Research On Person-post Matching Based On User Portrait

With the rapid development of Internet technology,my country’s Internet penetration rate has been increasing,providing good conditions for online recruitment.In the field of recruitment,the degree of fit between the quality of talents and job requirements,that is,the research on matching between people and positions,has become a hot topic.At present,various recommendation algorithms have been widely used in the field of human-post matching.However,due to the limitations of the recommendation algorithm itself,there are still many problems in human-post matching.This specific makes some improvements from different angles.The specific work is as follows:Firstly,the concept of "user portrait" is introduced into the person-post matching process,and the attributes that need to be considered for users and positions are systematically analyzed.Traditional content-based recommendation generally only considers the attributes that have a corresponding relationship between a few users and positions,ignores the characteristics of users and positions themselves,and lacks potential information mining.Based on deep learning,resume features and recruitment requirements are directly spliced and input into the prediction layer,which is not explanatory and lacks the information flow relationship between recruitment and requirements.To this end,this specific makes a detailed division of user and job attributes through the characterization of user portraits and job portraits,and analyzes more potential attributes from existing data.Secondly,user and post data are labeled,which effectively eliminates irrelevant text words and reduces the introduction of noise.In the existing person-post matching process,they all focus on the similarity calculation of unstructured text.The existing document-level and sentence-level matching are prone to introduce the interference of irrelevant words.Labeling the attribute data of users and positions can effectively eliminate irrelevant features and reduce the introduction of noise.In order to extract suitable labels,this specific constructs a user dictionary of this domain,trains the domain’s inverse document frequency IDF corpus to compute TF-IDF to extract labels.In order to enrich the label of the portrait,the Text CNN model is used to predict the user’s competent position label,and the position is graded according to work experience.Thirdly,the calculation of multi-attribute cross-features between users and positions realizes dimensionality reduction,and analyzes historical data to introduce delivery preference features.The attributes of users and positions are mostly text label types.Directly converting them to word vectors as machine learning brings the problem of too high dimensionality,and directly splicing the features of users and positions will lose potential semantics.For attributes with corresponding relationships,different matching calculation methods are proposed in combination with semantic relationships,so that the independent features of users and positions are replaced by the matching features of the two,and the input dimension is reduced.In addition,to further explore the potential characteristics of non-corresponding attributes between users and positions,this specific combines the historical matching data of users and positions to mine the delivery preferences of different professions and the delivery preferences of different professions,so as to introduce new features.Finally,the GBDT+LR fusion model is used to establish a person-post matching model.Using the feature that the GBDT model can combine features,generate new discrete feature vectors and explore the potential connections of features.Then,the new features are used as input to a logistic regression(LR)model to produce the final predictions.Through comparative experiments,the GBDT+LR fusion model has a certain improvement in various indicators compared with the single GBDT and LR models in the field of person-post matching.