenes with the aid of deep learning algorithms. Through the user portrait technology, researchers can mine rich user labels from the user portrait technology, analyze users and understand users by combining clear data, and further provide personalized and customized services for the users, so that the home equipment really realizes intelligence. In the past few years, user portraits have been applied to the ground in many fields, such as Webert and Syskill, and the satisfaction degree of users on websites is analyzed by means of statistical analysis, so that a user interest model is built. Along with the development of machine learning and deep learning theory, rahimi et al construct a user region label prediction model based on friend relations by grabbing interactive information in microblog texts and using Logistics Regression and LPA algorithm. Bhtacharyya et al analyzed the keyword text in FaceBook using NLP techniques and mined the effects of friend making relationships in social interactions. However, the existing technology for constructing the user portrait has some disadvantages, mainly including the following points: (1) User portrait technology has been widely used in the fields of e-commerce, libraries, travel and the like, but has been started later in the fields of medical health, intelligent home and the like, and related research and practice are less. (2) The traditional portrait method is built based on a large amount of sufficient data scenes, however, the situation that the data amount is small often occurs in actual use, and a solution to the problem of cold start is lacking. In the process of constructing a portrait of a home user, researchers often need to start from a device usage record and mine hidden information such as habit preferences of the user for a specific device in order to describe the behavior habits of the user at multiple angles as much as possible. For equipment types with longer marketing time, furniture manufacturers can obtain a large number of use record data of different users, and send the data into a specific network for training so as to obtain meaningful classification results. However, for newly marketed devices, the tag classification network cannot be sufficiently trained due to the small data size, which results in low fitting degree and poor classification performance. Disclosure of Invention The invention aims to provide a new equipment cold start method for intelligent family portrait construction, which aims to solve the problem that a usage record of new equipment for constructing a user portrait cannot be correctly labeled and classified under a small sample scene. In order to achieve the above object, the present invention provides a new device cold start method for smart home portrait construction, comprising the steps of: s1, dividing a historical record data set of household equipment to obtain an input data set; S2, combining the label classification model and the meta learning model, and then performing meta training on the training task set to obtain an optimal initialization parameter set; s3, generating a test task set from the input data set, and performing meta-test based on the optimal initialization parameter set to obtain a label classification result; S4, cold start is carried out based on the label classification result. The method for dividing the historical record data set of the household equipment to obtain an input data set comprises the following steps: And dividing the historical record data set of the equipment by taking the Task as a unit to obtain an input data set. The label classification model is Tsformer model, and the meta learning model is MAML model. Wherein, the Tsformer model comprises a time sequence feature processing module and an extraction channel feature processing module; The time sequence feature processing module comprises a position coding module, a Multi-Head Attention module, two Add & LayerNorm modules, a fully-connected neural network module, a merging module, a GRU (gate-controlled loop) neural network unit, a fully-connected network and a softmax classifier. After c