A method and system for matching people to jobs by integrating multimodal features and graph neural networks.
By constructing attribute relationship graphs and orthogonal projection operators to decouple representations, and combining time-stamped trajectory graphs and skill relationship data, the association strength and threshold conditions are dynamically adjusted. This solves the problems of multimodal representation entanglement, matching metric symmetry, and inability to characterize occupational evolution in existing human-job matching methods, and achieves more accurate and flexible human-job matching.
Patent Information
- Application Number
- CN202611477385.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for matching people to jobs suffer from problems such as entanglement of multimodal representations, symmetry of matching metrics, inability to characterize occupational evolution, literal skill matching, and the flattening of rigid thresholds, making them difficult to resist noise.
By constructing an attribute relationship graph, the verifiable qualification attributes and non-verifiable style attributes of candidates are separated. Orthogonal projection operators are used to decouple the representations. Combined with time-stamped trajectory graphs and skill relationship data, the association strength and threshold conditions are dynamically adjusted to achieve bidirectional matching satisfaction calculation.
It solves the problem of multimodal features entangled in dense vectors, dynamically portrays occupational evolution, avoids misjudgment in skill matching and the influence of noise, and improves the accuracy and flexibility of matching.
Smart Images

Figure CN123118053A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer data processing, and in particular to a method and system for matching people to jobs by integrating multimodal features and graph neural networks. Background Technology
[0002] In recruitment recommendation scenarios, existing job matching methods typically employ a multimodal feature concatenation and isomorphic interaction graph convolution processing path. These methods first use a pre-trained language model to encode the candidate's resume and job description into sentence vectors. Then, structured fields such as education, years of work experience, and expected salary are encoded using one-hot or binning encoding. The resume layout and portfolio images are encoded into visual vectors using a general visual backbone. These three vector paths are then concatenated end-to-end to obtain single dense profile vectors for both the candidate and the job. Next, a bipartite interaction graph is constructed using the candidate's and job's historical application and browsing records as undirected edges. Several layers of standard graph convolution are performed on this graph, and mean aggregation with degree normalization is applied to neighbor representations to enhance the collaborative signal. Finally, the inner product or cosine similarity of the final representation vectors on both sides is used as the matching score, and the recommendation list is output in descending order of score. Training is supervised by binary classification using historical applications and tags. However, the above methods have the following drawbacks in the specific recruitment scenario: 1. Multimodal features are concatenated and then mixed layer by layer into the same dense vector, making verifiable qualification dimensions and difficult-to-verify style dimensions inseparable in the vector space. This makes it impossible to apply differentiated weights to the different dimensions that recruiters and job seekers focus on. 2. Constructing a graph with undirected edges and ending with inner product or cosine similarity mathematically forces the matching score to be symmetrical about the two sides. However, job seekers' demands for salary, commuting, and growth opportunities are not from the same source as recruiters' demands for skill proficiency, stability, and labor costs. 3. The interaction graph is a static snapshot without a timestamp. Graph convolution treats neighbors equally and performs mean aggregation. This cannot characterize the rate and direction of advancement in a career path, nor can it respond to the short-term job-seeking intentions revealed by the candidate's recent behavior. This results in a low hit rate for candidates at a career inflection point.
[0003] Therefore, existing job matching methods suffer from problems such as multimodal representation entanglement, symmetric matching metrics, static snapshots failing to depict career evolution, skill matching remaining literal, and rigid thresholds being excessively smoothed out and unable to resist resume noise. Summary of the Invention
[0004] One of the objectives of this invention is to provide a job matching method that integrates multimodal features and graph neural networks to solve the problems in the prior art, such as entanglement of multimodal representations, symmetry of matching metrics, inability to characterize occupational evolution, literal skill matching, and the flattening of rigid thresholds and difficulty in resisting noise.
[0005] This invention is achieved through the following technical solution: a person-job matching method integrating multimodal features and graph neural networks, comprising the following steps: acquiring multimodal resume data of candidates and job description data of positions, and extracting modal features of the multimodal resume data under different modalities; determining a first representation component and a second representation component of the candidate based on the modal features, wherein the first representation component and the second representation component belong to two mutually orthogonal subspaces, the first representation component indicating verifiable qualification attributes, and the second representation component indicating non-verifiable style attributes; on the interaction graph, along the first directed edge from the candidate to the position, and using the second representation component... The first type of related attributes of the job position are aggregated to obtain a first satisfaction level. Then, along the second directed edge from the job position to the candidate, the second type of related attributes of the candidate are aggregated based on the qualification requirements of the job position to obtain a second satisfaction level. The first type of related attributes are different from the second type of related attributes. Based on the surplus of the first satisfaction level relative to the first retention level and the surplus of the second satisfaction level relative to the second retention level, the two-way matching satisfaction level between the candidate and the job position is determined. The two-way matching satisfaction level is taken as the minimum value when either of the two surpluses is not positive. The matching result is output according to the two-way matching satisfaction level.
[0006] Further, the extraction of modal features from the multimodal resume data in different modalities includes: extracting at least two of the following from the multimodal resume data: textual semantic features, structured field features, and layout visual features; and extracting cross-modal entities from the extracted modal features, wherein the cross-modal entity is a semantic unit that leaves evidence in two or more modalities; constructing an attribute relationship graph with the candidate as a node, each cross-modal entity as a hyperedge connecting multiple nodes, and the association strength between the candidate and the cross-modal entity as the hyperedge weight; and determining the first and second representation components of the candidate based on the modal features includes: performing hyperedge aggregation on the attribute relationship graph, and for each hyperedge... The hyperedge is described, and the representations of each node are normalized and weighted according to the association strength of the nodes it connects to, and then the self-embedding of the cross-modal entity corresponding to the hyperedge is superimposed to obtain the hyperedge representation. Node backpropagation is performed, and for each node, its hyperedges are grouped according to the category of the cross-modal entity. The mean of the hyperedge representations within the group is taken, and then the weighted sum is obtained according to the inter-group coefficients to obtain the update increment of the node. The inter-group coefficients are positive and their sum is one. The first projection operator and the second projection operator are applied to the sum of the node's representation and the update increment, respectively, to obtain the first representation component and the second representation component. The product of the first projection operator and the second projection operator is a zero operator, and the output dimension of both is the same as its input dimension.
[0007] Furthermore, the determination of the association strength includes: determining the occurrence strength of the cross-modal entity according to each modality, wherein the occurrence strength of each modality is converted to the same bounded value range; when the ratio of the maximum to the minimum occurrence strength of the same cross-modal entity in each modality exceeds the imbalance threshold, the occurrence strength of each modality is raised to a set proportion not lower than the maximum, and then weighted and summed according to the modality confidence coefficient corresponding to the category to which the cross-modal entity belongs; and scaling the association strength of the hyperedge with a normalization factor that monotonically decreases with the number of nodes connected by the hyperedge, wherein the normalization factor increases at a rate lower than the inverse growth rate of the number of nodes.
[0008] Furthermore, the person-job matching method also includes: adding a reconstruction term, a variance constraint term, and a decoupling penalty term to the training objective; the reconstruction term requires the first representation component and the second representation component to jointly reconstruct the multimodal resume data; the variance constraint term makes the values of the first representation component and the second representation component comparable; and the decoupling penalty term increases as the empirical cross-correlation between the first representation component and the second representation component increases. After each round of parameter update, the basis corresponding to the first projection operator and the second projection operator is orthogonalized so that their product remains zero.
[0009] Furthermore, the job matching method also includes: segmenting the job description data into semantic units, segmenting the multimodal resume data into modal local units, wherein the number of modal local units is greater than the number of semantic units; determining alignment weights using the semantic units as the query endpoints and the modal local units as the query endpoints, wherein the alignment weights are normalized only on the modal local units, and the alignment weights are also adjusted by the prior bias corresponding to the modality to which the modal local unit belongs; performing a weighted summation of the results obtained by weighting the alignment weights according to the demand rigidity weights of each semantic unit, wherein the sum of the demand rigidity weights is one; the demand rigidity weights are determined according to the type of qualifiers that each semantic unit is subject to in the job description data, wherein the demand rigidity weights corresponding to semantic units modified by non-concessible qualifiers are greater than the demand rigidity weights corresponding to semantic units modified by optional qualifiers.
[0010] Furthermore, the person-job matching method also includes: reading in skill relationship data, which includes directed relationships from prerequisite skills to derived skills and undirected relationships connecting related skills; constructing a hierarchical heterogeneous graph based on the candidate, skills, and job positions as nodes, according to the skill relationship data; determining the migration confidence from the second skill to the first skill based on the migration cost between the first skill and the candidate's existing second skill in the skill relationship data, as well as the hierarchical depth of their nearest common ancestor; and determining the candidate's proficiency level in the first skill based on the proficiency level of the second skill, using the migration confidence level as a decay factor.
[0011] Furthermore, the passage cost is differentiated by direction: when traveling from the prerequisite skill to the derived skill along the directed relationship, a learning cost is taken; when traveling in the opposite direction along the directed relationship, a directional penalty term is superimposed on the learning cost; when traveling along the undirected relationship, a cost that monotonically decreases with the source strength is taken. The step of determining the candidate's proficiency level in the first skill based on the proficiency level of the second skill includes: for each second skill whose migration confidence is not lower than the migration lower limit, multiplying the proficiency level of the second skill, the migration confidence from the second skill to the first skill, and the time factor that monotonically decreases with the time elapsed since the most recent practice of the second skill to obtain a candidate value; taking the maximum of all candidate values and the candidate's original proficiency level in the first skill as the candidate's proficiency level in the first skill; wherein, the decay rate of the time factor decreases as the time increases.
[0012] Furthermore, the access costs also include: determining the fault tolerance level for the position based on the urgency of recruitment and the dispersion of job level distribution within the team to which the position belongs; adjusting the weight ratio of core skills to peripheral skills in the hierarchical heterogeneous graph when aggregating skill nodes according to the fault tolerance level, so that the higher the fault tolerance level, the greater the weight of peripheral skills, and the weight of core skills is not lower than the lower limit of the weight; determining a compensability indicator for each skill required by the position but not directly possessed by the candidate, wherein the compensability indicator is zero when the skill is based on a qualification certificate, and the value decreases as the required accumulation time for the skill increases; only skills with a compensability indicator higher than the lower limit of compensability participate in the adjustment of the fault tolerance level.
[0013] Furthermore, the person-job matching method also includes: constructing a time-stamped trajectory map based on the candidate's historical resume, including the employing unit, job level, and tenure period; first aggregating the representation of neighboring nodes along the topological dimension on the trajectory map, and then performing convolution along the time dimension using a convolution kernel that only takes historical moments to obtain an advancement rate component and a potential component; incorporating the advancement rate component and the potential component into the candidate's representation; the advancement rate component is determined as follows: for two adjacent tenure periods in the historical resume, taking the ratio of their job level span to tenure duration, and adjusting for monotonically increasing gain according to the hierarchical changes of the employing unit corresponding to the two tenure periods, to obtain... The process involves progressively increasing the rate of advancement; summing each progressively increasing rate of advancement in a monotonically decreasing manner based on the time elapsed since the end of the candidate's tenure; constructing a behavioral relationship graph based on the candidate's interactions with the position within a set time window, where the interactions include at least positive and negative behaviors. In the behavioral relationship graph, the weight of the edge corresponding to the negative behavior is negative, and the weight of the edge corresponding to the positive behavior is positive, with the weight of each edge decreasing monotonically with the time elapsed since the occurrence of the interaction; performing multi-stage propagation on the behavioral relationship graph with an order not exceeding a set upper limit, and summing the results of each stage of propagation in a weighted manner based on a coefficient that sums to one, to obtain the short-term intent component. Furthermore, the difference between the job level center and function center of the position pointed to by the interaction behavior within the time window and the candidate's current job level and current function is used as the intention offset; the gating coefficient is determined by a bounded monotonically increasing mapping based on the magnitude of the intention offset; the short-term intention component and the candidate's long-term representation component are weighted and fused according to the gating coefficient, with the larger the magnitude, the greater the weight of the short-term intention component.
[0014] Furthermore, a smoothing term and a bias term are added to the training objective. The smoothing term increases as the second difference of the predicted matching result along the job level sequence increases, and the bias term biases the rate of change of the predicted matching result at the adjacent higher job level of the candidate's current job level in the direction of increasing.
[0015] Furthermore, by obtaining the threshold conditions for the positions, the degree of satisfaction of each threshold condition is determined according to a condition-by-condition bounded monotonically increasing function, and the product of each degree of satisfaction is taken as a gating mask. The difference between the value of each threshold condition and its threshold value is normalized according to the scale of the threshold condition before being substituted into the bounded monotonically increasing function. The gating mask is used to simultaneously attenuate the messages transmitted on the first directed edge and the second directed edge, as well as the gradients returned via the first directed edge and the second directed edge. The gating mask is used to scale the bidirectional matching satisfaction.
[0016] Furthermore, the threshold conditions include a first type of threshold conditions and a second type of threshold conditions; for the first type of threshold conditions, the difference between the candidate's value on the threshold condition and the threshold value is directly calculated; for the second type of threshold conditions, a deduction amount is added to the candidate's value on the threshold condition and then the difference is made with the threshold value. The deduction amount increases as the candidate's migration confidence in the skill corresponding to the threshold condition increases, and does not exceed a set proportion of the threshold value.
[0017] Further, the average pairwise direction consistency of the first representation component on all candidates is calculated layer by layer to obtain the collapse indicator; when the collapse indicator exceeds the collapse upper limit, the propagation stops and the output of the current layer is taken as the first representation component and the second representation component.
[0018] Furthermore, the person-job matching method further includes: constructing a first view based on the candidate's multimodal descriptive features, and constructing a second view based on the hierarchical and collaborative topology of the candidate's historical resume; using the first and second views of the same candidate as positive sample pairs and cross-view combinations between different candidates as negative sample pairs, training to make the representational similarity of positive sample pairs higher than that of negative sample pairs; determining a consistency confidence level based on the deviation of the representational similarity between the candidate's first and second views from the corresponding similarity distribution of candidates in the same job family, and scaling the first and second satisfaction levels with the consistency confidence level; wherein the deviation is determined based on the median and median absolute deviation of the distribution, and the consistency confidence level is not lower than the confidence lower limit; the first and second satisfaction levels are obtained by weighted summation of the satisfaction levels of multiple demands on their respective sides after marginal diminishing mapping, wherein the gain of the marginal diminishing mapping on the satisfaction level of each demand level decreases as the satisfaction level increases, and the diminishing strength corresponding to each demand level can be configured separately.
[0019] Furthermore, the first retention level is determined by the high quantile of the first satisfaction level corresponding to the positions that the candidate has historically explicitly rejected or withdrawn, and the second retention level is determined by the low quantile of the second satisfaction level corresponding to the candidates who have historically expressed their intention to hire for the position; when the number of corresponding historical records for the candidate or the position is lower than the lower limit of records, the median of the corresponding retention level of the position family to which it belongs is taken.
[0020] Furthermore, the step of outputting the matching result based on the two-way matching satisfaction includes: inducing the preference order of the candidate side and the preference order of the job side respectively with the two-way matching satisfaction, using the recruitment quota of each job as a capacity constraint, executing a delayed acceptance process until no candidate can submit an application, and outputting the obtained matching allocation scheme; wherein the pairing with the minimum value of the two-way matching satisfaction is considered unacceptable.
[0021] Furthermore, when the number of historical interaction edges of the candidate in the interaction graph is lower than the interaction lower limit, a proxy neighborhood is formed by at least one of the candidate set in the skill neighborhood of the candidate in the skill relationship data and the candidate set in the same trajectory segment in the trajectory graph. The aggregate input of the candidate on the first directed edge and the second directed edge is supplemented by the weighted average value of the candidate representation in the proxy neighborhood. When the threshold condition of the position is empty, the gating mask is set to a constant maximum value.
[0022] Another aspect of the present invention provides a job matching system that integrates multimodal features and graph neural networks, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the job matching methods that integrate multimodal features and graph neural networks as described above.
[0023] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. This invention solves the problem of weak modes being submerged by high-norm modes by constructing attribute relationship graphs with cross-modal entities as hyperedges and applying a multiplicative lower bound to truncate the intensity of unbalanced modes when determining the association strength. Furthermore, it solves the problem of high-order relations being diluted by giant hyperedges formed by ubiquitous entities by scaling the association strength with a normalization factor that decreases monotonically with the number of nodes connected by the hyperedge and whose growth rate is lower than the inverse growth rate.
[0024] 2. This invention solves the problem of entanglement between qualification-type attributes and style-type attributes in dense vectors by decomposing node representations using a pair of projection operators with a product of zero and the same output and input dimensions, supplemented by reconstruction terms, variance constraint terms, and decoupling penalty terms. At the same time, by distinguishing the passage cost by direction and determining the migration confidence by combining the nearest common ancestor level, and then using this as a basis for attenuation propagation of proficiency level, this invention solves the problem of skill matching remaining at the literal level and mistakenly eliminating homologous candidates, thus avoiding misjudging those who only master upper-level applications as having lower-level capabilities.
[0025] 3. This invention resolves the conflict between recalling cross-disciplinary candidates and ensuring that core requirements are not relaxed by determining the fault tolerance level based on the urgency of recruitment and the dispersion of job level distribution, and by limiting the scope of relaxation with a complementability indicator; it extracts the advanced rate component and potential component by using a time-stamped trajectory map and a temporal convolution that only takes historical moments, and performs fusion according to the modulus of the intended offset. Attached Figure Description
[0026] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention.
[0027] Figure 2 This is a diagram showing the influence of relative connection strength provided in Embodiment 1 of the present invention.
[0028] Figure 3 This is a comparison chart of the contribution percentage of the weakest mode provided in Embodiment 1 of the present invention.
[0029] Figure 4 This is a comparison chart of the convergence of the empirical cross-correlation of the two types of subspace components with the training rounds provided in Embodiment 1 of the present invention.
[0030] Figure 5 This is a comparison diagram of the attenuation of capability migration confidence along the positive and negative directions of the skill topology, as provided in Embodiment 1 of the present invention.
[0031] Figure 6 This is a precision-recall comparison chart provided in Embodiment 1 of the present invention.
[0032] Figure 7 This is a diagram showing the adjustment of the skill discontinuity tolerance on the aggregation weight of core and peripheral skills, as provided in Embodiment 1 of the present invention.
[0033] Figure 8 This is a comparison diagram of the offset distribution provided in Embodiment 1 of the present invention.
[0034] Figure 9 This is a comparison chart showing the change in hard threshold penetration rate with training rounds, as provided in Embodiment 1 of the present invention.
[0035] Figure 10 This is a comparison chart showing the evolution of subspace collapse degree with the number of propagation layers, as provided in Embodiment 1 of the present invention.
[0036] Figure 11 This is a comparison chart showing the impact of the resume watering ratio on matching accuracy provided in Embodiment 1 of the present invention.
[0037] Figure 12This is a two-sided utility surplus distribution and false high-match identification map provided in Embodiment 1 of the present invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0039] It should be noted that, in this application, multimodal resume data can originate from electronic resume files uploaded by candidates on recruitment platforms, personal information filled in by candidates in structured forms, historical application files synchronized from enterprise recruitment management systems, data from third-party talent pools authorized by candidates, or other data sources that provide evidence of candidates' work experience and abilities. Multimodal resume data includes at least: textual data that can represent work experience and project descriptions, structured field data that can represent educational level and years of experience, and layout visual data that can represent page organization and portfolio presentation. Optionally, multimodal resume data may also include: transcribed text of audio / video interview records, portfolio images, public contribution statistics from code hosting platforms, or derived indicator layers calculated from the aforementioned basic data. In this application, job description data refers to the collective term for textual and structured data published by the recruiter to describe the scope of responsibilities, ability requirements, entry barriers, and working environment of a specific position, typically including fields such as job responsibility section, job requirement section, salary range, work location, and team size. In this application, person-job matching refers to the process of measuring the suitability of any pair of candidates and jobs under their respective requirements, given a set of candidates and a set of jobs, and generating a recommendation list or matching and allocation scheme accordingly.
[0040] Example 1
[0041] This embodiment discloses a method for matching people to positions by fusing multimodal features and graph neural networks. Figure 1 The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps: Step 1: Obtain candidates' multimodal resume data and job description data, extract text semantic features, structured field features and layout visual features, extract cross-modal entities, and construct a multimodal attribute-entity hypergraph with cross-modal entities as hyperedges.
[0042] Among them, textual semantic features refer to the high-dimensional vector representation obtained by encoding the information such as job responsibilities, technical terms, and achievement descriptions carried in the natural language part of multimodal resume data. It reflects what the candidate has done and how he describes what he has done.
[0043] Structured field features refer to the encoding of fields with clearly defined value ranges, such as educational level, graduating institution level, years of work experience, job grade ordinal, expected salary range, and available start date. These features reflect the verifiable qualification attributes of candidates. It should be noted that for categorical fields with natural ordering relationships, such as educational level, language proficiency, and job grade, this embodiment uniformly maps them to ordered integer codes (e.g., educational level is mapped from 1 to 4 for associate's, bachelor's, master's, and doctoral degrees respectively), allowing for direct subtraction in subsequent threshold comparison calculations. For unordered categorical fields (e.g., city or industry label), indicator codes are used, and the threshold comparison degenerates into an equality judgment.
[0044] Layout visual features refer to the characteristics extracted from the page organization of a resume, including the division of page blocks, the area proportion and reading order of each block, the distribution of font size and emphasis methods such as bolding, and the composition and color scheme of portfolio images. These reflect a candidate's ability to organize their information and their expression preferences. Understandably, layout visual features are valuable because a well-structured and focused resume itself is weak evidence of a candidate's structured expression ability, and this type of evidence is of practical reference value in job categories such as design, product, and consulting.
[0045] A cross-modal entity refers to a semantic unit that can leave evidence in two or more modalities simultaneously. For example, the entity "high concurrency" may appear as a technical term in the text modality or as a technical tag in the structured field modality; the entity "project leader" may appear as a job description in the text modality or as a bold subheading in the layout visual modality; the entity "985 university" may appear as a university level value in the structured field modality or as a specific university name in the text modality. In this embodiment, cross-modal entities are divided into four categories according to a preset entity category table: technical capability, qualification certificate, organizational role, and style expression. Each category has an independent hyper-border type identifier.
[0046] A multimodal attribute-entity hypergraph is a high-order relational graph structure with candidates as nodes and each cross-modal entity as a hyperedge. Unlike a regular graph where an edge can only connect two nodes, a hyperedge in a hypergraph can connect any number of nodes at once. The reason for using a hypergraph instead of a conventional bipartite graph is that multiple candidates sharing the same cross-modal entity have a holistic similarity relationship, which cannot be equivalently expressed by a series of pairwise binary edges. Connecting all sharers with a hyperedge at once allows for the formation of a consensus signal during message passing regarding the overall nature of the entity shared by others, thus constituting true high-order relation propagation.
[0047] In this embodiment, multimodal resume data can be obtained by candidates actively uploading it on the client side and then parsing it with the server, or it can be obtained synchronously by the server from the enterprise recruitment management system according to the candidate identifier, or it can be obtained by the server in batches by searching a pre-built talent database according to job family tags. Job description data can be entered by the recruiter when posting the job, or it can be obtained by the server by parsing it from the enterprise's official recruitment page according to a structured template.
[0048] In this embodiment, to enable evidence belonging to three different modalities and with different dimensions to be measured and compared on the same intensity axis, thereby preventing any modality from overwhelming the others with its numerical magnitude at the beginning of graph construction, the specific implementation of constructing a multimodal attribute-entity hypergraph may include the following sub-steps: Sub-step 1: Divide each modal data into modal local units, and calculate the occurrence intensity of cross-modal entities according to modal type.
[0049] In this context, a modal local unit refers to the smallest processing unit with relatively complete semantics within each modality. In this embodiment, the text modality is segmented by experience segments and project segments, with each experience segment or project description constituting a local unit; the structured field modality is segmented by fields, with each valid field constituting a local unit; and the layout visual modality is segmented by connected blocks of the layout, with each layout block having an independent title or border constituting a local unit. After segmentation, a typical resume can yield dozens to hundreds of modal local units, while a typical job description, after semantic unit segmentation, typically yields only three to eight semantic units.
[0050] In this embodiment, to make the three intensities comparable, the intensity of the text modality is expressed as a normalized term frequency-inverse document frequency (IF-FRF) form, the intensity of the structured field modality is expressed as a range-normalized form, and the intensity of the layout visual modality is expressed as a product of area proportion and positional salience. For example, in this embodiment, the occurrence intensity of cross-modal entities can be calculated using the following formula:
[0051] in, As a candidate, For cross-modal entities, These are respectively text semantic modality, structured field modality, and layout visual modality; For entities In the candidates Number of times it appears in the resume text; For the complete list of candidates, Its size; as a candidate In entity Corresponding structured fields The value on, This is the set of valid values for this field. Its effective range, It is an indicator function; To be with the entity Related page areas, For the entire page area, This refers to the position of this region in the reading order. Let be a positional significance function that monotonically decreases with respect to the position. For example, we can take... ,in The saliency decay constant for the layout can be taken as the position constant in this embodiment. .
[0052] It is understandable that the first one is multiplied by an extra factor. This is because the upper bound of the regular inverse document frequency is . When the candidate pool reaches a scale of millions, this upper bound can reach over 13, while the values of the second and third candidates naturally fall within the range of... If this normalization is not performed, the text modality alone can numerically outperform the other two modalities by an order of magnitude. The so-called multimodal fusion has degenerated into text dominance from the very beginning of graph construction.
[0053] It should be noted that the positional saliency function in the third branch reflects the prior information unique to the layout modality: the information contained in the first screen of the resume and the bolded title area is usually the core qualifications that the candidate identifies as having, and its evidentiary value is significantly higher than that of the secondary listings on the last page; if the positional weight is ignored and only the area ratio is taken, then a resume that spends a lot of space listing general skill terms will have its invalid blocks gain high intensity, causing the superedge weight to be diluted by irrelevant information.
[0054] Sub-step 2: Determine whether the intensity of each mode is unbalanced. If unbalanced, apply a multiplicative lower bound truncation and calculate the hyperedge weights.
[0055] Modal imbalance refers to a situation where the intensity of the same entity's presence in the three modalities differs by orders of magnitude. For example, a candidate might repeatedly emphasize a certain technical skill in their text, but this skill is not checked in their structured skills list, nor is it emphasized in the layout. Multiplicative lower bound truncation refers to a process where, after confirming an imbalance, the intensity of each modality is increased to a fixed proportion no less than that of the strongest modality.
[0056] In this embodiment, to ensure that weak modes are not completely eliminated without interfering with already balanced entities, truncation is activated only when the range ratio exceeds the imbalance threshold. For example, in this embodiment, the hyperedge weights can be calculated using the following formula:
[0057]
[0058]
[0059] in, For entities The entity category to which it belongs; Modal confidence coefficients configured according to entity category, satisfying For example, qualification certificate entities primarily use structured fields as evidence, which can be taken as... For entities exhibiting stylistic expression, text and layout are the primary evidence, and this approach is acceptable. ; For modality In this embodiment, the lower limit of the proportion in the hyperedge weight can be uniformly taken as... ; The imbalance threshold can be taken as... ; The incidence matrix of the hypergraph. Indicates candidate Overtake Connected; For super-edge The order of It is the superedge order-rarity normalization factor.
[0060] It is understandable that the introduction The underlying reason is that ubiquitous entities, such as office software for undergraduate degrees, connect the vast majority of nodes in the candidate pool, forming giant hyperedges of extremely high order. Without suppression, subsequent hyperedge aggregation will degenerate into averaging across all nodes in the pool, resulting in a result that is almost identical for every candidate, completely diluting the discriminative information that high-order relationships should carry. The reason for using the reciprocal of the logarithm instead of the direct reciprocal of the order is to suppress giant hyperedges while preventing rare hyperedges connecting only two or three candidates from gaining overwhelming weight and introducing overfitting. Figure 2 This diagram illustrates the influence of the hyperedge order normalization factor on the relative connectivity strength of giant and rare hyperedges in this embodiment. Figure 2 As can be seen, as the number of candidates connected by a superedge increases, the order normalization factor causes its relative connection strength to monotonically decrease to 0.12, with a decrease of 87.96%, while the relative connection strength of rare superedges with extremely low order is basically unaffected. Figure 3 This diagram shows a comparison of the contribution percentage of the weakest mode under the modal intensity imbalance condition in this embodiment. Figure 3 As can be seen, as the intensity imbalance ratio between the strongest and weakest modes increases, the contribution ratio of the weakest mode in the control scheme is continuously reduced, reaching only 0.021 when the imbalance ratio reaches 40. However, this scheme stabilizes this ratio at 0.076 by using the multiplicative lower limit truncation, which is 3.6 times that of the control scheme, thus preventing multimodal input from mathematically degenerating into single-modal input from the source.
[0061] It should be noted that if the multiplicative lower bound truncation design is ignored, then in the case of a situation that is very common in real resumes, where one modality has extremely strong evidence while another modality is almost zero, the contribution of the weak modality will be completely reduced to zero in the weighted summation. Multimodal inputs are mathematically degenerated into single-modal inputs, and all subsequent designs based on modality distinguishability in this method will lose the support of the input side.
[0062] Step 2: Perform two-stage message passing of hyperedge aggregation and node backpropagation on the multimodal attribute-entity hypergraph. Decouple the node representation into hard capability index subspace components and soft style subspace components through orthogonal projection. Perform asymmetric cross attention alignment between job description and resume modal locality to obtain decoupled talent profile and job requirement profile.
[0063] Two-phase message passing refers to a round of propagation being split into two phases in opposite directions: first, nodes aggregate towards the superedge to form a superedge representation, and then the superedge propagates back to the nodes to form the node's update increment.
[0064] The hard competency indicator subspace refers to the representation subspace used to carry verifiable qualification attributes such as educational level, qualifications, skill proficiency, and years of work experience. The soft style subspace refers to the representation subspace used to carry attributes that are difficult to verify directly, such as organizational ability, collaborative tendencies, stable style, and cultural adaptation tendencies. The reason for separating these two types of attributes into two subspaces is that the two sides in the recruitment process pay drastically different levels of attention to these two types of attributes. Recruiters primarily rely on hard indicators when screening candidates, while candidates primarily rely on soft factors. If both are mixed in the same dense vector, the weighting of preferences on either side will inevitably contaminate the other side.
[0065] Orthogonal projection is an idempotent linear transformation induced by a set of row orthogonal basis matrices that maps a vector to a linear subspace.
[0066] In this embodiment, to ensure that decoupling is not a post-hoc split but a structural constraint during the propagation process, thereby preventing hard metrics and soft styles from re-mixing after propagation through several layers, the specific implementation of this step may include the following sub-steps: Sub-step 1: Aggregate nodes to hyperedges to generate hyperedge representations.
[0067] In this embodiment, the hyperedge representation is the weighted sum of the representations of all candidate nodes it connects to, normalized according to the hyperedge weight, and then superimposed with the entity's own embedding. For example, in this embodiment, the hyperedge representation can be calculated using the following formula:
[0068] in, For the first Layer candidate nodes The representation vector, For entities The self-embedded vector. In this embodiment, the initial value of the node representation. It can be obtained by a modal confidence-weighted average of the local units representing each modality of the candidate, with the entity having its own embedding. It can be initialized by the semantic encoding of the entity name.
[0069] It should be noted that, One essential element is indispensable. Without self-embedding, when a hyperedge connects only one candidate, its hyperedge representation will degenerate into a copy of that candidate's representation, creating a self-reinforcing echo during backpropagation. Rare entities not only fail to provide additional semantics but also amplify the noise of the candidate itself. With self-embedding, rare entities can still participate in propagation with their own semantic priors, thus maintaining the effectiveness of propagation on long-tail entities.
[0070] Sub-step 2: The superedge is propagated back to the node, and node update increments are generated by grouping them according to entity type.
[0071] In this embodiment, to prevent the large number of technical capability entities from overwhelming the small but crucial qualification certificate entities during mean aggregation, the backhaul employs a two-level aggregation: first, intra-type averaging, then inter-type weighted averaging. For example, in this embodiment, the node update increment can be calculated using the following formula:
[0072]
[0073] in, as a candidate Category: The set of superedges This represents the aggregation coefficient between types.
[0074] Understandably, taking the arithmetic mean within each type rather than summing them ensures that candidates who list numerous technical terms on their resumes cannot increase their technical prowess by simply piling on numbers; while normalizing the coefficients between types ensures that the relative importance of the four types of entities is determined by configuration or learning, rather than by the number of listed terms.
[0075] Sub-step 3: Perform subspace decomposition on the updated representation using a pair of mutually orthogonal projection operators, and apply variational reconstruction constraints and cross-correlation penalties.
[0076] For example, in this embodiment, the subspace decomposition can be calculated using the following formula:
[0077]
[0078] in, and Let be two row orthogonal projection basis matrices, i.e. , ; and Induced by it Orthogonal projection operator; and Let be the dimension of the two subspaces, satisfying In this embodiment, This condition is achieved by updating the concatenated matrix after each round of parameter updates. It is enforced by applying Gram-Schmidt orthogonalization.
[0079] It is understandable that this embodiment uses a projection operator. Instead of directly using a dimensionality reduction matrix This is to ensure that the two subspace components maintain the same dimension as the original representation. This allows it to be directly used as input for message passing to the next layer without any dimension adaptation, enabling iteration at any number of layers; if directly using Left multiplication will change the output dimension to or Multiple layers stacking do not hold true in terms of dimension. At the same time, since the orthogonality condition is enforced by the explicit orthogonalization process of each layer, its effectiveness does not depend on whether the training converges, and it is a hard constraint.
[0080] Orthogonal projection can only guarantee that two subspaces are disjoint in a linear algebraic sense, but it cannot guarantee that they are statistically uncorrelated. Therefore, in this embodiment, variational reconstruction constraints and cross-correlation penalties are simultaneously applied to the objective function. For example, in this embodiment, the decoupling objective can be calculated using the following formula:
[0081]
[0082]
[0083] in, as a candidate The original multimodal observations, and These are the variational coding distribution and the reconstruction distribution, respectively. For relative entropy, The prior is the standard normal distribution. The empirical cross-correlation matrix of the two subspace components is given. and Let its mean vector and standard deviation be denoted as . It is the Frobenius norm; As the decoupling penalty coefficient, in this embodiment, it can be taken as... .
[0084] Understandably, the reconstruction term requires that the two components combined be sufficient to reconstruct the original multimodal observations. Its purpose is to prevent decoupling from degenerating into one path carrying all the information while the other collapses into a trivial solution of constant value. This is the most likely failure mode to occur when only orthogonal constraints are applied. The relative entropy term constrains the variance scale within each subspace, making the two components numerically comparable, so that subsequent differential weighting is meaningful. The cross-correlation penalty term directly suppresses the statistical dependence between the two paths. Figure 4 This embodiment shows a comparison of the convergence of the empirical cross-correlation of the two types of subspace components with respect to training epochs; from Figure 4 As can be seen, the empirical cross-correlation norm of the two types of subspace components in this scheme drops to 0.056 and tends to plateau within more than ten rounds of training, while the control scheme that relies solely on the spontaneous uncorrelation of the reconstruction target still has a high value of 0.516 after the same number of rounds, a decrease of 89.11%.
[0085] Sub-step 4: Perform asymmetric cross-attention alignment using the job semantic unit as the query end and the resume modal local unit as the query end.
[0086] Asymmetric cross-attention refers to an attention operator where the query end and the queryed end are not interchangeable, and the attention weight is normalized only on the queryed end. Demand rigidity weight refers to the degree to which the demands expressed by a semantic unit in the job description are uncompromising. In this embodiment, demand rigidity weight is extracted according to the type of qualifiers modified by the semantic unit in the job description: those modified by qualifiers such as "must have" or "do not accept" are classified as rigid; those modified by qualifiers such as "required to have" are classified as regular; and those modified by qualifiers such as "priority bonus" are classified as flexible. These three categories are assigned original weights of 3, 2, and 1 respectively, and then normalized within the job description.
[0087] In this embodiment, to address the severe information asymmetry caused by a resume being segmented into dozens to hundreds of modal local units, while a job description can only be segmented into three to eight semantic units, and to enable short and abstract job requirements to proactively retrieve long and redundant resume segments instead of being overwhelmed by them, the attention operator used in this sub-step is asymmetric in three aspects: the query end is fixed on the job side and normalization is only performed on the resume side; the queried end is introduced with a prior bias based on modality; and the contribution of each semantic unit on the job side is weighted by the rigid weight of the normalized requirement. For example, in this embodiment, asymmetric cross-attention alignment can be calculated using the following formula:
[0088]
[0089] in, For the position The Each semantic unit is represented. Its quantity; as a candidate The Modal local unit characterization, For its quantity, usually ; , , These are the transformation matrices for the query, key, and value, respectively. The dimension of the key vector. A scaling factor to prevent the inner product from saturating as the dimension increases; For local units The mode to which it belongs, This is the prior bias for this mode; For job semantic units The normalization requirement is a rigid weight.
[0090] Understandably, for The significance of using normalized weights for weighted summation instead of direct summation lies in making... The norm of the job description does not drift with the number of semantic units in the job description; if the results are summed directly, the norm of the alignment vector generated by a job description that lists ten requirements will be about three times that of a job description that lists only three requirements, making the matching scores of the two jobs incomparable.
[0091] It should be noted that, since the decoupling in sub-step 3 has already been completed, each local unit of the resume focused on in this sub-step can be clearly classified into either hard indicator evidence or soft style evidence based on the ratio of its projection norms in the two subspaces. This makes the alignment results not only useful for calculating matching scores, but also for explaining to recruiters or candidates which part of the resume and what type of evidence supports a particular judgment. If attention alignment is performed directly without decoupling, the attention weights can still be visualized, but the local units they point to are semantically mixed and cannot support differentiated preference calculations.
[0092] Step 3: Read in the industry skills topology knowledge graph, construct a hierarchical heterogeneous graph of candidate-skill-prerequisite skills-related skills-position, calculate the ability migration confidence based on directed migration distance and the least common ancestor level, perform proficiency decay transmission to complete skills that candidates do not directly possess, and adaptively adjust the skill gap tolerance according to the recruitment urgency of the position and the team echelon distribution.
[0093] Among them, the industry skills topology knowledge graph refers to the external structured knowledge composed of industry skills as nodes and the prerequisite and association relationships between skills as edges. Its nodes are organized into a tree-like hierarchy according to the abstract level, such as from programming languages to statically typed compiled languages and then to a specific language, and so on, with each level being refined.
[0094] A prerequisite edge is a directed edge that points from a prerequisite skill to a derived skill. It signifies that mastering the prerequisite skill is a necessary foundation for learning the derived skill; for example, data structures and algorithms lead to distributed system design. An association edge is an undirected edge that connects two skills sharing a common engineering paradigm or toolchain; its edge weight is determined by the similarity of origins.
[0095] The confidence level of ability transfer refers to the degree of confidence that a candidate can reach the required level of another skill at a relatively low cost, given that the candidate already possesses a certain skill. The value ranges from 0 to 1.
[0096] A skills gap refers to the difference between the set of skills required for a job and the set of skills possessed by a candidate. Skills gap tolerance refers to the extent to which a specific job is willing to accept candidates with skills gaps.
[0097] A hierarchical heterogeneous graph refers to a graph structure obtained by nesting and fusing the above-mentioned skill topology knowledge graph with a job-talent graph, containing various node types such as candidate nodes, skill nodes, prerequisite skill nodes, related skill nodes, and job nodes.
[0098] In this embodiment, the industry skills topology knowledge graph can be constructed and maintained regularly by the recruitment platform based on industry occupational classification standards and job description corpora, or it can be provided by a third-party occupational competency standards organization, or it can be generated by the server from the enterprise's internal job competency model.
[0099] In this embodiment, to prevent candidates who do not literally contain job keywords but have accumulated extensive experience in adjacent skills from being mistakenly eliminated by literal comparison, and at the same time to avoid misjudging skills that are literally similar but actually unsupported as transferable, the specific implementation of this step may include the following sub-steps: Sub-step 1: Calculate the migration distance from the candidate's existing skills to the job-required skills using the directed travel cost.
[0100] In this embodiment, the directionality of the preceding edge is of substantial significance: mastering a preceding skill helps in learning its derived skills, but deriving a preceding skill from a derived skill does not automatically hold true. Therefore, the travel cost of a path must be differentiated by direction. For example, in this embodiment, the migration distance can be calculated using the following formula:
[0101] in, The skills required for the position. The skills that the candidate already possesses; For the reason arrive The complete set of all pathways; The cost of learning to travel along the front edge in the forward direction; An additional directional penalty imposed for traveling in the wrong direction, exemplarily, is preferable. ,in This represents the average cost of the front edges across the entire graph. The strength of the common origin of the related edges; the stronger the common origin, the shorter the equivalent path. To prevent small quantities with a denominator of zero.
[0102] Understandably, if we take the undirected shortest path as in the conventional approach, we will get... This means that migrating from underlying principles to upper-level applications is as difficult as deriving underlying principles from upper-level applications. This contradicts common sense in engineering practice and may cause the system to misjudge a large number of candidates who only know how to call high-level frameworks as having underlying capabilities. Figure 5This embodiment shows a comparison of the attenuation of capability transfer confidence along the positive and negative directions of the skill topology; from Figure 5 As can be seen, under the same skill topology travel cost, the migration confidence obtained by deriving the predecessor skill from the derived skill is always lower than that in the opposite direction. At the travel cost of 2, the two are 0.104 and 0.221 respectively, with the former being 52.76% lower than the latter. This avoids misjudging candidates who can only call high-level frameworks as having low-level capabilities.
[0103] Sub-step 2: Calculate the capability transfer confidence by multiplying the product of three factors: topological distance decay, common ancestor principle sharing degree, and homology correction.
[0104] In this embodiment, ability transfer is decomposed into three independent and non-substitutable resistance factors: the length of the learning path determines the additional investment required for transfer; the depth of the level where the two skills' nearest common ancestor is located determines how thick the underlying principles they share; and whether they belong to the same technology family is used to correct cross-family misjudgments that may be caused by the first two factors.
[0105] For example, in this embodiment, the capability transfer confidence score can be calculated using the following formula:
[0106]
[0107]
[0108] in, The lowest common ancestor of the two skills in the topology tree; This represents the node's hierarchical depth, with the root node having a depth of 0. For technology family identifiers; For the same source correction magnitude, in this embodiment, we can take... ; As the path attenuation scale, it can be taken as... .
[0109] Understandably, the intention behind multiplying the three factors together rather than using a weighted sum is to ensure that when any factor approaches zero, the overall value approaches zero: factors with extremely distant paths do not gain high confidence simply because they belong to the same class, and factors with very shallow shared levels are not misjudged as transferable due to accidental short paths. This property is precisely what the weighted summation form lacks. In the weighted summation form, the extremely low value of a factor can be linearly compensated for by the high values of the other factors, thus leading to the error of classifying completely unrelated factors as transferable simply because they belong to the same large class.
[0110] It should be noted that the denominator of the homology correction factor To ensure that the value of the algebraic expression does not exceed 1, thus It can be directly used as the attenuation factor in subsequent conduction; if written as commonly, it is directly taken as... Without normalization, the confidence level of skills from the same race will be greater than 1. In subsequent transmission, it will amplify rather than diminish the skill level, resulting in the completed skill level exceeding the original skill level, which contradicts the basic fact that migration always involves loss.
[0111] Sub-step 3: Use the maximum migration pathway to perform proficiency decay propagation to complete the skills that the candidate does not directly possess.
[0112] For example, in this embodiment, the proficiency level after completion can be calculated using the following formula:
[0113]
[0114] in, as a candidate The set of skills already possessed; For its skills Original proficiency level; As the lower bound of migration confidence, in this embodiment, we can take... Skills below this lower limit do not participate in the transmission; For skills The time since the most recent practice; The skill forgetting factor is in power-law form; and For its time scale and decay exponent, an example can be taken. Months .
[0115] Understandably, the key design of this formula is to take the maximum value of the outer layer instead of summing all transferable sources. If it were to sum instead, a candidate with more than ten weakly related skills could accumulate high proficiency by simply adding up the numbers, thus creating a new channel for inflating resumes, which directly contradicts the purpose of this method to suppress resume exaggeration. Taking the maximum value means that whether you can transfer to a certain skill depends on your strongest transfer path, which is consistent with the actual mechanism of ability transfer. Figure 6 This embodiment shows a precision-recall comparison chart for cross-domain candidate recall; from Figure 6 As can be seen, at the same precision level of 0.80, the recall rate of the control scheme is only 0.308, while that of this scheme is 0.457, which is a relative improvement of 48.49%. Moreover, the curve of this scheme is above that of the control scheme throughout the entire range.
[0116] It should be noted that using a power-law rather than an exponential forgetting factor can be understood as a characterization of the empirical law that skill decline is rapid at first and then slows down: skills that have not been used for a long time will become rusty in the first year or two, while the remaining theoretical understanding will remain relatively stable thereafter; if exponential forgetting is used, the contribution of experiences from ten years ago will be compressed to almost zero, and senior candidates who have laid a solid theoretical foundation in their early years and have recently turned to management will be systematically underestimated.
[0117] Sub-step 4: Calculate the skill gap tolerance coefficient based on recruitment urgency and team echelon distribution entropy, and limit the scope of relaxation according to the directional compensability criterion.
[0118] In this embodiment, to allow the tolerance level for skills gaps to be adaptively determined by observable indicators of the job and team itself, rather than by manual allocation on a job-by-job basis, the skills gap tolerance coefficient and core / peripheral skill weights can be calculated, for example, using the following formula:
[0119]
[0120]
[0121] in, For the position The recruitment urgency indicator can be obtained by normalizing factors such as the length of time a job posting has been posted, the completion rate of the recruitment plan, and the degree of narrowing of the interview conversion funnel. The first in the team where this position is located The percentage of employees at each job level; Shannon entropy for job level hierarchy distribution: the lower the entropy value, the more concentrated the personnel are in a few job levels and the more serious the hierarchical gap is. It is a logarithmic probability function; These are calibration coefficients; and These represent the weights applied to core skills and peripheral skills when aggregating skill nodes in a hierarchical heterogeneous graph; As the absolute lower limit of the core skill weight, in this embodiment, we can take... .
[0122] Furthermore, in this embodiment, to prevent the urgent recruitment signal from being misused as a relaxation of compliance requirements, it is also necessary to differentiate the directional recoverability of skill gaps. For example, the recoverability of a certain skill gap can be calculated using the following formula:
[0123] in, A subset of skills that are contingent upon certification; For skills The industry average number of years of experience required to achieve the required proficiency level for the position; The compensability decay scale; This is the transfer confidence threshold for complementarity determination. Only... Only faults exceeding the set value are included in the tolerance relaxation range.
[0124] It should be noted that a lower limit for the weight of core skills is set. The significance of this engineering approach is that, no matter how urgent the position or how disparate the talent pool, the aggregate weight of core skills must not fall below 60%. If this lower limit is ignored, the tolerance coefficient for extremely urgent positions will approach 1, and the weight of core skills will approach 0. The system will output the result of pushing completely unsuitable people into the position to fill the vacancy. This is far more harmful than the vacancy itself in recruitment practice. Figure 7 This embodiment shows a plot of how skill discontinuity tolerance affects the aggregation weights of core and peripheral skills; from Figure 7 As can be seen, as the urgency of recruitment increases, the weight of peripheral skills aggregation for positions with higher tier dispersion increases from 0.190 to 0.379, while the weight of core skills aggregation remains at a minimum of 0.621 under both tier conditions, never falling below the lower limit.
[0125] Step 4: Construct a time-stamped career trajectory evolution sub-graph based on the company, job level, and tenure period in the candidate's resume, and extract the career advancement rate component and development potential component; construct a real-time intent graph based on the behavior within a set time window and extract the short-term intent component; fuse the two components by gating the intent offset to obtain the spatiotemporal evolution of the person-job representation.
[0126] Among them, the career trajectory evolution subgraph refers to a time-stamped subgraph that uses each of the candidate's employment experiences as nodes, the temporal relationship between the experiences as directed edges, and labels the company level and job grade ordinal number on the nodes and the time interval on the edges.
[0127] Career advancement rate refers to the number of job levels a candidate traverses within a unit of time, adjusted for changes across corporate levels; it reflects the speed and direction of career development. Development potential component refers to the representational component extracted from the trajectory subgraph via spatiotemporal graph convolution, used to characterize the candidate's future achievable job level range.
[0128] A real-time intent graph is a behavioral graph that starts with the candidate, uses the job positions and job attributes they interact with within a set time window as relay nodes, and uses actions such as clicks, communications, favorites, and withdrawals as directed edges. Short-term intent components are representational components extracted from the real-time intent graph through multi-stage propagation with time decay, used to characterize the candidate's recent job-seeking inclinations.
[0129] Intent offset refers to the difference vector between the job level center and function center of the position to which the candidate's behavior is directed within a time window, relative to their current job level and function. Its magnitude reflects the degree to which the candidate is shifting their focus.
[0130] In this embodiment, to shift the recommendation from matching who the candidate is now to matching who the candidate wants to become next, this step can be implemented in the following sub-steps: Sub-step 1: Calculate the progressive rate of each segment and weight it according to time decay to obtain the long-term career advancement component.
[0131] For example, in this embodiment, the progressive rate and long-term career advancement component can be calculated using the following formula:
[0132]
[0133] in, , , Candidates No. Companies, job levels, and tenure during each period of experience; Company hierarchy mapping; For a cross-enterprise transition gain function that monotonically increases with the hierarchical span, an example can be taken as follows: ; For job level The embedding vector; For the long-term decay constant, in this embodiment, it can be taken as... Months; This refers to the current moment.
[0134] Understandably, using the end date of employment rather than the start date as the decaying independent variable is because the advanced conclusions of an experience are only fully valid at its end; while recent experiences are given higher weight, reflecting the recency effect of career trajectories. Introducing cross-firm career advancement gains. This is to distinguish between lateral transfers between companies at the same level and promotions that accompany upward movement within a company. The changes in job rank may be exactly the same for both, but the career dynamics they reflect are completely different.
[0135] Sub-step 2: First perform spatial aggregation on the trajectory subgraph, then perform causal temporal convolution to obtain the development potential component.
[0136] For example, in this embodiment, spatial aggregation can be calculated using the following formula:
[0137] Then, a causal convolution is performed along the time axis, which can be calculated as follows:
[0138]
[0139] in, For the first The trajectory nodes of the layer represent tensors; The adjacency matrix of the trajectory subgraph with added self-loops. Its degree matrix; This is the spatial transformation matrix; For length is The first of the temporal convolution kernels One tap; It is a non-linear activation function; This involves concatenating vectors. The time derivative of the advanced rate is the advanced acceleration. Read the matrix for potential; This represents the total number of spatiotemporal convolution layers.
[0140] It should be noted that the summation over the time dimension only applies to... The historical values are used, meaning the convolution kernel is a causal kernel. If the convolution kernel is allowed to access future moments, the model will peek at candidates' promotions that have not yet occurred during training, resulting in significantly inflated offline evaluation metrics. However, when deployed online, the actual performance will deteriorate drastically because future information is unavailable. This type of time leakage is extremely subtle in timestamped graph modeling. Therefore, this embodiment explicitly separates spatial aggregation and temporal convolution into two separate equations so that causal constraints can be applied and verified separately.
[0141] Sub-step 3: Perform truncated multi-stage propagation with time decay on the real-time intent graph to obtain short-term intent components.
[0142] For example, in this embodiment, the short-term intent component can be calculated using the following formula:
[0143]
[0144]
[0145] in, The real-time intention transition matrix is a row-randomized matrix; Weights for behavior types; The moment when the action occurs; It is a short-term decay constant and In this embodiment, it is possible to take Hour; For the first The contribution coefficient of the first-order propagation can be exemplarily taken as follows: After geometric decay, it is normalized; To be the maximum propagation order, we can take... The weighting of behavior types reflects the difference in the intensity of intent for different behaviors. For example, communication can be weighted at 1.0, favorite at 0.6, click at 0.3, and withdrawal at 0.4%. .
[0146] Understandably, assigning a negative weight to withdrawal behavior is a customized design in this embodiment specifically for the recruitment scenario. In general recommendation scenarios, withdrawal or unfollowing is usually considered an invalid event and discarded; however, in a recruitment scenario, a candidate withdrawing their application is highly informative evidence of negative intent, often indicating that the candidate has actively rejected the position after obtaining more information. Including this and assigning a negative weight allows the intent map to simultaneously express what they want and what they explicitly don't want.
[0147] It should be noted that, due to the presence of negative weights, the denominator of a certain row may approach zero or even be negative. Therefore, a lower bound protection must be applied to the denominator during implementation. Furthermore, when the sum of the weights of all outgoing edges of a node is not positive, the row is degenerated into a self-loop to ensure a well-defined transition matrix. Additionally, the use of truncated multi-order propagation instead of summing all traversal paths is because the length does not exceed [a certain threshold]. The number of paths varies It grows exponentially, and summing each path is not only computationally infeasible, but its total weight does not normalize with the path set, causing the norm of the short-term intention component to drift with the propagation order, making it numerically incomparable to the long-term component.
[0148] Sub-step 4: Calculate the gating coefficients according to the magnitude of the intended offset, fuse the long and short components, and add a career transition gradient regularization term to the training objective.
[0149] For example, in this embodiment, the intention offset, gating coefficient, and fusion result can be calculated using the following formula:
[0150]
[0151]
[0152] in, and These are the job level center and function center of the position to which the candidate's behavior points within the time window; and His / her current rank and functions; The gating coefficient; , These are the gate offset and steepness, respectively; , These are the transformation matrices for the two components, respectively.
[0153] Understandably, the gating factor is driven by the magnitude of the offset rather than its sign, meaning that regardless of whether a candidate moves upward into management or laterally into a new role, as long as the offset is large enough, short-term intentions will be amplified; while when the offset is close to zero... The smaller value is chosen, and long-term career logic takes over again. The underlying reason is that when recent browsing aligns with the existing career trajectory, short-term behavior is merely a noisy repetition of long-term logic. Increasing its weight at this point is not only unhelpful but also introduces conversational noise. Only when recent behavior significantly deviates from the existing trajectory does short-term behavior carry new information that the long-term trajectory does not contain.
[0154] Furthermore, in this embodiment, to ensure that the output not only matches the current skill level but also targets the candidate's next career inflection point, a career transition gradient regularization term needs to be added to the training objective. For example, it can be calculated using the following formula:
[0155] in, For the model to candidates With job level Predicted matching score for job placement; This represents the gradient bias intensity coefficient.
[0156] It's important to note that the first term is the second-order sum of squared differences along the job level sequence. Its purpose can be understood as requiring the matching score to change smoothly with job level, preventing jagged patterns that cannot be explained by professional logic, such as giving a high score to one job level, a very low score to an adjacent job level, and then a high score to a higher job level. The second term is negative, meaning the optimization process actively increases the positive gradient of the matching score at the candidate's next job level, thus shifting the recommendation focus from similar positions at the current job level to one level higher. The combined effect of these two terms gives the recommendation result both continuity along the career path and a slight forward-looking nature. If only the second term existed without the first, the model would tend to push scores to higher job levels indiscriminately, resulting in numerous recommendations that are completely out of reach for many candidates. Figure 8 This embodiment shows a comparison chart of the offset distribution of recommended positions relative to the candidate's current job level; from Figure 8 As can be seen, the recommended distribution of the comparison scheme has the candidate's current job level as the mode and the average deviation is -0.04, while the mode of this scheme shifts to the adjacent higher job level and the average deviation reaches 0.93, with an overall shift of 0.97 job levels forward.
[0157] Step 5: Using the spatiotemporal evolution of job-person representation and the ability vector completed by transfer as node representation, construct a bidirectional asymmetric heterogeneous graph with two types of directed edges: job seeker-to-job and job-to-candidate. Perform differentiated message passing along the two paths, apply attenuation penalty to the edge messages and gradients with hard constraint gating masks, and correct the demand satisfaction component with graph-multimodal comparison calibration.
[0158] In this context, a bidirectional asymmetric heterogeneous graph refers to a heterogeneous graph that simultaneously contains directed edges from candidates to job positions and directed edges from job positions to candidates, where the two types of directed edges do not share semantics, parameters, or aggregation neighborhood. Directed edges from candidates to job positions represent job seeker preferences, while directed edges from job positions to candidates represent employer screening preferences.
[0159] Environment and Development nodes refer to attribute nodes attached to the job node's neighborhood, used to describe the job's work environment and development prospects. In this embodiment, they include at least team size nodes, overtime culture label nodes, salary distribution nodes, commuting time topology nodes, reporting level nodes, and promotion channel nodes. Ability and Background nodes refer to attribute nodes attached to the candidate node's neighborhood, used to describe the candidate's abilities and professional background. In this embodiment, they include at least skill depth nodes, past company level nodes, stability indicator nodes, project scale nodes, and collaborative role nodes.
[0160] Hard threshold conditions refer to qualification requirements set by the recruiter that will disqualify applicants if not met. In this embodiment, hard threshold conditions are further divided into non-negotiable conditions and quasi-hard conditions: the former, such as holding a professional certificate, the city of employment, and legal professional qualifications, are either one or the other, and there is no partial fulfillment; the latter, such as years of work experience, educational level, and specific language proficiency, are thresholds that can be partially compensated for by other evidence.
[0161] A hard-constraint gating mask is a scalar with a value between 0 and 1, obtained by compiling the above hard threshold conditions into differentiable multiplicative gating functions and taking the conditional product.
[0162] In this embodiment, in order to eliminate the constraint imposed by symmetry metrics This structural constraint, and to prevent the hard threshold of a veto type from being smoothed out by the excessive smoothing of deep graph propagation, can be implemented in the following sub-steps: Sub-step 1: Perform differentiated message passing from a dual perspective along two directed paths, where parameters are not shared, neighborhoods do not overlap, and query sources are different.
[0163] For example, in this embodiment, dual-view message passing can be calculated using the following formula:
[0164] in, For the position A set of environment and development-related nodes within the neighborhood; as a candidate A set of capability and background class nodes within the neighborhood; and The soft style and hard index subspace components are respectively taken from the decoupling obtained in step 2; , These are the message transformation matrices for each of the two paths; , , , These are the query and key transformation matrices for the two paths, and they do not share parameters.
[0165] Understandably, the query sources for the two paths are taken from two orthogonal subspaces. The job seeker uses their soft style components to query the job's environment and development nodes, while the job posting uses its hard performance requirements to query the candidate's abilities and background nodes. This makes the decoupling in step 2 truly effective, rather than merely a separation of representational forms. Furthermore, since the neighborhood range, transformation parameters, and query sources of the two paths are all different... and Structurally they cannot be equal, and the structural constraints imposed by symmetry are thus completely removed.
[0166] Sub-step 2: Compile the hard threshold condition into a differentiable gated mask and apply it to both the forward message and the backward gradient.
[0167] For example, in this embodiment, the hard-constraint gating mask can be calculated using the following formula:
[0168]
[0169]
[0170]
[0171] in, and Each position The set of non-negotiable conditions and the set of quasi-hard conditions; as a candidate Under conditions The value of ; This is the threshold value for this condition; To be a robust measure of this condition across the entire database, the median absolute deviation can be taken. It is a logarithmic probability function; To follow the training rounds A linearly increasing steepness coefficient. Its initial value, The steepness growth constant; For candidates in terms of quasi-hard conditions The available deduction amount is supplemented by the directional compensability in step 3. Determined together with the maximum migration confidence, As a condition The corresponding skills To offset the upper limit coefficient, this embodiment can take... That is, the deduction shall not exceed 30% of the threshold value.
[0172] Understandably, this is done with a robust approach. Normalization is necessary: the ordered coding values for educational level range from approximately 1 to 5, while the values for years of work experience can range from 0 to 30. Without normalization, the same initial value would be lost. The degree of softness or hardness corresponding to the two types of conditions will differ by an order of magnitude, and the behavior of the mask will strongly depend on the dimension of the condition rather than its business nature.
[0173] It should be noted that the monotonically increasing steepness with each training epoch is a deterministic annealing method that strikes a balance between optimization feasibility and constraint strictness: in the early stages of training, the mask is nearly linear near the threshold, thus preserving usable gradients and allowing the model to learn the difference between near and far below the threshold; in the later stages of training, the mask approaches a step, thus preventing penetration. This timeline is entirely determined by the number of training epochs. The decision contains no learnable components, therefore its tightening process can be fully predicted and reproduced. Figure 9 This example shows a comparison chart of the hard threshold penetration rate changing with training rounds; from Figure 9 As can be seen, under the steepness annealing arrangement according to training rounds, the proportion of pairs that do not meet the hard threshold in the output list of this scheme drops rapidly from the early stage of training, reaching only 0.195% by the fiftieth round, while the control scheme without gating mask fluctuates steadily around 16.518%.
[0174] Furthermore, if all hard conditions are treated as steep masks, it will conflict with the capability transfer mechanism established in step 3: a cross-disciplinary candidate recalled through skill transfer often has a slight deficiency in conditions such as years of experience in the field. If rejected outright, the recall results of step 3 will be completely negated in this step. Therefore, this embodiment only opens the deduction channel determined by the transfer confidence for hard conditions, while the product term of non-negotiable conditions does not contain any deduction channel. No matter how excellent the candidate is in other aspects, as long as the non-negotiable conditions are not met, their mask value will approach zero, and the semantics of rejection will be strictly preserved.
[0175] Sub-step 3: Adaptively determine the number of propagation layers based on the subspace collapse degree.
[0176] Among them, the subspace collapse degree refers to the average pairwise cosine similarity of the representations of all nodes in a certain subspace. The closer its value is to 1, the more the node representations tend to be in the same direction, that is, the more serious the oversmoothing is.
[0177] For example, in this embodiment, the subspace collapse degree can be calculated using the following formula:
[0178] in, For the first Layer capacity is a hard indicator of subspace collapse. When When the set limit is exceeded, propagation stops and the next step is taken. The layer output is the final representation; in this embodiment, the upper limit can be taken. .
[0179] Understandably, this criterion transforms the hyperparameter of the number of propagation layers, which could only be set through repeated trial and error based on experience, into an adaptive quantity directly determined by the distinguishability of the representation. Its deeper meaning lies in the fact that candidate libraries of different sizes and densities can withstand different propagation depths. A fixed number of layers will either result in insufficient propagation on sparse libraries or excessive smoothing on dense libraries. Figure 10 This diagram illustrates the evolution of subspace collapse degree with the number of propagation layers in this embodiment; from Figure 5 As can be seen, the subspace collapse degree of the control scheme exceeds the set upper limit and quickly approaches saturation near the fourth layer, while the collapse degree of this scheme is only 0.628 at the eighth layer, while the control scheme at the same layer has reached 0.977. This allows this scheme to use a larger number of propagation layers to cover a wider skill neighborhood and trajectory neighborhood.
[0180] Sub-step 4: Calculate the resume consistency confidence score based on the comparative learning results of the description view and the resume topology view, and multiplicatively reduce the two-way claim satisfaction components.
[0181] The descriptive view refers to the view composed of the candidate's multimodal descriptive features; the resume topology view refers to the view composed of the structural features of the company hierarchy and project collaboration topology subgraph corresponding to the candidate's resume. It is understandable that candidates can arbitrarily exaggerate their written descriptions of their experiences, but the objective topological structure—such as the hierarchy of their former companies, the scale of their projects, their role in the collaboration network, and the job level distribution of their colleagues—is difficult to fabricate. Therefore, the consistency between the two views can serve as a natural criterion for identifying resume inflation.
[0182] For example, in this embodiment, the comparison calibration target can be calculated using the following formula:
[0183] in, To describe the view representation; Represented as a topological view of the resume; Cosine similarity; To compare the temperature coefficient, this embodiment can take... The objective takes two views of the same candidate as positive sample pairs and cross-view combinations between different candidates as negative sample pairs. Its optimal solution makes the two view representations of the same candidate move closer together and the different candidates push each other away, which is equivalent to maximizing a lower bound of the mutual information between the two views.
[0184] Based on this, the confidence level and reduction of resume consistency can be calculated using the following formula:
[0185] in, as a candidate The set of candidates for the corresponding job family; and These are the median and absolute deviation of the median for the set, respectively. As the lower limit of the consistency confidence level, in this embodiment, we can take... .
[0186] It's important to note that using the median and median absolute deviation instead of the mean and standard deviation is to prevent the criterion from being contaminated by a few extreme cases of inflated data. Under mean-standard deviation standardization, even a small number of extreme samples in the database will skew the center and scale of the overall distribution, causing genuinely inflated samples to fall into the normal range. The significance of setting a lower limit for the consistency confidence level is that consistency testing is only probabilistic evidence, not a factual determination. A candidate should not be completely eliminated from the system based solely on this test; the reduced candidates still have the opportunity to be recovered through hard indicators and transferability. Figure 11 A comparison chart showing the impact of the history water injection ratio on matching accuracy in this embodiment is shown; from Figure 11 As can be seen, as the proportion of inflated resume samples increases, the matching accuracy deteriorates rapidly without comparative calibration. When the proportion of inflated samples reaches 30%, the decrease is 0.210. However, after introducing graph-multimodal comparative calibration and consistency confidence reduction, the decrease is only 0.090, which is about 0.43 times that of the former, making this scheme robust in noisy and sparse scenarios.
[0187] Step 6: Substitute the satisfaction components of both sides into the job seeker's utility function and the job's utility function respectively, take the geometric mean of the surplus of the utility on both sides relative to their respective reserved utility, and construct a two-way matching satisfaction matrix. Sort the matching results according to the matrix and output the matching results. Optionally, perform stable matching allocation under capacity constraints.
[0188] Here, the utility function refers to a non-linear function that converts the degree of satisfaction of a subject on each demand item into the overall sense of gain for that subject. Retention utility refers to the opportunity cost that a subject on one side could obtain by not accepting the current pairing; it means that I would rather not have it than accept anything below this level.
[0189] A two-way matchmaking satisfaction matrix is a matrix with candidates as rows and positions as columns, where each element represents the matchmaking quality of the candidate and position provided that both parties have positive utility.
[0190] In this embodiment, to ensure that the matching criterion truly reflects whether both parties are willing, rather than simply conflating the linearly additive claims of both sides as in a vector dot product, the specific implementation of this step may include the following sub-steps: Sub-step 1: Construct a two-sided nonlinear utility function in an isoelastic form.
[0191] For example, in this embodiment, the two-sided utility can be calculated using the following formula:
[0192] in, The set of job seekers’ demands includes, in this embodiment, at least salary, commuting, growth opportunities and team atmosphere. This set of requirements from the recruiter's side includes, in this embodiment, at least, skill proficiency, stability, labor costs, and talent pool fit. To read the request item from the corresponding message vector The range of the satisfaction reading function is defined as follows: and with the smallest positive number A safety net is provided to ensure that exponentiation is defined and that the various utilities can be summed. For the claim item The risk aversion coefficient is a factor whose higher value indicates that the item is less replaceable. In this embodiment, commuting and labor costs can be considered... The growth potential and talent pool are suitable. ; and The demands of both sides are weighted and uniformly assigned; and This serves as a retention function for both sides.
[0193] Understandably, the utility of the isoelastic form exhibits a strict diminishing marginal utility property. The utility increment resulting from increasing salary satisfaction from 0.4 to 0.6 is significantly greater than the increment from increasing it from 0.8 to 1.0. This property is precisely what is essential in recruitment scenarios but lacks in vector dot products: no matter how high the salary, it cannot compensate for a three-hour one-way commute; no matter how strong the skills, they cannot compensate for hiring costs far exceeding the budget. It should be noted that... At that time, this term degenerates into logarithmic utility. When it degenerates into linear utility, this form actually includes linear additivity as a special case, and its expressive power is strictly stronger than that of the inner product.
[0194] Sub-step 2: Estimate the bilateral retention utility online from historical behavior.
[0195] In this embodiment, if the retention utility is taken as a global constant, the phenomenon of candidates not being interested cannot be individually characterized. The acceptable lower limit for a senior candidate with multiple offers from competitors is drastically different from that of a recent graduate urgently needing employment. For example, the retention utilities on both sides can be calculated using the following formula:
[0196]
[0197] in, as a candidate A collection of positions that were historically explicitly rejected or withdrawn; For the position The set of candidates who have actually received offers of employment in history; For quantile operators; For quantile margin, we can take This can be understood as follows: a candidate's retention utility is no lower than the best among the positions they have previously rejected, and the recruiter's retention utility is no higher than the weakest among the candidates they have previously been willing to offer a position to. For new users with insufficient historical records, the median retention utility of their job family is used as a safety net.
[0198] Sub-step 3: Take the geometric mean of the utility surpluses on both sides, construct a two-way matching satisfaction matrix, and output the sorting results.
[0199] For example, in this embodiment, the bidirectional matching satisfaction level can be calculated using the following formula:
[0200]
[0201] in, The hard-constraint gating mask obtained in step 5; To obtain the positive part operator.
[0202] Understandably, this formula takes the geometric mean of the utility surpluses on both sides, and its mathematical properties constitute the core of the matching criterion in this method: if the surplus on either side is zero or negative, that side would rather not accept this pairing. The score is directly zero. Pairings destined to fail—those where the candidate doesn't want them and those where the company can't afford them—are placed in the same high-scoring range as truly win-win pairings under the inner product metric, but are both pushed to zero under the geometric mean criterion. Multiplying this by a hard-constraint gating mask creates a triple progressive criterion: hard-condition rejection, simultaneously positive bilateral utility, and maximum surplus product. Figure 12 This embodiment illustrates the two-sided utility surplus distribution and the identification map for false high-matches; from Figure 12 As can be seen, since the utility surpluses on the job seeker side and the recruiter side are significantly negatively correlated, 60.67% of the top pairings selected by the linearly additive control criterion fall into the region where the one-sided surplus is not positive, which means that the candidates do not like them or the companies cannot afford them, which is a false high match. However, this criterion, which takes the geometric mean of the two-sided surpluses, sets all of these pairings to zero without exception.
[0203] Furthermore, in this embodiment, to ensure that the output goes beyond sorting and can be implemented as an executable allocation scheme, a delayed acceptance post-processing under capacity constraints can be performed on the bidirectional matching satisfaction matrix. This involves only processing according to... A practical drawback of outputting the recommendation list in descending order is that a few top candidates will appear at the top of a large number of job listings simultaneously, while these positions cannot actually hire the same person at the same time. This results in the duplicate use of recommendation resources and the continuous vacancy of long-tail positions. For example, the two-sided preference order can be induced as follows:
[0204] in, and Strict preference order is applied to both the candidate and job positions; based on the number of openings for each position. For capacity constraints, a delayed acceptance process is executed until no more applications can be submitted. Since the preference order on both sides is induced by the same satisfaction matrix, and this matrix has set unacceptable pairings on one side to zero and considered them unacceptable, the assignments converged by the delayed acceptance process are stable, i.e., there is no pair of candidate and position whose satisfaction is higher than their respective current assignments.
[0205] It should be noted that this post-processing only changes the final allocation format and does not alter the representation and criteria of the aforementioned layers. When there are sufficient job openings, or when the application scenario only requires a recommendation list and not an allocation scheme, this step can be skipped and the process can proceed directly. Output in descending order.
[0206] Furthermore, this embodiment also addresses the scenarios of sparsity and cold start. In this embodiment, when the number of historical interaction edges of a candidate in a bidirectional asymmetric heterogeneous graph is lower than a set threshold, message passing on both sides will fail due to the small neighborhood. To address this, a proxy neighborhood is formed by the candidate's skill neighborhood in the skill topology knowledge graph and the set of candidates in the same trajectory segment on the career trajectory evolution subgraph. The message passing input is supplemented by the weighted average of the candidate representations within the proxy neighborhood, with the weight being the product of the skill neighborhood similarity and the trajectory segment similarity. When the set of hard threshold conditions for a certain position is empty, the hard constraint gating mask is always set to 1, and the retention utility of that position is set to the median of the retention utility of its respective position family.
[0207] Regarding the overall training objective. In this embodiment, the overall training objective of the model is a weighted combination of the matching supervision term and three regularization terms:
[0208] in, The ordered regression loss is constructed using a three-level labeling system: historical delivery, approval, and acceptance. The weight coefficients for each regularization term can be exemplarily taken as follows: , , In terms of the training schedule, a two-stage arrangement can be adopted: the first stage freezes the steepness of the hard constraint gating mask to the initial value and mainly optimizes the decoupling and contrast calibration terms to allow the representation to take shape first; the second stage enables steep annealing and introduces the career transition gradient regularization term to gradually tighten the matching criteria.
[0209] Furthermore, regarding the deployment configuration, the method described above in this embodiment can be implemented using a two-tier architecture: The offline tier reconstructs multimodal attribute-entity hypergraph, hierarchical heterogeneous graph, and career trajectory evolution subgraph in batches on a daily or hourly basis, and calculates candidate-side and job-side representations in batches, writing them into the vector index; the online tier only performs incremental updates of the real-time intent graph, gating fusion, hard-constraint gating mask calculation, and bilateral utility aggregation, and uses approximate nearest neighbor retrieval to shrink the candidate set from the full database size to the thousand-level scale before refining. It is understood that the computational complexity of the hard-constraint gating mask is proportional to the number of conditions and independent of the database size; therefore, it can be used as a filtering condition in the retrieval stage, thereby significantly reducing the online computational load without sacrificing the rigor of the criteria.
[0210] In summary, after steps 1 to 6, for any pair of candidates and positions in the candidate set and position set, the system's final output is no longer a scalar similarity with an unclear direction obtained from vector dot products, but a bidirectional matching satisfaction level determined by a hard-constrained gating mask, the job seeker's utility surplus, and the recruiter's utility surplus. This satisfaction level can be read row-wise to obtain a list of recommended positions for a specific candidate, or column-wise to obtain a list of candidates for a specific position. Furthermore, under capacity constraints, it can be transformed into a stable matching allocation scheme through a delayed acceptance process. Simultaneously, since the candidate representation is constrained throughout the propagation process within two mutually orthogonal subspaces of hard ability indicators and soft style, the system can, while providing each recommendation, also read the contribution ratio of that recommendation in the hard indicator dimension and the soft style dimension, as well as the local resume unit focused on by attention, ensuring that each recommendation result has a traceable chain of evidence.
[0211] Example 2
[0212] This embodiment discloses a person-job matching system that integrates multimodal features and graph neural networks. This system is used to execute the method of Embodiment 1.
[0213] In this embodiment, the system described above can be integrated into an electronic device, which may be a terminal device, a server, a server cluster, or a cloud computing platform. The electronic device includes at least a processor and a memory. The memory stores a computer program that can be executed by the processor. When the processor executes the computer program, it implements the steps of the person-job matching method that integrates multimodal features and graph neural networks in Embodiment 1.
[0214] It is understandable that the above modules can be deployed on the same physical device, or they can be deployed on different physical devices or different computing clusters, respectively, for offline and online levels.
[0215] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method of Embodiment 1. The computer-readable storage medium may be a read-only memory, random access memory, magnetic disk, optical disk, or other medium capable of storing program code.
[0216] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for matching people to positions by integrating multimodal features and graph neural networks, characterized in that, The person-job matching method includes: Obtain multimodal resume data of candidates and job description data of positions, and extract modal features of the multimodal resume data under different modalities; The candidate's first and second representation components are determined based on the modal features. The first and second representation components belong to two mutually orthogonal subspaces. The first representation component indicates verifiable qualification attributes, and the second representation component indicates unverifiable style attributes. On the interaction graph, the first type of related attributes of the position are aggregated along the first directed edge from the candidate to the position, using the second representation component as the query basis, to obtain the first satisfaction level. Then, the second type of related attributes of the candidate are aggregated along the second directed edge from the position to the candidate, using the qualification requirements of the position as the query basis, to obtain the second satisfaction level. The first type of related attributes are different from the second type of related attributes. The two-way matching satisfaction of the candidate and the position is determined based on the surplus of the first satisfaction level relative to the first retention level and the surplus of the second satisfaction level relative to the second retention level, wherein the two-way matching satisfaction is the minimum value when either of the two surpluses is not positive. The matching result is output according to the bidirectional matching satisfaction level.
2. The person-job matching method integrating multimodal features and graph neural networks according to claim 1, characterized in that, The extraction of modal features from the multimodal resume data under different modalities includes: Extract at least two of the following from the multimodal resume data: textual semantic features, structured field features, and layout visual features; and extract cross-modal entities from the extracted modal features, wherein the cross-modal entities are semantic units that have evidence in two or more modalities. An attribute relationship graph is constructed using the candidates as nodes, each cross-modal entity as a hyperedge connecting multiple nodes, and the association strength between the candidates and the cross-modal entities as the hyperedge weight. The step of determining the first and second representational components of the candidate based on the modal features includes: Perform hyperedge aggregation on the attribute relationship graph. For each hyperedge, normalize and weight the representation of each node according to the association strength of the nodes it connects, and superimpose the self-embedding of the cross-modal entity corresponding to the hyperedge to obtain the hyperedge representation. Execute node backhaul. For each node, group its hyperedges according to the category of the cross-modal entity, take the mean of the hyperedge representations within the group, and then sum them up by weighted inter-group coefficients to obtain the update increment of the node. The inter-group coefficients are positive and their sum is one. The sum of the representation of the node and the update increment, which are applied by the first projection operator and the second projection operator respectively, yields the first representation component and the second representation component. The product of the first projection operator and the second projection operator is a zero operator, and their output dimensions are the same as their input dimensions.
3. The person-job matching method integrating multimodal features and graph neural networks according to claim 2, characterized in that, The methods for determining the correlation strength include: The occurrence intensity of the cross-modal entity is determined separately for each modality, and the occurrence intensity under each modality is converted to the same bounded value range; When the ratio of the maximum to the minimum occurrence intensity of the same cross-modal entity in each modality exceeds the imbalance threshold, the occurrence intensity in each modality is raised to a set proportion not lower than the maximum, and then weighted and summed according to the modal confidence coefficient corresponding to the category to which the cross-modal entity belongs; The association strength of the hyperedge is scaled by a normalization factor that decreases monotonically with the number of nodes connected to the hyperedge, wherein the normalization factor increases at a rate lower than the inverse rate with the number of nodes.
4. The method for person-job matching that integrates multimodal features and graph neural networks according to any one of claims 1 to 3, characterized in that, The job matching method further includes: segmenting the job description data into semantic units and segmenting the multimodal resume data into modal local units, wherein the number of modal local units is greater than the number of semantic units; The alignment weight is determined using the semantic unit as the query end and the modal local unit as the query end. The alignment weight is normalized only on the modal local unit, and the alignment weight is also adjusted by the prior bias corresponding to the modality to which the modal local unit belongs. The result obtained by weighting the alignment weights is weighted and summed according to the rigidity weights of each semantic unit, and the sum of the rigidity weights of the claims is one. The rigidity weight of the appeal is determined according to the type of qualifier that each semantic unit is subject to in the job description data. The rigidity weight of the appeal corresponding to a semantic unit modified by an uncompromising qualifier is greater than the rigidity weight of the appeal corresponding to a semantic unit modified by an optional qualifier.
5. The person-job matching method integrating multimodal features and graph neural networks according to claim 1, characterized in that, The person-job matching method also includes: Read in skill relationship data, which includes directed relationships from predecessor skills to derived skills and undirected relationships connecting skills of the same origin; Using the candidate, skill, and job position as nodes, a hierarchical heterogeneous graph is constructed based on the skill relationship data; For the first skill required for the position but not directly possessed by the candidate, the migration confidence from the second skill to the first skill is determined based on the travel cost of the first skill and the second skill already possessed by the candidate in the skill relationship data, as well as the hierarchical depth of their nearest common ancestor. Using the migration confidence as a decay factor, the candidate's proficiency level in the first skill is determined by the proficiency level of the second skill.
6. The person-job matching method integrating multimodal features and graph neural networks according to claim 5, characterized in that, The passage cost is differentiated by direction: when traveling from the prerequisite skill to the derived skill along the directional relationship, a learning cost is taken; when traveling in the opposite direction along the directional relationship, a directional penalty term is superimposed on the learning cost; when traveling along the undirected relationship, a cost is taken that monotonically decreases with the source strength. Determining the candidate's proficiency level in the first skill based on the proficiency level of the second skill includes: For each second skill whose migration confidence is not lower than the migration lower limit, the proficiency level of the second skill, the migration confidence from the second skill to the first skill, and the time factor that monotonically decreases with the time elapsed since the most recent practice of the second skill are multiplied together to obtain a candidate value. The maximum value between all the candidate values and the candidate's original proficiency level in the first skill is taken as the candidate's proficiency level in the first skill. The decay rate of the time factor decreases as the duration increases.
7. The person-job matching method integrating multimodal features and graph neural networks according to claim 1, characterized in that, The person-job matching method also includes: Construct a time-stamped trajectory map based on the candidates' past work units, job levels, and tenure periods; On the trajectory graph, the neighboring node representations are first aggregated along the topological dimension, and then convolution is performed along the time dimension using a convolution kernel that only takes historical moments to obtain the advanced rate component and the potential component. The advancement rate component and the potential component are incorporated into the candidate's representation; The advanced rate component is determined in the following manner: For two adjacent positions in the historical resume, the ratio of their rank span to their tenure is taken, and the gain is adjusted monotonically by the hierarchical change of the unit corresponding to the two positions to obtain the progressive rate of advancement. The progressive rates of advancement are then summed monotonically by weighting the time elapsed since the end of their respective positions. A behavioral relationship graph is constructed based on the candidate's interaction behavior with the position within a set time window. The interaction behavior includes at least positive behavior and withdrawal behavior. In the behavioral relationship graph, the weight of the edge corresponding to the withdrawal behavior is negative and the weight of the edge corresponding to the positive behavior is positive. Furthermore, the weight of each edge decreases monotonically with the time elapsed since the occurrence of the interaction behavior. Perform multi-stage propagation on the behavior graph with an order not exceeding a set upper limit, and sum the propagation results of each stage by weighted summation of coefficients that sum to one to obtain the short-term intent component.
8. The person-job matching method integrating multimodal features and graph neural networks according to claim 1, characterized in that, The person-job matching method also includes: The first view is constructed using the multimodal description features of the candidate, and the second view is constructed using the hierarchical level of the employing unit and the collaborative topology corresponding to the candidate's historical resume. Using the first and second views of the same candidate as positive sample pairs and cross-view combinations between different candidates as negative sample pairs, the training aims to make the representational similarity of positive sample pairs higher than that of negative sample pairs. The consistency confidence is determined based on the degree of deviation of the representational similarity between the first view and the second view of the candidate from the corresponding similarity distribution of candidates in the same job family, and the first satisfaction and the second satisfaction are scaled using the consistency confidence. The degree of deviation is determined based on the median and the absolute deviation of the median of the distribution, and the confidence level of consistency is not lower than the lower confidence limit. The first satisfaction level and the second satisfaction level are obtained by weighted summation of the satisfaction levels of multiple claims on their respective sides after marginal reduction mapping. The gain of the marginal reduction mapping on the satisfaction level of each claim decreases as the satisfaction level increases, and the reduction intensity corresponding to each claim can be configured separately.
9. The person-job matching method integrating multimodal features and graph neural networks according to claim 8, characterized in that, The first retention level is determined by the high quantile of the first satisfaction level corresponding to the positions that the candidate has explicitly rejected or withdrawn in the past, and the second retention level is determined by the low quantile of the second satisfaction level corresponding to the candidates who have expressed their intention to hire in the past; when the number of corresponding historical records of the candidate or the position is lower than the record lower limit, the median of the corresponding retention level of the position family to which it belongs is taken.
10. A person-job matching system integrating multimodal features and graph neural networks, characterized in that, The job matching system includes: processor; The memory stores a computer program that, when executed by a processor, implements the human-job matching method that integrates multimodal features and graph neural networks as described in any one of claims 1 to 9.
