CN113362034A - Position recommendation method - Google Patents
Position recommendation method Download PDFInfo
- Publication number
- CN113362034A CN113362034A CN202110661337.7A CN202110661337A CN113362034A CN 113362034 A CN113362034 A CN 113362034A CN 202110661337 A CN202110661337 A CN 202110661337A CN 113362034 A CN113362034 A CN 113362034A
- Authority
- CN
- China
- Prior art keywords
- job
- user
- browsing
- positions
- recommendation
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 239000013598 vector Substances 0.000 claims abstract description 58
- 238000012549 training Methods 0.000 claims abstract description 50
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 13
- 238000005295 random walk Methods 0.000 claims abstract description 11
- 238000012545 processing Methods 0.000 claims description 4
- 230000014759 maintenance of location Effects 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 230000007115 recruitment Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000009191 jumping Effects 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000004931 aggregating effect Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2457—Query processing with adaptation to user needs
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/067—Enterprise or organisation modelling
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Entrepreneurship & Innovation (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Economics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- General Business, Economics & Management (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Databases & Information Systems (AREA)
- Development Economics (AREA)
- Educational Administration (AREA)
- Game Theory and Decision Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a position recommendation method, which comprises the following steps: s0, acquiring all job position data, user historical behavior data and currently browsed jobs of target users, and acquiring job browsing sequences of all users according to all user historical behavior data; s1, generating a new job position sequence group by using the job position browsing sequences of all users and combining a random walk algorithm; s2, training a Word2vec model by adopting a Skip-gram frame to generate a feature vector of each position; and S3, calculating the cosine similarity of the feature vectors between the currently browsed positions of the target user and all positions, and forming a position candidate set. The job candidate set is obtained by utilizing the feature vector cosine similarity between the currently browsed job of the target user and all jobs and is used as the job recommendation set.
Description
Technical Field
The invention belongs to the technical field of recommendation systems, and particularly relates to a position recommendation method.
Background
Most of the current human resource communication platforms and job recommendation methods in the recruitment website are based on traditional recommendation algorithms, including collaborative filtering, content-based recommendation methods, hybrid recommendation methods and the like.
According to the Chinese patent with the authorized notice number of CN 105893641B and the name of a position recommendation method, the first scoring value of the position is calculated and obtained based on the multi-domain scoring value of the content according to the user preference model and the position model, and the position set is obtained in sequence; if a position has a delivery record and belongs to the position set, calculating the score value of the position based on the similarity of the user background information according to the user preference model and the position data to obtain a second score value of the position; obtaining a mixed score value of the position according to a first score value and a second score value of the position, and sequencing to obtain a recommendation list, wherein the process mainly comprises the steps of firstly matching according to the preference attribute and the position attribute of a user to obtain a position set; and then, taking an intersection from the position set and the delivery position set corresponding to all the users with similar backgrounds to the current user to obtain final position recommendation information. Although the process solves the problems of Martha effect and cold start due to the fact that the interest weight values can be set in the user preference model by self or calculated according to historical browsing records, and meanwhile, the personalized recommendation can be realized well by utilizing the group; however, when the job hunting intention of the user changes suddenly but the related attributes in the preference model are not changed in time, the conversion is only reflected in that the browsed job span is large, when the browsed number of jobs with the large span is small, the latest interest weight in the preference model obtained by using historical data is low, so that the recommended jobs are the jobs which the user has been interested in, and the recommended jobs do not tightly meet the current job hunting requirements of the user, so that the job recommendation method is poor in real-time performance and low in accuracy.
Disclosure of Invention
In order to solve the technical problems of poor instantaneity and low accuracy of the conventional position recommendation algorithm, the invention provides a position recommendation method, which utilizes the cosine similarity of a feature vector between a currently browsed position of a target user and all positions to obtain a position candidate set as a position recommendation set, so that the instantaneity and the accuracy of the recommendation method are improved; in addition, related positions with high heat from multiple dimensions are added in the position candidate set, so that the cold start problem is further solved, and the recommendation effect is improved; finally, an online recommendation model is obtained according to the historical browsing position time of all users and whether the users are collected as tags, positions in the position candidate set are predicted and scored by the online recommendation model, and the accuracy of the recommendation method is further improved; the position recommendation method provided by the invention avoids the problems of Martha effect and cold start, and has high real-time performance and accuracy.
In order to achieve the above purpose, the invention adopts a technical scheme as follows:
a job recommendation method comprises the following steps:
step S0, acquiring all position data, user historical behavior data and positions currently browsed by a target user, and acquiring position browsing sequences of all users according to the user historical behavior data;
step S1, generating a new job position sequence group by using the job position browsing sequences of all users and combining a random walk algorithm;
step S2, training a Word2vec model by adopting a Skip-gram frame to generate a feature vector of each position;
step S3, calculating the cosine similarity of the feature vectors between the currently browsed position of the target user and all positions, and forming a position candidate set.
Further, step S1 specifically includes the following steps:
step S11, generating a role weighted directed relationship graph by using the role browsing sequence of each user, wherein the direction of directed edges among the roles is determined according to the time sequence, and the weight of the directed edges is determined according to the user context information corresponding to the downstream roles associated with the directed edges;
and step S12, selecting any position node in the position weighted directed relationship graph to randomly walk to form a new position sequence group, wherein the probability of walking from the current position node to the next position node is related to the directed edge weight of the current position node.
Further, the user context information in step S11 includes the life cycle of the job and the manner of entering the job browsing page.
Further, step S2 specifically includes the following steps:
step S21, obtaining a word vector training sample set by adopting a Skip-gram frame, sequentially extracting position groups from each new position sequence by utilizing a sliding window, carrying out one-hot coding on a central position in the position groups to form a central position coding vector as input data of the word vector training sample, and carrying out Multi-hot coding on a plurality of edge positions in the position groups to form a mixed edge position coding vector as a label of the word vector training sample;
and step S22, training the Word2vec model by using the Word vector training sample set to obtain the feature vector of each position.
Further, the candidate set in step S3 further includes a plurality of popular positions selected from four dimensions of collection number, release time, average browsing time, and recent browsing number.
Further, the position recommendation method further comprises the following steps:
and step S4, deducing and scoring positions in the position candidate set by using an online recommendation model obtained from the user historical behavior data and the position information, and taking a plurality of positions meeting the conditions as a position recommendation set.
Further, step S4 includes the steps of:
step S41, extracting job ID, job title, release time, job category, collection rate and average browsing time from the job data; extracting a user ID, a browsing position ID, a staying time and whether collecting and browsing time stamps from the historical behavior data of the user; integrating user ID, position title, release time, position category, collection rate, average browsing time and browsing timestamp according to the position ID to construct input data of an online recommendation model training sample, and regarding the retention time corresponding to the same position ID and whether to collect the input data as a label of the online recommendation model training sample;
step S42, training the TensorFlow neural network model by using the on-line recommendation model training sample set to obtain an on-line recommendation model;
and step S43, utilizing an online recommendation model to infer and score positions in the position candidate set, and taking a plurality of positions meeting the conditions as a position recommendation set.
Further, the label of the on-line recommended model training sample in step S41 is obtained by the following method:
carrying out normalization processing on the stay time corresponding to the same position ID to obtain L1 data, wherein when the position command L2 is collected to be 0.5, the position command L2 is not collected to be 0;
and when the sum of L1 and L2 is greater than 1, the label of the on-line recommended model training sample is 1, otherwise, the label of the on-line recommended model training sample is 0.
Compared with the prior art, the technical scheme of the invention has the following advantages:
1. according to the position recommending method, a Word vector training sample set is obtained by using a random walk algorithm and a Skip-gram frame, a Word2Vec model is used for obtaining a feature vector of each position, and then a position candidate set is obtained by using cosine similarity of the feature vector between the currently browsed position of a target user and all positions and is used as a position recommending set, so that the historical position browsing data of the target user is weakened, cold start is avoided, and the real-time performance and the accuracy of the position recommending method are improved;
2. context information of a user is mined in a random walk algorithm, so that a Word2Vec model is more accurate;
3. related positions with high heat from multiple dimensions are added in the position candidate set, so that the cold start problem is further solved, and the recommendation effect is improved;
4. and obtaining an online recommendation model according to the historical browsing position time and whether the users collect, and further predicting and scoring positions in the position candidate set by using the online recommendation model, so that the accuracy of the recommendation method is further improved.
Drawings
FIG. 1 is a flowchart illustrating a job recommendation method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating a job browsing sequence of a user according to an embodiment of the present invention;
fig. 3 and 4 are a role weighted directed relationship diagram and a new role sequence group generated by a random walk algorithm according to an embodiment of the present invention, respectively;
FIG. 5 is a schematic diagram of training the Word2Vec model based on the Skip-gram framework according to the present invention.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments.
As shown in fig. 1, a job recommendation method includes the following steps:
step S0, acquiring all position data, user historical behavior data and positions currently browsed by a target user, and acquiring position browsing sequences of all users according to the user historical behavior data;
acquiring service data from a human resource communication platform or a recruitment website database, wherein the service data comprises user data, position data, user behavior data and the like, and data cleaning and data standardization are required before the data is taken for use, so that position data, user historical behavior data and positions currently browsed by a target user, which are required by the invention, are formed, wherein the position data comprises position ID, position titles, release time, position categories, collection rates, average browsing time and other information; the historical behavior data of the user comprises a user ID, a browsing position ID, a staying time, a collection time stamp and a browsing time stamp; the currently browsed position of the target user refers to currently browsed position data of a user needing to recommend the position;
and then, according to the historical behavior data of all the users and the job browsing sequences of all the users, as shown in fig. 2, one user corresponds to one job browsing sequence according to the time sequence.
Step S1, generating a new job position sequence group by using the job position browsing sequences of all users and combining a random walk algorithm;
random walk is a very basic network-based algorithm. The essence of the method is that starting from a node, one adjacent point is randomly selected, starting from the adjacent point to the next node, the step is repeated, and all nodes passing through are recorded. Variations of this algorithm are widely used in the Google search and finance fields. By random walk we can get a path from each node, and this path represents the structural information of this node.
Step S2, training a Word2vec model by adopting a Skip-gram frame to generate a feature vector of each position;
inputting a one-hot into a Skip-gram framework, and outputting a multi-hot vector formed by aggregating a plurality of one-hot; word2vec, a group of related models for generating Word vectors, is used for training to reconstruct the linguistic Word text. The network is represented by words and the input words in adjacent positions are guessed, and the order of the words is unimportant under the assumption of the bag-of-words model in word2 vec. After training is finished, the Word2vec model can be used for mapping each Word to a vector and representing the relation between Word and Word, and the vector is a hidden layer of a neural network, so that the Word2vec model can be trained by adopting a Skip-gram framework to obtain a feature vector corresponding to each position;
it should be noted that, in steps S0, S1 and S2, the job browsing sequence, the new job sequence group and the feature vector of each job of each user are sequentially obtained according to all the historical behavior data of the user, and this part of the job can be completed before the target user clicks the current browsing job, for example, all the historical data of a time point or the previous day is reached, so that calculating the feature vector of each job in advance improves the calculation efficiency of providing the recommended job at the later stage.
Step S3, calculating cosine similarity of feature vectors between the currently browsed positions of the target user and all positions, and forming a position candidate set, where positions with similarity higher than a certain threshold are generally selected to form a position recommendation set.
According to the job recommendation method, a Word vector training sample set is obtained by using a random walk algorithm and a Skip-gram frame, a Word2Vec model is used for obtaining a feature vector of each job, and then a job candidate set is obtained by using cosine similarity of feature vectors between the currently browsed job of a target user and all jobs and serves as the job recommendation set, so that historical job browsing data of the target user are weakened, cold start is avoided, and instantaneity and accuracy of the job recommendation method are improved.
Specifically, step S1 specifically includes the following steps:
step S11, generating a role weighted directed relationship graph by using the role browsing sequence of each user, wherein the direction of directed edges among the roles is determined according to the time sequence, and the weight of the directed edges is determined according to the user context information corresponding to the downstream roles associated with the directed edges; in the embodiment, the context information comprises the life cycle of the position and the mode of entering the position browsing page;
as shown in fig. 3, when the user Ui browses job a and job B one after the other, a directed edge from a to B is generated. If multiple identical directed edges are subsequently generated, the weight of the directed edges is strengthened. Each time a directed edge is generated, the weight of the directed edge is increased by 1. Meanwhile, when the user enters the professional page from the job search result page and the collection list page, the weight is increased by 0.2 on the basis of the previous weight. When the job of the user browsing the page is in an active life cycle (within 60 days after the job is released), the weight is increased by 0.2 on the previous basis, and the increased value of the weight value can be set without limitation.
And step S12, selecting any position node in the position weighted directed relationship graph to randomly walk to form a new position sequence group, wherein the probability of walking from the current position node to the next position node is related to the directed edge weight of the current position node.
The probability of jumping from one node to another during random walks is: the ratio of the weight of the jump directed edge to the sum of the weights of all directed edges starting from the current position node, for example, in fig. 3, the probability of the E node jumping to the F node is: the weight of the EF side/(the weight of the EF side + the weight of the EC side). The new position sequence set generated according to fig. 3 is shown in fig. 4.
Specifically, step S2 specifically includes the following steps:
step S21, obtaining a word vector training sample set by adopting a Skip-gram frame, sequentially extracting position groups from each new position sequence by utilizing a sliding window, carrying out one-hot coding on a central position in the position groups to form a central position coding vector as input data of the word vector training sample, and carrying out Multi-hot coding on a plurality of edge positions in the position groups to form a mixed edge position coding vector as a label of the word vector training sample;
for example, if the new position sequence is EDGFSG and the sliding window is 3, then the positions that can be generated are EDG, DGF, GFS and FSG, and the training samples are (D | E, G), (G | D, F), (F | G, S), (S | F, G) before encoding. (| front is input data, | rear is output data). The input vector (input data of the word vector training sample) is an One-hot encoded vector converted from the role ID (if there are five roles in total, then the input vectors of the five roles after the One-hot encoding are 00001, 00010, 00100, 01000, 10000, and five in total, so how many roles in total the input vector is, how many dimensions the input vector is), and the output vector (label of the word vector training sample) is a Multi-hot encoded vector converted from a plurality of output roles. Taking (D | E, G) in fig. 3 as an example, if the one-hot encoding of the E-position is (0001) and the one-hot encoding of the G-position is (0100), then the Multi-hot encoding vector of the output vector (training label) is the aggregation of the E, G-position one-hot encoding (0101).
And step S22, training the Word2vec model by using the Word vector training sample set to obtain the feature vector of each position.
The structure of the Word2vec model is set as a three-layer neural network, the output layer adopts softmax as an activation function, the dimensionalities of the input layer and the output layer are V, V is the number of positions, the dimensionality of the hidden layer is N, N is smaller than the size of V, and N is also the dimensionality of a position feature vector. Assuming a total of 10000 positions in the current data, V is equal to 10000.
As shown in FIG. 5, the activation function of hidden layer of Word2vec model is linear in nature, which means that no processing is done, and we need to train this neural network, and use back propagation algorithm, which is essentially chain derivation. After the model is trained, the finally obtained weights are actually weights of the neural network, for example, an x one-hot encoder is input to be [0,0, …,1, …,0], only the weight corresponding to the position of 1 is activated in the weights from the input layer to the hidden layer, the number of the weights is consistent with the number of nodes of the hidden layer, and therefore the weights form a vector vx to represent x, and the vector vx can be used for uniquely representing x because the positions of 1 in the one-hot encoder of each word are different.
Sometimes, when the target user browses the latest position and does not exist in the historical data, the target position cannot be recommended, so that in some embodiments, the candidate set in the step S3 further includes a plurality of popular positions selected from four dimensions of collection quantity, release time, average browsing time and the number of recently browsed people, the positions in the shown position candidate set are not repeated, the problem of cold start is further solved, and the recommendation effect is improved; the Martian effect is also solved because the candidate set comprises popular positions and positions with high similarity to the current browsing positions of the user.
The position recommendation method further comprises the following steps:
and step S4, deducing and scoring positions in the position candidate set by using an online recommendation model obtained from the user historical behavior data and the position information, and taking a plurality of positions meeting the conditions as a position recommendation set. The input data of the online recommendation model is the interest degree of all users in positions at a certain time in the past, and the cold start problem is further solved.
Step S4 includes the following steps:
step S41, extracting job ID, job title, release time, job category, collection rate and average browsing time from the job data; extracting a user ID, a browsing position ID, a staying time and whether collecting and browsing time stamps from the historical behavior data of the user; integrating user ID, position title, release time, position category, collection rate, average browsing time and browsing timestamp according to the position ID to construct input data of an online recommendation model training sample, and regarding the retention time corresponding to the same position ID and whether to collect the input data as a label of the online recommendation model training sample;
extracting all job position data from the human resource communication platform or the recruitment website database, and extracting the following characteristics: job ID, job title, release time, job category, collection rate, average browsing duration. A job information form is built from the above data, as shown in table 1:
TABLE 1
Extracting all user behavior data from the human resource communication platform or the recruitment website database, and extracting the following characteristics: user ID, ID of positions browsed by the user, page dwell time, whether to collect or not, and a timestamp for entering browsing. A user behavior table is established by the data, and table 2 shows:
| user ID | Browsing job ID | Residence time | Whether to store or not | Browsing time stamp |
| 1 | 2 | 35 | 0 | 1112486027 |
| 1 | 3 | 30 | 1 | 1112484328 |
| 2 | 32 | 47 | 1 | 1112484092 |
| 2 | 4 | 33 | 0 | 1112488778 |
| 3 | 2 | 22 | 0 | 1112486453 |
| *** | *** | *** | *** | *** |
TABLE 2
The input data of the recommended model training sample is obtained by combining the table 1 and the table 2 according to the job ID, which is specifically shown in table 3:
TABLE 3
The label of the online recommended model training sample in step S41 is obtained by the following method:
carrying out normalization processing on the stay time corresponding to the same position ID to obtain L1 data, wherein when the position command L2 is collected to be 0.5, the position command L2 is not collected to be 0;
and when the sum of L1 and L2 is greater than 1, the label of the on-line recommended model training sample is 1, otherwise, the label of the on-line recommended model training sample is 0.
Step S42, training the TensorFlow neural network model by using the on-line recommendation model training sample set to obtain an on-line recommendation model;
the neural network recommendation model is constructed by TensorFlow, the recommendation model is divided into four layers, an input layer and two 128-dimensional full-connection hidden layers, and finally a final predicted value is generated by a sigmoid output neuron.
In step S43, an online recommendation model is used to infer and score positions in the position candidate set, and a plurality of eligible positions are used as a position recommendation set, where positions with scores higher than a certain threshold are generally combined into a position recommendation set.
The deduction scoring shows the interest degree of the current user in the candidate centralized positions at the moment, an online recommendation model is obtained according to the historical browsing position time and whether the current user collects the candidate centralized positions, the online recommendation model is used for further predicting and scoring the positions in the candidate centralized positions, and the accuracy of the recommendation method is further improved.
The above examples are merely illustrative of several embodiments of the present invention, which are described in more detail and detail but are not to be construed as limiting the scope of the present invention. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (8)
1. A job recommendation method is characterized by comprising the following steps:
step S0, acquiring all position data, user historical behavior data and positions currently browsed by a target user, and acquiring position browsing sequences of all users according to the user historical behavior data;
step S1, generating a new job position sequence group by using the job position browsing sequences of all users and combining a random walk algorithm;
step S2, training a Word2vec model by adopting a Skip-gram frame to generate a feature vector of each position;
step S3, calculating the cosine similarity of the feature vectors between the currently browsed position of the target user and all positions, and forming a position candidate set.
2. The method of claim 1, wherein the step S1 comprises the following steps:
step S11, generating a role weighted directed relationship graph by using the role browsing sequence of each user, wherein the direction of directed edges among the roles is determined according to the time sequence, and the weight of the directed edges is determined according to the user context information corresponding to the downstream roles associated with the directed edges;
and step S12, selecting any position node in the position weighted directed relationship graph to randomly walk to form a new position sequence group, wherein the probability of walking from the current position node to the next position node is related to the directed edge weight of the current position node.
3. The method according to claim 2, wherein the user context information in step S11 includes a life cycle of the job and a way to enter a browsing page of the job.
4. The method of claim 3, wherein the step S2 comprises the following steps:
step S21, obtaining a word vector training sample set by adopting a Skip-gram frame, sequentially extracting position groups from each new position sequence by utilizing a sliding window, carrying out one-hot coding on a central position in the position groups to form a central position coding vector as input data of the word vector training sample, and carrying out Multi-hot coding on a plurality of edge positions in the position groups to form a mixed edge position coding vector as a label of the word vector training sample;
and step S22, training the Word2vec model by using the Word vector training sample set to obtain the feature vector of each position.
5. The position recommendation method according to claim 4, wherein the candidate set in step S3 further comprises hot positions selected from four dimensions of collection number, release time, average browsing time and recent browsing number.
6. The position recommendation method according to claim 4 or 5, further comprising the steps of:
and step S4, deducing and scoring positions in the position candidate set by using an online recommendation model obtained from the user historical behavior data and the position information, and taking a plurality of positions meeting the conditions as a position recommendation set.
7. The position recommendation method according to claim 6, wherein the step S4 comprises the steps of:
step S41, extracting job ID, job title, release time, job category, collection rate and average browsing time from the job data; extracting a user ID, a browsing position ID, a staying time and whether collecting and browsing time stamps from the historical behavior data of the user; integrating user ID, position title, release time, position category, collection rate, average browsing time and browsing timestamp according to the position ID to construct input data of an online recommendation model training sample, and regarding the retention time corresponding to the same position ID and whether to collect the input data as a label of the online recommendation model training sample;
step S42, training the TensorFlow neural network model by using the on-line recommendation model training sample set to obtain an on-line recommendation model;
and step S43, utilizing an online recommendation model to infer and score positions in the position candidate set, and taking a plurality of positions meeting the conditions as a position recommendation set.
8. The position recommendation method according to claim 7, wherein the label of the online recommendation model training sample in step S41 is obtained by:
carrying out normalization processing on the stay time corresponding to the same position ID to obtain L1 data, wherein when the position command L2 is collected to be 0.5, the position command L2 is not collected to be 0;
and when the sum of L1 and L2 is greater than 1, the label of the on-line recommended model training sample is 1, otherwise, the label of the on-line recommended model training sample is 0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110661337.7A CN113362034A (en) | 2021-06-15 | 2021-06-15 | Position recommendation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110661337.7A CN113362034A (en) | 2021-06-15 | 2021-06-15 | Position recommendation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113362034A true CN113362034A (en) | 2021-09-07 |
Family
ID=77534356
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110661337.7A Pending CN113362034A (en) | 2021-06-15 | 2021-06-15 | Position recommendation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113362034A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116881575A (en) * | 2023-09-08 | 2023-10-13 | 腾讯科技（深圳）有限公司 | Content pushing method, device, computer equipment and storage medium |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104933239A (en) * | 2015-06-09 | 2015-09-23 | 江苏大学 | Hybrid model based personalized position information recommendation system and realization method therefor |
| CN105893641A (en) * | 2016-07-01 | 2016-08-24 | 中国传媒大学 | Job recommending method |
| CN106250502A (en) * | 2016-07-28 | 2016-12-21 | 五八同城信息技术有限公司 | Determine the method and device of similar position |
| CN107341233A (en) * | 2017-07-03 | 2017-11-10 | 北京拉勾科技有限公司 | A kind of position recommends method and computing device |
| CN109347606A (en) * | 2018-11-30 | 2019-02-15 | 维沃移动通信有限公司 | A kind of data processing method, device, network side equipment and terminal device |
| CN110516162A (en) * | 2019-09-02 | 2019-11-29 | 上海喜马拉雅科技有限公司 | A kind of information recommendation method, device, equipment and storage medium |
| CN110929162A (en) * | 2019-12-04 | 2020-03-27 | 腾讯科技（深圳）有限公司 | Recommendation method and device based on interest points, computer equipment and storage medium |
| CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
| CN111310040A (en) * | 2020-02-11 | 2020-06-19 | 腾讯科技（北京）有限公司 | Artificial intelligence based recommendation method and device, electronic equipment and storage medium |
| CN111460085A (en) * | 2020-04-17 | 2020-07-28 | 支付宝(杭州)信息技术有限公司 | Knowledge point recommendation method and device based on graph structure |
| CN111858631A (en) * | 2020-07-16 | 2020-10-30 | 深圳刺猬教育科技有限公司 | Third-party platform delivery system |
| CN112559852A (en) * | 2019-09-10 | 2021-03-26 | 北京搜狗科技发展有限公司 | Information recommendation method and device |
| CN112559868A (en) * | 2020-12-17 | 2021-03-26 | 广州博冠信息科技有限公司 | Information recall method and device, storage medium and electronic equipment |
-
2021
- 2021-06-15 CN CN202110661337.7A patent/CN113362034A/en active Pending
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104933239A (en) * | 2015-06-09 | 2015-09-23 | 江苏大学 | Hybrid model based personalized position information recommendation system and realization method therefor |
| CN105893641A (en) * | 2016-07-01 | 2016-08-24 | 中国传媒大学 | Job recommending method |
| CN106250502A (en) * | 2016-07-28 | 2016-12-21 | 五八同城信息技术有限公司 | Determine the method and device of similar position |
| CN107341233A (en) * | 2017-07-03 | 2017-11-10 | 北京拉勾科技有限公司 | A kind of position recommends method and computing device |
| CN109347606A (en) * | 2018-11-30 | 2019-02-15 | 维沃移动通信有限公司 | A kind of data processing method, device, network side equipment and terminal device |
| CN110516162A (en) * | 2019-09-02 | 2019-11-29 | 上海喜马拉雅科技有限公司 | A kind of information recommendation method, device, equipment and storage medium |
| CN112559852A (en) * | 2019-09-10 | 2021-03-26 | 北京搜狗科技发展有限公司 | Information recommendation method and device |
| CN110929162A (en) * | 2019-12-04 | 2020-03-27 | 腾讯科技（深圳）有限公司 | Recommendation method and device based on interest points, computer equipment and storage medium |
| CN111127142A (en) * | 2019-12-16 | 2020-05-08 | 东北大学秦皇岛分校 | Article recommendation method based on generalized neural attention |
| CN111310040A (en) * | 2020-02-11 | 2020-06-19 | 腾讯科技（北京）有限公司 | Artificial intelligence based recommendation method and device, electronic equipment and storage medium |
| CN111460085A (en) * | 2020-04-17 | 2020-07-28 | 支付宝(杭州)信息技术有限公司 | Knowledge point recommendation method and device based on graph structure |
| CN111858631A (en) * | 2020-07-16 | 2020-10-30 | 深圳刺猬教育科技有限公司 | Third-party platform delivery system |
| CN112559868A (en) * | 2020-12-17 | 2021-03-26 | 广州博冠信息科技有限公司 | Information recall method and device, storage medium and electronic equipment |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116881575A (en) * | 2023-09-08 | 2023-10-13 | 腾讯科技（深圳）有限公司 | Content pushing method, device, computer equipment and storage medium |
| CN116881575B (en) * | 2023-09-08 | 2023-11-21 | 腾讯科技（深圳）有限公司 | Content pushing method, device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107992531B (en) | News personalized intelligent recommendation method and system based on deep learning | |
| CN112214685A (en) | Knowledge graph-based personalized recommendation method | |
| CN110942337A (en) | Accurate tourism marketing method based on internet big data | |
| CN103533390B (en) | The method and system of television program recommendations are carried out based on social network information | |
| CN111259133A (en) | Personalized recommendation method integrating multiple information | |
| Volkovs et al. | New learning methods for supervised and unsupervised preference aggregation | |
| CN103970873A (en) | Music recommending method and system | |
| CN110674407A (en) | Hybrid recommendation method based on graph convolution neural network | |
| CN107357793A (en) | Information recommendation method and device | |
| CN111309936A (en) | Method for constructing portrait of movie user | |
| CN104484380A (en) | Personalized search method and personalized search device | |
| CN112199608A (en) | Social media rumor detection method based on network information propagation graph modeling | |
| CN111666496A (en) | Group recommendation method based on comment text | |
| CN107391582A (en) | The information recommendation method of user preference similarity is calculated based on context ontology tree | |
| CN110321421B (en) | Expert recommendation method for website knowledge community system and computer storage medium | |
| Chughtai et al. | Retracted: E-Learning recommender systems based on goal-based hybrid filtering | |
| CN105677828A (en) | User information processing method based on big data | |
| CN112215629B (en) | Multi-target advertisement generating system and method based on construction countermeasure sample | |
| CN114077661A (en) | Information processing apparatus, information processing method, and computer readable medium | |
| CN104572915A (en) | User event relevance calculation method based on content environment enhancement | |
| Chuang et al. | TPR: Text-aware preference ranking for recommender systems | |
| CN113362034A (en) | Position recommendation method | |
| CN105868422A (en) | Collaborative filtering recommendation method based on elastic dimensional feature vector optimized extraction | |
| CN113741759B (en) | Comment information display method and device, computer equipment and storage medium | |
| CN109948665B (en) | Human activity type classification method and system based on long-time and short-time memory neural network |
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 |