A deep learning-based travel route recommendation method

Abstract

The application discloses a kind of tourism route recommendation method based on deep learning, the key words of the text such as tourist attraction title, tourist attraction description and tourist attraction comment are extracted, vectorization is carried out to tourist attraction text and user comment by using Word2Vec, then GRU bidirectional neural network model is introduced, and tourist attraction attribute features and user interest features are deeply extracted, factor decomposition machine is combined to excavate tourist attraction for recommendation, then user benefit is calculated according to positive attributes and negative attributes of tourist attraction, further filter out the optimal tourist attraction, and the optimal route is calculated based on user benefit using Floyd algorithm, improve the accuracy of recommended tourism route.

Description

Technical Field

[0001] This invention relates to the field of information recommendation, specifically to a deep learning-based method for recommending travel routes. Background Technology

[0002] Before traveling, gathering travel and attraction information and developing a travel itinerary are crucial for ensuring a comfortable trip. With the rise of the big data era, tourists often need to spend a lot of time searching for information on worthwhile attractions on various travel websites and online platforms, and develop travel plans and itineraries that suit their time and budget. Travel route recommendations can help tourists quickly find suitable travel routes.

[0003] Traditional travel route recommendation methods include those based on attraction popularity, graph models, and topic models. Attraction popularity-based methods simply calculate attraction popularity and rank them, resulting in a lack of personalization. Graph model-based methods use information from various attractions in the dataset to build a graph model and make recommendations by analyzing the correlations between attractions, but they do not consider the relationships between words within an attraction and their context. Topic model-based methods extract attraction topics by analyzing the probability distribution of words in the attraction text and recommend attractions based on those topics, but they lack in-depth analysis of user interests. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for recommending travel routes, which mines deep-level scenic spot attribute features and user interest features, calculates user benefits based on the positive and negative attributes of scenic spots, and makes further recommendations to improve the effectiveness of travel route recommendations.

[0005] To achieve the above functions, this invention designs a travel route recommendation method based on deep learning. For the text related to various attractions on the target travel website, the following steps S1-S6 are executed to complete targeted attraction recommendations for the target user:

[0006] Step S1: For each scenic spot, extract keywords from the scenic spot title, description text, and reviews based on the TF-IDF algorithm, and convert the extracted keywords into word vectors as scenic spot attribute word vectors, thus obtaining the scenic spot attribute word vectors for each scenic spot.

[0007] Step S2: Merge all comments made by target users on target travel websites regarding various attractions and convert them into word vectors, which will serve as user interest word vectors;

[0008] Step S3: Based on the GRU bidirectional neural network model, with the scenic spot attribute word vector and the user interest word vector as input, perform feature extraction on the user interest word vector and the scenic spot attribute word vector respectively, and output the user interest feature vector and the scenic spot attribute feature vector.

[0009] Step S4: Merge the user interest feature vector and the attribute feature vector of each attraction into a real-valued feature matrix. Using the real-valued feature matrix as input, output the predicted score of each attraction through the factorization machine model. Sort the attractions according to the predicted scores from high to low to obtain the first candidate attraction list.

[0010] Step S5: Based on the first candidate attraction list, and based on the division of the attraction attributes of each attraction into preset positive and preset negative attributes, introduce the user benefit function, calculate the benefit value between each attraction according to the positive and negative attributes of each attraction, and sort the attractions a second time according to the benefit value to obtain the optimal attraction list.

[0011] Step S6: Based on the ranking of attractions in the optimal attraction list, use the Floyd algorithm to calculate the shortest path between attractions in the optimal attraction list, form the optimal travel route using the shortest path between attractions, and send it to the target user.

[0012] As a preferred technical solution of the present invention, the specific steps of step S1 are as follows:

[0013] Step S11: Remove stop words from the text related to each attraction, including attraction titles, descriptions, and reviews.

[0014] Step S12: Extract keywords from attraction titles, attraction descriptions, and attraction reviews based on the TF-IDF algorithm;

[0015] Step S13: Remove duplicate words from all extracted keywords and use Word2Vec to vectorize each deduplicated keyword.

[0016] As a preferred technical solution of the present invention: Step S2 merges all comments published by the target user on the target tourism website for each attraction, converts them into word vectors, and uses them as user interest word vectors. The bidirectional encoder Transformer is then used to reduce the dimensionality of the user interest word vectors.

[0017] As a preferred technical solution of the present invention, the specific method of step S3 is as follows:

[0018] Using the word vectors of scenic spot attributes and the word vectors of user interests as inputs to the GRU bidirectional neural network model, the reset gate in the GRU bidirectional neural network model... and Update Gate The calculation is as follows:

[0019]

[0020]

[0021] In the formula, This represents the activation function. , These represent the weight matrices for the reset gate and the update gate, respectively. This represents the input at time t. This represents the hidden state information from the previous moment. , These represent the deviation parameters for resetting the door and updating the door, respectively;

[0022] Candidate hidden state information at time t The calculation is as follows:

[0023]

[0024] In the formula, Indicates the time-weighted parameter. This represents the hidden state information from the previous moment. Indicates the deviation parameter. This indicates resetting the gate neuron;

[0025] Based on the hidden state information of the previous moment Candidate hidden state information at time t Calculate the hidden state information at time t. As shown in the following formula:

[0026]

[0027] In the formula, This indicates the updating of the gate neuron;

[0028] The feature vector output after passing through the GRU bidirectional neural network model The expression is as follows:

[0029]

[0030] In the formula, This represents the weight matrix and eigenvectors in the output layer of a GRU bidirectional neural network model. The features to be retained include the feature vectors of attraction attributes and the feature vectors of user interests.

[0031] As a preferred technical solution of the present invention, the specific method of step S4 is as follows:

[0032] The feature vectors of scenic spot attributes and user interest are merged into a real-valued feature matrix. As input to the factorization machine model, the prediction expression of the factorization machine model is as follows:

[0033]

[0034] In the formula, , These are the real-valued eigenma matrices. The i-th and j-th elements in the middle, express The weighting coefficients, It is a vector matrix. , They are vector matrices The i-th and j-th vectors in the model are given, and n represents the number of vectors output by the GRU bidirectional neural network model. As shown in the following formula:

[0035]

[0036] In the formula, , They are vectors , The element at position f in the middle, For vectors , The dimension of.

[0037] As a preferred technical solution of the present invention: the positive attributes of each scenic spot in step S5 include the traffic distance between scenic spots, the seasonal popularity coefficient of the scenic spot, and the scenic spot level; the negative attributes of each scenic spot include the walking distance between scenic spots, the number of traffic lights, the road congestion coefficient, and the maximum average vehicle delay time.

[0038] As a preferred technical solution of the present invention, the specific method of step S5 is as follows:

[0039] The revenue value between each attraction is calculated using the following formula:

[0040]

[0041] In the formula, Indicates the current attraction Next attraction The profit value between them This represents the i-th positive attribute. Let represent the j-th negative attribute, l be the total number of positive attributes, and m be the total number of negative attributes. Indicates the current attraction Next attraction The distance between them This is the adjustment coefficient;

[0042] in, The calculation is as follows:

[0043]

[0044] In the formula, R represents the Earth's radius. , These represent the current attractions. Next attraction The difference in latitude and longitude between them Indicates the current attraction latitude, Indicates the next attraction Latitude.

[0045] As a preferred embodiment of the present invention, the specific steps of step S6 are as follows:

[0046] Step S61: Calculate the state transition equation according to the Floyd algorithm as follows:

[0047]

[0048] In the formula, This represents the shortest path from attraction i to attraction j after transferring through no more than k attractions.

[0049] Step S62: Let ,in Given the adjacency matrix from scenic spot i to scenic spot j, we use a rolling array approach to roll away the first dimension, thereby reducing the dimensionality of the state transition equation:

[0050]

[0051] Solving for That is, the shortest path from attraction i to attraction j, where, , .

[0052] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0053] This invention designs a deep learning-based travel route recommendation method. Compared with existing technologies, its significant advantages are: this invention uses Word2Vec to vectorize words, clearly representing semantic information; then it introduces a GRU bidirectional neural network model to mine deep-level scenic spot attribute features and user interest features, and calculates user benefits based on the positive and negative attributes of scenic spots, which can further filter the recommendation results and improve the accuracy of recommended travel routes. Attached Figure Description

[0054] Figure 1 It is a flowchart of a travel route recommendation method based on deep learning provided by an embodiment of the present invention. Detailed implementation manners

[0055] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and cannot be used to limit the protection scope of the present invention.

[0056] Refer to Figure 1 An embodiment of the present invention provides a travel route recommendation method based on deep learning. For the text related to each scenic spot on the target travel website, the following steps S1 - S6 are executed to complete the targeted scenic spot recommendation for the target user:

[0057] Step S1: For the text related to each scenic spot, based on the TF-IDF algorithm, extract the keywords of the scenic spot title, scenic spot description text, and scenic spot comments, and convert the extracted keywords into word vectors as the scenic spot attribute word vectors corresponding to the scenic spots, that is, obtain the scenic spot attribute word vectors corresponding to each scenic spot respectively;

[0058] The specific steps of Step S1 are as follows:

[0059] Step S11: For the text related to each scenic spot, remove the stop words in the scenic spot title, scenic spot description text, and scenic spot comments; the stop words are words filtered according to preset rules before processing natural language data, including punctuation marks and single Chinese characters with high usage frequency, such as the Chinese character "de". Such single Chinese characters cannot express the information related to the text alone. Removing stop words can improve the data processing efficiency and will not cause information loss.

[0060] Step S12: Based on the TF-IDF algorithm, extract the keywords of the scenic spot title, scenic spot description text, and scenic spot comments; where TF represents term frequency and IDF represents inverse document frequency.

[0061] Step S13: Remove duplicate words from all the extracted keywords, and use Word2Vec to vectorize each de-duplicated keyword respectively.

[0062] Step S2: Merge all the comments made by the target user on each scenic spot on the target travel website, convert them into word vectors as the user interest word vectors;

[0063] Step S2 merges all the comments made by the target user on each scenic spot on the target travel website, converts them into word vectors as the user interest word vectors, and uses a bidirectional encoder Transformer to reduce the dimension of the user interest word vectors.

[0064] The dimensionality reduction process is as follows:

[0065] The dimension of the original sentences input in batches after the embedding operation is calculated as follows, where the embedding operation uses the embedding method:

[0066]

[0067] in, This indicates the number of sentences in the batch input sequence. Indicates the word length of the original sentence. This represents the dimension of the original sentence after the embedding operation.

[0068] The input user interest word vector is multiplied by three weight matrices respectively. Three matrices are obtained. , representing the information to be queried, the content to be matched, and the input features, respectively, and their dimensions are calculated as follows:

[0069]

[0070] To each The three matrices undergo linear transformation through a multi-head self-attention mechanism, with the dimensionality changes as follows:

[0071]

[0072] in, This represents the dimensionality of each head in a multi-head self-attention mechanism. Based on the linearly transformed matrix... The output vector is obtained and calculated as follows:

[0073]

[0074] in, Representation matrix Dimensions.

[0075] Therefore, the dimension of the output vector is calculated as follows:

[0076]

[0077] in, This indicates the number of heads in a multi-head self-attention mechanism.

[0078] Step S3: Based on the GRU bidirectional neural network model, with the scenic spot attribute word vector and the user interest word vector as input, perform feature extraction on the user interest word vector and the scenic spot attribute word vector respectively, and output the user interest feature vector and the scenic spot attribute feature vector.

[0079] The specific method for step S3 is as follows:

[0080] Using the word vectors of scenic spot attributes and user interest as inputs to the GRU bidirectional neural network model, the model includes an update gate and a reset gate. The update gate outputs the degree to which the previous time-step state information is incorporated into the current time-step, while the reset gate removes invalid information from the previous time-step, outputting the extracted scenic spot attribute feature vectors and user interest feature vectors. The reset gate in the GRU bidirectional neural network model... and Update Gate The calculation is as follows:

[0081]

[0082]

[0083] In the formula, This represents the activation function. , These represent the weight matrices for the reset gate and the update gate, respectively. This represents the input at time t. This represents the hidden state information from the previous moment. , These represent the deviation parameters for resetting the door and updating the door, respectively;

[0084] Candidate hidden state information at time t The calculation is as follows:

[0085]

[0086] In the formula, Indicates the time-weighted parameter. This represents the hidden state information from the previous moment. Indicates the deviation parameter. This indicates resetting the gate neuron;

[0087] Based on the hidden state information of the previous moment Candidate hidden state information at time t Calculate the hidden state information at time t. As shown in the following formula:

[0088]

[0089] In the formula, This indicates the updating of the gate neuron;

[0090] The feature vector output after passing through the GRU bidirectional neural network model The expression is as follows:

[0091]

[0092] In the formula, This represents the weight matrix and eigenvectors in the output layer of a GRU bidirectional neural network model. The features to be retained include the feature vectors of attraction attributes and the feature vectors of user interests.

[0093] Step S4: Merge the user interest feature vector and the attribute feature vector of each attraction into a real-valued feature matrix. Using the real-valued feature matrix as input, output the predicted score of each attraction through the factorization machine model. Sort the attractions according to the predicted scores from high to low to obtain the first candidate attraction list.

[0094] The specific method for step S4 is as follows:

[0095] The feature vectors of scenic spot attributes and user interest are merged into a real-valued feature matrix. As input to the factorization machine model, the prediction expression of the factorization machine model is as follows:

[0096]

[0097] In the formula, , These are the real-valued eigenma matrices. The i-th and j-th elements in the middle, express The weighting coefficients, It is a vector matrix. , They are vector matrices The i-th and j-th vectors in the model are given, and n represents the number of vectors output by the GRU bidirectional neural network model. As shown in the following formula:

[0098]

[0099] In the formula, , They are vectors , The element at position f in the middle, For vectors , The dimension of.

[0100] Step S5: Based on the first candidate attraction list, and based on the division of the attraction attributes of each attraction into preset positive and preset negative attributes, introduce the user benefit function, calculate the benefit value between each attraction according to the positive and negative attributes of each attraction, and sort the attractions a second time according to the benefit value to obtain the optimal attraction list.

[0101] The positive attributes of each scenic spot mentioned in step S5 include the traffic distance between scenic spots, the seasonal popularity coefficient of the scenic spot, and the scenic spot level; the negative attributes of each scenic spot include the walking distance between scenic spots, the number of traffic lights, the road congestion coefficient, and the maximum average vehicle delay time.

[0102] The positive and negative attributes of scenic spot attributes are respectively used... and The attributes of the attractions are shown in Table 1:

[0103] Table 1

[0104]

[0105] The calculation method for the maximum average vehicle delay time is as follows:

[0106]

[0107] in, This indicates the maximum average delay time per vehicle, in seconds. ), This represents the average delay time of vehicles entering the i-th lane within a certain time interval, in seconds. ), This indicates the total number of entrance lanes for this road section.

[0108] The specific method for step S5 is as follows:

[0109] The revenue value between each attraction is calculated using the following formula:

[0110]

[0111] In the formula, Indicates the current attraction Next attraction The profit value between them This represents the i-th positive attribute. Let represent the j-th negative attribute, l be the total number of positive attributes, and m be the total number of negative attributes. Indicates the current attraction Next attraction The distance between them This is the adjustment coefficient;

[0112] in, The calculation is as follows:

[0113]

[0114] In the formula, R represents the Earth's radius, R = 6378.187 km. , These represent the current attractions. Next attraction The difference in latitude and longitude between them Indicates the current attraction latitude, Indicates the next attraction Latitude.

[0115] Step S6: Based on the ranking of attractions in the optimal attraction list, use the Floyd algorithm to calculate the shortest path between attractions in the optimal attraction list, form the optimal travel route using the shortest path between attractions, and send it to the target user.

[0116] The specific steps of step S6 are as follows:

[0117] Step S61: Calculate the state transition equation according to the Floyd algorithm as follows:

[0118]

[0119] In the formula, This represents the shortest path from attraction i to attraction j after transferring through no more than k attractions.

[0120] Step S62: Let ,in Given the adjacency matrix from scenic spot i to scenic spot j, we use a rolling array approach to roll away the first dimension, thereby reducing the dimensionality of the state transition equation:

[0121]

[0122] Solving for That is, the shortest path from attraction i to attraction j, where, , .

[0123] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A deep learning-based method for recommending travel routes, characterized in that, For the text related to various attractions on the target travel website, perform the following steps S1-S6 to complete targeted attraction recommendations for the target users: Step S1: For each scenic spot, extract keywords from the scenic spot title, description text, and reviews based on the TF-IDF algorithm, and convert the extracted keywords into word vectors as scenic spot attribute word vectors, thus obtaining the scenic spot attribute word vectors for each scenic spot. Step S2: Merge all comments made by target users on target travel websites regarding various attractions and convert them into word vectors, which will serve as user interest word vectors; Step S3: Based on the GRU bidirectional neural network model, with the scenic spot attribute word vector and the user interest word vector as input, perform feature extraction on the user interest word vector and the scenic spot attribute word vector respectively, and output the user interest feature vector and the scenic spot attribute feature vector. Step S4: Merge the user interest feature vector and the attribute feature vector of each attraction into a real-valued feature matrix. Using the real-valued feature matrix as input, output the predicted score of each attraction through the factorization machine model. Sort the attractions according to the predicted scores from high to low to obtain the first candidate attraction list. Step S5: Based on the first candidate attraction list, and based on the division of the attraction attributes of each attraction into preset positive and preset negative attributes, introduce the user benefit function, calculate the benefit value between each attraction according to the positive and negative attributes of each attraction, and sort the attractions a second time according to the benefit value to obtain the optimal attraction list. The positive attributes of each attraction include the travel distance between attractions, the seasonal popularity coefficient of the attraction, and the attraction level; the negative attributes of each attraction include the walking distance between attractions, the number of traffic lights, the road congestion coefficient, and the maximum average vehicle delay time. The specific method for step S5 is as follows: The revenue value between each attraction is calculated using the following formula: ； In the formula, Indicates the current attraction Next attraction The profit value between them This represents the i-th positive attribute. Let represent the j-th negative attribute, l be the total number of positive attributes, and m be the total number of negative attributes. Indicates the current attraction Next attraction The distance between them This is the adjustment coefficient; in, The calculation is as follows: ； In the formula, R represents the Earth's radius. , These represent the current attractions. Next attraction The difference in latitude and longitude between them Indicates the current attraction latitude, Indicates the next attraction Latitude; Step S6: Based on the ranking of attractions in the optimal attraction list, use the Floyd algorithm to calculate the shortest path between attractions in the optimal attraction list, form the optimal travel route using the shortest path between attractions, and send it to the target user.

2. The deep learning-based travel route recommendation method according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Remove stop words from the text related to each attraction, including attraction titles, descriptions, and reviews. Step S12: Extract keywords from attraction titles, attraction descriptions, and attraction reviews based on the TF-IDF algorithm; Step S13: Remove duplicate words from all extracted keywords and use Word2Vec to vectorize each deduplicated keyword.

3. The deep learning-based travel route recommendation method according to claim 1, characterized in that, Step S2 merges all comments made by the target user on the target tourism website for each attraction, converts them into word vectors, and uses them as user interest word vectors. Then, a bidirectional encoder Transformer is used to reduce the dimensionality of the user interest word vectors.

4. The deep learning-based travel route recommendation method according to claim 1, characterized in that, The specific method for step S3 is as follows: Using the word vectors of scenic spot attributes and the word vectors of user interests as inputs to the GRU bidirectional neural network model, the reset gate in the GRU bidirectional neural network model... and Update Gate The calculation is as follows: ； ； In the formula, This represents the activation function. , These represent the weight matrices for the reset gate and the update gate, respectively. This represents the input at time t. This represents the hidden state information from the previous moment. , These represent the deviation parameters for resetting the door and updating the door, respectively; Candidate hidden state information at time t The calculation is as follows: ； In the formula, Indicates the time-weighted parameter. This represents the hidden state information from the previous moment. Indicates the deviation parameter. This indicates resetting the gate neuron; Based on the hidden state information of the previous moment Candidate hidden state information at time t Calculate the hidden state information at time t. As shown in the following formula: ； In the formula, This indicates the updating of the gate neuron; The feature vector output after passing through the GRU bidirectional neural network model The expression is as follows: ； In the formula, This represents the weight matrix and eigenvectors in the output layer of a GRU bidirectional neural network model. The features to be retained include the feature vectors of attraction attributes and the feature vectors of user interests.

5. The deep learning-based travel route recommendation method according to claim 4, characterized in that, The specific method for step S4 is as follows: The feature vectors of scenic spot attributes and user interest are merged into a real-valued feature matrix. As input to the factorization machine model, the prediction expression of the factorization machine model is as follows: ； In the formula, , These are the real-valued eigenma matrices. The i-th and j-th elements in the middle, express The weighting coefficients, It is a vector matrix. , They are vector matrices The i-th and j-th vectors in the model are given, and n represents the number of vectors output by the GRU bidirectional neural network model. As shown in the following formula: ； In the formula, , They are vectors , The element at position f in the middle, For vectors , The dimension of.

6. The deep learning-based travel route recommendation method according to claim 5, characterized in that, The specific steps of step S6 are as follows: Step S61: Calculate the state transition equation according to the Floyd algorithm as follows: ； In the formula, This represents the shortest path from attraction i to attraction j after transferring through no more than k attractions. Step S62: Let ,in Given the adjacency matrix from scenic spot i to scenic spot j, we use a rolling array approach to roll away the first dimension, thereby reducing the dimensionality of the state transition equation: ； Solving for That is, the shortest path from attraction i to attraction j, where, , .

Images