A candidate job-seeking intention matching recommendation method fusing timing behavior similarity
Patent Information
- Application Number
- CN202611146024.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-10-13
AI Technical Summary
[0003]现有方案主要将浏览、收藏、投递、面试等行为统一编码进行为序列，未对招聘场景中大量存在的负向行为（如候选人屏蔽企业、拒绝面试邀约、重复浏览某类岗位却从不投递等）进行显式区分与建模；这些负向信号在现有LSTM/Transformer编码中被当作普通序列节点处理，导致候选人的“不感兴趣意向”被错误解读为“感兴趣意向”，进而向候选人反复推送其排斥的岗位，严重影响用户体验
1．该融合时序行为相似性的候选人求职意向匹配推荐方法，通过显式识别并惩罚候选人的负向行为（如屏蔽企业、拒绝邀约、重复浏览未投递等），能够在匹配打分阶段主动排除候选人不感兴趣的岗位，有效避免将负向信号误判为正向兴趣；同时，结合岗位侧历史交互行为的反向校验，使得推荐结果不仅考虑候选人自身的偏好，还参考了历史上与该岗位成功匹配的候选人特征，从而大幅提升人岗匹配的准确性，降低候选人对推荐结果的“不感兴趣”标记率。
Smart Images
Figure FT_1 
Figure SMS_2 
Figure SMS_21
Abstract
Description
Technical Field
[0001] This invention relates to the field of human resource information processing technology, specifically to a candidate job intention matching and recommendation method that integrates temporal behavioral similarity. Background Technology
[0002] With the popularization of online recruitment platforms, candidate job intention matching and recommendation systems have become a core infrastructure for improving recruitment efficiency. The existing job matching technology solutions mainly involve: the system collecting behavioral data such as browsing, saving, submitting, and interviewing of candidates from recruitment platforms, constructing behavioral sequences by sorting by timestamp, encoding behavioral features through time series models such as LSTM, and performing multimodal fusion with resume text features and job description text features. Finally, the system calculates job matching scores and ranks and recommends candidates using metrics such as cosine similarity.
[0003] Existing solutions primarily encode browsing, saving, applying, and interviewing behaviors as sequences, without explicitly distinguishing and modeling the numerous negative behaviors present in recruitment scenarios (such as candidates blocking companies, rejecting interview invitations, repeatedly browsing certain types of positions but never applying). These negative signals are treated as ordinary sequence nodes in existing LSTM / Transformer encoding, causing candidates' "disinterest intentions" to be misinterpreted as "interest intentions," leading to repeated pushes of positions they reject, severely impacting user experience. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a candidate job intention matching and recommendation method that integrates temporal behavioral similarity, thus solving the problems mentioned in the background section.
[0005] To achieve the above objectives, the present invention provides a candidate job intention matching and recommendation method that integrates temporal behavioral similarity, comprising the following steps: Step S1: Obtain the candidate's historical behavior data on the recruitment platform. Each behavior record in the historical behavior data carries a timestamp, behavior type identifier, and associated job identifier. Step S2: Based on the behavior type identifier, divide the historical behavior data into positive behavior subsets and negative behavior subsets. A candidate positive behavior sequence is constructed by arranging the positive behavior subset in ascending order of timestamps, and a candidate negative behavior sequence is constructed by arranging the negative behavior subset in ascending order of timestamps; the positive behavior sequence is used for subsequent encoding of the candidate comprehensive feature vector; The extraction rules for the negative behavior subset include at least one of the following: candidates marking the relevant job as "not interested", blocking company operations, failing to respond to company invitations after the application period expires, voluntarily withdrawing from the interview process, or repeatedly browsing the same job within a preset time window without applying. Step S3: For each negative action in the candidate's negative action sequence, calculate the temporal decay penalty term according to the following formula (1):
[0006] in This is a negative penalty coefficient. Let be the decay rate coefficient and take for explicit negative behavior. For implicit negative behaviors ,and , This is the current timestamp. For the first The timestamp of the occurrence of the negative behavior. This indicates the relationship between negative behavior and the candidate job position. Cosine similarity; Step S4: Targeting candidate positions Construct the reverse behavior sequence for this position. ,in, Indicates historical connection with positions The first interaction Candidate identifier, This indicates the timestamp of the interaction. This represents the weight of the final conversion result of this interaction, and its value satisfies... ; Step S5: Calculate the forward similarity and backward consistency scores respectively: Forward similarity is calculated using cosine similarity:
[0007] in The candidate's comprehensive feature vector is obtained by encoding the candidate's resume text, structured fields, and positive behavior sequences. The feature vector is obtained by encoding the job description text and the structured requirements fields of the job. The reverse consistency score is calculated according to the following formula (2):
[0008] in, The inverse sequence decay coefficient, Candidates for historical interaction The comprehensive feature vector, This is the job requirement alignment factor, with a value range of [0,1]. Step S6: Combine the forward similarity, backward consistency score, and temporal decay penalty term according to the following formula (3) to calculate the final matching score:
[0009] in ,and , It is a subset of negative behaviors; Step S7: Sort all candidate positions by their final matching scores in descending order, and take the Top-K positions as the job intention recommendation results.
[0010] Furthermore, the extraction of the negative behavior subset in step S2 further includes: When candidates apply for the same position within a preset time window Number of views within Exceeding browsing threshold Furthermore, if the number of submissions is zero, the browsing behavior will be marked as an implicit negative behavior; After receiving an interview invitation from a company, candidates should wait within the preset response timeframe. If the offer is neither confirmed nor rejected, the relevant job posting will be marked as an implicit negative behavior. When a candidate performs an action to "block" or "reduce such recommendations" for a company, all current positions within that company will be marked as an explicit negative action.
[0011] Furthermore, in step S3, the attenuation rate coefficient The value is dynamically adjusted based on the type of negative behavior:
[0012] This causes the penalty for explicit negative behavior to decay more quickly, meaning that recent explicit negative behavior has a more significant impact on the current match, while the penalty for long-term implicit negative behavior weakens gradually over time.
[0013] Furthermore, the rule for determining the weight of the transformation result in step S4 is as follows:
[0014] And satisfy This makes it relevant to the job. Candidate features that have generated positive conversions contribute positively to the reverse consistency score, while candidate features that have generated rejections contribute negatively; positive conversions result in hiring or interviews.
[0015] Furthermore, in step S5, the job requirement alignment factor... It is activated when a preset similarity threshold condition is met, and its calculation method further introduces a dynamic threshold determination mechanism:
[0016] in The cosine similarity function is used. To align with the preset job requirements threshold, when historical candidates... Characteristics and current position When the similarity of the requirements is lower than the threshold, the historical interaction record is considered invalid in the reverse consistency score calculation; Job requirement alignment factor Indicates historical interaction candidates Structural features With position demand The satisfaction function is used to filter out noise contributions from candidates who do not match the current job requirements.
[0017] Furthermore, the weighting coefficients in step S6 and According to the position Reverse behavior sequence density adaptive adjustment:
[0018] in This is the density adjustment coefficient, when the position... When the reverse behavior sequence density is higher, The larger the value, the higher the weight of the reverse consistency score.
[0019] Furthermore, the calculation of the final matching score in step S6 is further superimposed with a key capability asymmetric penalty term:
[0020] in The total number of capability dimensions. For the candidate in The gap between the required skills and the job requirements. This is the weight for this capability dimension. For asymmetric penalty functions:
[0021] in As the critical capability threshold, Ensure that critical capability gaps are subject to quadratic penalties; This is the penalty coefficient for key abilities, which is determined by... determination; This is the penalty coefficient for general abilities, which is determined by... determination.
[0022] Furthermore, it also includes: Step S8: Feed back the Top-K recommendation results to the candidate, collect the candidate's response to the recommendation results, and when a rejection or blocking response is collected, append the job identifier to the candidate's negative behavior subset. Then return to step S3 to recalculate the matching score, thereby realizing online iterative optimization of the matching model.
[0023] This invention provides a candidate job intention matching and recommendation method that integrates temporal behavioral similarity, which has the following beneficial effects: 1. This candidate job intention matching and recommendation method, which integrates temporal behavioral similarity, can proactively exclude positions that candidates are not interested in during the matching and scoring stage by explicitly identifying and penalizing negative behaviors of candidates (such as blocking companies, rejecting invitations, repeatedly browsing without applying, etc.), effectively avoiding misjudging negative signals as positive interests. At the same time, by combining reverse verification of historical interaction behavior on the job side, the recommendation results not only consider the candidate's own preferences, but also refer to the characteristics of candidates who have successfully matched the job in the past, thereby significantly improving the accuracy of person-job matching and reducing the rate of candidates marking the recommendation results as "not interested".
[0024] 2. This candidate job intention matching and recommendation method, which integrates temporal behavioral similarity, addresses the current limitations of existing technologies that only model behavioral sequences from the candidate's perspective, while the job position's side is only used as a static text feature in the calculation, leading to asymmetrical matching. In contrast, this invention constructs a reverse behavioral sequence on the job position side, using candidate behavioral patterns that have historically had positive interactions with the job position (such as hiring or interviewing) to reversely characterize the job position's intention profile. This makes the temporal behavioral similarity calculation on both the candidate and job positions symmetrical, and the recommendation results are more in line with the actual supply and demand relationship in the recruitment market. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the steps in Embodiment 1 of the present invention. Detailed Implementation
[0026] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.
[0027] Example 1: As Figure 1 As shown, the present invention provides a technical solution: a candidate job intention matching and recommendation method that integrates temporal behavioral similarity, comprising the following steps: Step S1: Behavioral Data Collection The system extracts full behavioral logs of candidates from the recruitment platform's business database. Behavioral types include, but are not limited to: View job details (view), save job (favorite), submit resume (apply), accept interview invitation (interview_accept), reject interview invitation (interview_reject), mark "not interested" (not_interested), block company (block_company), reduce similar recommendations (reduce_similar), receive interview invitation (interview_invite); each behavior record is structured as a five-tuple (uid, job_id, action_type, timestamp, context), where uid is the candidate identifier, job_id is the associated job identifier, action_type is the behavior type identifier, timestamp is the timestamp of the behavior, and context is the behavior context (such as browsing duration, application channel, etc.).
[0028] Step S2: Negative Behavior Extraction and Sequence Construction The system categorizes behaviors into positive and negative subsets based on action_type. The positive behavior subset includes behaviors that indicate a candidate's positive intention, such as viewing, favoring, applying for, and accepting interview invitations. The sequence of positive behaviors is used to encode the candidate's comprehensive feature vector. To capture candidates' interests and preferences; The rules for extracting explicit negative behavior are as follows: Behaviors with action_type ∈ {not_interested, block_company, reduce_similar, interview_reject} are directly categorized as follows: ; The rules for extracting implicit negative behaviors are as follows: Rule 1 (Repeated Viewing Without Submission): When a candidate views the same job ID (job_id) within a time window... Number of views within Furthermore, if the number of apply times within this time window is zero, the job_id will be marked as an implicit negative behavior. Rule 2 (No Response After Invitation Timeout): After receiving the interview_invite, the candidate must respond within the specified time limit. If neither the accept nor reject operation is executed, the job posting associated with the offer will be marked as an implicit negative behavior. Rule 3 (Enterprise-level masking): When a candidate performs the block_company or reduce_similar operation on an enterprise, all current positions within that enterprise will be marked as explicit negative behavior; After extraction, the positive behavior subset is sorted in ascending order by timestamp to construct a candidate positive behavior sequence. Construct a sequence of negative behaviors for candidates, sorted in ascending order by timestamp.
[0029] Step S3: Calculation of timing decay penalty term negative behavior sequence Each negative behavior in Calculate its compatibility with the candidate positions according to formula (1). The penalty value; in this embodiment, ; Explicit negative behavior Implicit negative behavior , Cosine similarity is used for calculation, i.e. in The feature vector of the job posting associated with negative behavior. For candidate positions eigenvectors.
[0030] Step S4: Construction of Reverse Behavior Sequence on the Job Side For candidate positions The system retrieves all historical candidates related to a particular position from the job-candidate interaction database. Construct a reverse behavior sequence from candidate records that have generated interactions. ; Weight of conversion results The rules for determining the value are as follows: (Holded) (interview), (No response) (Refuse to vote).
[0031] Step S5: Bidirectional Consistency Calculation Forward similarity calculation: ,in The candidate's comprehensive feature vector is obtained by encoding the candidate's resume text, structured fields, and the positive behavior sequence. The feature vector is obtained by encoding the job description text and the structured requirements fields of the job. Reverse consistency score calculation: Calculated according to formula (2). In this embodiment, Job requirement alignment factor The calculation method is as follows: statistically analyze historical interaction candidates. Structural features Meets job requirements demand The proportion of the number of dimensions to the total number of dimensions in the requirements, with a value range of [0,1]; for example, if the job requires a master's degree or above, more than 3 years of work experience, and Python skills, and the candidate... If two of them are satisfied, then .
[0032] Step S6: Merge Scoring The forward similarity, backward consistency score, and temporal decay penalty term are fused according to formula (3), where the weight coefficients are... and According to the position Reverse behavior sequence density adaptive adjustment:
[0033] In this embodiment ,when hour, 1, At this point, it degenerates into pure forward matching; In a preferred embodiment, the calculation of the final matching score is further superimposed with a critical capability asymmetric penalty term:
[0034] in The total number of capability dimensions. For the candidate in The gap between the required skills and the job requirements. This is the weight for this capability dimension. For asymmetric penalty functions:
[0035] In this embodiment, , , That is, for critical capability dimensions with a weight exceeding 0.7, the gap value is penalized by a square multiple; for non-critical capability dimensions, the gap value is penalized by a linear ratio.
[0036] Step S7: Sorting and Recommendation Output The system's final matching score for all candidate positions Sort the results in descending order and output the Top-K (K=20) job postings as the job preference recommendations to the candidates.
[0037] Step S8: Closed-loop feedback optimization The system collects candidates' responses to recommendation results (including clicks, favorites, applications, rejections, and blocks); when a rejection or block response is collected, the job posting identifier is appended to the candidate's negative behavior subset. This triggers incremental recalculation, returning to step S3 to recalculate the matching score, thereby achieving online iterative optimization of the matching model. The trigger conditions for incremental recalculation are set as follows: when the time interval between collecting strong negative behaviors of candidates towards a specific position (such as explicitly refusing to apply or blocking companies) meets a preset threshold (e.g., more than 24 hours since the last full recalculation), or when the core requirement tags of the position change. The scope of incremental recalculation is limited to: only updating the weights of a subset of positions affected by the negative behavior in the candidate's historical behavior subset (such as positions with the same industry or job category tags), rather than re-scoring all positions in the database, thereby reducing computing power consumption while ensuring the real-time nature of recommendations.
[0038] Example 2: Explanation of Feature Vector Encoding Method; In this example, the candidate comprehensive feature vector is explained in detail. Job feature vector and historical interaction candidate feature vectors Encoding method; Candidate comprehensive feature vector The encoding process employs a three-way feature fusion encoding. First, the candidate's resume text is semantically encoded using a pre-trained language model BERT, outputting a 768-dimensional text feature vector. Second, structured fields in the resume information, such as education, work experience, professional skills, desired location, expected salary, and available time, are embedded and encoded, outputting a 256-dimensional structured feature vector. Third, the positive behavior sequence is temporally modeled using a Transformer encoder, outputting a 256-dimensional behavior feature vector. The three features are concatenated and projected onto a 512-dimensional unified vector space using a multilayer perceptron (MLP) to obtain... .
[0039] Job Feature Vector Encoding: The job description (JD) text was encoded using BERT to obtain 768-dimensional text features. The structured requirement fields of the job (education, years of work experience, skills, location, salary range, etc.) were embedded to obtain 256-dimensional structured features. These were then concatenated and projected onto a 512-dimensional vector space using an MLP to obtain... .
[0040] Historical interaction candidate feature vector Encoding: Reusing candidate composite feature vectors The encoding method, for Each historical interaction candidate in Calculate its eigenvectors .
[0041] After each interaction (such as application, interview, hiring), the system immediately calculates the candidate's comprehensive feature vector and stores it in the candidate feature database, while simultaneously updating the reverse behavior sequence for the job position. When calculating the reverse consistency score in subsequent recommendations, historical interaction candidates are directly read from the feature library. eigenvectors This eliminates the need for repetitive coding.
[0042] In summary, this candidate job intention matching and recommendation method, which integrates temporal behavioral similarity, can proactively exclude positions that candidates are not interested in during the matching and scoring stage by explicitly identifying and penalizing negative behaviors of candidates (such as blocking companies, rejecting invitations, and repeatedly browsing without applying). This effectively avoids misjudging negative signals as positive interests. At the same time, by combining reverse verification of historical interaction behavior on the job side, the recommendation results not only consider the candidate's own preferences but also refer to the characteristics of candidates who have successfully matched the job in the past. This significantly improves the accuracy of job matching and reduces the rate of candidates marking the recommendation results as "not interested".
[0043] Currently, existing technologies only model behavioral sequences from the candidate's perspective, while the job position is only used as a static text feature in the calculation, resulting in asymmetrical matching. However, this invention constructs a reverse behavioral sequence for the job position, using the candidate's behavioral patterns that have historically had positive interactions with the job (such as recruitment and interviews) to reversely depict the job's intention profile. This makes the temporal behavioral similarity calculation of both the candidate and job positions symmetrical, and the recommendation results are more in line with the actual supply and demand relationship in the recruitment market.
[0044] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.
Claims
1. A candidate job intention matching and recommendation method that integrates temporal behavioral similarity, characterized in that: Includes the following steps: Step S1: Obtain the candidate's historical behavior data on the recruitment platform. Each behavior record in the historical behavior data carries a timestamp, behavior type identifier, and associated job identifier. Step S2: Based on the behavior type identifier, divide the historical behavior data into positive behavior subsets and negative behavior subsets. A candidate positive behavior sequence is constructed by arranging the positive behavior subset in ascending order of timestamps, and a candidate negative behavior sequence is constructed by arranging the negative behavior subset in ascending order of timestamps; the positive behavior sequence is used for subsequent encoding of the candidate comprehensive feature vector; The extraction rules for the negative behavior subset include at least one of the following: candidates marking the relevant job as "not interested", blocking company operations, failing to respond to company invitations after the application period expires, voluntarily withdrawing from the interview process, or repeatedly browsing the same job within a preset time window without applying. Step S3: For each negative action in the candidate's negative action sequence, calculate the temporal decay penalty term according to the following formula (1): in This is a negative penalty coefficient. Let be the decay rate coefficient and take for explicit negative behavior. For implicit negative behaviors ,and , This is the current timestamp. For the first The timestamp of the occurrence of the negative behavior. This indicates the relationship between negative behavior and the candidate job position. Cosine similarity; Step S4: Targeting candidate positions Construct the reverse behavior sequence for this position. ,in, Indicates historical connection with positions The first interaction Candidate identifier, This indicates the timestamp of the interaction. This represents the weight of the final conversion result of this interaction, and its value satisfies... ; Step S5: Calculate the forward similarity and backward consistency scores respectively: Forward similarity is calculated using cosine similarity: in The candidate's comprehensive feature vector is obtained by encoding the candidate's resume text, structured fields, and positive behavior sequences. The feature vector is obtained by encoding the job description text and the structured requirements fields of the job. The reverse consistency score is calculated according to the following formula (2): in, The inverse sequence decay coefficient, Candidates for historical interaction The comprehensive feature vector, This is the job requirement alignment factor, with a value range of [0,1]. Step S6: Combine the forward similarity, backward consistency score, and temporal decay penalty term according to the following formula (3) to calculate the final matching score: in ,and , It is a subset of negative behaviors; Step S7: Sort all candidate positions by their final matching scores in descending order, and take the Top-K positions as the job intention recommendation results.
2. The candidate job intention matching and recommendation method based on temporal behavioral similarity as described in claim 1, characterized in that: The extraction of the negative behavior subset in step S2 further includes: When candidates apply for the same position within a preset time window Number of views within Exceeding browsing threshold Furthermore, if the number of submissions is zero, the browsing behavior will be marked as an implicit negative behavior; After receiving an interview invitation from a company, candidates should wait within the preset response timeframe. If the offer is neither confirmed nor rejected, the relevant job posting will be marked as an implicit negative behavior. When a candidate performs an action to "block" or "reduce such recommendations" for a company, all current positions within that company are marked as explicit negative behavior.
3. The candidate job intention matching and recommendation method based on temporal behavioral similarity according to claim 1, characterized in that: The attenuation rate coefficient in step S3 The value is dynamically adjusted based on the type of negative behavior: This causes the penalty for explicit negative behavior to decay more quickly, meaning that recent explicit negative behavior has a more significant impact on the current match, while the penalty for long-term implicit negative behavior weakens gradually over time.
4. The candidate job intention matching and recommendation method integrating temporal behavioral similarity according to claim 1, characterized in that: The rules for determining the weight of the transformation result in step S4 are as follows: And satisfy This makes it relevant to the job. Candidate features that have generated positive conversions contribute positively to the reverse consistency score, while candidate features that have generated rejections contribute negatively; positive conversions result in hiring or interviews.
5. The candidate job intention matching and recommendation method integrating temporal behavioral similarity according to claim 1, characterized in that: Job demand alignment factor in step S5 It is activated when a preset similarity threshold condition is met, and its calculation method further introduces a dynamic threshold determination mechanism: in The cosine similarity function is used. To align with the preset job requirements threshold, when historical candidates... Characteristics and current position When the similarity of the requirements is lower than the threshold, the historical interaction record is considered invalid in the reverse consistency score calculation; Job requirement alignment factor Indicates historical interaction candidates Structural features With position demand The satisfaction function is used to filter out noise contributions from candidates who do not match the current job requirements.
6. The candidate job intention matching and recommendation method integrating temporal behavioral similarity according to claim 5, characterized in that: The weighting coefficient in step S6 and According to the position Reverse behavior sequence density adaptive adjustment: in This is the density adjustment coefficient, when the position... When the reverse behavior sequence density is higher, The larger the value, the higher the weight of the reverse consistency score.
7. The candidate job intention matching and recommendation method integrating temporal behavioral similarity according to claim 1, characterized in that: In step S6, the calculation of the final matching score is further superimposed with a key capability asymmetric penalty term: in The total number of capability dimensions. For the candidate in The gap between the required skills and the job requirements. This is the weight for this capability dimension. For asymmetric penalty functions: in As the critical capability threshold, Ensure that critical capability gaps are subject to quadratic penalties; This is the penalty coefficient for key abilities, which is determined by... determination; This is the penalty coefficient for general abilities, which is determined by... determination.
8. The candidate job intention matching and recommendation method based on temporal behavioral similarity according to claim 1, characterized in that: Also includes: Step S8: Feed back the Top-K recommendation results to the candidate, collect the candidate's response to the recommendation results, and when a rejection or blocking response is collected, append the job identifier to the candidate's negative behavior subset. Then return to step S3 to recalculate the matching score, thereby realizing online iterative optimization of the matching model.