Detailed Description
Various embodiments of the present invention will be described below with reference to the accompanying drawings.
FIG. 1 is a schematic flow chart diagram illustrating one embodiment of a resume evaluation method of the present invention.
As shown in fig. 1, the resume evaluation method according to the embodiment of the present invention includes:
s101, acquiring a plurality of first resumes corresponding to a plurality of first feedback data comprising first feedback action items;
the first feedback action item indicates that the first resume is associated with the first feedback action, and the first feedback data may be feedback data actively or automatically provided based on the first feedback action made by the resume filter on the resume, for example, if the resume filter considers that the details need to be viewed after seeing a summary of the resume, the first feedback data on the resume may be actively provided when or after performing the action of viewing the details of the resume, or the first feedback data may be automatically generated by the resume data platform based on the resume filter performing the action of viewing the details of the resume, and the first feedback data may include the first feedback action item indicating the viewed details. The first feedback data may also be feedback data actively submitted by the job seeker providing the resume to the resume data platform after confirming the feedback action of the resume screener, such as invited interviews, passing dates of use, and the like, and the resume data platform generates feedback data including items representing the feedback action such as inviting interviews, passing interviews, and the like based on the feedback of the job seeker.
The first feedback action items may be, in addition to data items associated with feedback actions such as viewing resume details, inviting interviews, passing interviews, etc., feedback actions associated with, for example, revenue candidate talent databases made by resume screeners on resumes, revenue candidate interview libraries, etc., without notifying job seekers.
It should be noted that, for the same resume, the feedback data generally includes a feedback action item, and when the resume subjected to a stage feedback action passes through the selection of the next stage and is associated with the feedback action of the next stage, the feedback action item associated with the previous stage in the feedback data of the resume is covered by the feedback action item of the new stage, so that the feedback data item in the feedback data of the resume can well reflect the selection stage of the resume at present.
In the embodiment of the invention, when the first resume is obtained, the concerned feedback action is taken as the first feedback action to obtain the first resume corresponding to the first feedback data of the first feedback action item related to the first feedback action, so that the obtained resumes are resumes in the same selection phase, for example, the resumes are also in the phase of being checked for the details of the resumes or in the phase of being invited to interview.
S102, constructing feature vectors of the first resumes as a plurality of positive samples to train an evaluation model to learn the mapping relation between the feature vectors and the first feedback action items;
after obtaining a plurality of first resumes indicating that the job seekers are all in the same selection stage, a feature vector may be constructed for each of the first resumes, where the feature vector may be composed of features of multiple dimensions, and some of the features may be, for example, a correlation between a text of the first resume and an industry feature, a correlation between a text of the first resume and a culture of the job seeker enterprise, a correlation between a text of the first resume and resumes of other job seekers who seek jobs in the same selection stage within the same industry, a correlation between a text of the first resume and resumes of other job seekers who seek jobs in the same selection stage within different industries and the same selection stage within the same industry, and the like. Here, the resumes of other job seekers in the same selection phase are identified by the feedback data of the other resumes also containing the first feedback action item, i.e., the other resumes are also associated with the first feedback action item. As an example, when the enterprise culture or job demand of the recruiter is a job seeker familiar with a certain overseas national culture, the relevance of the job seeker embodied in the resume for the elapsed time of the national study life is higher, and the value of the characteristic of the resume is higher; when there is more similarity between the text of the first resume and the resumes of other job seekers in the same selection stage, for example, the academic calendar, the work experience, and the like, the value of the feature of the resume is higher.
By analogy, a number of features that can explain the first feedback action can be built into the first resume. Each feature is mapped to a fixed dimension, and the value of the feature is used as the value of the dimension, so that a feature vector can be constructed for the first resume corresponding to each piece of first feedback data, the feature vector is determined as a positive sample based on the first feedback action item associated with the concerned feedback action, and a plurality of positive samples can be obtained as training samples for a plurality of first resumes.
After the training samples are constructed, a plurality of machine learning models can be selected as evaluation models to be trained on the training samples, so that the evaluation models can learn the mapping relation between the feature vectors and the first feedback action items according to a large number of training samples. Machine learning models such as GBDT, Logistic regression, Factorization Machine, DNN, and the like. These models can learn the mapping between feature vectors and sample identities/attributes based on a large number of training samples.
S103, calculating an evaluation score of the resume to be evaluated relative to the first feedback action item according to the mapping relation based on the feature vector of the resume to be evaluated through the evaluation model.
After the evaluation model is trained, any resume to be evaluated can be evaluated by the evaluation model. Specifically, a feature vector in the same mode as the training sample can be constructed for a resume to be evaluated, the feature vector of the resume to be evaluated is input into the evaluation model, and the evaluation model evaluates values for each feature in the feature vector of the resume to be evaluated according to the learned mapping relationship, so that an evaluation score of the resume to be evaluated with respect to the first feedback action item is obtained.
According to the embodiment of the invention, the evaluation models are respectively established aiming at different feedback action items, so that the targeted reference scores can be provided for the resume screeners, different requirements of the resume screeners can be met, and the working efficiency is improved. For example, for an HR that wishes to see more of a few candidate resumes, more attention may be paid to the recommendation scores given by the evaluation model trained based on the feedback action "see details", whereas for an HR that wishes to develop an interview as soon as possible, more attention may be paid to the recommendation scores given by the evaluation model trained based on the feedback action "invite interview" or "go through interview".
According to the embodiment of the invention, the evaluation model is established by utilizing the prior resume feedback data, the mapping relation between the concerned feedback action and the characteristics of the resume can be learnt in a targeted manner, and for any resume to be evaluated, the evaluation model can give out a relatively accurate evaluation score based on the learnt mapping relation according to the characteristics of the resume, so that powerful technical assistance is provided for resume screeners by effectively utilizing the resume feedback data, and the workload of enterprise HR and first-line business supervisor in brief duration screening is greatly reduced.
FIG. 2 is a schematic flow chart diagram illustrating another embodiment of a resume evaluation method of the present invention.
As shown in fig. 2, the method of the embodiment of the present invention includes:
s201, obtaining a plurality of first resumes corresponding to a plurality of first feedback data comprising first feedback action items, and constructing feature vectors for the first resumes to serve as a plurality of positive samples;
s202, a plurality of second resumes corresponding to a plurality of second feedback data comprising second feedback action items are obtained, and feature vectors are constructed for the plurality of second resumes to serve as a plurality of negative samples;
s203, training an evaluation model to learn a mapping relation between the feature vector and the first feedback action item by using the positive sample and the negative sample;
and S204, calculating the evaluation score of the resume to be evaluated relative to the first feedback action item according to the learned mapping relation based on the feature vector of the resume to be evaluated through the evaluation model.
Steps S201 and S204 in the embodiment of the present invention are the same as those in the embodiment shown in fig. 1, and only steps S202 and S203 will be described here.
Specifically, in the embodiment of the present invention, the feature vector constructed for the first resume is used as a positive sample, and the feature vector constructed for the second resume is also used as a negative sample to train the evaluation model together with the positive sample. Here, the second resume corresponds to the second feedback data in order to distinguish from the feedback data of the first resume, and the first feedback data and the second feedback data may have the same source and form, except that the second feedback data cannot include the first feedback action item but includes the second feedback action item, the second feedback action item indicates that the second resume is associated with the second feedback action, and the second feedback action indicates the feedback action of any previous stage of the first feedback action.
In the embodiment of the present invention, in order to enable the evaluation model to identify the positive sample and the negative sample, a label (label) with a value of 1 may be assigned to the positive sample, and a label with a value of 0 may be assigned to the negative sample. The value of the flag is not limited to 0 and 1, and will be specifically described later.
As an example, when the first feedback action is an invitation interview, the second feedback action may be a feedback action at any stage prior to the invitation interview stage, such as viewing a resume summary or viewing resume details; when the first feedback action is to pass the interview, the second feedback action can be a feedback action to pass any stage prior to the interview stage, such as viewing a summary of the resume, viewing details of the resume, or inviting the interview. In other words, the second resume is a job seeker resume that has not been considered by the resume screener in the previous selection phase or that was considered by the resume screener to not belong to the appropriate person selection, which failed to reach the first resume reached phase.
In the embodiment of the invention, the evaluation model is trained by taking the feature vectors of the second resumes as the negative samples together with the positive samples of the feature vectors of the first resumes, so that the learning progress and the learning accuracy of the evaluation model can be accelerated.
In some embodiments of the present invention, based on the embodiment shown in fig. 2, different weighting coefficients are set for the feedback actions at different stages in the second feedback action, and the negative samples are adjusted according to the set weighting coefficients. For example, when a first feedback data item indicates passing of the interview, among different second feedback data items indicating feedback actions at different stages, those indicating that only the summary of the resume is viewed and the details of the resume are not viewed are relatively low in reference value for evaluation, and those indicating that the interview is invited and the review is not passed have a high reference value for evaluation, a higher weighting coefficient may be set for the second feedback action with a high reference value to indicate that it is more important, and a lower weighting coefficient may be set for the second feedback action with a low reference value to indicate that it is less important, and corresponding negative samples may be adjusted according to the weighting coefficients, such as setting different importance marks for the negative samples, or duplicating the negative samples to correspond to multiple shares of the weighting coefficients, and so on. By the embodiment of the invention, the learning accuracy of the evaluation model can be further improved.
In each of the above embodiments of the present invention, after the third feedback data of the resume to be evaluated is obtained, the mapping relationship learned by the evaluation model may be modified according to whether the third feedback data includes the first feedback action item and the evaluation score that the evaluation model has given for the resume to be evaluated. For example, when the evaluation model provides a higher evaluation score for a resume to be evaluated with respect to a predetermined feedback action item, that is, the evaluation result is that the resume to be evaluated has a higher probability of being subjected to a corresponding feedback action of the resume screener, no matter whether the feedback action actually made by the resume screener on the resume to be evaluated is the feedback action for which the evaluation score is directed, the mapping relationship learned by the evaluation model can be corrected according to third feedback data correspondingly obtained by the resume to be evaluated and the evaluation score previously obtained by the resume to be evaluated, because the feature vector constructed for the resume generally consists of a plurality of features, and it is difficult for the situation that each feature of a value-taking feature vector is an upper evaluation limit value or a lower evaluation limit value to occur. According to the embodiment of the invention, the evaluation result and the actual feedback action are continuously learned, so that an excellent training sample can be provided for the evaluation model, the learning accuracy of the evaluation model is further improved, and the evaluation reference degree is improved.
In some embodiments of the present invention, the first feedback data of the first resume may include, in addition to the feedback action item representing the feedback action, other related auxiliary information to assist in the evaluation, so that the evaluation model can give an evaluation score to the resume to be evaluated based on the auxiliary information at the same time, and the given evaluation score can be associated with the auxiliary information, thereby meeting different screening requirements of the resume screener.
In one embodiment of the invention, the auxiliary information in the first feedback data of the first resume may include first position information. For example, when the resume screener needs to screen the resume of the job seeker for a specific job and does not consider the specific job (e.g., jobs such as accounting and foreground required by each job), the feedback information of the processed first resume can include the first job information, so that the training samples used by the evaluation model in learning the mapping relationship are all based on the job information, and the requirement of the resume screener for selecting the specific job can be met in a targeted manner.
In an embodiment of the present invention, the auxiliary information in the first feedback data of the first resume may include first industry information and first position information, so that training samples used for learning by the evaluation model are based on a specific position of a specific industry, and therefore, a resume screener can be specifically helped to select a suitable person with experience of the industry for the position.
In one embodiment of the present invention, the auxiliary information in the first feedback data of the first resume may further include feedback time information, so that when the first resume is selected, the selection may be performed according to a time range of interest, so that training samples used by the evaluation model for learning are all based on the feedback data in a specific time period, and therefore, the resume screener can be pertinently helped to select candidate resumes in the specific time period.
It should be noted that, although the above describes an example in which the feedback data of the first resume includes a plurality of auxiliary information, in the embodiment of the present invention, the feedback data of the first resume may not include any auxiliary information, or the auxiliary information is included but is not considered when the first resume is selected, so that the universality score of the job seeker can be evaluated. On the other hand, it may also be the case that: the resumes recommended by the resume data platform to the resume screener have been adapted to, for example, specific positions, so that the selected first resume is actually associated with the specific positions, but is not represented in the first feedback data thereof.
The embodiment of the invention can be applied to various scenes. For example, for job seekers, their resumes can be scored by using the trained evaluation model of the resume data platform; for the resume screener, multiple candidate resumes primarily selected based on text matching can be scored by utilizing an evaluation model trained by the screener, an enterprise or a resume data platform, and then candidate resumes with scores above a certain threshold value can be selected from the candidate resumes for selection based on the ranking of the candidate resume scores. The resume screener can also screen the candidate resumes by combining other screening conditions such as the current working place, salary requirements, working age and the like of the candidate, and sort the screening results according to the scores of the evaluation model to ensure that the resumes arranged in the front are more in line with the requirements of the resume screener. Therefore, the embodiment of the invention can greatly improve the working efficiency of the resume screener, reduce the workload and improve the referability of the resume recommendation.
FIG. 3 is a schematic flow chart diagram illustrating one embodiment of a resume evaluation method of the present invention.
As shown in fig. 3, the resume evaluation method according to the embodiment of the present invention includes:
s301, acquiring a plurality of first resumes and first feedback data thereof;
s302, constructing a feature vector of a part of first resumes, corresponding to first feedback data comprising positive feedback action items, in the plurality of first resumes to serve as a plurality of positive samples;
s303, constructing a feature vector of another part of the first resumes corresponding to the first feedback data comprising the negative feedback action item in the plurality of first resumes to serve as a plurality of negative samples;
s304, training an evaluation model to learn a mapping relation between the feature vector and the positive feedback action item by using a plurality of positive samples and a plurality of negative samples; and
s305, calculating the evaluation score of the resume to be evaluated relative to the positive feedback action item according to the mapping relation based on the feature vector of the resume to be evaluated through the evaluation model.
In the embodiment of the present invention shown in fig. 3, S304-S305 are similar to S203-S204 of the embodiment shown in fig. 2, except that the evaluation model has different feedback action items concerned, and the training samples are constructed in different ways due to the different feedback action items concerned, which will be described in detail below.
In general, in the embodiment of the present invention, rather than focusing on a specific feedback action associated with the resume as in the embodiment shown in fig. 1-2, all feedback actions associated with the resume are divided into two categories, one being positive feedback actions and one being negative feedback actions. The positive feedback action may include, for example, actions of the resume screener to check resume details, invite interview, accept the job seeker to select a suitable person to a corresponding extent through interview, and the like; the negative feedback action is, for example, an action of the resume screener to view the summary of the resume on the first resume, and the condition of the job seeker is considered to be not in accordance with the minimum requirement of the resume screener because the resume screener cannot be interested in viewing the complete resume in detail.
Therefore, in S301, the first resume is selected, instead of determining whether the first feedback data includes a certain feedback action item to determine whether the first feedback data is a satisfactory resume, a corresponding number of resumes with feedback data may be obtained according to the required total amount of the training samples, then in S302-S303, the first resumes are divided into two parts of first resumes according to whether the first feedback data includes a positive feedback action item representing a positive feedback action or a negative feedback action item representing a negative feedback action, a part of the first resumes corresponding to the positive feedback action is constructed with feature vectors as a plurality of positive samples, another part of the first resumes corresponding to the negative feedback action is constructed with feature vectors as a plurality of negative samples, and then in S304, the positive samples and the negative samples are used to train and evaluate the mapping relationship between the model learning feature vectors and the positive feedback action items, and in S305, a feature vector input evaluation model is constructed for the resume to be evaluated, and an evaluation score of the resume to be evaluated with respect to the positive feedback action item is calculated by the evaluation model based on the feature vector of the resume to be evaluated according to the learned mapping relationship.
Similar to the foregoing embodiments with respect to the construction of the feature vector of the resume and the training of the evaluation model, the feature vector of the first resume may be constructed based on, for example, a correlation between the first resume associated with the positive feedback action and other resumes associated with any one of the positive feedback actions as one of the features, and a detailed description thereof is omitted.
In the embodiment of the invention, the feedback actions are divided into two types in the construction of the training sample, namely positive feedback actions and negative feedback actions, so that the trained evaluation model is used for learning the mapping relation aiming at all the positive feedback actions, and the characteristics of the resume associated with the positive feedback actions in different stages can be comprehensively considered for comprehensively scoring the resume to be evaluated, thereby providing a reference value based on comprehensive consideration for resume screeners.
In some embodiments of the present invention, similar to the previous embodiments, different weighting coefficients may be set for different positive feedback operations, and the positive samples may be adjusted according to the set weighting coefficients. For example, in different positive feedback data items representing positive feedback actions at different stages, those positive feedback action items representing that resume details are viewed have relatively low reference value for evaluation, and those positive feedback action items representing that interviews are invited or that interviews are passed have higher reference value for evaluation, a higher weighting coefficient may be set for the positive feedback action with high reference value to represent that the positive feedback action with high reference value has higher importance, and a lower weighting coefficient is set for the positive feedback action with low reference value to represent that the positive feedback action with low reference value has lower importance, and corresponding negative samples may be adjusted according to the weighting coefficients. By the embodiment of the invention, the learning accuracy of the evaluation model can be further improved.
Regarding the utilization of the weighting coefficients, if a machine learning model supporting label as a real number, such as GBDT, is selected, a sample weight corresponding to the weighting coefficient may be set for each positive sample, for example, when the original label of a positive sample is 1 and the weighting coefficient of the positive sample is 5, the original label 1 of the positive sample may be multiplied by the weighting coefficient 5 to obtain 5, which is used as a new label of the positive sample. If a machine learning model supporting only 0 and 1 as a label, such as Logistic Regression (Logistic Regression), is selected, the number of copies of a positive sample can be copied according to the value of the weighting coefficient corresponding to the positive sample, such as 10 copies of the positive sample when the weighting coefficient is 10. Both methods can enable the evaluation model to attach more importance to important forward feedback data in the learning process, and if a resume is recommended in the future, feedback with higher importance can be obtained, and the obtained evaluation score is higher.
Similarly, in the embodiment of the present invention, after the second feedback data of the resume to be evaluated is obtained, the mapping relationship learned by the evaluation model may also be corrected according to whether the second feedback data includes a positive feedback action item and the evaluation score that the evaluation model has given for the resume to be evaluated.
Similarly, in the embodiment of the present invention, the first feedback data of the first resume may or may not include auxiliary information, and if the auxiliary information is included, the evaluation score given by the evaluation model is associated with the auxiliary information. Here, the auxiliary information may also include first position information, first business information, and/or feedback time information, etc.
Similarly, the evaluation model in the embodiment of the invention can also be used for giving evaluation scores to a plurality of candidate resumes of the resume screener respectively, and the resume screener can select the candidate resumes with the scores above a certain threshold value from the candidate resumes for selection based on the ranking of the candidate resume scores, so that the working efficiency of the resume screener is improved.
FIG. 4 is a schematic flow chart diagram illustrating another embodiment of a resume evaluation method of the present invention.
As shown in fig. 4, the resume evaluation method according to the embodiment of the present invention includes:
s401, acquiring a plurality of first resumes corresponding to a plurality of first feedback data comprising first feedback action items, and constructing a feature vector for each first resume as a negative sample;
s402, obtaining a plurality of second resumes corresponding to a plurality of second feedback data comprising second feedback action items, and constructing a feature vector for each second resume as a positive sample;
s403, training an evaluation model by using positive samples and negative samples to learn the mapping relation between the feature vectors and the second feedback action items;
s404, calculating the probability that the resume to be evaluated is associated with the second feedback action under the condition that the resume to be evaluated is associated with the first feedback action according to the mapping relation based on the feature vector of the resume to be evaluated through the evaluation model.
Unlike the previous embodiments, the first feedback action item represents that the first resume is associated with the first feedback action, the second feedback action item represents that the second resume is associated with the second feedback action, and the first feedback action is defined as a feedback action of a previous stage of the second feedback action, i.e. the first feedback action and the second feedback action are feedback actions of two adjacent stages, in other words, the present embodiments focus on a relationship between the feedback actions of the two adjacent stages, or further, focus on a possibility or probability of being subjected to a feedback action of a next stage after being subjected to the feedback action of the previous stage, and thus, the present embodiments are different from the previous embodiments in construction of a training sample.
Specifically, in order to learn the relationship between the feedback actions of two adjacent stages, it is first necessary to determine a feedback action of a subsequent stage as an attention target, for example, attention is paid to a feedback action of an invitation interview made by a resume screener for a resume, the invitation interview is taken as a second feedback action, the feedback action of viewing details of the resume at a previous stage of the invitation interview is taken as a first feedback action, the resume with feedback data is searched, a first resume corresponding to the feedback data including the feedback action item of viewing details and a second resume corresponding to the feedback data including the feedback action item of inviting interview are obtained from the feedback action, a feature vector constructed for the first resume is taken as a negative sample, a feature vector constructed for the second resume is taken as a positive sample, and the positive sample and the negative sample are taken as training samples to train the evaluation model, to learn the mapping relationship between the feature vector and the second feedback action item as the attention target. After the evaluation model is trained, for the resume to be evaluated which is already associated with the first feedback action, a feature vector input evaluation model can be constructed to obtain the probability that the resume to be evaluated is associated with the second feedback action. The probability here is also an evaluation score from another point of view, for example, if the evaluation model has a score upper and lower limit range of 0-100 in the present embodiment, when the evaluation model gives a score of 60 to a resume, the probability that the resume is associated with the second feedback action on the basis of the first feedback action can be considered to be 60%.
The method for constructing the feature vector of the resume in the embodiment of the present invention and the method for training the evaluation model are similar to the aforementioned embodiments, for example, the feature vector of the first resume or the second resume is constructed by using the correlation between the first resume or the second resume and other resumes associated with the second feedback action as one of the features. One difference between the embodiment of the present invention and the foregoing embodiment is that the weighting coefficients of the samples are not considered, and the whole resume feedback data does not need to be considered, for example, when training an evaluation model based on the feedback action of interview invitation, only the feedback data including the feedback data item representing the details of the review resume and the feedback data including the feedback data item representing the interview invitation need to be considered, the former is used for designing the negative samples, and the latter is used for designing the positive samples, and correspondingly training the evaluation model. According to the setting, the trained evaluation model can predict the probability of 'sending out interview invitation under the condition that the resume screener checks the details of the resume' of the resume to be evaluated.
Therefore, through the embodiment of the invention, the recommended value of the next stage can be dynamically given according to the stage action of the resume screener on the candidate resumes, for example, after a predetermined number of candidate resumes to be invited to interview are selected from the candidate resumes recommended by the model and considering the view details, the model respectively gives the corresponding probability of the job seeker to be invited to interview, and can sort the resumes from high to low based on the given probability value, so that an accurate reference value is provided for the resume screener, a part of workload of further screening is avoided, the work efficiency of the resume screener is improved, and the method is suitable for the condition that the resume screener slightly depends on the evaluation model.
Similarly, in the embodiment of the present invention, after the third feedback data of the resume to be evaluated is obtained, the mapping relationship learned by the evaluation model may also be modified according to whether the third feedback data includes the second feedback action item as the attention target and according to the probability associated with the second feedback action item, which is previously evaluated by the evaluation model for the resume to be evaluated.
Similarly, in the embodiment of the present invention, the first feedback data of the first resume may or may not include auxiliary information, and if the auxiliary information is included, the evaluation probability given by the evaluation model is associated with the auxiliary information. Here, the auxiliary information may also include first position information, first business information, and/or feedback time information, etc.
Fig. 5 is a schematic flow chart of an embodiment of the resume evaluation method of the present invention, which is a modified embodiment based on the embodiment shown in fig. 4.
As shown in fig. 5, the evaluation method according to the embodiment of the present invention includes:
s501, selecting a plurality of groups of resumes, wherein each group of resumes comprises one part of resumes corresponding to a plurality of feedback data comprising one feedback action item and the other part of resumes corresponding to a plurality of feedback data comprising another feedback action item;
s502, for each group of resumes, constructing feature vectors for one part of resumes as negative samples, and constructing feature vectors for the other part of resumes as positive samples;
s503, training an evaluation model to learn a mapping relation between the feature vector and another feedback action item by using the negative sample and the positive sample;
s504, establishing a fusion evaluation model based on a plurality of evaluation models of a plurality of groups of resumes;
and S505, calculating the probability of the resume to be evaluated being associated with the target feedback action item based on the feature vector of the resume to be evaluated through the fusion evaluation model.
In the embodiment of the invention, an evaluation model is built between every two feedback actions of a plurality of continuous stages possibly made on the resume by the resume screener. For each set of resumes, the feedback action represented by the one feedback action item is a feedback action of a previous stage of the feedback action represented by the other feedback action item, and for the plurality of sets of resumes, the one feedback action item in one set of resumes represents a feedback action of an initial stage in the feedback actions of the consecutive stages, and the other feedback action item in the other set of resumes represents a feedback action of a last stage in the feedback actions of the consecutive stages, the feedback action of the last stage corresponding to the target feedback action item. Meanwhile, besides the feedback action items representing the initial stage and the final stage, the feedback actions represented by the other feedback action items of different resume groups have pairwise continuous relationship.
In the embodiment of the present invention, in S501, in selecting each group of resumes, two resumes including feedback action items representing feedback actions belonging to two consecutive stages respectively are classified into the same group, so that feedback data of a part of resumes included in the group of resumes includes a feedback action of a previous stage of the two consecutive stages, and feedback data of another part of resumes included in the group of resumes includes a feedback action of a next stage of the two consecutive stages.
In S502, for each set of resumes, a feature vector is constructed for a part of the resumes associated with the feedback action of the previous stage in the two consecutive stages as a negative sample, a feature vector is constructed for another part of the resumes associated with the feedback action of the next stage in the two consecutive stages as a positive sample, and when the feature vector is constructed for the part of the resumes or the another part of the resumes, each feature vector can be constructed by taking the correlation between each of the part of the resumes or the another part of the resumes and the other resumes associated with the another feedback action item as one of the features.
In S503, for each set of resumes, training an evaluation model corresponding to the set of resumes with corresponding positive and negative samples, and learning a mapping relationship between a feature vector of the resume and a feedback action item representing a feedback action at a later stage.
Then, as a key of the embodiment of the present invention, it is necessary to fuse the trained evaluation models of each group of resumes through a certain algorithm to establish a fused evaluation model, and then perform probability evaluation on the target feedback action on the resumes to be evaluated by using the fused evaluation model.
For example, the view details evaluation model F1 may be trained for the resume group with a view resume summary as a later stage feedback action, the invitation interview evaluation model F2 may be trained for the resume group with an invitation interview as a later stage feedback action, and the pass interview evaluation model F3 may be trained for the resume group with a pass interview as a later stage feedback action. Referring to the embodiment shown in fig. 4, each of these evaluation models F1, F2, and F3 may be used independently to predict, for example, the probability that a resume screener will click on and view the full information of a resume with a view to the summary of the resume, the probability that the resume screener will initiate an invitation to interview after viewing the details of the resume, and the probability that a candidate will pass the interview, for a given position, respectively. In the embodiment of the present invention shown in fig. 5, after the three evaluation models F1, F2, and F3 are fused, a fused evaluation model F (F1(x), F2(x), F3(x)) can be generated, where x is a feature vector constructed for the resume.
In the embodiment of the invention, the integral acceptance probability for the candidate resume can be given to the resume screener as the integral recommendation value of all feedback stages, for example, after a predetermined number of candidate resumes are recommended from a resume data platform or other ways, the fusion evaluation model can directly give the corresponding probability of the job seeker passing the interview to each candidate resume, and can sort the resumes from high to low based on the given probability value, thereby providing a recommendation scheme with higher reference value for the resume screener, avoiding most of the workload of the resume screener, greatly improving the working efficiency of the resume screener, and being applicable to the condition that the resume screener highly depends on the evaluation model. In fact, after the model is trained by a large number of a priori training samples and a posteriori training sample, the evaluation accuracy of the model is greatly improved, so that even for a resume screener who slightly depends on the evaluation model before, the highly dependent evaluation model can be considered to reduce the workload.
In embodiments of the present invention, the feature vector may be constructed in the same manner for each resume in each set of resumes, such that the feature vector of each resume includes features of the same dimension.
In the embodiment of the present invention, there may be a plurality of fusion manners for combining the plurality of evaluation models into the fusion evaluation model, for example, combining the plurality of evaluation models into the fusion evaluation model by performing weighted average on each evaluation model, or combining the plurality of evaluation models into the fusion evaluation model by performing weighted average after taking the logarithm of each evaluation model, and so on.
Similarly, in the embodiment of the present invention, after the feedback data of the resume to be evaluated is obtained, the mapping relationship learned by the evaluation model may also be corrected according to whether the feedback data includes the target feedback action item and according to the probability associated with the target feedback action item, which is previously evaluated for the resume to be evaluated by fusing the evaluation model.
Similarly, in the embodiment of the present invention, the feedback data of each resume in each set of resumes may also include or not include auxiliary information, and if the feedback data includes auxiliary information, the evaluation probability given by the evaluation model is associated with the auxiliary information. Here, the auxiliary information may also include first position information, first business information, and/or feedback time information, etc.
In addition, in the embodiment of the present invention, the feedback data of each resume in each set of resumes may further include information of a screener as auxiliary information, and since the training samples used are all constructed based on a specific screener, the thus trained evaluation model will become an evaluation model specific to the specific screener, and can give a higher evaluation value to a candidate resume meeting the requirements of the specific screener with higher accuracy, while the resume may obtain a lower evaluation value according to the requirements of another screener.
A scheme for training the evaluation model based on the screener information will be described in detail below with reference to fig. 6.
FIG. 6 is a schematic flow chart diagram illustrating another embodiment of a resume evaluation method of the present invention.
As shown in fig. 6, the resume evaluation method according to the embodiment of the present invention may include:
s601, acquiring a plurality of first resumes corresponding to a plurality of first feedback data comprising a first feedback action item and first screener information;
s602, constructing feature vectors of the first resumes as a plurality of positive samples to train an evaluation model to learn the mapping relation between the feature vectors and the first feedback action items; and
s603, calculating the evaluation score of the resume to be evaluated relative to the first feedback action item according to the mapping relation based on the feature vector of the resume to be evaluated through the evaluation model.
This embodiment of the invention is a variation of the embodiment shown in fig. 1. On the basis of the embodiment shown in fig. 1, the embodiment of the present invention further considers the first filter information included in the first feedback data of the first resume, that is, the first feedback action represented by the first feedback action item is an action made on the first resume by the first filter. In the embodiment of the present invention, the processing procedure is basically the same as that of the embodiment shown in fig. 1, except that the selection of the first resume is based on the resume of the feedback action made by the same filter on the resume.
According to the embodiment of the invention, different evaluation models are trained for different resume screeners according to different requirements or personal preferences of the different resume screeners, and the different evaluation models may give different recommendation values to the same candidate resume, so that the different resume screeners can obtain personalized recommendations for the candidate resume according to personalized requirements.
For example, in practical applications, sometimes there may be multiple resume screeners for the same job, such as for an IOS high-level engineer, there may be supervisors from two different departments to screen resumes, and there may be some difference between the preferences of two candidates, such as for a person with a supervisor that is more prone to having an excessively large company history. Theoretically, one position can be established for the two resume screeners respectively, but from the perspective of the HR of the enterprise, the division of the positions is too fine to facilitate management, so that it is also a common practice to combine the requirements of each business department into one position. In the embodiment of the invention, because the characteristic vector is constructed for the resume, the personalized recommendation can be realized based on the personal data of a specific resume screener, namely, for two service supervisors who look over the candidate resumes of the positions of IOS senior engineers, different candidate resumes can be recommended for the two service supervisors by the correspondingly trained evaluation model, which cannot be reached by the conventional resume recommendation method based on text matching.
In the embodiment of the present invention, there are some differences in the construction of the feature vector of the first resume from the other embodiments described above. Specifically, when constructing the feature vector for the first resume, the feature vector for each first resume may be constructed by taking the correlation between the first resume and the other resumes associated with the first feedback action of the same filter as one of the features, and the feature vector for each first resume may be constructed by taking the correlation between the first resume and the other resumes associated with the first feedback actions of different filters as one of the features. By considering the characteristics related to the screeners when constructing the characteristic vector for the resume, the personal preference of a specific screener can be captured, and the personal preference of the screener is taken into account when evaluating the candidate resume, so that personalized accurate recommendation is provided for the specific resume screener.
In one embodiment of the present invention, similar to the previous embodiment, negative examples may also be created for training of the evaluation model. For example, a plurality of second resumes corresponding to a plurality of second feedback data including the second feedback action items and the first filter information may be acquired, and feature vectors may be constructed as a plurality of negative samples for the plurality of second resumes, so as to train the evaluation model learning mapping relationship together with the positive samples. The second feedback action item indicates that the second resume is associated with a second feedback action, and the second feedback action may be a feedback action in a previous stage of the first feedback action (see the embodiment shown in fig. 4), a feedback action in each previous stage of the first feedback action (see the embodiment shown in fig. 2), or a feedback action with an opposite attribute to the first feedback action (for example, the first feedback action and the second feedback action are positive and negative feedback actions, respectively, and particularly refer to the embodiment shown in fig. 3).
Similarly, in the embodiment of the present invention, after the third feedback data of the resume to be evaluated is obtained, the mapping relationship learned by the evaluation model may also be corrected according to whether the third feedback data includes the first feedback action item and the evaluation value that the evaluation model has given for the resume to be evaluated.
Similarly, in the embodiment of the present invention, the first feedback data of the first resume may or may not include auxiliary information, and if the auxiliary information is included, the evaluation score given by the evaluation model is associated with the auxiliary information. Here, the auxiliary information may also include first position information, first business information, and/or feedback time information, etc.
Similarly, the evaluation model in the embodiment of the invention can also be used for giving evaluation values such as evaluation scores or evaluation probabilities for a plurality of candidate resumes of the resume screener respectively, and the resume screener can select candidate resumes with scores above a certain threshold value from the candidate resumes for selection based on the ranking of the candidate resume evaluation values, so that the work efficiency of the resume screener is improved.
In the following description, a cloud recruitment system is described as an example of a resume feedback data acquisition method in the embodiment of the present invention, it should be noted that the present invention is not limited to this example, and other feedback data acquisition methods, such as job seekers or enterprises actively submitting feedback data to a resume data platform or an enterprise to which a resume screener belongs, construct a resume database and automatically record feedback action data after the resume database is constructed on the resume data platform or the resume data is downloaded, and automatically construct the resume database and automatically record feedback action data are also within the scope of the present invention.
Different from the traditional enterprise internal recruitment system, the cloud recruitment system is an SAAS (software association) platform. The platform is developed and maintained by a third party and can be accessed to the Internet to serve all enterprises with recruitment management requirements. As a user, an enterprise using the platform does not need to purchase, deploy and maintain the system, and only needs to pay according to the number of users and the use time. The cloud recruitment systems have two ways of establishing and maintaining talent resume libraries, and the first situation is that the cloud recruitment systems are developed by large-scale internet recruitment websites and can directly share talent resume resources of parent companies; the second case is an independent cloud recruitment system, which requires the enterprise to authorize the system to download all resumes related to the enterprise post on the large internet recruitment website using the enterprise account. In the second case, once the cloud recruitment system has enough enterprise users, a relatively comprehensive talent resume library can be accumulated.
An enterprise user can newly release a job requirement in the cloud recruitment system or send a new employment application for an existing job, and for a job, descriptions such as a job name, a working year requirement, a academic requirement, a regional requirement, a skill requirement, a work responsibility and the like need to be provided, and a registered mailbox or system ID of one or more resume screeners also needs to be provided. Multiple resume screeners may be considered for the same job.
When the resume screener browses talent resumes recommended by the system, which resumes are only used for simply seeing the summaries, which resumes are clicked and examined in detail, and which resumes are marked as being capable of inviting interviews are automatically recorded by the cloud recruitment system. After the subsequent candidates participate in the interview, the resume screener also needs to fill in the interview evaluation and scoring of the candidates in the system. Here, the browsing situation, the detailed survey data and the interview invitation data of the resume can be acquired by a traditional internet recruitment website, but generally cannot be accurately corresponding to a specific resume screener, and can only be associated with an enterprise account. And the evaluation information of the candidate interviews cannot be obtained by the traditional internet recruitment website.
Further, in the cloud recruitment system, the resume screener can record whether the candidate receives the offer after interviewing and whether the candidate passes the trial period after enrollment.
Some examples of data recorded by the system are as follows:
<uid:xiaoming@Atech.com,position_id:45678,cv_id:23653490012,action:go_detail,time_stamp:1510578275>
<uid:limei@Atech.com,position_id:45678,cv_id:23653490012,action:invitation,time_stamp:1510576491>
<uid:xiaoming@byteScience.com,position_id:31682,cv_id:23653490012,action:invitation,time_stamp:1510590021>
the uid records an enterprise mailbox of the resume screener (which is an example of the screener information in the above embodiment), the position _ ID records an ID of a specific position, the cv _ ID records a talent resume ID, the action records actions of the resume screener, including actions of viewing details (go _ detail), inviting interview (invitation), and the like, and the time _ stamp records a timestamp of the occurrence of the actions.
After resume feedback data generated by the cloud recruitment platform user is collected, a corresponding resume evaluation model can be trained according to the data according to the scheme of any embodiment.
Embodiments of the present invention further provide a resume evaluation apparatus, which is implemented as, for example, a computer apparatus including a processor, and the computer apparatus further includes a memory, where computer-executable instructions may be stored, and the corresponding computer-executable instructions may be designed and stored in the memory as needed, so that the processor can execute the resume evaluation method in any embodiment of the present invention when executing the corresponding computer instructions.
While the embodiments of the present invention have been described in detail, the present invention is not limited to these specific embodiments, and those skilled in the art can make various modifications and modifications of the embodiments based on the concept of the present invention, which fall within the scope of the present invention as claimed.