WO2022088661A1 - 基于注意力机制的群体旅游路线推荐方法 - Google Patents

基于注意力机制的群体旅游路线推荐方法 Download PDF

Info

Publication number
WO2022088661A1
WO2022088661A1 PCT/CN2021/095764 CN2021095764W WO2022088661A1 WO 2022088661 A1 WO2022088661 A1 WO 2022088661A1 CN 2021095764 W CN2021095764 W CN 2021095764W WO 2022088661 A1 WO2022088661 A1 WO 2022088661A1
Authority
WO
WIPO (PCT)
Prior art keywords
group
scenic spot
representation vector
model
feature representation
Prior art date
Application number
PCT/CN2021/095764
Other languages
English (en)
French (fr)
Inventor
曹杰
陈蕾
申冬琴
王有权
陶海成
朱桂祥
Original Assignee
云境商务智能研究院南京有限公司
南京理工大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 云境商务智能研究院南京有限公司, 南京理工大学 filed Critical 云境商务智能研究院南京有限公司
Publication of WO2022088661A1 publication Critical patent/WO2022088661A1/zh

Links

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
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • 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
    • G06Q50/10—Services
    • G06Q50/14—Travel agencies

Definitions

  • the invention relates to the field of tourism applications, in particular to a group tourism route recommendation method based on an attention mechanism.
  • tourism has become the main way of entertainment for people to spend time with their family and friends. Economically speaking, tourism is one of the largest and fastest growing service industries in the world, providing a large number of employment opportunities and making a significant contribution to the world economy. Since traveling in groups is a common phenomenon in tourism, it is of great significance to study group travel route recommendation. In real life, group members often have different ages, genders, occupations and personalities. How to integrate the different interests of group members to obtain common interests and preferences of group members, so that the recommended results can meet the needs of most group members as much as possible, is group tourism. Key issues that need to be addressed in route recommendation methods. At the same time, it is also necessary to consider the spatial and temporal constraints of the route (departure, destination and travel route time, etc.) when formulating a group tourism route.
  • the present invention provides a group travel route recommendation method based on an attention mechanism, which considers the common preference of the group and the time and space constraints of the travel route, and recommends a satisfactory travel route for group tourists.
  • Step 1 Obtain the mobile traffic time between different scenic spots and the average visiting time of each scenic spot;
  • Step 2 Establish a group preference model according to the user's historical travel route and the group's historical travel route;
  • Step 3 According to the group preference model and the scenic spot heat model, construct the scenic spot utility function model;
  • Step 4 According to the scenic spot utility function model, the group travel route constraint obtains the travel route that best matches the group's common preference as the final recommended travel route to the group.
  • step 2 specifically includes the following sub-steps:
  • Step 2.1 Perform one-hot encoding for each scenic spot, user, and group to obtain a one-hot vector, and then go through the embedding layer to obtain the feature representation vector of the scenic spot, user, and group;
  • Step 2.2 Use the self-attention mechanism to learn the interaction between group members, and obtain the updated group member feature representation vector.
  • the specific calculation process is as follows:
  • F l represents the similarity between members of the group
  • softmax( ) is the normalized exponential function
  • Z l represents the aggregation of the user feature representation vector corresponding to the group into a matrix
  • D u represents the dimension of the user feature representation vector
  • X l the feature representation vector representing the updated group member
  • Step 2.3 Use the attention mechanism to calculate the weight of each user in the group, and based on the weight of each user, perform preference fusion on the users in the group to obtain an updated group feature representation vector.
  • the specific calculation process is as follows:
  • g l represents the updated group feature representation vector
  • x i is the updated feature representation vector of group member i obtained in the step 2.2
  • q l is the initial group feature representation vector
  • ⁇ i is the weight of group member i
  • Step 2.4 Use the attention mechanism to calculate the weight of the word in the description text of the scenic spot, and obtain the updated representation vector of the scenic spot description based on the weight of each word.
  • the specific calculation process is as follows:
  • t j represents the feature representation vector of the scenic spot description text
  • ⁇ t is the weight of the t-th word in the scenic spot description text
  • x′ t is the word representation after the word embedding layer
  • Step 2.5 Use the attention mechanism to calculate the weight of the scenic spot description, the scenic spot category, and the scenic spot ID in the scenic spot representation. Based on the weight, the updated scenic spot representation vector is obtained.
  • the specific calculation process is as follows:
  • P j represents the updated scenic spot feature representation vector
  • ⁇ j , ⁇ ′ j , ⁇ ′′ j are the weights of the scenic spot text description, scenic spot category, and scenic spot ID in the scenic spot representation
  • t j is the feature representation vector of the description text
  • t′ j is the feature representation vector embedded by category
  • t′′ j is the feature representation vector embedded by ID
  • Step 2.6 Create a group-attraction interaction pair according to the group's historical travel route, create a user-attraction interaction pair according to the user's historical travel route, and calculate the interaction between the user feature representation vector, the updated group feature representation vector, and the updated scenic spot feature representation vector , using a multi-objective learning mechanism to build a user preference model and a group preference model.
  • step 3 includes the following sub-steps:
  • Step 3.1 According to the number of visits to the scenic spot, build a hot spot model
  • Step 3.2 Weighted group preference model and scenic spot popularity model to construct a scenic spot utility model, specifically:
  • Q(g l , p j ) represents the group preference model
  • Pop(p j ) represents the scenic spot popularity model
  • ⁇ , 1- ⁇ represent the group preference model and scenic spot popularity model, respectively The weight of the model.
  • step 4 includes the following sub-steps:
  • Step 4.1 Determine candidate attractions
  • Step 4.2 Randomly select scenic spots from the candidate scenic spots, and obtain the group tourism route S through the local search algorithm
  • Step 4.3 Randomly delete intermediate scenic spots from the group tourism route S, and obtain the group tourism route S' through the local search algorithm described in step 4.2;
  • Step 4.5 Go to step 4.3 until the preset number of iterations is reached;
  • Step 4.6 Obtain the travel route S that best matches the group's common preference as the final recommended travel route to the group.
  • the method for recommending a group travel route based on the attention mechanism provided by the present invention learns the interaction relationship between group members through the self-attention mechanism, and obtains the feature representation vector of the group members.
  • the attention mechanism is used to calculate the weight of each user in the group, and based on the weight of each user, the preference fusion of the users in the group is performed to obtain the feature representation vector of the group.
  • a user preference model and a group preference model are constructed according to the scenic spot feature representation vector, the group feature representation vector, and the user feature representation vector.
  • the scenic spot utility function model is constructed.
  • the group travel route constraint obtains the travel route that best matches the group's common preference as the final recommended travel route to the group.
  • the present invention provides a group travel route recommendation method based on an attention mechanism. Since the present invention adopts an attention mechanism, the interaction relationship between group members can be learned, and weights can be automatically assigned to each group member to obtain more accurate information. group preference. The invention maximizes the group preference under the space-time constraints of the group travel route, and recommends a satisfactory travel route for the group members.
  • FIG. 1 is a flowchart of an embodiment of a group travel route recommendation method based on an attention mechanism of the present invention.
  • Fig. 2 is the user preference aggregation model diagram based on attention mechanism in the present invention
  • FIG. 3 is a model diagram of a scenic spot feature representation based on an attention mechanism in the present invention
  • Fig. 4 is the interactive learning model diagram based on multi-task learning in the present invention.
  • Fig. 5 is the flow chart of generating group travel route in the present invention.
  • FIG. 6 is a flow chart of the local search algorithm in the present invention.
  • FIG. 1 is a flowchart of an embodiment of a group travel route recommendation method based on an attention mechanism of the present invention, and the method includes:
  • Step 1 Obtain the moving traffic time between different scenic spots and the average visiting time of each scenic spot, specifically, including the following sub-steps:
  • Step 1.1 Obtain the mobile traffic time between different scenic spots
  • the obtained two scenic spots are divided into three ways: walking, cycling and driving according to the length of the distance. If the distance between the two scenic spots is less than 2 kilometers, the default is the walking time, and the distance is between 2 kilometers and 5 kilometers. The time is the riding time by default, and the driving time is the default when it is greater than 5 kilometers. The time is obtained through the Baidu Map API to obtain the mobile traffic time between two different scenic spots.
  • the travel time between the scenic spots p i and the scenic spots p j is denoted as T( pi , p j ) .
  • Step 1.2 Obtain the average visit time of each attraction
  • the average visit time of each scenic spot is counted.
  • the attraction visit time is the time difference between arriving at the attraction and leaving the attraction.
  • the average visiting time of the scenic spot p i is recorded as D( pi ).
  • Step 2 Establish a group preference model according to the user's historical travel route and the group's historical travel route. Specifically, the following sub-steps are included:
  • Step 2.1 Perform one-hot encoding for each scenic spot, user, and group to obtain a one-hot vector, and then go through the embedding layer to obtain the feature representation vector of the scenic spot.
  • the scenic spots, users, and group data are numbered to obtain the one-hot encoding (one-hot) of the scenic spots p j , the user ui , and the group q l to obtain the one-hot vector, and then go through the embedding layer to obtain
  • the feature of the scenic spot represents a vector p j
  • the feature of the user represents a vector ui
  • the feature of the group represents a vector q l
  • the value of the position corresponding to the scenic spot number is set to 1, and all other positions are set to 0. For example, there are 5 attractions and the one-hot vector for the third is [0,0,1,0,0].
  • Step 2.2 Adopt the self-attention mechanism to learn the interaction relationship between the group members, and obtain the updated feature representation vector of the group members.
  • the self-attention score matrix F l is created by the scale dot product, F l represents the similarity between group members, and reflects the interaction information between group members.
  • the calculation method is as follows:
  • D u represents the dimension of the user feature representation vector
  • softmax( ) is a normalized exponential function.
  • vi represents the ith element in V
  • the softmax function of this element is the value of where e( ⁇ ) is an exponential function.
  • the initial population feature representation Z l is multiplied by F l , and the population feature representation vector is updated by incorporating information from population members.
  • the updated feature representation X l of the group members is obtained by means of residual connection, where x i represents the updated feature representation vector of group member i, which is calculated as follows:
  • Step 2.3 Using the attention mechanism, calculate the weight of each user in the group, and based on the weight of each user, perform preference fusion on the users in the group to obtain an updated group feature representation vector.
  • the attention mechanism is used to learn the weight of each user in the group, and the updated group feature representation vector is obtained.
  • g l represents the updated group feature representation vector
  • x i is the updated feature representation vector of group member i obtained in the step 2.2
  • q l is the initial group feature representation vector
  • ⁇ i is the weight of group member i.
  • the weight of the user in the group is calculated, specifically:
  • h t , V t , and W t represent the weight parameters of the attention network
  • b t is the bias vector.
  • the normalized exponential function is used to calculate the user weight.
  • Step 2.4 Use the attention mechanism to calculate the weight of the word in the description text of the scenic spot, and obtain the updated representation vector of the scenic spot description based on the weight of each word.
  • the attention mechanism is used to learn the weight of each word in the description text of the scenic spot, and the updated representation vector of the description text of the scenic spot is obtained.
  • the calculation method is as follows:
  • t j represents the feature representation vector of the scenic spot description text
  • ⁇ t is the weight of the t-th word in the scenic spot description text.
  • the weight of the word in the description text is calculated, specifically:
  • V s , W s represent the weight parameters of the attention network
  • b s is the bias vector
  • the normalized exponential function is used to calculate the word weight
  • x′ t is the word feature representation vector after the word embedding layer
  • q l is the initial population feature representation vector.
  • Step 2.5 Using the attention mechanism, calculate the weight of the description of the scenic spot, the category of the scenic spot, and the ID of the scenic spot in the scenic spot representation, and obtain the updated feature representation vector of the scenic spot based on the weight.
  • the attention mechanism is used to learn the weights of scenic spot description, scenic spot category, and scenic spot ID in the feature representation of scenic spots.
  • the calculation method is as follows:
  • P j represents the updated scenic spot feature representation vector
  • ⁇ j , ⁇ ′ j , ⁇ ′′ j are the weights of the scenic spot text description, scenic spot category, and scenic spot ID in the scenic spot representation
  • t j is the feature representation vector of the description text
  • t′ j is the feature representation vector embedded in the category
  • t′′ j is the feature representation vector embedded in the ID.
  • the weight of the scenic spot text description, the scenic spot category, and the scenic spot ID in the scenic spot feature representation is calculated, specifically:
  • V r , W r represent the weight parameters of the attention network
  • br is the bias vector
  • q l is the initial group feature representation vector
  • Step 2.6 Create a group-attraction interaction pair according to the group's historical travel route, create a user-attraction interaction pair according to the user's historical travel route, and calculate the interaction between the user feature representation vector, the updated group feature representation vector, and the updated scenic spot feature representation vector , using a multi-objective learning mechanism to build a user preference model and a group preference model.
  • a group-attraction interaction pair (g l , p j ) is created according to the group's historical travel route
  • a user-scenic spot interaction pair (u i , p j ) is created according to the user's historical travel route.
  • the preference model Q(g l , p j ) of the group to the scenic spot is output, and the calculation method is as follows:
  • W 1 and b 1 represent the weight matrix and bias vector of the first layer of the multilayer perceptron, respectively, and W 2 and b 2 respectively represent the weight matrix and bias vector of the second layer of the multilayer perceptron,
  • Sigmoid( ⁇ ) is the logistic regression function
  • ReLU( ⁇ ) is a linear rectification function
  • ReLU(x) max(0, x).
  • the user's preference model R(u i , p j ) for the scenic spot is output, and the calculation method is as follows:
  • Step 3 According to the group preference model and the scenic spot heat model, construct the scenic spot utility function model, specifically, including the following sub-steps:
  • Step 3.1 According to the number of visits to the scenic spot, build a hot spot model.
  • the scenic spot heat model Pop(p j ) is constructed.
  • Step 3.2 Weighted group preference model and attraction popularity model to construct attraction utility model.
  • Q(g l , p j ) represents the group preference model
  • Pop(p j ) represents the attraction model of the scenic spot
  • ⁇ , 1- ⁇ represent the group preference model, respectively and the weight of the attraction heat model.
  • Step 4 the group travel route constraint obtains the travel route that best matches the group's common preference as the final recommended travel route for the group. Specifically, the following sub-steps are included:
  • Step 4.1 Identify candidate attractions.
  • the candidate scenic spots refer to the recommended visiting scenic spots in the set area, which can be expressed as the points of interest of tourism resources in a certain city.
  • the points of interest of tourism resources in Nanjing are: Sun Yat-sen Mausoleum, Confucius Temple, Presidential Palace, Qixia Mountain, etc.
  • Step 4.2 Randomly select scenic spots from the candidate scenic spots, and obtain the group tourism route S through local search.
  • the group travel route S 0 ⁇ p 1 , p N >, where p 1 is the predefined departure place of the group travel route, and p N is the predefined destination of the group travel route.
  • the scenic spots are randomly selected from the candidate scenic spots, and the group travel route S is obtained through local search.
  • the specific calculation method of the totalCost is as follows
  • T(p k , p k+1 ) is the moving traffic time between the scenic spot p k calculated in the step 1.1 and the scenic spot p k+1
  • D(p k ) is the scenic spot calculated in the step 1.2.
  • Average access time for p k is the moving traffic time between the scenic spot p k calculated in the step 1.1 and the scenic spot p k+1
  • D(p k ) is the scenic spot calculated in the step 1.2.
  • the scenic spot pi is randomly selected from the group travel route S, and it is ensured that the selected scenic spot pi is not the destination p N .
  • Spots p j are randomly selected from the candidate spots, and it is ensured that the selected spots p j are not included in the group travel itinerary S.
  • Determine whether the updated travel route time obtained after inserting the scenic spot p j exceeds the time budget, and the calculation process of the total time of the updated travel route is as follows:
  • updateCost totalCost+T(pi , p j ) +D(p j )
  • updateCost is the total time of the updated travel route
  • modt time budget If updateCost ⁇ buget, that is, the total time of the updated travel route does not exceed the preset time budget, then the scenic spot p j can be inserted after the scenic spot p i in the travel route S, and the value of totalCost is updated to updateCost.
  • maxLoop times where maxLoop is a predefined number of iterations. After the iteration, the group travel route S is obtained.
  • Step 4.3 Randomly delete intermediate scenic spots from the group travel route S, and obtain the group travel route S' through the local search algorithm described in step 4.2.
  • Step 4.5 Go to step 4.3 until the preset number of iterations is reached.
  • Step 4.6 Obtain the travel route S that best matches the group's common preference as the final recommended travel route to the group.
  • the experimental data came from six public datasets, and the first four datasets were extracted from Flickr, which included tourists' travel route trajectories in Italian cities. Because no explicit group information is included.
  • Similar groups are obtained by aggregating users who have visited similar attraction categories in historical records, while random groups are not subject to the above restrictions.
  • the latter two datasets come from the location-based social networking platform Gowalla, where users can check in to visited attractions and build social relationships with friends.
  • COM Classical model of group recommendation based on probabilistic graphical model. COM obtains the group's overall preference for attractions by fusing group member preferences with different weights.
  • PIT The classical model of group recommendation based on topic model. PIT selects the most influential members of the group to represent the group.
  • PersTour-AVE Set the same weight value for group members to calculate the group's overall preference for scenic spots.
  • PersTour-LM adopt the least pain strategy, and calculate the group's overall preference for scenic spots according to the lowest weight value of group members.
  • PersTour-MS The maximum satisfaction strategy is adopted, and the group's overall preference for scenic spots is calculated according to the highest weight value of group members.
  • the word embedding layer adopts the word2vec model, and the word embedding dimension is set to 100. Embedding dimensions for groups, group members, and attractions are set to 128.
  • the optimal batch-size selects the optimal value of 64 from [16, 32, 64, 128, 256].
  • the learning rate is selected from [0.0001, 0.0005, 0.001, 0.005], according to the experimental results of the validation set, FLOR-Rand and ROME-Simi are set to 0.0005, while the other four datasets are set to 0.001 for better results.
  • Table 1 shows the comparison of the recommended recall rate, accuracy rate and F value between the group travel route recommendation method of the present invention and other recommendation methods.
  • PersTour-based methods ie, PersTour-AVG, PersTour-LM, and PersTour-MS.
  • PersTour-AVG outperforms PersTour-LM in the FLOR-Rand dataset, but not PersTour-LM in the LA-Social dataset.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于注意力机制的群体旅游路线推荐方法，包括如下步骤：获取不同景点之间的移动交通时间，每个景点的平均访问时间；根据用户历史旅游路线、群体历史旅游路线建立群体偏好模型；根据群体偏好模型，景点热度模型，构建景点效用函数模型；根据景点效用函数模型，群体旅游路线约束得到群体共同偏好最匹配的旅游路线作为最终向群体推荐的旅游路线。由于一种基于注意力机制的群体旅游路线推荐方法，采用了注意力机制，可以学习群体成员之间的交互关系，并为每个群体成员自动分配权重，获得更准确的群体偏好，在群体旅游路线的时空约束条件下，最大化群体偏好，为群体成员推荐满意的旅游路线。

Description

基于注意力机制的群体旅游路线推荐方法 技术领域
本发明涉及旅游应用领域，尤其涉及一种基于注意力机制的群体旅游路线推荐方法。
背景技术
随着全球经济的飞速发展，旅游已成为人们与家人、朋友休闲度假的主要娱乐方式。从经济上说，旅游业是世界上最大，增长最快的服务行业之一，提供了大量的就业机会，为世界经济做出了重大贡献。由于结伴出行是旅游中的常见的现象，研究群体旅游路线推荐具有重要的意义。在实际生活中，群体成员往往具有不同年龄、性别、职业和个性，如何融合群体成员不同的兴趣以获取群体成员的共同兴趣偏好，使推荐结果尽可能满足大部分群体成员的需求，是群体旅游路线推荐方法需要解决的关键问题。同时，制定群体旅游路线还需要考虑路线的时空约束条件(出发地、目的地和旅游路线时间等)。
尽管传统的旅游推荐系统可以为群体游客提供标准的旅游套餐，但这些套餐通常只包含热门景点，而无法满足群体游客的个性化需求。现有的主流方法试图将群体旅游路线推荐问题转化成定向问题的变体形式，并同时考虑群体偏好和时空约束，然后利用优化算法推荐群体路线。具体来说，群体偏好通过预定义的偏好聚合策略进行融合(例如，平均法、最小痛苦法、最大满意度法)。然而，在实际生活中，由于群体是由年龄、性别、职业、个性等差异巨大的成员组成的异构集合。个体的差异影响群体旅游路线的选择。而传统融合策略是数据独立的，缺乏动态调整群体权重的灵活性，同时，也忽略了群体成员之间的交互关系。
为了获取群体的偏好信息，大量的研究将景点划分成不同类别(例如，河流、山川、人文等)，并从群体的历史旅游路线记录中提取群体对景点类别的偏好。然而，由于一些景点类别不存在于历史记录中，造成基于景点类别的方法效果不佳。为了解决这一问题，一些研究者尝试融合景点类别信息和景点文本信息，预测更细粒度的群体偏好。但是，他们忽略了相同的景点文本信息可能对不同群体具有不同的影响。例如，群体1可能会更关注景点描述中的“适合亲子游”，而群体2可能被“海岛”一词吸引。因此，建模景点信息对不同群体偏好的影响有利于捕获更细粒度的群体偏好。
发明内容
发明目的：针对现有技术的缺点，本发明提供一种基于注意力机制的群体旅游路线推荐方法，考虑群体共同偏好和旅游路线的时空约束，为群体游客推荐满意的旅游路线。
技术方案：一种基于注意力机制的群体旅游路线推荐方法，其特征在于，包括如下步骤：
步骤1.获取不同景点之间的移动交通时间，每个景点的平均访问时间；
步骤2.根据用户历史旅游路线、群体历史旅游路线建立群体偏好模型；
步骤3.根据群体偏好模型，景点热度模型，构建景点效用函数模型；
步骤4.根据景点效用函数模型，群体旅游路线约束得到群体共同偏好最匹配的旅游路线作为最终向群体推荐的旅游路线。
进一步地，所述步骤2具体包括如下子步骤：
步骤2.1对每个景点、用户、群体进行独热编码得到独热向量，再经过嵌入层，得到景点、用户、群体的特征表示向量；
步骤2.2.采用自注意力机制，学习群体成员之间的交互关系，得到更新的群体成员特征表示向量，具体计算过程为：
Figure PCTCN2021095764-appb-000001
X l＝Z lF l+Z l
其中，F l表示群体成员之间的相似性，softmax(·)为归一化指数函数，Z l表示群体对应的用户特征表示向量聚合成矩阵，D u表示用户特征表示向量的维度，X l表示更新的群体成员的特征表示向量；
步骤2.3.采用注意机制，计算每个用户在群组内的权重，基于每个用户的权重，对群体中的用户进行偏好融合，得到更新的群体特征表示向量，具体计算过程为：
Figure PCTCN2021095764-appb-000002
其中，g l表示更新的群体特征表示向量，x i为所述步骤2.2中得到的更新的群体成员i的特征表示向量，q l为初始群体特征表示向量，α i为群体成员i的权重；
步骤2.4.采用注意力机制，计算单词在景点描述文本中的权重，基于每个单词的权重，得到更新后的景点描述表示向量，具体计算过程为：
Figure PCTCN2021095764-appb-000003
其中，t j表示景点描述文本的特征表示向量，β t为第t个单词在景点描述文本中的权重，x′ t为经过单词嵌入层后的单词表示；
步骤2.5.采用注意力机制，计算景点描述、景点类别、景点ID在景点表示中的权重，基于权重，得到更新后的景点表示向量，具体计算过程为：
p j＝γ jt j+γ′ jt′ j+γ″ jt″ j
其中，P j表示更新的景点特征表示向量，γ j，γ′ j，γ″ j分别为景点文本描述、景点类别、景点ID在景点表示中的权重，t j为描述文本的特征表示向量，t′ j为经过类别嵌入的特征表示向量，t″ j为经过ID嵌入的特征表示向量；
步骤2.6.根据群体历史旅游路线创建群体-景点交互对，根据用户历史旅游路线创建用户-景点交互对，计算用户特征表示向量、更新的群体特征表示向量以及更新的景点特征表示向量之间的交互，采用多目标学习机制，构建用户偏好模型和群体偏好模型。
进一步地，所述步骤3包括如下子步骤：
步骤3.1.根据景点被访问次数，构建景点热度模型；
步骤3.2.加权群体偏好模型和景点热度模型，构建景点效用模型，具体为：
U(g l，p j)＝ηQ(g l，p j)+(1-η)Pop(p j)建立所述景点的效用函数模型。
其中，0＜η＜1，Q(g l，p j)表示所述群体偏好模型，Pop(p j)表示所述景点热度模型，η，1-η分别表示所述群体偏好模型和景点热度模型所占权重。
进一步地，所述步骤4包括如下子步骤：
步骤4.1.确定候选景点；
步骤4.2.从候选景点中随机挑选景点，通过局部搜索算法得到群体旅游路线S；
步骤4.3.从群体旅游路线S中随机删除中间景点，通过步骤4.2所述局部搜索算法得到群体旅游路线S′；
步骤4.4.利用景点效用函数U(·)，分别计算群体旅游路线S和S′中包含的景点总效用值，如果U(S′)＞U(S)，那么以一定概率接受群体旅游路线S′，并设置S＝S′；
步骤4.5.跳转至步骤4.3，直至达到预先设置的迭代的次数；
步骤4.6.得到群体共同偏好最匹配的旅游路线S作为最终向群体推荐的旅游路线。
有益效果：本发明提供的基于注意力机制的群体旅游路线推荐方法，通过自注意力机制，学习群体成员之间的交互关系，得到群体成员的特征表示向量。采用注意机制，计算每个用户在群组内的权重，基于每个用户的权重，对群体中的用户进行偏好融合，得到群体的特征表示向量。通过多目标学习机制根据所述景点特征表示向量，群体特征表示向量，用户特征表示向量，构建用户偏好模型和群体偏好模型。通过所述群体偏好模型和景点热度模型，构建景点效用函数模型。根据景点效用函数模型，群体旅游路线约束得到群体共同偏好最匹配的旅游路线作为最终向群体推荐的旅游路线。本发明提供了一种基于注意力机制的群体旅游路线推荐方法，由于本发明采用了注意力机制，可以学习群体成员之间的交互关系，并为每个群体成员自动分配权重，获得更准确的群体偏好。本发明在群体旅游路线的时空约束条件下，最大化群体偏好，为群体成员推荐满意的旅游路线。
附图说明
图1为本发明基于注意力力机制的群体旅游路线推荐方法的一个实施例的流程图。
图2为本发明中基于注意力机制的用户偏好聚合模型图；
图3为本发明中基于注意力机制的景点特征表示模型图；
图4为本发明中基于多任务学习的交互学习模型图；
图5为本发明中生成群体旅游路线流程图；
图6为本发明中局部搜索算法流程图。
具体实施方式
下面结合附图对本发明作进一步说明：
图1为本发明基于注意力力机制的群体旅游路线推荐方法的一个实施例的流程图，则该方法包括：
步骤1.获取不同景点之间的移动交通时间，每个景点的平均访问时间，具体地，包括以下子步骤：
步骤1.1获取不同景点之间的移动交通时间；
获取到的两个景点之间根据距离的长短分为步行、骑行与车行三种方式，如果两个景点之间的距离小于2公里时默认为步行时间、距离在2公里到5公里之间时默认为骑行时间，大于5公里时默认为车行时间。时间通过百度地图API获取到不同的两个景点之间的移动交通时间。景点p i与景点p j之间的移动交通时间记做T(p i，p j)。
步骤1.2.获取每个景点的平均访问时间；
根据用户历史旅游路线，群体历史旅游路线，统计每个景点的平均访问时间。所述景点访问时间为到达景点和离开景点之间的时间差。景点p i的平均访问时间记做D(p i)。
步骤2.根据用户历史旅游路线、群体历史旅游路线建立群体偏好模型，具体地，包括以下子步骤：
步骤2.1.对每个景点、用户、群体进行独热编码(one-hot)，得到独热向量，再经过嵌入层(embedding layer)，得到景点的特征表示向量。
具体地，对景点、用户、群体数据进行编号，得到景点p j，用户u i，群体q l的独热编码(one-hot)，得到独热向量，再经过嵌入层(embedding layer)，得到景点的特征表示向量p j，用户的特征表示向量u i，群体特征表示向量q l。所述独热编码是对景点编号对应位置的值设为1，其他位置全为0。例如，有5个景点，第三个景点的独热向量为[0,0,1,0,0]。而嵌入层可以表示为p＝W Tx,其中W为嵌入层的权重向量，x为独热向量，p为特征表示向量。
步骤2.2.采用自注意力机制，学习群体成员之间的交互关系，得到更新的群体成员的特征表示向量。
如图2所示，将所述步骤2.1中群体q l对应的用户特征表示向量聚合成矩阵Z l，Z l＝[z 1，z 2，...，z Nl]，其中，N l表示所述群体中的群体成员个数，z i对应于第i个用户的特征表示向量。通过尺度点积创建自注意力得分矩阵F l，F l表示群体成员之间的相似性，反映群体成员之间的交互信息，计算方式如下：
Figure PCTCN2021095764-appb-000004
其中，D u表示用户特征表示向量的维度，softmax(·)为归一化指数函数，假设有一个N维数组V，v i表示V中的第i个元素，那么这个元素的softmax函数值为
Figure PCTCN2021095764-appb-000005
其中e(·)为指数函数。
进一步地，将F l乘以初始群体特征表示Z l，通过合并来自群体成员的信息来更新群体特征表示向量。此外，考虑到原始成员的偏好不能通过相邻成员的融合来表示，所以通过残差连接的方式，得到更新的群体成员的特征表示X l，
Figure PCTCN2021095764-appb-000006
其中x i代表更新的群体成员i的特征表示向量，计算方式如下：
X l＝Z lF l+Z l
步骤2.3.采用注意机制，计算每个用户在群体内的权重，基于每个用户的权重，对群体中的用户进行偏好融合，得到更新的群体特征表示向量。
如图2所示，采用注意力机制，学习每个用户在群体内的权重，得到更新的群体特征表示向量。群体成员权重越高，表示群体成员在群决策中重要程度越高。计算方式如下：
Figure PCTCN2021095764-appb-000007
g l表示更新的群体特征表示向量，x i为所述步骤2.2中得到的更新的群体成员i的特征表示向量，q l为初始群体特征表示向量，α i为群体成员i的权重。
所述步骤2.3中，计算用户在群体内的权重，具体为：
Figure PCTCN2021095764-appb-000008
Figure PCTCN2021095764-appb-000009
其中，h t,V t,W t表示注意力网络的权值参数，b t为偏置向量，使用归一化指数函数，计算用户权重。
步骤2.4.采用注意力机制，计算单词在景点描述文本中的权重，基于每个单词的权重，得到更新后的景点描述表示向量。
如图3所示，采用注意力机制，学习每个单词在景点描述文本中的权重，得到更新的景点描述文本表示向量。权重越高，表示单词在景点文本中的重要程度越高。计算方 式如下：
Figure PCTCN2021095764-appb-000010
其中，t j表示景点描述文本的特征表示向量，β t为第t个单词在景点描述文本中的权重。
所述步骤2.4中，计算单词在描述文本中的权重，具体为：
Figure PCTCN2021095764-appb-000011
Figure PCTCN2021095764-appb-000012
其中，V s,W s表示注意力网络的权值参数，b s为偏置向量，使用归一化指数函数，计算单词权重，x′ t为经过单词嵌入层后的单词特征表示向量，q l为初始群体特征表示向量。
步骤2.5.采用注意力机制，计算景点描述、景点类别、景点ID在景点表示中的权重，基于权重，得到更新后的景点特征表示向量。
如图3所示，采用注意力机制，学习景点描述、景点类别、景点ID在景点特征表示中的权重。权重越高，表示特征的重要程度越高，计算方式如下：
p j＝γ jt j+γ′ jt′ j+γ″ jt″ j
P j表示更新的景点特征表示向量，γ j，γ′ j，γ″ j分别为景点文本描述、景点类别、景点ID在景点表示中的权重，t j为描述文本的特征表示向量，t′ j为经过类别嵌入的特征表示向量，t″ j为经过ID嵌入的特征表示向量。
所述步骤2.5中，计算景点文本描述、景点类别、景点ID在景点特征表示中的权重，具体为：
Figure PCTCN2021095764-appb-000013
Figure PCTCN2021095764-appb-000014
其中，V r,W r表示注意力网络的权值参数，b r为偏置向量，使用归一化指数函数，计算景点文本描述、景点类别、景点ID的权重，q l为初始群体特征表示向量。
步骤2.6.根据群体历史旅游路线创建群体-景点交互对，根据用户历史旅游路线创建用户-景点交互对，计算用户特征表示向量、更新的群体特征表示向量以及更新的景点特征表示向量之间的交互，采用多目标学习机制，构建用户偏好模型和群体偏好模型。
如图4所示，根据群体历史旅游路线创建群体-景点交互对(g l，p j)，根据用户历史旅游路线创建用户-景点交互对(u i，p j)。通过乘积操作，捕捉群体与景点之间的交互g l⊙p j和用户与景点之间的交互u i⊙p j，其中⊙为哈达玛(Hadamard)积，为矩阵中对应元素相乘。
根据所述群体与景点之间的交互g l⊙p j，通过多层感知机，输出群体对景点的偏好模型Q(g l，p j)，计算方式如下：
Q(g l，p j)＝Sigmoid(W 2ReLU(W 1(g l⊙p j)+b 1)+b 2)
其中，W 1和b 1分别代表多层感知机第一层的权重矩阵和偏置向量，W 2和b 2分别代表多层感知机第二层的权重矩阵和偏置向量，其中Sigmoid(·)为逻辑回归函数，
Figure PCTCN2021095764-appb-000015
ReLU(·)为线性整流函数，ReLU(x)＝max(0，x)。
同理，根据所述用户与景点之间的交互u i⊙p j，通过多层感知机，输出用户对景点的偏好模型R(u i，p j)，计算方式如下：
R(u i，p j)＝Sigmoid(W 2ReLU(W 1(u i⊙p j)+b 1)+b 2)
步骤3.根据群体偏好模型，景点热度模型，构建景点效用函数模型，具体地，包括以下子步骤：
步骤3.1.根据景点被访问次数，构建景点热度模型。
根据景点在群体历史旅游路线和用户历史旅游路线中被访问的总次数，构建景点热度模型Pop(p j)。
步骤3.2.加权群体偏好模型和景点热度模型，构建景点效用模型。
采用U(g l，p j)＝ηQ(g l，p j)+(1-η)Pop(p j)建立所述景点的效用函数模型。
其中，0＜η＜1，Q(g l，p j)表示所述群体偏好模型，Pop(p j)表示所述景点所述景点热度模型，η，1-η分别表示所述群体偏好模型和景点热度模型所占权重。
步骤4.如图5所示，根据景点效用函数模型，群体旅游路线约束得到群体共同偏好最匹配的旅游路线作为最终向群体推荐的旅游路线，具体地，包括以下子步骤：
步骤4.1.确定候选景点。
本实施例中，候选景点指设定区域内的推荐访问景点，可以表示为某城市内旅游资源的兴趣点，如在南京市的旅游资源的兴趣点为：中山陵、夫子庙、总统府、栖霞山等。
步骤4.2.从候选景点中随机挑选景点，通过局部搜索得到群体旅游路线S。
具体地，初始化群体旅游路线S 0＝<p 1，p N>，其中p 1为群体旅游路线预定义的出发地，p N为群体旅游路线预定义的目的地。随机从候选景点中选择景点，通过局部搜索得到群体旅游路线S。
所述局部搜索算法如图6所示，输入群体旅游路线S＝<p 1，...，p N>，旅游路线已花费时间totalCost和预先设定的群体旅游路线时间预算buget，其中，所述totalCost具体计算方式如下
Figure PCTCN2021095764-appb-000016
其中，T(p k，p k+1)为所述步骤1.1中计算的景点p k和景点p k+1之间的移动交通时间，D(p k)为所述步骤1.2中计算的景点p k的平均访问时间。
进一步地，从群体旅游路线S中随机选择景点p i，并且确保选择的景点p i不是目的地p N。从候选景点中随机选择景点p j，并且确保选择的景点p j不包含在群体旅游路线S中。判断插入景点p j后得到的更新的旅游路线时间是否超过时间预算，更新的旅游路线总时间计算过程如下：
updateCost＝totalCost+T(p i，p j)+D(p j)
其中，updateCost为更新的旅游路线总时间，buget为预先设定的时间预算。如果updateCost≤buget，即更新的旅游路线总时间不超过预先设定的时间预算，则可以在旅游路线S中的景点p i后插入景点p j，并将totalCost值更新为updateCost。重复以上操 作maxLoop次，maxLoop为预定义的迭代次数。迭代结束后，得到群体旅游路线S。
步骤4.3.从群体旅游路线S中随机删除中间景点，通过步骤4.2所述局部搜索算法得到群体旅游路线S′。
步骤4.4.利用景点效用函数U(·)，分别计算群体旅游路线S和S′中包含的景点总效用值，如果U(S′)＞U(S)，那么以一定概率接受群体旅游路线S′，并设置S＝S′。
利用步骤3.2所述景点效用模型U(·)，分别计算群体旅游路线S和S′中包含的景点总效用值，如果U(S′)＞U(S)，那么以概率acceptPro接受群体旅游路线S′，并设置S＝S′。
步骤4.5.跳转至步骤4.3，直至达到预先设置的迭代的次数。
步骤4.6.得到群体共同偏好最匹配的旅游路线S作为最终向群体推荐的旅游路线。
为了验证算法的效果，进行了如下实验，实验数据来自六个公开数据集，前四个数据集抽取自Flickr，包含游客在意大利城市中的旅游路线轨迹。由于没有包含明确的群体信息。我们按照文献中的方法，提取了两种隐式群体：相似群体(FLOR-Simi、ROME-Simi)和随机群体(FLOR-Rand、ROME-Rand)。通过聚合历史记录中访问了相似景点类别的用户得到相似群体，而随机群体不受上述限制。后两个数据集来自基于位置的社交网络平台Gowalla，用户可以在Gowalla上签到访问过的景点，并且可以与朋友建立社交关系。对于Gowalla数据集，我们只抽取来自奥斯汀和洛杉矶两个城市的签到数据，并聚合在社交网络上有关联的并且旅游路线具有相同出发地和目的地的用户作为社交群体(AUS-Social、LA-Social)。这种分组方法模拟了实际生活中朋友结伴旅游的应用场景。为了验证项目推荐的性能，我们随机选择20％、10％的群体旅游路线(包括对应的用户-景点交互和群体-景点交互信息)作为测试集和验证集，剩下的数据作为训练集。S g、S r分别表示真实的和算法推荐的群体旅游路线，而p g(p r)对应于旅游路线S g(S r)中间访问景点。算法推荐结果的准确性由准确率、召回率、F值来衡量。准确率、召回率、F值的定义如下：
准确率(Precision)
Figure PCTCN2021095764-appb-000017
召回率(Recall)
Figure PCTCN2021095764-appb-000018
F值(F-measure)
Figure PCTCN2021095764-appb-000019
同时，采用了如下模型作为对比实验：
1)AGREE:最早提出的采用注意力机制学习偏好融合策略用以解决群体推荐问题的模型。
2)COM：基于概率图模型的群体推荐经典模型。COM通过融合具有不同权重的群体成员偏好获取群体对景点的总体偏好。
3)PIT：基于主题模型的群体推荐经典模型。PIT挑选群体中影响力较大的成员代表群体。
4)PersTour-AVE：为群体成员的设置相同的权重值计算群体对景点的整体偏好。
5)PersTour-LM：采用最小痛苦策略，根据群体成员最低的权重值计算群体对景点的整体偏好。
6)PersTour-MS：采用最大满意度策略，根据群体成员最高的权重值计算群体对景点的整体偏好。
对于所述对比方法，我们在最佳参数设置情况下进行实验。对于我们提出的方法(AMT-IRE)，单词嵌入层采用word2vec模型，并且将单词嵌入维度设置为100。群体、群体成员和景点的嵌入维度设置为128。最优批尺寸(batch-size)从[16，32，64，128，256]中挑选出最优值64。从[0.0001、0.0005、0.001、0.005]中选择学习率，根据验证集的实验结果，FLOR-Rand和ROME-Simi设置为0.0005，而其他四个数据集则设置为0.001效果更好。负采样率ρ从0到10的整数中进行选择，结果表明，对于FLOR-Rand数据集，ρ＝3效果更好，而对于其他数据集，ρ＝4效果更好。在所有梯度下降中选择Adam算法，并且使用dropout策略避免过拟合，dropout比率设置为λ＝0.2。
表1为本发明群体旅游路线推荐方法与其他推荐方法在推荐的召回率、准确率和F值方面的实验结果比较。
Figure PCTCN2021095764-appb-000020
根据表1中的实验结果，我们可以发现提出的AMT-IRE算法在六个数据集的召回率上优于所有对比方法，同时与最优对比方法相比，平均召回率提高了11.1％。在36次比较实例中，AMT-IRE 35次在准确率和F值方面优于对比方法。其次，AMT-IRE和AGREE方法在所有情况下均优于COM和PIT方法，显示出神经网络在群体、群体成员、景点交互建模中的优越性。同时AMT-IRE方法优于PersTour-AVG，PersTour-LM和PersTourMS方法，这些方法之间的主要区别在于后三个对比方法既没有充分考虑群体成员的个体的差异，也没有利用群体成员之间的相互影响，而AMT-IRE以动态方式为不同的群体成员分配了不同的权重。此外，基于PersTour的方法(即PersTour-AVG，PersTour-LM和PersTour-MS)中没有明显的优劣之分。例如，在FLOR-Rand数据集中，PersTour-AVG优于PersTour-LM，但在LA-Social数据集中却不及PersTour-LM。这些结果表明使用简单的群体融合策略不足以拟合群体决策的复杂性和动态性。
以上所述仅是本发明的优选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本发明的保护范围。

Claims (4)

  1. 一种基于注意力机制的群体旅游路线推荐方法，其特征在于，包括如下步骤：
    步骤1.获取不同景点之间的移动交通时间，每个景点的平均访问时间；
    步骤2.根据用户历史旅游路线、群体历史旅游路线建立群体偏好模型；
    步骤3.根据群体偏好模型，景点热度模型，构建景点效用函数模型；
    步骤4.根据景点效用函数模型，群体旅游路线约束得到群体共同偏好最匹配的旅游路线作为最终向群体推荐的旅游路线。
  2. 根据权利要求1所述的基于注意力机制的群体旅游路线推荐方法，其特征在于，所述步骤2具体包括如下子步骤：
    步骤2.1对每个景点、用户、群体进行独热编码得到独热向量，再经过嵌入层，得到景点、用户、群体的特征表示向量；
    步骤2.2.采用自注意力机制，学习群体成员之间的交互关系，得到更新的群体成员特征表示向量，具体计算过程为：
    Figure PCTCN2021095764-appb-100001
    X l＝Z lF l+Z l
    其中，F l表示群体成员之间的相似性，softmax(·)为归一化指数函数，Z l表示群体对应的用户特征表示向量聚合成的矩阵，D u表示用户特征表示向量的维度，X l表示更新的群体成员的特征表示向量；
    步骤2.3.采用注意机制，计算每个用户在群组内的权重，基于每个用户的权重，对群体中的用户进行偏好融合，得到更新的群体特征表示向量，具体计算过程为：
    Figure PCTCN2021095764-appb-100002
    其中，g l表示更新的群体特征表示向量，x i为所述步骤2.2中得到的更新的群体成员i的特征表示向量，q l为初始群体特征表示向量，α i为群体成员i的权重；
    步骤2.4.采用注意力机制，计算单词在景点描述文本中的权重，基于每个单词的权重，得到更新后的景点描述表示向量，具体计算过程为：
    Figure PCTCN2021095764-appb-100003
    其中，t j表示景点描述文本的特征表示向量，β t为第t个单词在景点描述文本中的权 重，x′ t为经过单词嵌入层后的单词表示向量；
    步骤2.5.采用注意力机制，计算景点描述、景点类别、景点ID在景点表示中的权重，基于权重，得到更新后的景点表示向量，具体计算过程为：
    p j＝γ jt j+γ′ jt′ j+γ″ jt″ j
    其中，P j表示更新的景点特征表示向量，γ j，γ′ j，γ″ j分别为景点文本描述、景点类别、景点ID在景点表示向量中的权重，t j为描述文本的特征表示向量，t′ j为经过类别嵌入的特征表示向量，t″ j为经过ID嵌入的特征表示向量；
    步骤2.6.根据群体历史旅游路线创建群体-景点交互对，根据用户历史旅游路线创建用户-景点交互对，计算用户特征表示向量、更新的群体特征表示向量以及更新的景点特征表示向量之间的交互，采用多目标学习机制，构建用户偏好模型和群体偏好模型。
  3. 根据权利要求1所述的基于注意力机制的群体旅游路线推荐方法，其特征在于，所述步骤3包括如下子步骤：
    步骤3.1.根据景点被访问次数，构建景点热度模型；
    步骤3.2.加权群体偏好模型和景点热度模型，构建景点效用模型，具体为：
    U(g l，p j)＝ηQ(g l，p j)+(1-η)Pop(p j)建立所述景点的效用函数模型。
    其中，0＜η＜1，Q(g l，p j)表示所述群体偏好模型，Pop(p j)表示所述景点热度模型，η，1-η分别表示所述群体偏好模型和景点热度模型所占权重。
  4. 根据权利要求1所述的基于注意力机制的群体旅游路线推荐方法，其特征在于，所述步骤4包括如下子步骤：
    步骤4.1.确定候选景点；
    步骤4.2.从候选景点中随机挑选景点，通过局部搜索算法得到群体旅游路线S；
    步骤4.3.从群体旅游路线S中随机删除中间景点，通过步骤4.2所述局部搜索算法得到群体旅游路线S′；
    步骤4.4.利用景点效用函数U(·)，分别计算群体旅游路线S和S′中包含的景点总效用值，如果U(S′)＞U(S)，那么以一定概率接受群体旅游路线S′，并设置S＝S′；
    步骤4.5.跳转至步骤4.3，直至达到预先设置的迭代的次数；
    步骤4.6.得到群体共同偏好最匹配的旅游路线S作为最终向群体推荐的旅游路线。
PCT/CN2021/095764 2020-11-01 2021-05-25 基于注意力机制的群体旅游路线推荐方法 WO2022088661A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011199320.6A CN112149010B (zh) 2020-11-01 2020-11-01 基于注意力机制的群体旅游路线推荐方法
CN202011199320.6 2020-11-01

Publications (1)

Publication Number Publication Date
WO2022088661A1 true WO2022088661A1 (zh) 2022-05-05

Family

ID=73955142

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/095764 WO2022088661A1 (zh) 2020-11-01 2021-05-25 基于注意力机制的群体旅游路线推荐方法

Country Status (2)

Country Link
CN (1) CN112149010B (zh)
WO (1) WO2022088661A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114781625A (zh) * 2022-06-13 2022-07-22 支付宝(杭州)信息技术有限公司 一种网络模型训练、推送内容确定方法及装置
CN115096324A (zh) * 2022-05-25 2022-09-23 科大讯飞股份有限公司 路线推荐方法及相关装置、车机、车辆和存储介质
CN115344777A (zh) * 2022-07-15 2022-11-15 重庆工商大学 一种基于大数据的多人定制化旅游的方法及装置
CN116304362A (zh) * 2022-12-21 2023-06-23 山东省人工智能研究院 一种基于相关性注意力的群组推荐方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112149010B (zh) * 2020-11-01 2024-05-24 云境商务智能研究院南京有限公司 基于注意力机制的群体旅游路线推荐方法
CN112784153B (zh) * 2020-12-31 2022-09-20 山西大学 融合属性特征注意力与异质类型信息的旅游景点推荐方法
CN113051461B (zh) * 2021-03-31 2023-04-28 郑州旅游职业学院 基于人工智能和云平台的旅游热度分析方法与系统
CN113343127B (zh) * 2021-04-25 2023-03-21 武汉理工大学 一种旅游路线推荐方法、系统、服务器及存储介质
CN113343119B (zh) * 2021-05-24 2023-05-23 武汉大学 一种基于层次注意力机制的群组推荐方法
CN113515697A (zh) * 2021-05-27 2021-10-19 武汉理工大学 一种基于用户多意图的群组动态旅游路线推荐方法及系统

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103870604A (zh) * 2014-04-04 2014-06-18 北京航空航天大学 旅游推荐方法和装置
CN103984721A (zh) * 2014-05-13 2014-08-13 中国矿业大学 基于交互式进化优化的图书个性化搜索方法
CN103995837A (zh) * 2014-04-25 2014-08-20 西北工业大学 一种基于群体足迹的个性化旅游路线规划方法
US20170023370A1 (en) * 2015-07-24 2017-01-26 International Business Machines Corporation Journey Planning
CN107194503A (zh) * 2017-05-08 2017-09-22 厦门大学 一种基于启发式算法的景区个性化线路设计方法
CN108534791A (zh) * 2017-03-01 2018-09-14 马云阁 全息旅游路径动态规划方法、装置、电子设备及系统
CN108829852A (zh) * 2018-06-21 2018-11-16 桂林电子科技大学 一种个性化旅游路线推荐方法
CN111177572A (zh) * 2020-01-16 2020-05-19 西北大学 一种基于用户动态兴趣的个性化旅游路线推荐方法
US20200334732A1 (en) * 2019-04-17 2020-10-22 Ford Global Technologies, Llc Adaptive vehicle feature matching system
CN112149010A (zh) * 2020-11-01 2020-12-29 云境商务智能研究院南京有限公司 基于注意力机制的群体旅游路线推荐方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463730A (zh) * 2014-12-29 2015-03-25 广州神马移动信息科技有限公司 基于旅游目标地域来挖掘旅游路线的方法和设备
CN105069717A (zh) * 2015-07-29 2015-11-18 陕西师范大学 一种基于游客信任度的个性化旅游线路推荐方法
CN110175192A (zh) * 2019-04-10 2019-08-27 南京灵澈信息技术有限公司 一种基于主题序列模式的旅游产品推荐方法
CN110276008A (zh) * 2019-06-13 2019-09-24 南京邮电大学 一种基于用户旅行决策过程的景点推荐方法及装置
CN111143680B (zh) * 2019-12-27 2024-03-26 上海携程商务有限公司 路线的推荐方法、系统、电子设备及计算机存储介质
CN111241306B (zh) * 2020-01-21 2022-04-29 浙江大学 一种基于知识图谱和指针网络的路径规划方法

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103870604A (zh) * 2014-04-04 2014-06-18 北京航空航天大学 旅游推荐方法和装置
CN103995837A (zh) * 2014-04-25 2014-08-20 西北工业大学 一种基于群体足迹的个性化旅游路线规划方法
CN103984721A (zh) * 2014-05-13 2014-08-13 中国矿业大学 基于交互式进化优化的图书个性化搜索方法
US20170023370A1 (en) * 2015-07-24 2017-01-26 International Business Machines Corporation Journey Planning
CN108534791A (zh) * 2017-03-01 2018-09-14 马云阁 全息旅游路径动态规划方法、装置、电子设备及系统
CN107194503A (zh) * 2017-05-08 2017-09-22 厦门大学 一种基于启发式算法的景区个性化线路设计方法
CN108829852A (zh) * 2018-06-21 2018-11-16 桂林电子科技大学 一种个性化旅游路线推荐方法
US20200334732A1 (en) * 2019-04-17 2020-10-22 Ford Global Technologies, Llc Adaptive vehicle feature matching system
CN111177572A (zh) * 2020-01-16 2020-05-19 西北大学 一种基于用户动态兴趣的个性化旅游路线推荐方法
CN112149010A (zh) * 2020-11-01 2020-12-29 云境商务智能研究院南京有限公司 基于注意力机制的群体旅游路线推荐方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHEN XIUJUN, WANG QING: "Tour Route Recommendation Begins with Multimodal Classification", JOURNAL OF MULTIMEDIA, ACADEMY PUBLISHER, FINLAND, vol. 7, no. 1, Finland , XP055926293, ISSN: 1796-2048, DOI: 10.4304/jmm.7.1.21-30 *
LI FANGWEN: "Group Travel Recommendation Research Based on User Generated Spatial Data", CHINESE MASTER’S THESES FULL-TEXT DATABASE (INFORMATION SCIENCE AND TECHNOLOGY), no. 1, 15 January 2020 (2020-01-15), XP055926297 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115096324A (zh) * 2022-05-25 2022-09-23 科大讯飞股份有限公司 路线推荐方法及相关装置、车机、车辆和存储介质
CN114781625A (zh) * 2022-06-13 2022-07-22 支付宝(杭州)信息技术有限公司 一种网络模型训练、推送内容确定方法及装置
CN115344777A (zh) * 2022-07-15 2022-11-15 重庆工商大学 一种基于大数据的多人定制化旅游的方法及装置
CN115344777B (zh) * 2022-07-15 2023-06-20 重庆工商大学 一种基于大数据的多人定制化旅游的方法及装置
CN116304362A (zh) * 2022-12-21 2023-06-23 山东省人工智能研究院 一种基于相关性注意力的群组推荐方法
CN116304362B (zh) * 2022-12-21 2023-12-15 山东省人工智能研究院 一种基于相关性注意力的群组推荐方法

Also Published As

Publication number Publication date
CN112149010A (zh) 2020-12-29
CN112149010B (zh) 2024-05-24

Similar Documents

Publication Publication Date Title
WO2022088661A1 (zh) 基于注意力机制的群体旅游路线推荐方法
US20230162071A1 (en) Autonomous learning platform for novel feature discovery
Yao et al. POI recommendation: A temporal matching between POI popularity and user regularity
Chen et al. Personalized itinerary recommendation: Deep and collaborative learning with textual information
WO2019184833A1 (zh) 旅游信息推荐方法和装置
Zheng et al. GeoLife: A collaborative social networking service among user, location and trajectory.
Bin et al. A travel route recommendation system based on smart phones and iot environment
Chen et al. CEM: A convolutional embedding model for predicting next locations
WO2022134478A1 (zh) 路线推荐方法、装置、电子设备和存储介质
Chen et al. Next POI recommendation based on location interest mining with recurrent neural networks
Chen et al. Attentive multi-task learning for group itinerary recommendation
Lin et al. A stacking model for variation prediction of public bicycle traffic flow
Sarkar et al. gTour: Multiple itinerary recommendation engine for group of tourists
Xue et al. Forecasting hourly attraction tourist volume with search engine and social media data for decision support
Xu et al. Deep convolutional recurrent model for region recommendation with spatial and temporal contexts
CN110008411A (zh) 一种基于用户签到稀疏矩阵的深度学习兴趣点推荐方法
Wang et al. DeepFM-Based taxi pick-up area recommendation
Karimi et al. A comparative assessment of machine learning methods in extracting place functionality from textual content
CN115827898A (zh) 一种面向用户意图的时空知识图谱构建方法
Smirnov et al. Tourist attraction recommendation service: an approach, architecture and case study
Yabe et al. Unsupervised translation via hierarchical anchoring: functional mapping of places across cities
Panagiotakis et al. The Tourist Trip Design Problem with POI Categories via an Expectation-Maximization Based Method.
Wang [Retracted] Self‐Driving Tour Travel Route Planning Model Construction Using Recurrent Neural Network
Lu Personalized Recommendation Algorithm of Smart Tourism Based on Cross‐Media Big Data and Neural Network
Zhou et al. POI Route Recommendation Model Based on Symmetrical Naive Bayes Classification Spatial Accessibility and Improved Cockroach Swarm Optimization Algorithm

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21884404

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21884404

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21884404

Country of ref document: EP

Kind code of ref document: A1