CN112948716B - Continuous interest point package recommendation method based on multi-head attention mechanism - Google Patents
Continuous interest point package recommendation method based on multi-head attention mechanism Download PDFInfo
- Publication number
- CN112948716B CN112948716B CN202110245049.3A CN202110245049A CN112948716B CN 112948716 B CN112948716 B CN 112948716B CN 202110245049 A CN202110245049 A CN 202110245049A CN 112948716 B CN112948716 B CN 112948716B
- Authority
- CN
- China
- Prior art keywords
- user
- sequence
- card punching
- model
- week
- 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
- 230000007246 mechanism Effects 0.000 title claims abstract description 32
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000004080 punching Methods 0.000 claims abstract description 65
- 238000012549 training Methods 0.000 claims abstract description 25
- 238000011156 evaluation Methods 0.000 claims abstract description 6
- 238000004806 packaging method and process Methods 0.000 claims abstract description 4
- 230000008859 change Effects 0.000 claims abstract description 3
- 239000013598 vector Substances 0.000 claims description 21
- 230000006399 behavior Effects 0.000 claims description 19
- 238000012360 testing method Methods 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 8
- 238000012795 verification Methods 0.000 claims description 6
- 238000004140 cleaning Methods 0.000 claims description 5
- 238000013135 deep learning Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 4
- 238000011161 development Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000002354 daily effect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 230000009182 swimming Effects 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
- G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
-
- 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
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Business, Economics & Management (AREA)
- Software Systems (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Primary Health Care (AREA)
- Marketing (AREA)
- Human Resources & Organizations (AREA)
- General Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a continuous interest point package recommendation method based on a multi-head attention mechanism, which comprises the following steps of: 1. obtaining user card punching history information, and forming a sequence by card punching data; 2. forming the card punching data into a sequence group; 3. pushing the sequence group to a model for training; obtaining interest point change information in the sequence through an encoder and a multi-head attention mechanism, inputting the obtained information into a decoder for fitting, wherein the decoder is symmetrical to the encoder in structure, and learning the optimal hyper-parameter adaptive to the region user in a pipeline searching mode; 4. deploying the training model in a plurality of servers, inputting an access sequence of a user in a region in a last week, and packaging and pushing a scheduling suggestion in a subsequent week to the user; 5. recording the sequence of the card punching places and the evaluation of the access points of interest of the user in the next week. The invention can provide the follow-up scheduling suggestion in one week for the user.
Description
Technical Field
The invention relates to the technical field of interest point recommendation, in particular to a continuous interest point package recommendation method based on a multi-head attention mechanism.
Background
With the rapid development of mobile social platforms, people share their own travel on various platforms every day. Such behavior allows us to obtain a large amount of user punch-card data. The card punching data is helpful for establishing a point of interest recommendation system, and the experience of a user in a city is improved.
An efficient recommendation system not only can bring flow and profit to the facilitator, but also can help the user plan the schedule for a subsequent period of time. The traditional recommendation interest point recommendation algorithm does not take the influence of recently accessed interest points into consideration, which results in a great promotion space for system performance. In the prior art, interest points which are not visited by a user are recommended through the preference of the user, but on one hand, the cold start and data sparsity processing capability of the traditional collaborative filtering idea is not good enough. On the other hand it only takes into account the influence of user preferences on the non-visited points of interest and not the influence of historical visited points of interest on the next point of interest visit. There is also a technique to find personalized recommendations by best matching the context, advertising, and audience. Although this technique takes many factors into consideration, it does not know the relationship between the point of interest and the time for point of interest recommendation. Since from a large scale of data, users have a significant periodicity in accessing points of interest. And the two technologies are used for performing cutting recommendation on the interest points of the user in the next days or even months. The recommended interest points of the users are single and do not meet the current requirements of the users, and the practicability is not strong in practical application and the method cannot adapt to rapid social development.
Therefore, how to provide continuous point of interest package recommendations for a user becomes a problem to be solved by those skilled in the art.
Disclosure of Invention
It is an object of the present invention to provide a method for continuous point of interest package recommendation based on a multi-head attention mechanism that overcomes some or all of the deficiencies of the prior art.
The invention discloses a continuous interest point package recommendation method based on a multi-head attention mechanism, which comprises the following steps of:
acquiring historical card punching information of users, and forming a sequence of card punching data of each user in a week;
step two, forming the card punching data of all the users in one week into a sequence group, wherein each sequence in the sequence group is a sequence card punching interest point of one user in one week, namely the sequence group comprises the card punching sequences of all the effective users in one week in the area;
step three, pushing the sequence group to a model for training; obtaining interest point change information in the sequence through an encoder and a multi-head attention mechanism, inputting the obtained information into a decoder for fitting, wherein the decoder is symmetrical to the encoder in structure, and learning the optimal hyper-parameter adaptive to the region user in a pipeline searching mode;
deploying the training model in a plurality of servers, dividing real-time user card punching data sequences into a plurality of parts according to a certain frequency in one week, transmitting the parts into different servers for prediction, recommending interest point sequences of a subsequent journey in one week for the user, and packaging prediction results according to the user ID and transmitting the prediction results back to the user side in real time;
and step five, recording a card punching place sequence and evaluation on access interest points of the user in the next week, obtaining user feedback data for optimizing the model, and updating the model periodically through user feedback and a new user card punching sequence.
Preferably, in the first step, the user card punching history information is acquired from a big data platform of the mobile application operator, and the user card punching history information comprises a user ID, a card punching place, a card punching time, card punching longitude and latitude and a user score.
Preferably, in the second step, after the sequence group is formed, the sequence group is cleaned, and sequences with a small number of times of card punching within one week are removed.
Preferably, in the second step, the cleaning is to remove new users whose card punching times are less than 5, remove interest points whose card punching times are less than 5, and remove sequences whose card punching times are less than 10 in one week.
Preferably, in step three, the model adopts a deep learning mode, and model training is carried out on a Tensorflow framework.
Preferably, in the third step, the sequence components are divided into a training set, a test set and a verification set, the training set is pushed to the model for training, the model is encoded in an encoder based on multi-head attention, and a vector is embedded in the user ID for splicing; inputting the obtained vector into a decoder with a structure similar to that of the encoder, and decoding to obtain a packet sequence interest point; then comparing the obtained packet sequence interest point with a target interest point sequence to reduce loss; and after the validity of the model is ensured through the verification set, inputting the test set for testing, retraining the model if an overfitting phenomenon exists, and temporarily deploying the model if the overfitting phenomenon does not exist.
Compared with the prior art, the invention has the following advantages:
(1) The continuous interest point package recommendation system adopting the multi-head attention mechanism can perform continuous interest point package recommendation on subsequent behaviors of the user according to the user card punching sequence in one week, comprehensively considers the user preference, the geographical influence of the interest points and the influence of the interest points in the long sequence, recommends the interest point sequence which is most likely to be visited by the user in one week to the user, and recommends all subsequent interest point sequences of the user in one week. By adopting a mode of an encoder and a decoder, an encoding and decoding infrastructure adopts a multi-head attention mechanism and a residual error network to model user sequence characteristics, and finally, a learned model is deployed on a line, and an interest point access sequence in a subsequent week of a user is recommended in time, so that the past prediction of interest point preference of the user after several days or even several months is changed, and an optimal sequence for accessing interest points in the subsequent week is recommended for the user.
(2) The invention deploys the model on the server, and intensively predicts the interest point sequences of the next stage of all the card punching users in a region so as to achieve the aim of assisting the user schedule arrangement. The model can be deployed on the mobile equipment, so that the more personalized fitting of the behavior characteristics of the user is facilitated, the model can be deployed on the mobile equipment, the size and the hyper-parameters of the model can be adjusted according to actual requirements for the user who frequently checks the card of the interest point, and the experience of the user on schedule behavior arrangement is improved.
(3) The method ensures that the model can be quickly fitted by using a position coding mode to process the sequence data in parallel and adopting a multi-head attention mechanism and a residual error network mode which are simple in structure and effective in coding, and the fitting speed is accelerated to help the model to process larger-scale data volume and user groups, so that the data are more quickly iterated on model deployment and updating. The invention aims to plan the schedule of the user in the following week, the length of the user history sequence is longer, the hidden information content is large, and therefore the model is used for improving the fitting speed so as to have excellent deployment capability.
(4) The invention adopts a symmetrical decoding mode, is embedded and spliced with the user before decoding, and fits the subsequent behaviors of the specific user within a week, and the mode ensures that the personalized package recommendation of the travel arrangement of the user within a week is realized, i.e. a series of subsequent personalized travel arrangement suggestions are provided for the user.
(5) The model is adjusted by collecting the card punching conditions of the user on the interest point sequence recommended by the model and the evaluation conditions of each interest point. Through the fine tuning operation, even the model is iterated to adapt to the rapid development of the society and the migration of the user preference.
Drawings
Fig. 1 is a flowchart of a continuous poi packet recommendation method in embodiment 1 based on a multi-head attention mechanism.
Detailed Description
For a further understanding of the invention, reference should be made to the following detailed description taken in conjunction with the accompanying drawings and examples. It is to be understood that the examples are illustrative of the invention and not limiting.
Example 1
As shown in fig. 1, the present embodiment provides a continuous interest point package recommendation method based on a multi-head attention mechanism, which includes the following steps.
The method comprises the steps of firstly, obtaining user card punching history information from a big data platform of a mobile application operator, wherein the user card punching history information comprises a user ID, a card punching place, card punching time, card punching longitude and latitude and user evaluation, and forming a sequence of card punching data of each user within one week.
Step two, forming the card punching data of all the users in one week into a sequence group, wherein each sequence in the sequence group is a card punching sequence interest point of one user in one week, namely the sequence group comprises the card punching sequences of the effective users in the region in one week; and cleaning the sequence group, and removing sequences with few card punching times in one week. Cleaning is to remove new user data with less than 5 times of card punching and eliminate interest points with less than 5 times of card punching.
And step three, dividing the sequence components into a training set, a test set and a verification set, pushing the training set to a model for training, modeling a long card punching sequence of the user within a week through a coding and decoding structure, a multi-head attention mechanism and a residual error network structure, coding the card punching behaviors of the user within a week through the structure, and compressing the user behavior semantics within a week. And the influence of the interest points in the sequence at different positions can be learned, which has practical comparative significance for subsequent recommendation. Finally, splicing the user ID embedded vector and the semantic vector to obtain preference conditions of different users in the region; the model adopts a deep learning mode and is trained on a Tensorflow framework. Directly dividing the subsequently collected user data into a training set and a testing set, taking the historical model as a training model, carrying out fine tuning operation, and updating the model; and then adjusting the model according to the feedback of the user and the reasonability of the influence generated by the model. And after the validity of the model is ensured through the verification set, inputting the test set for testing, retraining the model if an overfitting phenomenon exists, and deploying the model temporarily if the overfitting phenomenon does not exist.
The model training specifically comprises: and coding the card punching behavior of each user within one week, wherein a multi-head attention mechanism and a residual error network are adopted as a coding mode. The specific implementation comprises the following steps: and carrying out embedded position coding processing on interest points of each user in a week, and then learning a coding vector of a card punching behavior of the user in the week through a coding layer.
Wherein, the above formula i is the position of the interest point in all interest points, pos is the position occupied by the interest point in the sequence, d model For the length of the required code, the position code of each interest point in a sequence is obtained from the length, and the code is as follows:
the operation can carry out position coding on a segment of sequence, so that the operation can greatly reduce the operation amount by utilizing a multi-head attention mechanism to process sequence data in parallel, and is suitable for the conditions of large user amount and long user card punching sequence. The sequences within a week of the user are then encoded using a multi-head attention mechanism and a residual network to obtain the behavior characteristics of each user and the influence of each interest point at different positions. The learned influence vector is:
MultiHead(Q,K,V)＝Concat(head 1 ,...,head h )W O ；
head i ＝Attention(QW i Q ,KW i K ,VW i V )；
wherein Q is a query vector, K is a key vector, V is a value vector, and we adopt a self-attention mechanism, the three are generally equal, head i Representing the number of heads of the multi-head attention mechanism, splicing the learned vectors of each head in Concat, and then passing through a weight parameter W O Obtaining the vector coded by the multi-head attention mechanism, wherein the code is as follows:
deploying the training model in a plurality of servers, dividing real-time user card punching data sequences into a plurality of parts according to a certain frequency in one week, transmitting the parts into different servers for prediction, recommending interest point sequences of a subsequent journey in one week, and packaging prediction results according to user IDs and transmitting the prediction results to a user side in real time; the trained models are deployed in a plurality of servers, package-type recommendation is timely performed on subsequent behaviors of the users within a week according to a certain frequency, convenience is provided for the travel planning of the users, personalized recommendation is performed on specific users according to user IDs, and results are timely transmitted back to the user side.
And step five, recording the card punching location and the score of the interest points of the user in the next stage, obtaining user feedback data for optimizing the model, and regularly updating the model through user feedback and a new user card punching sequence.
The embodiment focuses on the fact that the model is deployed on a server, but the model is not limited to the server and includes mobile terminals of various brands. After the card punching sequence of the user within one week is collected, the sequence can be timely input into the product model for prediction, and the model is periodically updated according to user feedback.
In this embodiment, the updated model is redeployed in the server, and then the feedback condition of the user within a period of time after updating is observed, if the effect of the user feedback model is obviously reduced, the third step is repeated, and a suitable recommendation model is found.
The embodiment flexibly uses the position coding and the multi-head attention mechanism to calculate the influence of the interest point in the middle and long term, and the larger the influence is, the user has great influence on the behavior in the next week after the user cards the interest point. Such as: when a user climbs a mountain in a certain scenic spot, the user may not go to a place with vigorous exercise, such as a swimming pool, a gymnasium, etc., in the following week. Through the model, the influence degree of interest points in a region on the medium and long-term behaviors of the user can be obtained, and the ranking is sorted out to be used for optimizing a recommendation list by combining the evaluation of the user on the interest points to be punched, so that the recommendation system is more suitable for the production activities of the deployed region.
Daily life tells us that most users usually move in a week period, and the experimental method helps us to better learn the behavior characteristics of the users in the middle and long periods, and the characteristics can help us to analyze the big data condition of the activities of urban residents and help us to detect public sentiments.
The data used for the offline experiments was Instagram, which is one of the most popular cell phone based social networks. The Instagram data not only comprises the point-of-interest check-in information of the user, but also comprises the written text content of the user. The data set deleted fewer than 5 embedded points of interest and fewer than 10 commented users. After preprocessing, our dataset has 2216631 sign-ins, 13187 points of interest, 78233 users, and 337073 daily punch sequences were extracted from it for use as training models.
The online data is obtained by collecting active user data of the big data platform, and the data collected regularly is processed in a streaming mode, so that the online data can have the function of timely recommending appropriate interest points for each user in a user group within a week, and a recommendation system can be updated timely.
The method for recommending the packets of the interest points for the activity in the next week of the user is designed, the method has the capability of updating in time, the training time of the model is greatly reduced by processing parallel processing data through a multi-head attention mechanism, and the packet recommendation of the subsequent interest points is performed on the user by occupying little storage space, so that the recommendation of the access travel of the interest points in the next week is recommended to the user according to the historical data of the user. The specific advantage of the embodiment is that under the condition that huge users are available and the number of user card punching data sequences is very large, a recurrent neural network is not used for serially processing data, on the contrary, the position coding processing is carried out on the sequences within one week of the users, and the sequences are parallelly input into the model through a series of operations for processing sequence information, so that the calculation time and the prediction time are greatly reduced under the condition that the capability of a recommendation system is reserved. This can help us to predict the user's behavior more timely in such a middle-to-far period of a week in practical applications. The method has strong practical application capability, and can provide convenience for the routing of a large number of users in time without deploying the model in too many servers. The embodiment is mainly characterized in that follow-up behaviors of the user in one week are recommended, the method can adapt to tourists and residents living for a long time, an optimal visiting route can be designed for the tourists, and personalized suggestions can be provided for the recent journey arrangement of the residents living for a long time.
The present invention and its embodiments have been described above schematically, without limitation, and what is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Therefore, if the person skilled in the art receives the teaching, without departing from the spirit of the invention, the person skilled in the art shall not inventively design the similar structural modes and embodiments to the technical solution, but shall fall within the scope of the invention.
Claims (4)
1. A continuous interest point package recommendation method based on a multi-head attention mechanism is characterized by comprising the following steps: the method comprises the following steps:
acquiring historical card punching information of users, and forming a sequence of card punching data of each user in a week;
step two, forming the card punching data of all the users in one week into a sequence group, wherein each sequence in the sequence group is a sequence card punching interest point of one user in one week, namely the sequence group comprises the card punching sequences of all the effective users in one week in the area;
step three, pushing the sequence group to a model for training; obtaining interest point change information in a sequence through an encoder, a multi-head attention mechanism and a residual error network, inputting the obtained information into a decoder for fitting, wherein the decoder is symmetrical to the encoder in structure, and learning the optimal hyper-parameter adapting to a regional user in a pipeline searching mode;
in the third step, the sequence components are divided into a training set of 70%, a testing set of 20% and a verification set of 10%, the training set is pushed to a model for training, coding is carried out in a coder based on a multi-head attention mechanism and a residual error network, and the user ID is embedded into a vector and a semantic vector for splicing; inputting the obtained vector into a decoder with a structure similar to that of the encoder, and decoding to obtain a packet sequence interest point; then comparing the obtained packet sequence interest point with a target interest point sequence to reduce loss; after the validity of the model is ensured through the verification set, inputting the test set for testing, retraining the model if an overfitting phenomenon exists, and deploying the model temporarily if the overfitting phenomenon does not exist;
the model adopts a deep learning mode and is trained on a Tensorflow framework; directly dividing the subsequently collected user data into a training set and a testing set, taking the historical model as a training model, carrying out fine tuning operation, and updating the model; then, adjusting the model according to the feedback of the user and the reasonability of the influence force generated by the model;
the model training specifically comprises: coding the card punching behavior of each user within one week, wherein the coding mode adopts a multi-head attention mechanism and a residual error network, and the specific implementation comprises the following steps: embedding position coding processing is carried out on interest points of each user within one week, and coding vectors of card punching behaviors of the user within one week are learned through a coding layer;
wherein, the above formula i is the position of the interest point in all interest points, pos is the position occupied by the interest point in the sequence, d model Obtaining the position code of each interest point in a sequence for the length needing coding;
coding sequences of the users in a week by using a multi-head attention mechanism and a residual error network form to obtain behavior characteristics of each user and influence of each interest point at different positions; the learned influence vector is:
MultiHead(Q,K,V)＝Concat(head 1 ,...,head h )W O ；
head i ＝Attention(QW i Q ,KW i K ,VW i V )；
wherein Q is query vector, K is key vector, V is value vector, self-attention mechanism is adopted, the three are equal, head i Representing the number of heads of the multi-head attention mechanism, splicing the learned vectors of each head in Concat, and then passing through a weight parameter W O Obtaining a vector coded by a multi-head attention mechanism;
deploying the training model in a plurality of servers, dividing real-time user card punching data sequences into a plurality of parts according to a certain frequency in one week, transmitting the parts into different servers for prediction, recommending interest point sequences of a subsequent journey in one week, and packaging prediction results according to user IDs and transmitting the prediction results to a user side in real time;
and fifthly, recording the card punching place sequence and the evaluation of the access interest points of the user in the next week, obtaining user feedback data for optimizing the model, and updating the model regularly through user feedback and the new user card punching sequence.
2. The method of claim 1, wherein the method for recommending the consecutive interest point packets based on the multi-head attention mechanism comprises: in the first step, obtaining user card punching history information from a mobile application operator big data platform, wherein the user card punching history information comprises a user ID, a card punching place, card punching time, card punching longitude and latitude and user score.
3. The method of claim 2, wherein the method for recommending the consecutive interest point packets based on the multi-head attention mechanism comprises: and step two, after the sequence group is formed, cleaning the sequence group, and removing sequences with few card punching times in one week.
4. The method of claim 3, wherein the method comprises: in the second step, cleaning is to remove new users with the card punching times less than 5, remove interest points which are punched with the card for less than 5 times, and remove sequences with the card punching times less than 10 times in one week.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110245049.3A CN112948716B (en) | 2021-03-05 | 2021-03-05 | Continuous interest point package recommendation method based on multi-head attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110245049.3A CN112948716B (en) | 2021-03-05 | 2021-03-05 | Continuous interest point package recommendation method based on multi-head attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112948716A CN112948716A (en) | 2021-06-11 |
| CN112948716B true CN112948716B (en) | 2023-02-28 |
Family
ID=76247815
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110245049.3A Active CN112948716B (en) | 2021-03-05 | 2021-03-05 | Continuous interest point package recommendation method based on multi-head attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112948716B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110083770A (en) * | 2019-04-29 | 2019-08-02 | 苏州市职业大学 | It is a kind of based on feature level deeper from the sequence of recommendation method of attention network |
| CN110929164A (en) * | 2019-12-09 | 2020-03-27 | 北京交通大学 | Interest point recommendation method based on user dynamic preference and attention mechanism |
| CN111429234A (en) * | 2020-04-16 | 2020-07-17 | 电子科技大学中山学院 | Deep learning-based commodity sequence recommendation method |
| CN111932336A (en) * | 2020-07-17 | 2020-11-13 | 重庆邮电大学 | Commodity list recommendation method based on long-term and short-term interest preference |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11138392B2 (en) * | 2018-07-26 | 2021-10-05 | Google Llc | Machine translation using neural network models |
| US20200342968A1 (en) * | 2019-04-24 | 2020-10-29 | GE Precision Healthcare LLC | Visualization of medical device event processing |
-
2021
- 2021-03-05 CN CN202110245049.3A patent/CN112948716B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110083770A (en) * | 2019-04-29 | 2019-08-02 | 苏州市职业大学 | It is a kind of based on feature level deeper from the sequence of recommendation method of attention network |
| CN110929164A (en) * | 2019-12-09 | 2020-03-27 | 北京交通大学 | Interest point recommendation method based on user dynamic preference and attention mechanism |
| CN111429234A (en) * | 2020-04-16 | 2020-07-17 | 电子科技大学中山学院 | Deep learning-based commodity sequence recommendation method |
| CN111932336A (en) * | 2020-07-17 | 2020-11-13 | 重庆邮电大学 | Commodity list recommendation method based on long-term and short-term interest preference |
Non-Patent Citations (2)
| Title |
|---|
| POI Recommendation Based on First-Order Collaborative Filtering Tree;Jinghua Zhu 等;《2019 15th International Conference on Mobile Ad-Hoc and Sensor Networks (MSN)》;20200416;第265-266页 * |
| 基于动态和静态偏好的兴趣点推荐算法;杨丽 等;《计算机应用》;20200916;第41卷(第2期);第398-399页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112948716A (en) | 2021-06-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Learning to estimate the travel time | |
| Alam et al. | Domain adaptation with adversarial training and graph embeddings | |
| CN109034448B (en) | Trajectory prediction method based on vehicle trajectory semantic analysis and deep belief network | |
| CN110929164A (en) | Interest point recommendation method based on user dynamic preference and attention mechanism | |
| CN112199608B (en) | Social media rumor detection method based on network information propagation graph modeling | |
| CN107357899B (en) | Short text sentiment analysis method based on sum-product network depth automatic encoder | |
| CN111369299A (en) | Method, device and equipment for identification and computer readable storage medium | |
| CN112085048A (en) | Vehicle frequent track identification method and device based on LSTM self-encoding | |
| CN114780866B (en) | Personalized intelligent recommendation method based on spatio-temporal context interest learning model | |
| Kawamura et al. | A hybrid approach for optimal feature subset selection with evolutionary algorithms | |
| CN112948716B (en) | Continuous interest point package recommendation method based on multi-head attention mechanism | |
| Xu et al. | A taxi dispatch system based on prediction of demand and destination | |
| CN112948709B (en) | Continuous interest point real-time recommendation method driven by influence perception | |
| CN116824851A (en) | Path-based urban expressway corridor traffic jam tracing method | |
| Chen et al. | Improving the forecasting and classification of extreme events in imbalanced time series through block resampling in the joint predictor-forecast space | |
| Jiang et al. | Adversarial neural trip recommendation | |
| CN116295409A (en) | Route processing method, route processing device, computer readable medium and electronic equipment | |
| CN113741759B (en) | Comment information display method and device, computer equipment and storage medium | |
| CN116541755A (en) | Financial behavior pattern analysis and prediction method based on time sequence diagram representation learning | |
| CN113821985B (en) | Traffic state prediction method and device and electronic equipment | |
| CN105653686A (en) | Domain name network address activeness statistics method and system | |
| CN115393098A (en) | Financing product information recommendation method and device | |
| CN111274359B (en) | Query recommendation method and system based on improved VHRED and reinforcement learning | |
| CN113032688A (en) | Method for predicting access position of social network user at given future time | |
| CN112632404A (en) | Multi-granularity self-attention-based next interest point recommendation method |
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 |