Machine learning based demand forecasting method and system

Abstract

The application provides a demand prediction method and system based on machine learning, first, a user travel-related text data set containing a plurality of natural language statements marked by continuous timestamps is obtained, then the text data set is subjected to demand level analysis to obtain explicit demand elements and implicit demand clues of each user content expression unit, then the explicit demand elements and the implicit demand clues are subjected to demand association modeling to generate a demand association feature set containing semantic connection relationship and association strength information, then demand evolution analysis is carried out based on the demand association feature set to generate a demand evolution feature set containing demand element change rule with time and implicit demand clue conversion rule, finally, demand prediction results containing user demand priority distribution of different categories of tourism products and demand change trend description are generated according to the demand evolution feature set, so that user demand can be comprehensively and accurately mined, and accurate demand prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a demand forecasting method and system based on machine learning. Background Technology

[0002] In the tourism product sector, accurately predicting user demand for tourism products is crucial for tourism enterprises' market planning, product development, and service optimization. Currently, traditional demand forecasting methods mainly rely on historical sales data, market research questionnaires, and the experience and judgment of industry experts. While historical sales data can reflect the sales performance of tourism products over a past period, it cannot capture the dynamic changes in user demand or potential demand. Market research questionnaires have limited sample size, and users may have subjective biases when completing the questionnaires, resulting in inaccurate and incomplete demand information. The experience and judgment of industry experts are greatly influenced by personal knowledge and experience, lacking objectivity and universal applicability.

[0003] With the rapid development of internet technology, users have generated a large amount of travel-related text data on platforms such as social media and travel forums. This text data contains rich information about user needs. However, most existing text mining techniques only focus on explicit information in the text, and are insufficient in mining implicit needs cues conveyed through semantic associations that users do not directly state, making it difficult to comprehensively and accurately grasp user needs. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a demand forecasting method based on machine learning, the method comprising:

[0005] Obtain a set of text data related to user travel, the set of text data contains multiple user expression content units, each of which consists of natural language statements marked with consecutive timestamps;

[0006] The text data set is subjected to demand hierarchy parsing to obtain the explicit demand elements and implicit demand cues of each user's expressed content unit. The explicit demand elements are the user's clearly stated functional requirements for tourism products, and the implicit demand cues are the user's potential demand tendencies that are not directly stated but conveyed through semantic association.

[0007] The explicit demand elements and the implicit demand clues are subjected to demand association modeling to generate a demand association feature set, which includes the semantic connection relationship and association strength information between explicit and implicit demands.

[0008] Based on the aforementioned set of demand-related features, demand evolution analysis is performed to generate a set of demand evolution features. The set of demand evolution features includes the continuity of demand elements over time and the association rules for the transformation of implicit demand cues into explicit demand elements.

[0009] Demand forecasting is performed based on the set of demand evolution characteristics to generate tourism product demand forecasting results. The demand forecasting results include the distribution of user demand priorities for different categories of tourism products and a description of demand change trends.

[0010] In another aspect, embodiments of the present invention also provide a demand forecasting system based on machine learning, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention, by acquiring a set of user travel-related text data and performing demand hierarchy analysis, can comprehensively and accurately obtain explicit demand elements and implicit demand cues in each user's expressed content unit. This effectively overcomes the shortcomings of traditional methods that only focus on explicit information and fail to adequately mine implicit demands. Demand association modeling is performed on explicit demand elements and implicit demand cues to generate a demand association feature set containing semantic connection relationships and association strength information. This helps to deeply understand the intrinsic connections between user demands and uncover potential demand patterns. Based on the demand association feature set, demand evolution analysis is performed, and the generated demand evolution feature set can reveal the continuous pattern of demand elements changing over time and the association rules for the transformation of implicit demand cues into explicit demand elements. Finally, the demand prediction results generated based on the demand evolution feature set include the distribution of user demand priorities for different categories of tourism products and a description of demand change trends. This provides tourism enterprises with a scientific and accurate decision-making basis for formulating market strategies, developing new products, and optimizing services. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the demand prediction method based on machine learning provided in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the demand forecasting system based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a machine learning-based demand forecasting method according to an embodiment of the present invention. The machine learning-based demand forecasting method will be described in detail below.

[0015] Step S110: Obtain a set of text data related to user travel. The set of text data contains multiple user expression content units, each of which consists of natural language statements marked with consecutive timestamps.

[0016] In this embodiment, the user's travel-related text data set can originate from various user-generated content platforms, such as posts and replies in travel forums, chat logs in travel-related social groups, user inquiries on travel booking platforms, and user comments and feedback within travel apps. When acquiring this user travel-related text data set, it is crucial to protect users' sensitive privacy data. For example, anonymizing users' personal identification information and contact details is necessary. This involves removing identifying information such as names, phone numbers, and ID numbers from the text to ensure that data collection complies with relevant laws and regulations and does not infringe on user privacy.

[0017] Each user-generated content unit (UGC) refers to a continuous set of natural language statements posted by a user within a specific time period related to travel. To accurately record the temporal attributes of this content, consecutive timestamps can be added to each UGC. Timestamps can be recorded according to the specific moment the content was generated, such as in seconds, clearly reflecting the chronological order of the user's expressions. For example, a post containing multiple statements on a travel forum, from the first statement to the last, would be timestamped. These statements together constitute a UGC. Multiple UGCs are combined to generate the collected travel-related text data set.

[0018] Step S120: Perform demand hierarchy parsing on the text data set to obtain the explicit demand elements and implicit demand cues of each user's expressed content unit. The explicit demand elements are the user's clearly stated functional requirements for tourism products, and the implicit demand cues are the user's potential demand tendencies that are not directly stated but conveyed through semantic association.

[0019] After obtaining the text data set, the next step is to perform demand hierarchy analysis. The purpose is to distinguish the explicitly expressed tourism product functional requirements from the user's natural language expression, namely explicit demand elements, as well as the potential demand tendencies that are not directly stated but can be inferred through semantic connections, namely implicit demand cues.

[0020] Step S121: Perform natural language segmentation on the user expression content unit, and use preset domain segmentation rules to segment continuous natural language sentences into basic segmentation units with independent semantics, to obtain a segmentation result set containing multiple basic segmentation units.

[0021] In this embodiment, during the demand hierarchy parsing process, the user-expressed content units are first processed using natural language word segmentation. The preset domain-specific word segmentation rules are formulated based on the characteristics of the tourism industry, taking into account the large number of professional terms, common expressions, and specific terminology combinations present in the tourism sector. For example, the natural language statement "I want to find a hotel near the sea, preferably with a swimming pool and a gym," according to the preset domain-specific word segmentation rules, can be segmented into basic word segmentation units with independent semantics, such as "I," "want to find," "one," "near," "seaside," "of," "hotel," "preferably," "has," "swimming pool," "and," and "gym." These basic word segmentation units accurately reflect the semantics of each part of the statement. All the segmented basic word segmentation units are aggregated to generate a word segmentation result set. During the word segmentation process, some tourism-specific terms, such as "sea view room" and "self-guided tour," can be segmented as a whole basic word segmentation unit rather than broken into smaller parts to ensure semantic integrity.

[0022] Step S122: Perform explicit requirement element extraction processing on the word segmentation result set. By traversing the predefined functional requirement lexicon and matching basic word segmentation units, extract basic word segmentation units that are directly related to the functions of tourism products as explicit requirement candidate elements. The functional requirement lexicon contains semantic units corresponding to the core functions of tourism products.

[0023] After word segmentation, the segmentation results are then processed to extract explicit requirement elements. The predefined functional requirement lexicon is constructed in advance based on the core functions of various tourism products, containing various semantic units directly related to the functions of the tourism products. These semantic units cover functional descriptions of different types of tourism products, such as hotels, transportation, attractions, and catering. For example, hotels might include "swimming pool," "gym," "free WiFi," and "breakfast service"; transportation might include "direct access," "air conditioning," and "seat reservation"; and attractions might include "ticket reservation" and "guide service."

[0024] During the extraction process, each basic word segmentation unit in the word segmentation result set can be traversed and matched one by one with the semantic units in the functional requirement terminology. When a basic word segmentation unit is completely consistent with or highly matched with a semantic unit in the functional requirement terminology, the basic word segmentation unit will be extracted as an explicit requirement candidate element. For example, in the word segmentation result set of the above example, the basic word segmentation units such as "hotel", "swimming pool", and "gym" match the semantic units of hotel-related functions in the functional requirement terminology, and therefore will be extracted as explicit requirement candidate elements.

[0025] Step S1221: Construct a functional appeal terminology library, which contains multi-level semantic units corresponding to the core functions of tourism products. The multi-level semantic units contain first-level functional categories and second-level functional subcategories.

[0026] When constructing a functional appeal terminology database, it is necessary to first categorize and organize the core functions of tourism products into a multi-level structure. The first-level functional categories provide a macro-level classification of tourism product functions, such as "accommodation," "transportation," "attraction services," and "dining." Each first-level functional category is further subdivided into second-level functional subcategories. Taking "accommodation" as an example, its second-level functional subcategories can include "basic accommodation facilities," "leisure and entertainment facilities," "dining services," and "convenience services." Each second-level functional subcategory contains specific semantic units, such as "single room," "double room," and "king-size bed" under "basic accommodation facilities," and "swimming pool," "gym," and "SPA center" under "leisure and entertainment facilities." Through this multi-level structure, the functional appeal terminology database can comprehensively and systematically cover the semantic units corresponding to various core functions of tourism products.

[0027] Step S1222: Perform multi-round matching processing on the word segmentation result set. In the first round, match the semantic units corresponding to the primary function category and extract the basic word segmentation units related to the basic service function. In the second round, match the semantic units corresponding to the secondary function sub-category and further extract the basic word segmentation units related to the trip planning function from the primary function category matching results.

[0028] When performing multi-round matching on the word segmentation result set, the first round of matching is performed, which matches the semantic units corresponding to the first-level functional categories. For example, the semantic units corresponding to the first-level functional category "accommodation function" may include "hotel", "inn", "homestay", etc. When these basic word segmentation units appear in the word segmentation result set, they will be extracted as matching results related to basic service functions.

[0029] Building upon the first round of matching, a second round is conducted, further extracting semantic units corresponding to the second-level functional subcategories from the first round's matching results. For example, in the first-level matching results for "accommodation function," semantic units such as "swimming pool" and "gym" corresponding to the second-level functional subcategory "leisure and entertainment facilities" are matched. The basic word segmentation units belonging to these semantic units in the word segmentation results are extracted. These are the basic word segmentation units related to the specific functional requirements related to accommodation in itinerary planning. Through multiple rounds of matching, basic word segmentation units related to different levels of functions can be more accurately selected.

[0030] Step S1223: Perform contextual verification processing on the multi-round matching results, and combine the semantic information of the context of the basic word segmentation unit to exclude mismatched basic word segmentation units caused by grammatical errors or ambiguities.

[0031] After multiple rounds of matching, the matching results need to be verified against the context. This is because during word segmentation and matching, there may be some mismatches due to errors in sentence segmentation or polysemy. For example, in the sentence "This person is very interesting, likes swimming, and the water in the pool is very clean," "swimming" and "pool" might be mismatched as semantic units related to "swimming pool," but considering the context, it can be seen that "pool" here does not refer to the hotel's swimming pool.

[0032] Therefore, it is necessary to verify each basic word segmentation unit by combining the semantic information of its context within the original sentence. By analyzing the meanings expressed by the words before and after the basic word segmentation unit, it is determined whether it truly belongs to the semantics related to the tourism product function. If it is found that the contextual semantics of a certain basic word segmentation unit does not match the functional scenario corresponding to the semantic unit in the functional demand lexicon, the basic word segmentation unit will be excluded from the matching results to ensure the accuracy of explicit demand candidate elements.

[0033] Step S1224: Perform duplicate item merging processing on the verified matching results, merging duplicate basic word segmentation units under the same functional subclass into a single explicit demand candidate element.

[0034] After contextual validation, duplicate entries need to be merged for the valid matching results. Users may mention the same functional requirement multiple times in their statements, such as "I want a hotel with a swimming pool, preferably an outdoor one." Here, "swimming pool" appears twice, both times belonging to the "leisure and entertainment facilities" subclass.

[0035] In the above case, the basic word segmentation units that appear repeatedly under the same functional subclass can be merged into a single explicit requirement candidate element, that is, only "swimming pool" is retained as an explicit requirement candidate element, avoiding repeated calculations and redundant information, making the explicit requirement candidate elements more concise and focused.

[0036] Step S1225: Perform functional category labeling on the explicit requirement candidate elements, labeling each explicit requirement candidate element with its corresponding first-level functional category and second-level functional subclass, generating explicit requirement candidate elements containing functional hierarchy information.

[0037] After merging duplicates, the explicit requirement candidate elements need to be labeled with functional categories. Based on the multi-level structure of the functional requirement terminology, the primary functional category and secondary functional subcategory of each explicit requirement candidate element are determined. For example, the primary functional category for "swimming pool" is "accommodation function," and the secondary functional subcategory is "leisure and entertainment facilities"; the primary functional category for "free WiFi" is "accommodation function," and the secondary functional subcategory is "convenience services."

[0038] By labeling, each explicit requirement candidate element carries corresponding functional level information, which helps to conduct more detailed analysis and processing of explicit requirement elements in the future, and to clarify the position and relationship of different explicit requirement elements in the functional system.

[0039] Step S123: By analyzing the sentiment attributes, logical modification relationships, and scene relevance of the remaining basic word segmentation units in the unmatched functional demand word library, extract basic word segmentation units that reflect the potential demand tendencies of users as implicit demand candidate clues. The sentiment attributes include positive or negative sentiment tendencies, the logical modification relationships include comparison, hypothesis, or contrast relationships, and the scene relevance includes scene descriptions related to the tourism experience.

[0040] After extracting explicit demand candidate elements, the word segmentation result set will contain some basic word segmentation units that do not match the functional demand terminology. For these remaining basic word segmentation units, it is necessary to extract implicit demand candidate clues by analyzing their sentiment attributes, logical modification relationships, and scene relevance. Although these remaining basic word segmentation units do not directly correspond to the functional demands of tourism products, they contain potential user demand tendencies.

[0041] For example, the word "very nice" in "I think the environment there must be very nice" expresses a positive emotional attribute; the phrase "if...then..." in "If the weather is nice, I'll go to the beach" reflects a hypothetical logical modification relationship; and the word "on the beach" in "It's very comfortable to walk on the beach" describes a scene related to the travel experience, demonstrating scene relevance. By analyzing these aspects, we can uncover potential needs that users haven't directly stated.

[0042] Step S1231: Perform sentiment attribute analysis on the remaining basic word segmentation units, identify basic word segmentation units with sentiment tendencies by traversing the predefined sentiment lexicon, and generate a sentiment polarity tag for each identified basic word segmentation unit. The sentiment polarity tag contains positive or negative sentiment tendency information.

[0043] When performing sentiment attribute analysis on the remaining basic word segmentation units, a predefined sentiment lexicon can be used. The sentiment lexicon contains a large number of words with clear sentiment tendencies, which are divided into two categories: positive sentiment words and negative sentiment words. Positive sentiment words include "good," "not bad," "comfortable," and "like," while negative sentiment words include "bad," "terrible," "dislike," and "inconvenient."

[0044] Each remaining basic word segmentation unit is iterated and matched against words in the sentiment lexicon. When a basic word segmentation unit belongs to a positive sentiment word in the sentiment lexicon, a positive sentiment polarity tag is generated for it; when it belongs to a negative sentiment word, a negative sentiment polarity tag is generated. For example, the basic word segmentation unit "comfortable" matches a positive sentiment word in the sentiment lexicon and is therefore tagged with a positive sentiment polarity; "terrible" is tagged with a negative sentiment polarity. In this way, the sentiment-biased parts of the remaining basic word segmentation units can be identified.

[0045] Step S1232: Perform logical modification relationship analysis on the remaining basic word segmentation units. Identify basic word segmentation units with comparison, hypothesis, or transition relationships by traversing the predefined logical relationship lexicon. Generate a logical relationship tag for each identified basic word segmentation unit. The logical relationship tag contains logical relationship information of comparison, hypothesis, or transition.

[0046] Logical modification relationship analysis is performed using a predefined logical relation lexicon. This lexicon contains various words and phrases used to express logical relationships such as comparison, hypothesis, and contrast. Examples of words expressing comparison include "better than," "more than," and "not as good as"; examples of words expressing hypothesis include "if...then..." and "suppose...then..."; and examples of words expressing contrast include "but," "however," and "however."

[0047] Traverse the remaining basic word segmentation units, identify the words or phrases belonging to the logical relation thesaurus among them, and then generate corresponding logical relation tags according to the type of logical relation they express. For example, the "than" in "quieter than..." expresses a comparison relation, so a comparison logical relation tag is generated; the "but" in "but there are too many people" expresses a turning relation, and a turning logical relation tag is generated. The logical relation tags can help understand the logical thinking of the user in the expression process, and then dig out potential demand tendencies.

[0048] Step S1233: Perform scene relevance analysis and processing on the remaining basic word segmentation units. By traversing the predefined scene related word thesaurus, identify the basic word segmentation units describing scene descriptions related to tourism experience, and generate scene association tags for each identified basic word segmentation unit. The scene association tags contain scene description information related to tourism experience.

[0049] The scene relevance analysis and processing rely on the predefined scene related word thesaurus. This scene related word thesaurus contains a large number of scene description words related to tourism experience. These words cover multiple aspects such as natural landscapes, human environments, activity places, etc., such as "seaside", "mountaintop", "ancient town", "night market", "museum", etc.

[0050] In the analysis process, traverse the remaining basic word segmentation units and match them with the words in the scene related word thesaurus. When a basic word segmentation unit matches the words in the scene related word thesaurus, it means that the basic word segmentation unit describes a scene related to tourism experience, and a scene association tag can be generated for it. The scene association tag contains the specific description information of this scene. For example, the basic word segmentation unit "seaside" matches the words in the scene related word thesaurus, and the generated scene association tag contains the scene description information of "seaside"; "ancient town" generates a tag containing the scene description information of "ancient town". Through the scene association tags, the tourism scenes involved in the remaining basic word segmentation units can be clarified, which helps to further analyze the potential demands of users in these scenes.

[0051] Step S1234: Perform multi-dimensional feature fusion processing on the emotional polarity tags, the logical relation tags and the scene association tags, and merge the multiple tags corresponding to the same basic word segmentation unit into a composite implicit demand clue. The composite implicit demand clue simultaneously contains potential demand information of emotional tendency, logical relation and scene association.

[0052] After obtaining the sentiment polarity marker, logical relationship marker, and scene association marker, these markers need to undergo multi-dimensional feature fusion processing. For the same basic word segmentation unit, if it is simultaneously marked with multiple types of markers, such as a basic word segmentation unit having both sentiment polarity markers, logical relationship markers, and scene association markers, these markers need to be merged together to form a composite implicit demand cue.

[0053] For example, in the phrase "A walk on the beach is comfortable, but there are too many people," the "beach" is marked with a scene association, "comfortable" with a positive emotional polarity, "but" with a transitional logical relationship, and "too many people" with a negative emotional polarity. Combining these markers creates a composite implicit need cue that simultaneously includes the beach scene, both positive and negative emotional tendencies, and the transitional logical relationship—all latent need information. Through this multi-dimensional integration, we can more comprehensively capture all aspects of users' latent needs, making the implicit need cues richer and more accurate.

[0054] Step S1235: Perform domain relevance filtering on the composite implicit demand cues, retain composite implicit demand cues related to tourism product demand, remove implicit semantic information unrelated to tourism scenarios, and generate the final implicit demand candidate cues.

[0055] After generating composite implicit demand cues, domain relevance filtering is required. This is because composite implicit demand cues may contain implicit semantic information unrelated to tourism product demand. For example, some composite implicit demand cues may involve personal life matters unrelated to tourism; this information is not helpful in predicting tourism product demand and needs to be removed.

[0056] The filtering process involves determining whether the contextual, emotional, and logical information contained in composite implicit need cues is relevant to the user's needs for tourism products. If the information in a composite implicit need cue revolves around tourism scenarios, experiences, and potential expectations for the product, it is retained; otherwise, it is discarded. The remaining composite implicit need cues after filtering constitute the final candidate implicit need cues, accurately reflecting users' latent tendencies related to their needs for tourism products.

[0057] Step S124: Classify explicit requirement candidate elements into a unified functional category according to the preset functional classification system, and classify implicit requirement candidate clues into a unified potential requirement category according to the preset potential requirement classification system, thereby generating explicit requirement elements and implicit requirement clues with unified classification standards.

[0058] After extracting explicit demand candidate elements and implicit demand candidate clues, they need to be categorized. For explicit demand candidate elements, they are classified according to a pre-defined functional classification system. This system is a unified classification standard based on the functional characteristics of tourism products. For example, explicit demand candidate elements related to accommodation can be categorized under "Accommodation Function," and those related to transportation can be categorized under "Transportation Function," ensuring that each explicit demand candidate element belongs to a unified functional category.

[0059] Implicit demand candidate clues are categorized according to a pre-defined latent demand classification system. This system is a unified classification standard based on users' potential travel needs and inclinations, and can be categorized into areas such as "Environmental Preferences," "Experience Expectations," "Social Needs," and "Convenience Needs." The "Environmental Preferences" category may include implicit demand candidate clues related to the natural environment and cultural atmosphere; the "Experience Expectations" category may include latent demand information regarding the richness and uniqueness of travel activities; the "Social Needs" category may include latent demand clues related to interaction and sharing with others; and the "Convenience Needs" category may include latent expectations regarding various convenient conditions during the travel process.

[0060] For example, the previously mentioned complex implicit need cue, "A walk on the beach is comfortable, but there are too many people," after analysis, reveals that it includes the beach scene, positive and negative emotions, and a transitional logical relationship. This reflects the user's preference for the beach environment and their expectation of fewer people. Therefore, it can be categorized under the latent need category of "environmental preference." Through the above classification process, both explicit need candidate elements and implicit need candidate cues have a unified classification standard, forming explicit need elements and implicit need cues respectively.

[0061] Step S125: Associate explicit requirement elements and implicit requirement clues within the same user expression content unit with the corresponding timestamps according to the order of their appearance in the statements, and generate a requirement hierarchy parsing result containing time context information. The time context information is used to record the sequential relationship between explicit requirement elements and implicit requirement clues in the user expression process.

[0062] After obtaining explicit requirement elements and implicit requirement cues with a unified classification standard, they need to be associated with their corresponding timestamps. Explicit requirement elements and implicit requirement cues within the same user expression unit are generated during the user's continuous expression process. Each element and cue corresponds to a part of the original statement, and the original statement carries a corresponding timestamp.

[0063] Based on the order in which the statements appear, explicit requirement elements and implicit requirement cues are associated with the timestamps of the statements they appear in. For example, in a user's content unit, the statements in the order of their timestamps are "I want to find a hotel near the sea," "Preferably with a swimming pool," and "The environment there should be very nice." Here, "hotel" and "swimming pool" are explicit requirement elements, while "seaside" and "very nice environment" are implicit requirement cues. They are associated with the timestamps of their respective statements.

[0064] Through the above associations, the generated requirement hierarchy analysis results include temporal context information, which clearly records the sequential relationship between explicit requirement elements and implicit requirement cues in the user's expression process. This sequential relationship reflects the logical order of user requirement expression, helping to understand the generation and development process of user requirements.

[0065] Step S130: Perform demand association modeling on the explicit demand elements and the implicit demand clues to generate a demand association feature set, which includes the semantic connection relationship and association strength information between explicit and implicit demands.

[0066] After completing the requirement hierarchy analysis and obtaining explicit requirement elements and implicit requirement cues, the next step is to perform requirement association modeling on these two. The purpose of requirement association modeling is to uncover the inherent connections between explicit requirement elements and implicit requirement cues, clarifying their semantic connections and the tightness of the association, i.e., the association strength information. Through the above modeling process, the scattered explicit and implicit requirements can be integrated, thereby gaining a deeper understanding of the structure and characteristics of user needs.

[0067] Step S131: Construct a demand association graph model, taking the explicit demand elements as the first type of nodes in the graph model and the implicit demand clues as the second type of nodes in the graph model. The first type of nodes and the second type of nodes correspond to different demand feature types.

[0068] In the demand association graph model, nodes represent different types of demand characteristics. Explicit demand elements are defined as the first type of node, and implicit demand cues are defined as the second type of node.

[0069] For example, explicit demand elements such as "hotel," "swimming pool," and "gym" are each treated as independent first-type nodes; implicit demand cues such as "seaside environment," "few people," and "comfortable environment" are each treated as independent second-type nodes. In this way, the two different types of demand characteristics are presented as nodes in the graph model, clearly distinguishing their types.

[0070] Step S132: In the demand association graph model, establish edge connections between the first type of nodes and the second type of nodes based on the semantic logical relationship of the user's expressed content units. The semantic logical relationship includes modification, causal or parallel relationship. The weight value of the edge connection is determined by the co-occurrence frequency of explicit demand elements and implicit demand clues in the same semantic context.

[0071] In the demand association graph model, the edges connecting nodes represent the relationships between explicit demand elements and implicit demand cues. These edges are established based on the semantic and logical relationships in the user-expressed content units, including modification relationships, causal relationships, and parallel relationships.

[0072] A modifying relationship refers to one requirement feature describing or defining another. For example, in "a hotel by the sea," the implicit requirement cue "by the sea" (a second-type node) modifies the explicit requirement element "hotel" (a first-type node), thus establishing an edge connection between them. A causal relationship refers to one requirement feature being caused by another. For example, in "I need a hotel near the sea because I want to take a walk on the beach," the implicit requirement cue "walk on the beach" (a second-type node) and the explicit requirement element "hotel near the sea" (a first-type node) have a causal relationship, establishing a corresponding edge connection. A parallel relationship refers to two requirement features being of equal importance, jointly expressing a user's need. For example, in "I want a hotel with a swimming pool and a quiet surrounding environment," the explicit requirement element "swimming pool" (a first-type node) and the implicit requirement cue "quiet environment" (a second-type node) are parallel, establishing an edge connection.

[0073] The weight of an edge connection reflects the strength of the association between explicit requirement elements and implicit requirement cues, and its magnitude is determined by the co-occurrence frequency of the two in the same semantic context. Co-occurrence frequency refers to the number of times explicit requirement elements and implicit requirement cues appear simultaneously in the same semantic context (such as the same statement or adjacent statements) across multiple user-expressed content units. A higher co-occurrence frequency indicates a stronger association between the two, and a larger edge connection weight; conversely, a lower co-occurrence frequency results in a smaller weight.

[0074] Step S1321: Perform semantic dependency analysis on the user expression content unit, and extract the modification relationship, causal relationship and parallel relationship between the basic word segmentation units as semantic logical relationship. The semantic logical relationship is used to determine the association direction between demand features.

[0075] Semantic dependency analysis of user-expressed content units aims to accurately extract the semantic logical relationships between basic word segmentation units. Semantic dependency analysis determines the dependency relationships between words in a sentence by analyzing the grammatical and semantic connections between them, thereby identifying modification relationships, causal relationships, and parallel relationships.

[0076] For example, in the statement "I want a hotel near the sea and with a swimming pool," semantic dependency analysis reveals that both "near the sea" and "with a swimming pool" modify "hotel," indicating a modifying relationship. Simultaneously, "near the sea" and "with a swimming pool" are parallel relationships. These extracted semantic logical relationships will be used to determine the direction of association between explicit demand elements and implicit demand cues, that is, to clarify which demand feature is associated with which demand feature.

[0077] Step S1322: Based on the semantic logic relationship, determine the semantic association direction between explicit requirement elements and implicit requirement clues. The association direction includes the guiding relationship between explicit requirement elements and implicit requirement clues or the supplementary relationship between implicit requirement clues and explicit requirement elements.

[0078] Based on the extracted semantic logic relationships, the direction of association between explicit demand elements and implicit demand cues is further determined. If it is a modification relationship, the implicit demand cue usually plays a supplementary role to the explicit demand element, that is, the direction of association is from the implicit demand cue (second type node) to the explicit demand element (first type node). For example, in "hotel by the sea", "by the sea" (implicit demand cue) supplements the description of "hotel" (explicit demand element).

[0079] If it's a causal relationship, there might be two directions of association. If the explicit demand element is caused by the implicit demand cues, the association direction is from the implicit demand cues to the explicit demand element; if the implicit demand cues are caused by the explicit demand element, the association direction is from the explicit demand element to the implicit demand cues. For example, in the statement "Because the hotel has a swimming pool, I feel very satisfied," "having a swimming pool" (explicit demand element) leads to "being very satisfied" (implicit demand cues), and the association direction is from the explicit demand element to the implicit demand cues.

[0080] In a parallel relationship, the association direction can be bidirectional because both are of equal status and interconnected. By determining the association direction, the edge connections in the demand association graph model become directional, more accurately reflecting the semantic relationships between demand features.

[0081] Step S1323: Based on the association direction, establish directed edge connections from the first type of node to the second type of node or from the second type of node to the first type of node in the demand association graph model.

[0082] Based on the determined association direction, directed edges are established in the demand association graph model. If the association direction is a complementary relationship between implicit demand cues and explicit demand elements, a directed edge is established from the second type of node (implicit demand cues) to the first type of node (explicit demand elements); if the association direction is a guiding relationship between explicit demand elements and implicit demand cues, a directed edge is established from the first type of node to the second type of node; if it is a bidirectional parallel relationship, a bidirectional directed edge is established between the two nodes.

[0083] For example, in the phrase "a hotel by the sea," the association direction is "seaside" (a second-type node) supplementing "hotel" (a first-type node), thus a directed edge is established from the "seaside" node to the "hotel" node. In the phrase "the hotel has a swimming pool, so I'm very satisfied," the association direction is "swimming pool" (a first-type node) guiding "very satisfied" (a second-type node), thus a directed edge is established from the "swimming pool" node to the "very satisfied" node. In the phrase "the hotel has a swimming pool and a quiet environment," "swimming pool" (a first-type node) and "quiet environment" (a second-type node) are parallel, thus a bidirectional directed edge is established between them. The establishment of directed edges allows the demand association graph model to more accurately reflect the semantic association direction between demand features.

[0084] Step S1324: Calculate the time interval between the appearance of the explicit demand elements and the implicit demand clues in the user's expressed content unit, and generate a time correlation parameter. The time correlation parameter is used to measure the degree of correlation between demand features in the time dimension.

[0085] The time interval between the appearance of explicit requirement elements and implicit requirement cues in user-expressed content units is calculated using their associated timestamps. For example, if the timestamp of a statement containing an explicit requirement element is t1, and the timestamp of the statement containing the corresponding implicit requirement cue is t2, then the time interval is the absolute value of the difference between t2 and t1.

[0086] The time correlation parameter is inversely proportional to the time interval. The smaller the time interval, the closer the explicit demand elements and implicit demand cues appear in the user's expression process, the closer their correlation in the time dimension, and the larger the value of the time correlation parameter. Conversely, the larger the time interval, the smaller the value of the time correlation parameter.

[0087] Step S1325: Calculate the weight value of the directed edge connection by combining the time correlation parameter and the co-occurrence frequency. The weight value increases as the time correlation parameter decreases and the co-occurrence frequency increases.

[0088] The weight values ​​of directed edge connections are calculated by combining the temporal correlation parameter and the co-occurrence frequency. The calculation method involves standardizing the temporal correlation parameter and the co-occurrence frequency separately to bring them into the same numerical range, then assigning a certain weight ratio to each, and finally combining the standardized temporal correlation parameter and the co-occurrence frequency according to this ratio to obtain the final weight value.

[0089] For example, assuming the standardized value of the time correlation parameter is A, the standardized value of the co-occurrence frequency is B, the weight assigned to the time correlation parameter is 0.4, and the weight assigned to the co-occurrence frequency is 0.6, then the weight value is equal to A multiplied by 0.4 plus B multiplied by 0.6. This calculation method allows the weight value to simultaneously reflect the impact of time correlation and co-occurrence frequency on the degree of correlation between demand features. Furthermore, the weight value increases with the increase of the time correlation parameter (i.e., the time interval decreases) and the increase of the co-occurrence frequency, thus more accurately reflecting the strength of the correlation between demand features.

[0090] Step S133: Perform community detection processing on the demand association graph model, and identify node clusters connected by high-weight edges through clustering algorithm. The node clusters represent combinations of explicit demand elements and implicit demand clues with strong correlation.

[0091] Community detection aims to find sets of nodes with strong relationships, i.e., node clusters, within a demand association graph model. The clustering algorithm used can be a graph-based clustering algorithm, such as the label propagation algorithm or the Louvain algorithm.

[0092] Clustering algorithms analyze the weight values ​​of directed edges, grouping nodes connected by high-weight edges together to form node clusters. For example, nodes like "hotel," "beach," and "swimming pool" have high weight values ​​for directed edges, indicating a strong correlation between them, and the clustering algorithm will identify them as one node cluster. Similarly, nodes like "gym" and "quiet environment" that are also connected by high-weight edges will form another node cluster. Each node cluster represents a combination of explicit and implicit demand cues with strong correlations, reflecting a user's comprehensive needs in a particular area.

[0093] Step S134: Extract the connection path features between the first type of nodes and the second type of nodes in the node cluster, and generate the semantic connection relationship between explicit requirements and implicit requirements. The semantic connection relationship includes the logical transmission path between requirement features.

[0094] The connection path characteristics between the first type of nodes (explicit demand elements) and the second type of nodes (implicit demand clues) in a node cluster refer to the path patterns formed by the directed edges connecting the nodes. For example, in a certain node cluster, there exists a directed edge connection path such as "Seaside" (second type node) → "Hotel" (first type node) → "Swimming Pool" (first type node), or a path such as "Swimming Pool" (first type node) ← "Hotel" (first type node) ← "Comfortable Environment" (second type node).

[0095] Extracting these connection path features allows us to generate semantic connections between explicit and implicit requirements, including logical transmission paths between requirement features. These logical transmission paths reflect the semantic order and logical relationships between requirement features. For example, the path from "seaside" to "hotel" and then to "swimming pool" reflects a user's logical requirement for a hotel near the sea with a swimming pool.

[0096] Step S135: Statistically analyze the weight distribution of edge connections in the node cluster to generate association strength information between explicit and implicit requirements. The association strength information includes the degree of correlation between requirement features.

[0097] The weights of all directed edges in the node cluster are statistically analyzed to determine their distribution, including the range, average, median, and mode of the weights. These statistical results constitute information about the strength of the association between explicit and implicit demands.

[0098] For example, if the weight values ​​of edges in a node cluster are mostly concentrated in a high range with a large average, it indicates a high overall correlation between explicit demand elements and implicit demand cues within that node cluster. Conversely, if the weight values ​​in another node cluster are more dispersed with a smaller average, it indicates a relatively low correlation between demand features within that cluster. Correlation strength information quantifies the degree of correlation between demand features and is an important component of the demand correlation feature set.

[0099] Step S136: Merge the semantic connection relationship and the association strength information to generate a set of demand association features.

[0100] The extracted semantic connections and statistically obtained association strength information are combined to form a set of demand association features. Semantic connections describe the logical transmission path and semantic association mode between explicit and implicit demands, while association strength information quantifies the degree of their association. The two complement each other and comprehensively reflect the association characteristics between explicit demand elements and implicit demand cues.

[0101] For example, a set of demand-related features may contain not only the semantic connection relationship of "seaside → hotel → swimming pool" mentioned above, but also information on the strength of association, such as the distribution of weight values ​​of each edge connection along the path.

[0102] Step S140: Based on the demand association feature set, perform demand evolution analysis and processing to generate a demand evolution feature set. The demand evolution feature set includes the continuity of demand elements changing over time and the association rules for the transformation of implicit demand cues into explicit demand elements.

[0103] After obtaining the set of demand-related features, it is necessary to perform demand evolution analysis. The purpose of demand evolution analysis is to uncover the changing patterns of demand features over time, including the changes in explicit demand elements and implicit demand cues over time, and how implicit demand cues transform into explicit demand elements, and how explicit demand elements transform into implicit demand cues. Through the above analysis, we can grasp the dynamic changing trends of user needs.

[0104] Step S141: Arrange the demand association feature set in timestamp order to construct a time series demand association feature matrix, wherein the time series demand association feature matrix contains demand association feature information at different time points.

[0105] First, the features in the demand association feature set are arranged in chronological order according to the timestamps they are associated with. Each time point corresponds to a set of demand association features, which include explicit demand elements, implicit demand clues, semantic connections between them, and information on the strength of association at that time point.

[0106] Then, these demand-related features arranged chronologically are organized into a matrix, namely the time-series demand-related feature matrix. The rows of this matrix represent different time points, and the columns represent different dimensions of demand-related features. Each element in the matrix is ​​a specific feature value corresponding to the given time point and feature dimension. For example, a row corresponding to time t might contain information such as the semantic connection and strength of the explicit demand element "hotel" and the implicit demand cue "seaside" at time t. The time-series demand-related feature matrix illustrates the distribution of demand-related features over time.

[0107] Step S142: Perform sliding window processing on the time series demand association feature matrix to extract a subset of demand association features within each time window, wherein the time window contains a preset number of consecutive time points.

[0108] Sliding window processing refers to setting a window containing a preset number of consecutive time points, and then sliding this window sequentially along the timeline. Each slide captures the demand-related features within the window, forming a corresponding subset of demand-related features. In this embodiment, the preset number of consecutive time points can be determined based on the time span and data volume of the text data set. For example, if timestamps are recorded in days and the text data set covers a long time range, the window can be set to contain seven consecutive time points, which is one week.

[0109] In practice, the starting time point of the sliding window is first determined. This starting time point becomes the first time point of the window. A predetermined number of consecutive time points are selected to form the first window, and all demand-related features within this window are extracted to form the first subset of demand-related features. Next, the window is slid forward one time point, at which point the starting time point becomes the original second time point. Similarly, a predetermined number of consecutive time points are selected to form a new window, and demand-related features are extracted from it to form the second subset of demand-related features. This process of sliding the window continues until the last time point of the window reaches the last time point of the time series demand-related feature matrix.

[0110] For example, a time-series demand correlation feature matrix contains demand correlation feature information from time point 1 to time point 30, and the sliding window is set to contain five consecutive time points. Then, the first window contains time points 1 to 5, and the demand correlation features of these five time points are extracted to form the first subset of demand correlation features; the second window contains time points 2 to 6, and the second subset is formed; and so on, until the last window contains time points 26 to 30, which forms the last subset of demand correlation features. Through this sliding window processing, continuous time-series data can be divided into multiple subsequences with overlapping parts, each subsequence representing the demand correlation features within a specific time period.

[0111] Step S143: Perform feature change rate calculation on the subset of demand-related features, statistically analyze the addition, disappearance, and frequency changes of explicit demand elements and the appearance, transfer, and intensity changes of implicit demand clues, and generate demand element change features. The demand element change features contain dynamic change information of demand features in the time dimension.

[0112] After obtaining the subset of demand-related features within each time window, it is necessary to calculate the feature change rate. First, for explicit demand elements, it is necessary to count the additions between different windows, i.e., an explicit demand element exists in the current window but not in the previous window; this is a newly added explicit demand element. It is also necessary to count the disappearances, i.e., an explicit demand element existed in a previous window but not in the current window; this is a disappeared explicit demand element. Simultaneously, it is also necessary to count the frequency changes of explicit demand elements appearing in the windows, i.e., whether the number of times the same explicit demand element appears in the current window has increased or decreased compared to the number of times it appears in the previous window, and by what magnitude.

[0113] For implicit demand cues, we need to track their occurrence, i.e., when an implicit demand cue first appears in the current window; track their transfer, i.e. when an implicit demand cue is transferred in different scenarios or associated with different explicit demand elements; and track their intensity changes, i.e., whether the strength of the association between the implicit demand cue and the explicit demand element changes in different windows, whether it has increased or decreased.

[0114] By statistically analyzing and calculating these situations, the characteristic change rate is obtained, thereby generating the demand element change characteristics. For example, in the first window, the explicit demand element "swimming pool" appears three times; in the second window, the frequency of "swimming pool" increases to five times, and the explicit demand element "hot spring" is newly added, while the explicit demand element "gym" disappears. Regarding implicit demand cues, the strength of the association between "quiet environment" and "hotel" is constant in the first window, but increases in the second window. Simultaneously, the implicit demand cue "close to attractions" shifts from being associated with "hotel" to being associated with "homestay." This information collectively constitutes the demand element change characteristics, which comprehensively record the dynamic changes of demand characteristics over time.

[0115] Step S144: Analyze the transmission pattern of the changing characteristics of the demand elements in a continuous time window, and identify the evolution path between explicit demand elements and implicit demand cues. The evolution path includes the process of implicit demand cues continuously appearing and then transforming into explicit demand elements, or the process of explicit demand elements weakening over time and then transforming into implicit demand cues.

[0116] After obtaining the characteristics of changes in demand elements, it is necessary to analyze their transmission patterns within consecutive time windows to identify the evolutionary paths between explicit demand elements and implicit demand cues. First, observe the changes in explicit and implicit demand elements across multiple consecutive time windows to find their correlations and transmission patterns.

[0117] For example, if an implicit need cue, "I want a relaxing environment," appears continuously in multiple consecutive time windows and its intensity gradually increases, then in subsequent windows, an explicit need element, "SPA services," may appear. This could form an evolutionary path where an implicit need cue is transformed into an explicit need element, that is, after "I want a relaxing environment" appears continuously, it transforms into an explicit need for "SPA services."

[0118] On the other hand, the frequency and intensity of the explicit demand element "long-distance bus" appearing in multiple windows gradually decrease, while the intensity of the implicit demand cue "convenient transportation" gradually increases. This may form an evolutionary path in which an explicit demand element is transformed into an implicit demand cue, that is, the explicit demand element "long-distance bus" weakens over time and is transformed into an implicit demand cue for "convenient transportation".

[0119] By analyzing the changing characteristics of demand elements in these continuous windows, various evolutionary paths can be clearly identified, reflecting the dynamic transformation relationship between explicit demand elements and implicit demand cues.

[0120] Step S1441: Perform time series correlation analysis on the changes in the demand elements, calculate the Pearson correlation coefficient between the changes in explicit demand elements and the changes in implicit demand clues, and determine the correlation strength between the two in the time dimension.

[0121] When analyzing the transmission patterns of changes in demand factors within a continuous time window, the first step is to perform time-series correlation analysis on these changes. Specifically, this involves treating changes in explicit demand factors and implicit demand cues as two separate time series, and then calculating the Pearson correlation coefficient between these two time series.

[0122] The Pearson correlation coefficient measures the degree of linear correlation between two variables, with values ​​ranging from negative one to positive one. When the correlation coefficient is close to one, it indicates a strong positive correlation between the two time series, meaning that the changes in explicit demand factors and implicit demand clues are basically in the same trend; when the correlation coefficient is close to negative one, it indicates a strong negative correlation, meaning that the changes in the two are opposite; when the correlation coefficient is close to zero, it indicates that there is almost no linear correlation between the two.

[0123] For example, we can calculate the Pearson correlation coefficient between the time series of changes in the explicit demand element "swimming pool" and the implicit demand cue "summer retreat". If the result is 0.8, it indicates a strong positive correlation between the two over time, meaning that the change trend of the implicit demand cue "summer retreat" is relatively consistent with the change trend of the explicit demand element "swimming pool". By calculating the Pearson correlation coefficient, we can quantitatively determine the strength of the correlation between changes in explicit demand elements and changes in implicit demand cues.

[0124] Step S1442: Select explicit demand elements and implicit demand clue pairs with a correlation strength greater than the set correlation strength threshold as candidate objects for evolution paths.

[0125] After calculating the Pearson correlation coefficient between changes in explicit demand elements and changes in implicit demand cues, a correlation strength threshold needs to be set. This threshold can be determined based on the actual analytical needs and data characteristics, with the aim of filtering out pairs of explicit demand elements and implicit demand cues with strong correlations.

[0126] For example, the correlation strength threshold can be set to 0.6. For all calculated Pearson correlation coefficients, when the correlation coefficient between an explicit demand element and an implicit demand clue is greater than 0.6, it indicates that the correlation between them is strong. This pair of explicit demand elements and implicit demand clues is then selected as candidate evolutionary paths.

[0127] For example, the correlation coefficient between the explicit demand element "hot springs" and the implicit demand cue "winter heating" is 0.75, which is greater than the set threshold of 0.6. Therefore, this pair is selected as a candidate for an evolutionary path. However, the correlation coefficient between the explicit demand element "gym" and the implicit demand cue "parent-child activities" is 0.4, which is less than the threshold, so it is not selected as a candidate. Through the above screening, we can focus on demand element pairs with strong correlations, reduce the workload of subsequent analysis, and improve the efficiency and accuracy of identifying evolutionary paths.

[0128] Step S1443: Perform longitudinal tracking analysis on the candidate objects, extract the starting time point, duration and intensity change curve of the continuous appearance of implicit demand clues in the continuous time window, and at the same time extract the time point of addition or enhancement and intensity change curve of the corresponding explicit demand elements after the appearance of implicit demand clues.

[0129] For the selected evolutionary path candidates, longitudinal tracking analysis is required. Longitudinal tracking analysis refers to tracking and observing continuous time windows along the time axis, and recording in detail the changes in implicit demand cues and explicit demand elements in the candidate objects.

[0130] For implicit demand clues, we need to extract the starting time point of their continuous appearance, that is, the time point when the implicit demand clue first appears in the window; extract the duration, that is, the number of windows that the implicit demand clue goes through from the starting time point to the time point when the implicit demand clue no longer appears; at the same time, we need to draw its intensity change curve, with the time window as the horizontal axis and the intensity value of the implicit demand clue as the vertical axis, to intuitively show the change of the intensity of the implicit demand clue in the continuous time window.

[0131] At the same time, for the corresponding explicit demand elements, we need to extract the new time point after the appearance of the implicit demand clue, that is, the time point when the explicit demand element first appears, which is after the starting time point of the implicit demand clue; or extract the enhancement time point, that is, the time point when the intensity of the explicit demand element begins to increase significantly; similarly, we should draw its intensity change curve to show the change of the intensity of the explicit demand element in a continuous time window.

[0132] For example, the candidate objects are the implicit demand cue of "close to natural landscapes" and the explicit demand element of "hiking equipment". Through longitudinal tracking, it was found that the starting point of the implicit demand cue of "close to natural landscapes" is window three, and its duration is five windows. Its intensity change curve gradually rises from window three to window five, and gradually declines from window five to window seven. The corresponding explicit demand element of "hiking equipment" newly appears in window five, and its intensity change curve continuously rises from window five to window seven. Through the above extraction and analysis, the temporal relationship between implicit demand cues and explicit demand elements can be understood.

[0133] Step S1444: Compare the change curves of the implicit demand cues and explicit demand elements, identify the time delay parameters and intensity mapping relationship that trigger the addition or enhancement of explicit demand elements after the continuous appearance of implicit demand cues, and generate a positive evolution path for the transformation of implicit demand cues into explicit demand elements.

[0134] After extracting the intensity change curves of implicit demand cues and explicit demand elements, these two curves are compared and analyzed. By comparing them, we can observe when explicit demand elements appear or increase after the implicit demand cues have been continuously appearing, thereby determining the time delay parameter, that is, the number of time windows from the beginning of the continuous appearance of the implicit demand cues to the addition or increase of explicit demand elements.

[0135] At the same time, we analyze the mapping relationship between the changes in the intensity of implicit demand cues and the changes in the intensity of explicit demand elements during this period. For example, how much does the intensity of explicit demand elements increase correspondingly when the intensity of implicit demand cues increases by a certain amount? The above mapping relationship can be determined by the changing trends and magnitudes of the two curves.

[0136] Based on the established time delay parameters and intensity mapping relationships, a forward evolutionary path is generated to transform implicit demand cues into explicit demand elements. For example, the implicit demand cue "want to try local specialties" appears continuously starting from window two. In window four, the explicit demand element "local restaurant recommendations" appears, with a time delay parameter of two windows. By comparing the intensity change curves of the two, it is found that for every unit increase in the intensity of "want to try local specialties," the intensity of "local restaurant recommendations" increases by 0.8 units. This is the intensity mapping relationship between them. Based on this information, a forward evolutionary path is generated to transform the implicit demand cue "want to try local specialties" into the explicit demand element "local restaurant recommendations."

[0137] Step S1445: Extract the starting time point, duration, and intensity decay curve of the decrease in frequency of explicit demand elements in a continuous time window. At the same time, extract the time point and intensity change curve of the appearance or enhancement of the corresponding implicit demand cues after the decay of explicit demand elements. Identify the time delay parameters and intensity mapping relationship of the appearance or enhancement of implicit demand cues triggered by the weakening of explicit demand elements. Generate the reverse evolution path of the transformation of explicit demand elements into implicit demand cues.

[0138] In addition to the forward evolution path, it is also necessary to identify the reverse evolution path. The reverse evolution path refers to the process by which explicit demand elements weaken and transform into implicit demand cues. In this process, firstly, the starting point of the decrease in the frequency of explicit demand elements is extracted, that is, the first time window in which the frequency of explicit demand elements begins to decline; secondly, the duration is extracted, that is, the number of time windows from the starting point to the time window in which the frequency of explicit demand elements drops to a low level or disappears; finally, an intensity decay curve is plotted, with the time window as the horizontal axis and the intensity value of the explicit demand element as the vertical axis, showing the gradual decline in its intensity.

[0139] Simultaneously, extract the time point when the corresponding implicit demand cues appear after the decay of explicit demand elements, that is, the time window when the implicit demand cues first appear during or after the decay of explicit demand elements; or extract the enhancement time point, that is, the time window when the intensity of the implicit demand cues begins to increase significantly; and draw its intensity change curve to show the change in the intensity of the implicit demand cues in a continuous time window.

[0140] By comparing the intensity decay curves of explicit demand elements and the intensity change curves of implicit demand cues, we can identify the time delay parameter that triggers the appearance or enhancement of implicit demand cues after the weakening of explicit demand elements, i.e., the number of time windows from the starting point of the decrease in the frequency of explicit demand elements to the appearance or enhancement of implicit demand cues; and the intensity mapping relationship, i.e., the correspondence between the magnitude of the intensity decay of explicit demand elements and the magnitude of the increase in the intensity of implicit demand cues.

[0141] For example, the explicit demand factor for "long-distance buses" decreases in frequency starting from window ten, lasting for four windows, and its intensity decay curve shows a gradually decreasing trend. In window twelve, the implicit demand cue for "convenient transportation" begins to strengthen, and its intensity change curve shows an upward trend. The time delay parameter is two windows, and for every unit decrease in the intensity of "long-distance buses," the intensity of "convenient transportation" increases by 0.6 units. Based on this information, a reverse evolution path is generated to transform the explicit demand factor for "long-distance buses" into the implicit demand cue for "convenient transportation."

[0142] Step S1446: Merge the forward evolution path and the reverse evolution path to generate a set of evolution paths between explicit demand elements and implicit demand cues.

[0143] After identifying the forward and reverse evolution paths, these two types of evolution paths are merged. The merging process involves aggregating all forward and reverse evolution paths together to form a complete set of evolution paths between explicit demand elements and implicit demand cues.

[0144] This set encompasses various possible transformation relationships of demand elements, including paths where implicit demand cues transform into explicit demand elements, and vice versa. For example, forward evolution paths include "desire for a relaxing environment" transforming into "SPA services," and "desire to try local cuisine" transforming into "local restaurant recommendations." Reverse evolution paths include "long-distance bus" transforming into "convenient transportation," and "standard room" transforming into "comfortable accommodation." By merging these paths, the evolution path set comprehensively reflects the dynamic transformation between explicit and implicit demand cues over time.

[0145] Step S145: Summarize the triggering conditions and transmission rules of the evolution path, generate the continuity pattern of demand elements changing over time and the association rules for the transformation of implicit demand cues into explicit demand elements. The continuity pattern includes the stable change pattern of demand features in the time dimension, and the association rules include the preconditions and transmission methods for the transformation of demand features.

[0146] After obtaining the set of evolutionary paths, it is necessary to summarize the triggering conditions and transmission rules of these paths. Triggering conditions refer to the specific situations or factors that cause an evolutionary path to begin, while transmission rules refer to the laws and methods followed by the transformation between demand elements in the evolutionary path.

[0147] To understand the continuous patterns of demand factors changing over time, we can analyze the common characteristics of multiple evolutionary paths within a set of evolutionary paths to identify stable patterns of change in demand features over time. For example, analysis reveals that explicit demand factors related to "accommodation booking" gradually increase before the peak tourist season and gradually decrease after the peak season ends. This is a continuous pattern, reflecting a stable pattern of change in the demand characteristics of "accommodation booking" over time.

[0148] For the association rules governing the transformation of implicit demand cues into explicit demand elements, we summarize the preconditions and transmission methods for each transformation path. Preconditions refer to the conditions that must be met for an implicit demand cue to be transformed into an explicit demand element. For example, the prerequisite for the implicit demand cue "wanting a relaxing environment" to be transformed into the explicit demand element "SPA service" might be that the intensity of the implicit demand cue reaches a certain level and persists for a certain duration. Transmission methods refer to the way demand elements interact during the transformation process. For example, through repeated mentions of relevant scenarios by users, implicit demands gradually become explicit demands.

[0149] Based on the above summary, we can generate continuity patterns and correlation rules, which can help us understand the evolution mechanism of demand elements.

[0150] Step S146: Merge the continuity pattern and the association rule to generate a set of demand evolution features.

[0151] The continuous patterns of demand elements changing over time and the correlation rules for the transformation of implicit demand cues into explicit demand elements are combined to form a set of demand evolution characteristics. This set integrates the stable change patterns of demand elements over time, as well as the preconditions and transmission methods of demand characteristic transformation, comprehensively reflecting the evolutionary characteristics of demand.

[0152] For example, continuity patterns include "increased demand for accommodation before the peak tourist season" and "fluctuations in transportation demand during holidays"; association patterns include "the continuous demand for relaxing environments triggers demand for SPA services" and "repeated mentions of specialty food scenes lead to demand for restaurant recommendations".

[0153] Step S150: Perform demand forecasting processing based on the demand evolution feature set to generate tourism product demand forecasting results. The demand forecasting results include the distribution of user demand priorities for different categories of tourism products and a description of demand change trends.

[0154] After obtaining the set of demand evolution characteristics, demand forecasting can be performed based on this set. The purpose of demand forecasting is to predict users' future demand for various types of tourism products based on existing demand evolution patterns. In this embodiment, demand forecasting is carried out in a series of orderly steps. Through in-depth analysis and processing of the demand evolution characteristic set, a prediction result is finally obtained, which includes a description of the priority distribution and changing trends of users' demand for different categories of tourism products.

[0155] Step S151: Perform pattern matching processing on the continuity patterns and association rules in the set of demand evolution features, identify the degree of matching between the demand features at the current time point and the historical evolution patterns, and generate a pattern matching similarity value.

[0156] In this embodiment, pattern matching is performed first. The core of this step is to compare the current demand characteristics with historical demand evolution patterns to determine their similarity. Continuity patterns reflect stable changes in demand characteristics over time; for example, certain explicit demand elements may show a continuous strengthening or weakening trend within a specific time period. Association rules, on the other hand, reflect the preconditions and transmission methods for the transformation between demand characteristics. For instance, when a certain implicit demand clue appears continuously for a certain period, it can transform into a specific explicit demand element.

[0157] When performing pattern matching, it is necessary to first extract the continuity patterns and association rules from the set of demand evolution characteristics at the current point in time to form a description of the current demand characteristics. Simultaneously, various historical evolution patterns are extracted from historical data, each containing corresponding continuity patterns and association rules. Then, by comparing the current demand characteristics with each historical evolution pattern in terms of the manifestation of continuity patterns and the triggering conditions and transmission methods of association rules, the degree of matching between them is calculated. This degree of matching is presented in the form of a pattern matching similarity value; the higher the similarity value, the more similar the current demand characteristics are to that historical evolution pattern.

[0158] Step S1511: Construct a historical evolution pattern library, which contains a set of demand evolution characteristics and their corresponding actual demand results in different historical time periods.

[0159] Building a historical evolution pattern library is fundamental to pattern matching processing. During this process, sets of demand evolution characteristics from different historical time periods can be collected. These sets are obtained by processing user travel-related text data from different past periods using the same demand hierarchy analysis, association modeling, and evolutionary analysis methods as described above. Simultaneously, each historical time period's demand evolution characteristic set corresponds to a record of the actual travel product demand results generated by users after that time period ended. These actual demand results can include information such as the types of travel products actually booked by users and their preferences for various types of travel products.

[0160] For example, user text data from a past quarter is processed to obtain a set of demand evolution characteristics for that quarter. Simultaneously, the actual demand results—such as the type of hotel booked, the mode of transportation chosen, and the attractions visited—are recorded after the quarter ends. These two sets are then linked and stored in a historical evolution pattern library. Through this method, the historical evolution pattern library will accumulate a large number of demand evolution patterns from different historical periods and their corresponding actual results.

[0161] Step S1512: Perform feature extraction processing on the demand evolution feature set at the current time point to generate a current feature vector containing continuous regularity parameters and correlation rule parameters.

[0162] When extracting features from the current set of demand evolution characteristics, it is important to focus on continuity patterns and association rules. Continuity pattern parameters are quantitative representations of these patterns, reflecting various attributes of stable change patterns in demand characteristics over time, such as the frequency, magnitude, and duration of these changes. Association rule parameters are quantitative representations of these rules, reflecting the specific characteristics of the preconditions for demand feature transformation and the characteristics of their transmission methods. Examples include the triggering conditions required for implicit demand cues to transform into explicit demand elements, and the transmission speed during the transformation process.

[0163] By extracting and organizing these parameters, they are arranged and combined in a certain order to form the current feature vector. The current feature vector is a multi-dimensional vector, with each dimension corresponding to a specific continuous pattern parameter or association rule parameter, which can comprehensively and accurately represent the required characteristics at the current time point.

[0164] Step S1513: Perform feature extraction processing on each historical evolution pattern in the historical evolution pattern library to generate a historical feature vector containing continuity regularity parameters and association rule parameters.

[0165] Similar to the method used to extract the current feature vector, feature extraction is also performed on each historical evolution pattern in the historical evolution pattern library. For the set of required evolutionary features contained in each historical evolution pattern, the continuity pattern parameters corresponding to the continuity patterns and the association rule parameters corresponding to the association rules are extracted. The extraction criteria and methods for these parameters are consistent with those in the current feature vector to ensure comparability between the two.

[0166] The continuous pattern parameters and association rule parameters of each extracted historical evolution pattern are arranged and combined in the same order as the current feature vector to generate a historical feature vector corresponding to each historical evolution pattern. Each historical feature vector is also multi-dimensional, and its number of dimensions and the meaning of the parameters represented by each dimension are completely consistent with the current feature vector.

[0167] Step S1514: Calculate the cosine similarity between the current feature vector and each historical feature vector, and generate a pattern matching similarity value between each historical evolution pattern and the required feature at the current time point.

[0168] After obtaining the current feature vector and each historical feature vector, the cosine similarity between them is calculated. Cosine similarity is an indicator that measures the degree of similarity between two vectors in a direction. It is calculated by taking the ratio of the dot product of the two vectors to the product of their magnitudes. In this embodiment, the cosine similarity between the current feature vector and each historical feature vector is calculated in the above manner. The result is the pattern matching similarity value between each historical evolution pattern and the required feature at the current time point.

[0169] The pattern matching similarity value ranges from 0 to 1. The closer the value is to 1, the more similar the current demand characteristics are to the historical evolution pattern; the closer the value is to 0, the lower the similarity. These calculated pattern matching similarity values ​​can reflect the similarity between current demand characteristics and different historical evolution patterns.

[0170] Step S152: Based on the pattern matching similarity value, perform weighted fusion processing on the historical evolution pattern to generate a demand feature evolution prediction model for the future time period. The demand feature evolution prediction model includes the probability distribution of changes in demand elements and the probability distribution of implicit demand clue transformation.

[0171] In this embodiment, after obtaining the pattern matching similarity value, a weighted fusion process is performed. The core of the weighted fusion process is to determine the weight of each historical evolution pattern in the fusion process based on its pattern matching similarity value with the current demand feature. The higher the pattern matching similarity value of a historical evolution pattern, the higher its similarity to the current demand feature, and the greater its predictive reference value for future demand evolution; therefore, it will be assigned a higher weight. Conversely, historical evolution patterns with lower pattern matching similarity values ​​will have relatively lower weights.

[0172] After determining the weights, the various historical evolution patterns are merged according to their corresponding weights. During the fusion process, the demand evolution patterns contained in different historical evolution patterns can be comprehensively considered, and these patterns can be integrated and adjusted to form a predictive model that can reflect the possible evolution of demand characteristics in the future time period, i.e., the demand characteristic evolution prediction model.

[0173] The model includes a probability distribution of changes in demand elements, which describes the likelihood of different demand elements undergoing various changes (such as enhancement, weakening, appearance, disappearance, etc.) in the future; it also includes a probability distribution of the transformation of implicit demand cues, which reflects the likelihood of different implicit demand cues transforming into specific explicit demand elements in the future, as well as the likelihood of explicit demand elements transforming into implicit demand cues.

[0174] Step S153: Calculate the demand priority distribution of users for different categories of tourism products using the demand feature evolution prediction model. The demand priority distribution is determined by the frequency of occurrence and correlation strength of demand elements.

[0175] When calculating the demand priority distribution using the demand feature evolution prediction model, the first focus is on the demand elements related to different categories of tourism products in the model. For each category of tourism product, the frequency of occurrence of its corresponding demand element in the future time period can be statistically analyzed. The higher the frequency, the stronger the user's demand for that type of tourism product in terms of that demand element is likely to be.

[0176] Simultaneously, the strength of the correlation between these demand elements will be analyzed. A high degree of correlation between demand elements often indicates a more prominent overall demand from users for related tourism products. By comprehensively considering the frequency of occurrence and the strength of correlation of demand elements, different categories of tourism products will be prioritized. The prioritization method involves comprehensively evaluating the frequency of occurrence and the strength of correlation according to certain rules to obtain a comprehensive score for each category of tourism products, and then determining its position in the demand priority distribution based on the comprehensive score.

[0177] For example, for hotel-related tourism products, the corresponding demand elements might include "swimming pool," "free Wi-Fi," and "breakfast service." If a demand feature evolution prediction model indicates that these demand elements will appear frequently in the future and have strong interrelationships, then hotel-related tourism products are likely to rank higher in the demand priority distribution. Conversely, for certain transportation-related tourism products, if the corresponding demand elements appear less frequently and have weaker interrelationships, then their ranking in the priority distribution might be relatively lower. The resulting demand priority distribution clearly reflects the order in which users prioritize different categories of tourism products.

[0178] Step S154: Analyze the changing trends of demand elements in the demand characteristic evolution prediction model and generate a demand change trend description, which includes information on the direction of demand element enhancement, weakening or shift.

[0179] When analyzing the changing trends of demand elements, it is possible to delve into the probability distribution of changes in demand elements and the probability distribution of implicit demand cues transformation in the demand characteristic evolution prediction model. For each demand element, by observing its changes in the probability distribution, it can be determined whether it exhibits an increasing trend (i.e., a gradual increase in frequency and intensity), a decreasing trend (i.e., a gradual decrease in frequency and intensity), or a shifting trend (i.e., a transformation from one demand element to another related demand element, such as a gradual shift from demand for "budget hotels" to demand for "luxury hotels").

[0180] Simultaneously, we will also monitor the shift in demand elements between different categories of tourism products, such as whether there is a trend of users shifting their demand elements from scenic spot-related tourism products to dining-related tourism products. These analytical results will be organized and described to form a demand change trend description, which details the future direction of change for various demand elements.

[0181] Step S155: Verify the demand priority distribution and the demand change trend description, merge the verified demand priority distribution and demand change trend description, and generate tourism product demand forecast results.

[0182] Validating the demand priority distribution and demand change trend description is crucial to ensuring the accuracy and reliability of the forecast results. This validation process can involve comparing the currently obtained demand priority distribution and demand change trend description with recent historical data to check whether they match the actual demand results under similar historical circumstances.

[0183] For example, examine whether the demand priority distribution and trend descriptions obtained using similar forecasting methods during periods with similar demand characteristics in the past are consistent with the actual demand at those times. If there is a significant discrepancy, the current forecast results need to be adjusted. Additionally, the reasonableness of the forecast results can be judged by incorporating known tourism market dynamics and trends, eliminating those that are clearly inconsistent with the overall market trend.

[0184] After verification and necessary adjustments, the demand priority distribution and demand change trend descriptions are combined to form the final tourism product demand forecast. This result comprehensively reflects users' demand for different categories of tourism products over a future period, including the order of demand and the direction of change.

[0185] Figure 2 The illustration shows exemplary hardware and software components of a machine learning-based demand forecasting system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 may be used in the machine learning-based demand forecasting system 100 and to perform the functions described in this application.

[0186] For example, the machine learning-based demand forecasting system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the machine learning-based demand forecasting system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The machine learning-based demand forecasting system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0187] Furthermore, embodiments of the present invention also provide a readable storage medium having computer-executable instructions pre-set therein. When the processor executes the computer-executable instructions, the above-mentioned demand prediction method based on machine learning is implemented.

[0188] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A demand forecasting method based on machine learning, characterized in that, The method includes: Obtain a set of text data related to user travel, the set of text data contains multiple user expression content units, each of which consists of natural language statements marked with consecutive timestamps; The text data set is subjected to demand hierarchy parsing to obtain the explicit demand elements and implicit demand cues of each user's expressed content unit. The explicit demand elements are the user's clearly stated functional requirements for tourism products, and the implicit demand cues are the user's potential demand tendencies that are not directly stated but conveyed through semantic association. The explicit demand elements and the implicit demand clues are subjected to demand association modeling to generate a demand association feature set, which includes the semantic connection relationship and association strength information between explicit and implicit demands. Based on the aforementioned set of demand-related features, demand evolution analysis is performed to generate a set of demand evolution features. The set of demand evolution features includes the continuity of demand elements over time and the association rules for the transformation of implicit demand cues into explicit demand elements. Demand forecasting is performed based on the set of demand evolution characteristics to generate tourism product demand forecasting results. The demand forecasting results include the distribution of user demand priorities for different categories of tourism products and a description of demand change trends. The process of performing demand association modeling on the explicit demand elements and the implicit demand clues to generate a demand association feature set includes: Construct a demand association graph model, with the explicit demand elements as the first type of nodes in the graph model and the implicit demand clues as the second type of nodes in the graph model. The first type of nodes and the second type of nodes correspond to different demand feature types, respectively. In the demand association graph model, edge connections between the first type of nodes and the second type of nodes are established based on the semantic logical relationship of the user's expressed content units. The semantic logical relationship includes modification, causal or parallel relationship. The weight value of the edge connection is determined by the co-occurrence frequency of explicit demand elements and implicit demand clues in the same semantic context. The demand association graph model is subjected to community detection processing, and the clustering algorithm is used to identify node clusters connected by high-weight edges. The node clusters represent the combination of explicit demand elements and implicit demand clues with strong correlation. Extract the connection path features between the first type of nodes and the second type of nodes in the node cluster to generate the semantic connection relationship between explicit requirements and implicit requirements. The semantic connection relationship includes the logical transmission path between requirement features. The weight distribution of edge connections in the node cluster is statistically analyzed to generate association strength information between explicit and implicit requirements. The association strength information includes the degree of correlation between requirement features. The semantic connection relationship and the association strength information are merged to generate a set of demand association features; The step of performing demand evolution analysis based on the demand association feature set to generate a demand evolution feature set includes: Arrange the set of demand-related features in timestamp order to construct a time-series demand-related feature matrix, which contains demand-related feature information at different time points. The time series demand correlation feature matrix is ​​processed by a sliding window to extract a subset of demand correlation features within each time window, wherein the time window contains a preset number of consecutive time points; The feature change rate is calculated for the subset of demand-related features. The addition, disappearance, and frequency changes of explicit demand elements and the appearance, transfer, and intensity changes of implicit demand clues are statistically analyzed to generate demand element change features. The demand element change features contain dynamic change information of demand features in the time dimension. The transmission pattern of the changing characteristics of the demand elements in a continuous time window is analyzed, and the evolution path between explicit demand elements and implicit demand cues is identified. The evolution path includes the process of implicit demand cues continuously appearing and then transforming into explicit demand elements, or the process of explicit demand elements weakening over time and then transforming into implicit demand cues. The triggering conditions and transmission rules of the evolution path are summarized to generate the continuous pattern of demand elements changing over time and the association rules for the transformation of implicit demand cues into explicit demand elements. The continuous pattern includes the stable change pattern of demand characteristics in the time dimension, and the association rules include the preconditions and transmission methods for the transformation of demand characteristics. The continuity pattern and the association rule are combined to generate a set of demand evolution characteristics; The step of performing demand forecasting processing based on the demand evolution feature set to generate tourism product demand forecasting results includes: Pattern matching processing is performed on the continuity patterns and association rules in the set of demand evolution features to identify the degree of matching between the demand features at the current time point and the historical evolution patterns, and to generate a pattern matching similarity value. Based on the pattern matching similarity value, the historical evolution pattern is weighted and fused to generate a demand feature evolution prediction model for the future time period. The demand feature evolution prediction model includes the probability distribution of changes in demand elements and the probability distribution of transformation of implicit demand clues. The demand feature evolution prediction model is used to calculate the demand priority distribution of users for different categories of tourism products. The demand priority distribution is determined by the frequency of occurrence and correlation strength of demand elements. Analyze the changing trends of demand elements in the demand feature evolution prediction model, and generate a demand change trend description, which includes information on the direction of demand element enhancement, weakening or shift. The demand priority distribution and the demand change trend description are verified, and the verified demand priority distribution and demand change trend description are merged to generate tourism product demand forecast results.

2. The demand forecasting method based on machine learning according to claim 1, characterized in that, The process of performing a demand hierarchy parsing on the text data set to obtain the explicit demand elements and implicit demand cues for each user's expressed content unit includes: The user-expressed content unit is processed by natural language word segmentation. The continuous natural language sentence is divided into basic word segmentation units with independent semantics by using preset domain word segmentation rules, resulting in a word segmentation result set containing multiple basic word segmentation units. The word segmentation result set is subjected to explicit requirement element extraction processing. By traversing the predefined functional requirement word library and matching basic word segmentation units, basic word segmentation units directly related to the functions of tourism products are extracted as explicit requirement candidate elements. The functional requirement word library contains semantic units corresponding to the core functions of tourism products. By analyzing the sentiment attributes, logical modification relationships, and scene relevance of the remaining basic word segments in the unmatched functional demand word library, basic word segments reflecting potential user demand tendencies are extracted as implicit demand candidate clues. The sentiment attributes include positive or negative sentiment tendencies, the logical modification relationships include comparison, hypothesis, or contrast relationships, and the scene relevance includes scene descriptions related to the tourism experience. Explicit requirement candidate elements are classified into unified functional categories according to a preset functional classification system, and implicit requirement candidate clues are classified into unified potential requirement categories according to a preset potential requirement classification system, generating explicit requirement elements and implicit requirement clues with unified classification standards. Explicit requirement elements and implicit requirement clues within the same user expression content unit are associated with corresponding timestamps according to the order of their appearance in the statements, generating a requirement hierarchy parsing result containing time context information. The time context information is used to record the sequential relationship between explicit requirement elements and implicit requirement clues in the user expression process.

3. The demand forecasting method based on machine learning according to claim 2, characterized in that, The explicit requirement element extraction process for the word segmentation result set involves traversing a predefined functional requirement lexicon to match basic word segmentation units and extracting basic word segmentation units directly related to the functions of tourism products as explicit requirement candidate elements, including: Construct a functional appeal terminology library, which contains multi-level semantic units corresponding to the core functions of tourism products. The multi-level semantic units include first-level functional categories and second-level functional subcategories. The word segmentation result set is subjected to multiple rounds of matching processing. The first round matches the semantic units corresponding to the primary function category and extracts the basic word segmentation units related to the basic service function. The second round matches the semantic units corresponding to the secondary function subclass and further extracts the basic word segmentation units related to the trip planning function from the primary function category matching results. Contextual verification is performed on the multi-round matching results. By combining the semantic information of the context of the basic word segmentation unit, mismatched basic word segmentation units caused by grammatical errors or ambiguities are excluded. The verified matching results are processed to merge duplicate items, and duplicate basic word segmentation units under the same functional subclass are merged into a single explicit requirement candidate element. The explicit requirement candidate elements are labeled with functional categories. Each explicit requirement candidate element is labeled with its corresponding first-level functional category and second-level functional subclass, generating explicit requirement candidate elements containing functional hierarchy information.

4. The demand forecasting method based on machine learning according to claim 2, characterized in that, The process involves analyzing the sentiment attributes, logical modification relationships, and scenario relevance of the remaining basic word segments in the unmatched functional demand terminology database to extract basic word segments reflecting potential user needs as implicit demand candidate clues, including: The remaining basic word segmentation units are subjected to sentiment attribute analysis. By traversing a predefined sentiment lexicon, basic word segmentation units with sentiment tendencies are identified. A sentiment polarity tag is generated for each identified basic word segmentation unit. The sentiment polarity tag contains positive or negative sentiment tendency information. The remaining basic word segmentation units are subjected to logical modification relationship analysis and processing. By traversing the predefined logical relationship lexicon, basic word segmentation units with comparison, hypothesis, or contrast relationships are identified. Logical relationship tags are generated for each identified basic word segmentation unit. The logical relationship tags contain logical relationship information of comparison, hypothesis, or contrast. The remaining basic word segmentation units are subjected to scene association analysis. By traversing the predefined scene association word library, basic word segmentation units directly related to the tourism experience scene are identified. Scene association tags are generated for each identified basic word segmentation unit. The scene association tags contain scene description information related to the tourism experience. The emotional polarity marker, the logical relationship marker, and the scene association marker are subjected to multi-dimensional feature fusion processing to merge multiple markers corresponding to the same basic word segmentation unit into a composite implicit demand clue. The composite implicit demand clue simultaneously contains potential demand information of emotional tendency, logical relationship, and scene association. The composite implicit demand cues are subjected to domain relevance filtering to retain composite implicit demand cues related to tourism product demand and remove implicit semantic information unrelated to the tourism scenario, thereby generating the final implicit demand candidate cues.

5. The demand forecasting method based on machine learning according to claim 1, characterized in that, In the aforementioned demand association graph model, establishing edge connections between the first type of nodes and the second type of nodes based on the semantic and logical relationships of user-expressed content units includes: Semantic dependency analysis is performed on the user-expressed content units to extract the modification, causal, and parallel relationships between basic word segmentation units as semantic logical relationships. These semantic logical relationships are used to determine the association direction between demand features. Based on the semantic logic relationship, the semantic association direction between explicit demand elements and implicit demand clues is determined. The association direction includes the guiding relationship between explicit demand elements and implicit demand clues or the supplementary relationship between implicit demand clues and explicit demand elements. Based on the association direction, directed edge connections are established in the demand association graph model from the first type of node to the second type of node or from the second type of node to the first type of node. Calculate the time interval between the appearance of the explicit demand elements and the implicit demand cues in the user's expressed content unit, and generate a time correlation parameter. The time correlation parameter is used to measure the degree of correlation between demand features in the time dimension. The weight value of the directed edge connection is calculated by combining the time correlation parameter and the co-occurrence frequency. The weight value increases as the time correlation parameter decreases and the co-occurrence frequency increases.

6. The demand forecasting method based on machine learning according to claim 1, characterized in that, The analysis of the transmission patterns of the changing characteristics of the demand elements within a continuous time window, and the identification of the evolutionary paths between explicit demand elements and implicit demand cues, includes: Time series correlation analysis was performed on the changes in the aforementioned demand elements to calculate the Pearson correlation coefficient between changes in explicit demand elements and changes in implicit demand cues, and to determine the strength of their correlation over time. Explicit demand elements and implicit demand clue pairs with a correlation strength greater than a set correlation strength threshold are selected as candidate objects for evolutionary paths. The candidate objects are subjected to longitudinal tracking analysis to extract the starting time point, duration and intensity change curve of the continuous appearance of implicit demand cues in the continuous time window. At the same time, the corresponding explicit demand elements are extracted as the time point of addition or enhancement and intensity change curve after the appearance of implicit demand cues. By comparing the change curves of the implicit demand cues and the explicit demand elements, the time delay parameters and intensity mapping relationship of the explicit demand elements triggered by the continuous appearance of the implicit demand cues are identified, and a positive evolution path for the transformation of implicit demand cues into explicit demand elements is generated. Extract the starting time point, duration, and intensity decay curve of the decrease in frequency of explicit demand elements in a continuous time window. At the same time, extract the time point and intensity change curve of the corresponding implicit demand cues after the decay of explicit demand elements. Identify the time delay parameters and intensity mapping relationship of the implicit demand cues triggered by the weakening of explicit demand elements. Generate the reverse evolution path of the transformation of explicit demand elements into implicit demand cues. The forward evolution path and the reverse evolution path are merged to generate a set of evolution paths between explicit demand elements and implicit demand cues.

7. The demand forecasting method based on machine learning according to claim 1, characterized in that, The process of performing pattern matching on the continuity patterns and association rules in the demand evolution feature set, identifying the degree of matching between the demand features at the current time point and the historical evolution patterns, and generating a pattern matching similarity value includes: Construct a historical evolution pattern library, which contains a set of demand evolution characteristics and their corresponding actual demand results in different historical time periods; The current time point's demand evolution feature set is processed by feature extraction to generate a current feature vector containing continuous regularity parameters and correlation rule parameters; For each historical evolution pattern in the historical evolution pattern library, feature extraction processing is performed to generate a historical feature vector containing continuous regularity parameters and association rule parameters; Calculate the cosine similarity between the current feature vector and each historical feature vector to generate a pattern matching similarity value between each historical evolution pattern and the required feature at the current time point.

8. A demand forecasting system based on machine learning, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the machine learning-based demand prediction method according to any one of claims 1-7.

Images