CN115062220B - Attention merging-based recruitment recommendation system - Google Patents
Attention merging-based recruitment recommendation system Download PDFInfo
- Publication number
- CN115062220B CN115062220B CN202210679910.1A CN202210679910A CN115062220B CN 115062220 B CN115062220 B CN 115062220B CN 202210679910 A CN202210679910 A CN 202210679910A CN 115062220 B CN115062220 B CN 115062220B
- Authority
- CN
- China
- Prior art keywords
- user
- recruitment
- module
- information
- attention
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000007115 recruitment Effects 0.000 title claims abstract description 108
- 239000013598 vector Substances 0.000 claims abstract description 83
- 238000011176 pooling Methods 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000000034 method Methods 0.000 claims description 14
- 238000012549 training Methods 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 7
- 230000000694 effects Effects 0.000 abstract description 10
- 238000012512 characterization method Methods 0.000 abstract description 4
- 238000012545 processing Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 4
- 238000007477 logistic regression Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000000052 comparative effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000009428 plumbing Methods 0.000 description 1
- 238000009423 ventilation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/10—Office automation; Time management
- G06Q10/105—Human resources
- G06Q10/1053—Employment or hiring
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Computational Linguistics (AREA)
- Entrepreneurship & Innovation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a recruitment recommendation system based on attention merging, which comprises a preprocessing module, a processing module and a processing module, wherein the preprocessing module is used for extracting user information, user recruitment expected information and recruitment information in original recruitment data; the embedding module is used for respectively acquiring a first user feature vector, a user recruitment expected feature vector and a recruitment feature vector; the attention merging module is used for merging the first user characteristic vector and the user recruitment expected characteristic vector into a second user characteristic vector; the prediction output module is used for obtaining the call-through rate representing the probability of user dialing through the recruitment information by taking the spliced second user characteristic vector and the recruitment characteristic vector as input. The invention improves the effectiveness of user characterization learning in the strong correlation field and the weak correlation field through the attention merging module; the pooling module is used for averagely pooling the feature vectors with different lengths into the feature vectors with the same length, so that the problem that long-tail content cannot be effectively recommended due to the Martai effect is solved.
Description
Technical Field
The invention relates to the field of interest recommendation, in particular to a recruitment recommendation system based on attention merging.
Background
At present, the mobility of technical workers in the building field and the mechanical manufacturing field in China is high, the workers and enterprises or work heads are in flexible employment form, namely, the enterprises or work heads have activities which are short-term employment workers, and the workers also find activities when in idle state. The demands of both parties are high in flexibility, and the conventional recruitment websites and systems can not completely meet the demands of both parties. In reality, more of the two are recruited and found by telephone, but because of the asymmetry of information, the two parties cannot know the demands of the other party before communicating, enterprises or workers cannot find workers meeting the demands, and the workers cannot find the heart instrument, so that the recruitment efficiency is low, the time of the two parties is wasted intangibly, and the normal operation of the work is delayed.
The recommendation of recruitment based on a dial-up record of a worker user is a new business recently emerging. The method relies on a mature recommendation algorithm to recommend the enterprise or the work head telephone which is expected to be matched with the worker to the worker, so that accurate recruitment recommendation is realized. However, the current recommendation algorithm has the following problems:
the traditional deep learning technology has the problems of strong feature over-fitting and weak feature under-fitting and the Martai effect, so that the recommendation algorithm based on the deep learning technology has poor effect. The martai effect is a two-pole differentiation phenomenon, namely, most existing recommendation algorithms can only provide recommendation for popular goods or items of the main stream, and the probability of recommending cold goods or items at the long tail to a user is relatively low. The long-tail content of interest to the user is difficult to obtain wide and timely recommendation due to lack of behavior characteristics, so that the quality and diversity of the recommended content are low.
Overfitting of strong features and underfilling of weak features are problems with conventional deep learning based recommendation algorithms. The main reason is that the recommendation results are poor in accuracy due to the fact that the recommendation results are based on an Embedding & MLP (multilayer perceptron) model. When an enterprise or a worker performs a recruitment, these overfitting fields are closely related to the interests of most workers in the recruitment information, i.e., the recommendation is determined by a few fields representing strongly related fields. However, some workers are not interested in the overfitting field, which makes it difficult for the model to make efficient recommendations for that part of the user.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a method for solving the problems of strong feature overfitting and weak feature underfitting when facing the multi-field features, improving the effectiveness of a system in the characterization learning of workers and solving the problem that long-tail content cannot be effectively recommended due to the 'Martai effect'.
In order to achieve the above object, the present invention provides a technical solution comprising:
the recruitment recommendation system based on attention merging comprises a preprocessing module, an embedding module, an attention merging module and a prediction output module which are sequentially connected;
the preprocessing module is used for extracting user information, user recruitment expected information and recruitment information in the original recruitment data;
the embedding module is configured to process the user information, the user recruitment desired information and the recruitment information into a first user feature vector, a user recruitment desired feature vector and a recruitment feature vector respectively;
the attention merging module is configured to merge the first user feature vector and a user recruitment desired feature vector into a second user feature vector;
the prediction output module is used for obtaining the dialing rate representing the probability of user dialing through the recruitment information by taking the spliced second user characteristic vector and the recruitment characteristic vector as input, and recommending the recruitment information corresponding to the dialing rate to the user with the highest dialing rate.
In some preferred embodiments, the user information includes user id, user dial-up record, user search record, user job, user location, and user desired workplace;
the user recruitment desired information comprises a user search record, a user work type, a user place and a user desired work place;
the recruitment information comprises a recruitment information id, a recruitment information title, release time, a work kind and a region where the recruitment information is located.
In some preferred embodiments, the prediction output module is further configured to recommend the recruitment information corresponding to the dialing rate to other users matched with the user with the highest dialing rate; the matching includes: the user is the same as the user.
In some preferred embodiments, a pooling module is further disposed between the embedding module and the attention merging module; the pooling module is configured to average and pool feature vectors of different lengths into feature vectors of the same length.
In some preferred embodiments, the attention combining module further includes dividing feature fields for the first user feature vector and the user recruiting desired feature vector, respectively, and assigning attention weights and activating units to the feature fields.
In some preferred embodiments, the method of combining the first user feature vector with the user recruiting desired feature vector into the second user feature vector comprises:
the first user characteristic vector and the user recruitment expected characteristic vector are connected in series from the second dimension to obtain a matrix E epsilon R n×m The method comprises the steps of carrying out a first treatment on the surface of the Wherein m is the dimension of the vectors after series connection, and n is the number of the feature domains;
calculating an attention weight vector h:
a＝tanh(W 1 E)
wherein W is 1 ∈R k×n And W is 2 ∈R k For the weight matrix, k represents the size of the attention unit, aεR n An activation unit representing the feature field, h E R n ，i∈[0,n)，j∈[0,n)；
Combining the attention weight vector h and the matrix E into a second user feature vector Z:
Z＝hE
wherein Z is E R m 。
In some preferred embodiments, the objective function of the prediction output module is set as a log likelihood function:
wherein,,
is a training set with the size of N, x is the input of a prediction output module, y is epsilon { 0}1, indicating that y=1 when the recruitment information is a positive sample, otherwise y=0; p (x) is the output of the prediction output module and represents the probability that the predicted recruitment information is dialed.
Advantageous effects
The attention merging module distributes reasonable weights according to the user characteristics and the recruitment characteristics of the context and activates different nerve units, so that the effectiveness of user characterization learning in the strong correlation field and the weak correlation field is improved; the pooling module is used for averagely pooling the feature vectors with different lengths into the feature vectors with the same length, so that the problem that long-tail content cannot be effectively recommended due to the Martai effect is solved.
Drawings
FIG. 1 is a schematic diagram of a system architecture according to a preferred embodiment of the present invention;
FIG. 2 is a schematic diagram of a system architecture according to another preferred embodiment of the present invention;
FIG. 3 is a schematic diagram of a data flow and system configuration according to another preferred embodiment of the present invention;
FIG. 4 is a graph showing the results of comparison of AUC and loss for a preferred embodiment of the invention and a control experimental example;
Detailed Description
The present invention will be further described with reference to the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent. In the description of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "top," "bottom," "inner," "outer," and the like indicate or are based on the orientation or positional relationship shown in the drawings, merely to facilitate description of the present invention and to simplify the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and thus should not be construed as limiting the present invention.
Examples
As shown in fig. 1, the embodiment discloses a recruitment recommendation system based on attention merging, which comprises a preprocessing module, an embedding module, an attention merging module and a prediction output module which are sequentially connected;
the preprocessing module is used for extracting user information, user recruitment expected information and recruitment information in the original recruitment data; in this embodiment, the original recruitment data mainly includes data provided by the mobile operator, and mainly includes recruitment information, user dial-up records, user job area information, user expected workplace, user search records, job index and region index, and the following is briefly described about fields included in each information:
the recruiting information includes: pid (recruitment information id), pro_region (region where recruitment information is located), wlt _id (work id), create_time (creation time), title (recruitment information title);
the user dial-up record includes: a user id, a type (type=1, if type=1, it is described that the dial-up record is a worker-found record, if type=2, it is described that the dial-up record is a work head or enterprise recruitment record, only a record with type 1 is selected, an object_id (recruitment information id of user dial-up), a call_time (user dial-up time);
the user work area information comprises: uid (user id), wlt _id (job id), city_code (city id where the user is located);
the user desires to include, at work: uid (user id), city_code (user desired work city id);
the user search record contains: uid (user id), keyword (user search term);
the work index includes: wlt _id (seed id), wlt _name (seed name);
the area index includes: city_code (city id), city_name (city name);
it should be appreciated that due to the faster database update rate, a significant portion of the user's dialed recruitment records cannot be found in the recruitment information table, and many users do not have search records and desired job records. If the part of users with less information content is directly put into the model training, the model effect is affected, so that the embodiment needs to screen out seed users with more information content. If a user's pick-up record is found in the recruitment information table, he is selected as a seed user.
The preprocessing module extracts user information, user recruitment expected information and recruitment information required by subsequent steps from the original recruitment data; in some preferred embodiments, the user information includes user ID, user dial-up record, user search record, user job, user location and user desired job, and for each dial-up record of the user, record the dial-up time and make a difference between the user dial-up time and the time of release of the recruitment information, namely call_time-create_time, recorded as age, which is used to learn the user's preference for the new and old recruitment information, for example, { user ID:1721, user dial-up record: [ (2193,17289378973), (8324,1984739084) ], the user searches for records: [ installation door, day knot ], user job: 54, user location: 370200, the user desires to work: 440100}; further, in order to further mine the user information so that the subsequent attention combination can reflect more characteristic domain features, the user recruitment expected information is extracted on the basis of the user information, wherein the user information comprises a user search record, a user work type, a user place and a user expected work place; the recruitment information includes a recruitment information ID, a recruitment information title, a release time, a work kind, and a region where the work is located, for example, { the recruitment information ID:7354, information title: "economic development area project recruits installer/air conditioner/plumbing/ventilation", release time: 1641282268, work species: 375, area of: 21574264}.
The embedding module is configured to process the user information, the user recruitment desired information and the recruitment information into a first user feature vector, a user recruitment desired feature vector and a recruitment feature vector respectively; in some preferred embodiments, a Bert-base-Chinese version of the Chinese pre-training Bert model is selected as the embedding module, and is put into the model to train the adjustment parameters together in order to adapt to the recruitment dataset. It should be appreciated that because different users have different numbers of dial-up records and search records, the number of features included in a user feature vector may be different in different instances, which may result in different lengths of the user feature vector having multiple features in series. Whereas the subsequent fully-connected network can only handle input vectors of fixed length, in some preferred embodiments, it is contemplated that pooling operations are utilized to convert feature vectors of different lengths to feature vectors of the same length, thereby ensuring that the data can be handled by the fully-connected layer. Specifically, as shown in fig. 2, a pooling module is further disposed between the embedding module and the attention merging module; the pooling module is configured to average and pool feature vectors of different lengths into feature vectors of the same length.
The attention merging module is configured to merge the first user feature vector and a user recruitment desired feature vector into a second user feature vector; it should be appreciated that the conventional recommendation algorithm is based on a model of the tandem layer. And the serial layers connect different characteristic domains in series to complete parameter learning. During the training process, there are some specific domain overfitting problems, and these overfitting domains are closely related to the interests of most users in the recruitment information, i.e. the recommendation result is determined by a few fields representing strongly related domains. However, some users have not high interest in the overfitting field, which makes it difficult for the model to make efficient recommendations for that part of the users. The invention replaces the serial layers in the traditional algorithm by the attention merging module, merges the user characteristics and the user recruitment expected characteristics as more comprehensive user characterization. In some preferred embodiments, the attention merging module further includes a feature domain dividing unit for dividing the first user feature vector and the user recruiting desired feature vector, and an attention weight and activation unit for assigning attention weight and activation unit to the feature domain, and the specific method includes:
the first user characteristic vector and the user recruitment expected characteristic vector are connected in series from the second dimension to obtain a matrix E epsilon R n×m The method comprises the steps of carrying out a first treatment on the surface of the Wherein m is the dimension of the vectors after series connection, n is the number of the feature domains, and R is a matrix; it will be appreciated that the first row of two-dimensional vectors (i.e. the first dimension) represents one feature vector, and that stitching from the first dimension is not meaningful, so stitching from the second dimension is to join all feature vectors of one user together.
Calculating an attention weight vector h:
a＝tanh(W 1 E)；
wherein W is 1 ∈R k×n And W is 2 ∈R k For the weight matrix, k represents the size of the attention unit, aεR n An activation unit representing the feature field, h E R n I e [0, n ], j e [0, n); the above formula is softmax formula, with the purpose of normalizing, i representing the i-th feature domain, j on the denominator requires all traversal of i e 0, n). It should be appreciated that the weight matrix W 1 And W is 2 The matrix is randomly initialized at the beginning of training, and is automatically adjusted along with the training process to obtain the optimal weight matrix. The activation unit learns the weight of each feature domain in different user samples in the training process, and continuously and automatically adjusts in the training process to obtain the optimal activation unit.
Combining the attention weight vector h and the matrix E into a second user feature vector Z:
z=he; wherein Z is E R m 。
The prediction output module is used for obtaining the dialing rate representing the probability of user dialing through the recruitment information by taking the spliced second user characteristic vector and the recruitment characteristic vector as input, and recommending the recruitment information corresponding to the dialing rate to the user with the highest dialing rate. It should be appreciated that the stitching of the second user feature vector and the recruiting feature vector in this embodiment is performed by the concatate layer stitching.
On the other hand, as mentioned above, since there is a user whose information is not enough to complete in the original recruitment data, in some preferred embodiments, for this part of users, the prediction output module is further configured to recommend the recruitment information corresponding to the dial rate to the other users matching the user with the highest dial rate; the matching includes: the user is the same as the user.
In some preferred embodiments, a loss function is used to determine whether training of the recommendation system is complete, and specifically, the objective function is set as a log likelihood function:
wherein,,
the training set is a training set with the size of N, x is the input of a prediction output module, y epsilon {0,1}, which means that when the recruitment information is a positive sample, y=1, otherwise y=0; p (x) is the output of the prediction output module and represents the probability that the predicted recruitment information is dialed.
Comparative examples
In the comparison test, the Bert model of the embedded module adopts a Chinese pre-training model of the Bert-base-Chinese version, the hidden layer size (hidden size) is 768, the training sample size (batch size) is set to 4, the learning rate (learning rate) is set to 1e-5, and the SGD is used as a model optimizer. The activation function used is ReLU, the learning framework used is Pytorch, and the experimental operation platform is a computer with Intel Xeon@E5-2678v3cpu, GTX2080TI video memory of 11G and memory size of 32G. The attention combination-based recruitment recommendation system disclosed by the invention is named RecRec (Recruitment Recommendation Model based on Attention Merging), and the circulation and system structure of data are shown in fig. 3.
The contrast algorithm system comprises:
LR: logistic Regression (LR) is a shallow model that was widely used before in deep networks for CTR prediction tasks.
YouTube DNN: youTube DNN follows the Embedding & MLP architecture, and is the basis for most of the later developed deep networks for CTR modeling. It provides a powerful baseline for our model comparison.
Wide & Deep: in practical industrial applications, the Wide & Deep model has been widely accepted. It is composed of two parts: 1) A Wide model for processing the characteristics of the artificially designed cross products; 2) The Deep model automatically extracts nonlinear relations among features.
PNN: PNN introduces a product layer after the embedding layer to capture the interaction of higher-order features.
Deep fm: on the basis of a Wide & Deep model, the Deep FM takes a factorizer as a 'Wide' module so as to save characteristic engineering operation.
The evaluation index includes:
AUC is an index of wide application in the field of CTR prediction. AUC is the area under the ROC curve, representing the probability that the predictive score for a positive sample is greater than the predictive score for a negative sample. Specifically, if M represents the number of positive samples, N represents the number of negative samples, calculate all M N positive and negative sample pairs, how many groups of positive sample scores are greater than the negative sample score, and then divide by MN. The time complexity of the calculation method is O (n 2 ) Where n=m+n, AUC is calculated as follows:
wherein pred pos >pred neg Representing a positive and negative sample pair, the value is 1 if the positive sample score is greater than the negative sample score, otherwise the value is 0.
Furthermore, a RelaImpr index was introduced to measure the relative improvement to the model. For a random guesser, the AUC value was 0.5. As a basis for the CTR modeled depth network, we choose the YouTube DNN model as the base model. Thus, the definition of Relaimpr is as follows:
experimental results
All experiments were repeated 5 times and the average results are reported as shown in table 1. It is clear that all deep networks significantly defeat the LR system, which does demonstrate the strength of deep learning. PNN and Deep FM with special design structures perform better than Wide & Deep. The RecRec proposed by the present invention performs best in all comparative systems.
TABLE 1 comparison of System predictions across recruiting datasets
Comparison of attention combining layer and attention series layer
As shown in fig. 4, the attention combining layer in the model proposed by the invention was replaced with a conventional attention series layer, which was used as a control experiment example, and it can be seen that the attention combining layer performed better than the series layer in terms of AUC and loss of the test set. When the feature vector passes through the series layers and feeds forward, the same portion of all the user's neurons are activated. When passing through the attention merging layer, different neurons will be activated facing different user features, which means that interactions between different features are learned, which improves the performance of the system.
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (4)
1. Attention merging-based recruitment recommendation system, which is characterized in that: the device comprises a preprocessing module, an embedding module, an attention merging module and a prediction output module which are sequentially connected;
the preprocessing module is used for extracting user information, user recruitment expected information and recruitment information in the original recruitment data;
the embedding module is configured to process the user information, the user recruitment desired information and the recruitment information into a first user feature vector, a user recruitment desired feature vector and a recruitment feature vector respectively;
the attention merging module is configured to merge the first user feature vector and a user recruitment desired feature vector into a second user feature vector;
the prediction output module is used for obtaining a dialing rate representing the probability of user dialing through the recruitment information by taking the spliced second user characteristic vector and the recruitment characteristic vector as input, and recommending the recruitment information corresponding to the dialing rate to the user with the highest dialing rate;
the user information comprises user id, user dial-up record, user search record, user work, user place and user expected place;
the user recruitment desired information comprises a user search record, a user work type, a user place and a user desired work place;
the recruitment information comprises a recruitment information id, a recruitment information title, release time, a work type and a region where the recruitment information is located;
the attention merging module further comprises a feature domain dividing unit for dividing the first user feature vector and the user recruitment expected feature vector respectively, and an attention weight and activation unit for distributing the feature domain;
the method for combining the first user feature vector and the user recruitment expected feature vector into the second user feature vector comprises the following steps:
starting from a second dimension, connecting the first user characteristic vector and a user recruitment expected characteristic vector in series to obtain a matrix; wherein m is the dimension of the vectors after series connection, and n is the number of the feature domains;
calculating an attention weight vector
：
Wherein,,
and->
Is a weight matrix>
Indicating the size of the attention unit, +.>
An activation unit representing the characteristic field, +.>
，/>
，/>
；
The attention weight vector
Sum matrix->
Merging into a second user feature vector Z:
wherein,,
。
2. the attention-based merge recruitment recommendation system of claim 1, wherein: the prediction output module is further configured to recommend recruitment information corresponding to the dial rate to other users matched with the user with the highest dial rate; the matching includes: the user is the same as the user.
3. The attention-based merge recruitment recommendation system of claim 1, wherein: a pooling module is arranged between the embedding module and the attention merging module; the pooling module is configured to average and pool feature vectors of different lengths into feature vectors of the same length.
4. The attention-based merge recruitment recommendation system of claim 1, wherein: the objective function of the prediction output module is set as a log likelihood function:
wherein,,
is of size +.>
Training set, I/O (18)>
For predicting the input of the output module, < >>
Indicating +.f when the recruitment information is positive sample>
Otherwise->
；/>
The output of the prediction output module indicates the probability that the predicted recruitment information is dialed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210679910.1A CN115062220B (en) | 2022-06-16 | 2022-06-16 | Attention merging-based recruitment recommendation system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210679910.1A CN115062220B (en) | 2022-06-16 | 2022-06-16 | Attention merging-based recruitment recommendation system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115062220A CN115062220A (en) | 2022-09-16 |
| CN115062220B true CN115062220B (en) | 2023-06-23 |
Family
ID=83200083
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210679910.1A Active CN115062220B (en) | 2022-06-16 | 2022-06-16 | Attention merging-based recruitment recommendation system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115062220B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115481220B (en) * | 2022-09-20 | 2023-07-25 | 深圳今日人才信息科技有限公司 | Intelligent matching method and system for comparison learner post based on post and resume content |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20210047284A (en) * | 2020-05-08 | 2021-04-29 | 바이두 온라인 네트웍 테크놀러지 (베이징) 캄파니 리미티드 | Recruitment position description text generation method, device, apparatus and medium |
| CN113641904A (en) * | 2021-08-16 | 2021-11-12 | 上海花千树信息科技有限公司 | Method and device for recommending target customers for stores under Internet marriage line |
| CN114169869A (en) * | 2022-02-14 | 2022-03-11 | 北京大学 | Attention mechanism-based post recommendation method and device |
| CN114358657A (en) * | 2022-03-09 | 2022-04-15 | 北京大学 | Post recommendation method and device based on model fusion |
Family Cites Families (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105488662B (en) * | 2016-01-07 | 2021-09-03 | 北京华品博睿网络技术有限公司 | Online recruitment system based on bidirectional recommendation |
| CN107578214A (en) * | 2017-09-05 | 2018-01-12 | 四川民工加网络科技有限公司 | A kind of method and system of building trade recruitment |
| CN107679757A (en) * | 2017-09-30 | 2018-02-09 | 四川民工加网络科技有限公司 | The matching process and device of services dispatch |
| US20190114593A1 (en) * | 2017-10-17 | 2019-04-18 | ExpertHiring, LLC | Method and system for managing, matching, and sourcing employment candidates in a recruitment campaign |
| CN108960787A (en) * | 2018-08-14 | 2018-12-07 | 安徽网才信息技术股份有限公司 | A method of it is issued for enterprise and updates recruitment casual labour's information |
| CN109960759B (en) * | 2019-03-22 | 2022-07-12 | 中山大学 | Recommendation system click rate prediction method based on deep neural network |
| CN109949012A (en) * | 2019-03-25 | 2019-06-28 | 贵州爱唐文化网络科技有限公司 | It works in short term on a kind of line the method and interaction platform of process dynamic interaction |
| CN111062775B (en) * | 2019-12-03 | 2023-05-05 | 中山大学 | Recommendation system recall method based on attention mechanism |
| CN111339415B (en) * | 2020-02-25 | 2023-06-20 | 中国科学技术大学 | Click rate prediction method and device based on multi-interactive attention network |
-
2022
- 2022-06-16 CN CN202210679910.1A patent/CN115062220B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20210047284A (en) * | 2020-05-08 | 2021-04-29 | 바이두 온라인 네트웍 테크놀러지 (베이징) 캄파니 리미티드 | Recruitment position description text generation method, device, apparatus and medium |
| CN113627135A (en) * | 2020-05-08 | 2021-11-09 | 百度在线网络技术（北京）有限公司 | Method, device, equipment and medium for generating recruitment post description text |
| CN113641904A (en) * | 2021-08-16 | 2021-11-12 | 上海花千树信息科技有限公司 | Method and device for recommending target customers for stores under Internet marriage line |
| CN114169869A (en) * | 2022-02-14 | 2022-03-11 | 北京大学 | Attention mechanism-based post recommendation method and device |
| CN114358657A (en) * | 2022-03-09 | 2022-04-15 | 北京大学 | Post recommendation method and device based on model fusion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115062220A (en) | 2022-09-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110119467B (en) | Project recommendation method, device, equipment and storage medium based on session | |
| CN111080400B (en) | Commodity recommendation method and system based on gate control graph convolution network and storage medium | |
| WO2021139415A1 (en) | Data processing method and apparatus, computer readable storage medium, and electronic device | |
| CN112307332B (en) | Collaborative filtering recommendation method and system based on user portrait clustering and storage medium | |
| CN111209475A (en) | Interest point recommendation method and device based on space-time sequence and social embedded ranking | |
| Sun et al. | Self-attention network for session-based recommendation with streaming data input | |
| CN111563770A (en) | Click rate estimation method based on feature differentiation learning | |
| CN111581520A (en) | Item recommendation method and system based on item importance in session | |
| CN115062220B (en) | Attention merging-based recruitment recommendation system | |
| CN115510322A (en) | Multi-objective optimization recommendation method based on deep learning | |
| Chen et al. | Session-based recommendation: Learning multi-dimension interests via a multi-head attention graph neural network | |
| CN113704438B (en) | Conversation recommendation method of abnormal picture based on layered attention mechanism | |
| Zhang et al. | Emotion attention-aware collaborative deep reinforcement learning for image cropping | |
| CN115203529A (en) | Deep neural network recommendation model and method based on multi-head self-attention mechanism | |
| CN113344648B (en) | Advertisement recommendation method and system based on machine learning | |
| CN113590965B (en) | Video recommendation method integrating knowledge graph and emotion analysis | |
| Chen et al. | Generative adversarial reward learning for generalized behavior tendency inference | |
| CN110378269A (en) | Pass through the movable method not previewed in image query positioning video | |
| CN112364245A (en) | Top-K movie recommendation method based on heterogeneous information network embedding | |
| Chen et al. | Combine temporal information in session-based recommendation with graph neural networks | |
| CN116797799A (en) | Single-target tracking method and tracking system based on channel attention and space-time perception | |
| Sang et al. | Position-aware graph neural network for session-based recommendation | |
| CN115982373A (en) | Knowledge graph recommendation method combining multi-level interactive contrast learning | |
| CN112559905B (en) | Conversation recommendation method based on dual-mode attention mechanism and social similarity | |
| CN115689639A (en) | Commercial advertisement click rate prediction method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |