Intelligent recommendation method and electronic device

Abstract

Embodiments of the present application disclose a kind of intelligent recommendation method and electronic equipment, the method comprises: connect the call of user;Extract the portrait feature of the user and identify the intention information of the user;Based on the complementarity of commodity, the similarity of the portrait feature and commodity attribute and the coincidence of commodity and the historical behavior of the user, construct multiple commodity combinations and calculate the combination generation score of each commodity combination;Based on the intention information of the user and the combination generation score of each commodity combination, calculate the individualized score of each commodity combination;Based on commodity recommendation algorithm model, calculate the composite recommendation score of each commodity combination;According to the combination generation score, individualized score and composite recommendation score, the corresponding commodity combination is recommended to the user.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent recommendation method and electronic device. Background Technology

[0002] Currently, human-computer interaction intelligent decision-making has been applied to some extent in the field of activity recommendation, and a preliminary application architecture has been formed. This application architecture can convert user call content into text using speech-to-text technology, use keyword matching to initially identify user needs, and then generate recommended content according to preset activity recommendation rules. When an agent answers a user's call, the generated recommended content will be displayed in the form of a recommendation list, and the agent can manually select recommended content from the list and recommend it to the user.

[0003] However, the aforementioned activity recommendation rules are mostly single algorithms or static rules such as collaborative filtering algorithms or knowledge graphs, which are insufficient in handling semantic ambiguity. The recommendation rules are relatively rigid, resulting in a high rate of recommendation repetition and a relatively one-sided recommendation content, thus affecting the accuracy of the recommendation. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent recommendation method, apparatus, and electronic device to solve the problems in the prior art.

[0005] In a first aspect, an intelligent recommendation method provided in the embodiments of this application includes: Connect the user's call; Extract the user's profile features and identify the user's intent information; Based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior, multiple product combinations are constructed and a combined score for each product combination is calculated. Based on the user's intent information and the combination of each product combination, a score is generated, and a personalized score for each product combination is calculated. The composite recommendation score for each product combination is calculated based on the product recommendation algorithm model; Based on the combined rating, personalized rating, and composite recommendation rating, corresponding product combinations are recommended to the user.

[0006] Secondly, embodiments of this application provide an intelligent recommendation device, comprising: The connection module is used to connect a user's call; The processing module is used to extract the user's profile features and identify the user's intent information; The combined rating generation module is used to construct multiple product combinations and calculate the combined rating of each product combination based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior. A personalized rating module is used to generate a rating based on the user's intent information and the combination of each product combination, and to calculate the personalized rating of each product combination. The composite recommendation scoring module is used to calculate the composite recommendation score for each combination of products based on the product recommendation algorithm model; The recommendation module is used to generate a rating, a personalized rating, and a composite recommendation rating based on the combination, and recommend corresponding product combinations to the user.

[0007] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described intelligent recommendation method.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent recommendation method.

[0009] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described intelligent recommendation method.

[0010] As can be seen from the technical solutions provided by the embodiments of this application above, the embodiments of this application, by connecting a user's call, extracting the user's profile features and identifying the user's intent information, constructing multiple product combinations based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior, and calculating the combination generation score of each product combination, calculating the personalized score of each product combination based on the user's intent information and the combination generation score of each product combination, calculating the composite recommendation score of each product combination based on the product recommendation algorithm model, and recommending the corresponding product combination to the user based on the combination generation score, personalized score, and composite recommendation score, can realize multi-mode collaborative recommendation, solve the problem of recommendation limitations, and the recommendation results obtained by referring to the user's profile features and intent information can better adapt to the user's needs, improve the accuracy of recommendation and user satisfaction. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating an intelligent recommendation method provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the process of calculating the combined scores of various product combinations provided in this application embodiment; Figure 3 A flowchart illustrating the calculation of personalized scores for each product combination provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the sorting of multiple product combinations provided in an embodiment of this application. Figure 5 A flowchart illustrating the calculation of personalized scores for each product combination provided in an embodiment of this application; Figure 6 This is a schematic diagram illustrating the process of recommending product combinations to users according to an embodiment of this application; Figure 7 A flowchart illustrating another intelligent recommendation method provided in an embodiment of this application; Figure 8 This is a schematic diagram illustrating the application process of the intelligent recommendation method provided in the embodiments of this application; Figure 9 A schematic diagram of the structure of the intelligent recommendation device provided in the embodiments of this application; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] This application provides an intelligent recommendation method, apparatus, and electronic device.

[0014] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0015] The intelligent recommendation method and apparatus provided in this application can be applied to electronic devices, such as servers. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server capable of cloud computing; the specific application is not limited.

[0016] The application scenarios of the aforementioned intelligent recommendation method and device include: recommending relevant products to users based on their incoming calls on a server. Through this method and device, multi-mode collaborative recommendation can be achieved, overcoming the limitations of traditional recommendation methods. Furthermore, by referencing user profile features and intent information, the recommendation results are more tailored to user needs, improving recommendation accuracy and user satisfaction.

[0017] like Figure 1 As shown in the figure, this application provides an intelligent recommendation method applied to a server, which may include the following steps: S102: Connect the user's call.

[0018] The user's call can be initiated by the user to the operator's service hotline, such as for business inquiries, business processing, or complaints. The server can provide services to the user using intelligent voice, or it can transfer the call to a live agent for assistance.

[0019] In this embodiment of the application, after the server connects a user's call, it can collect user information, including but not limited to one or more of the following: collecting the user's voiceprint information, location information, channel information, intent information, etc. Among these, the voiceprint information can be used to extract the user's identity identifier, and can also be converted into text information for intent recognition, etc.

[0020] S104: Extract user profile features and identify user intent information.

[0021] User profile features reflect user habits and preferences, and can therefore be used to analyze potential user needs. User profile features can be represented in multiple dimensions, including but not limited to: data usage sensitivity or package sensitivity, video preference or gaming preference, high-value user or ordinary user, etc.

[0022] In this embodiment, entity extraction can be used to automatically identify and extract key information, such as identity, products, and behaviors, from unstructured text, and transform them into structured data, thereby building or improving customer profiles.

[0023] In this embodiment of the application, the user's voice information during a call can be converted into a text message, then semantically segmented, and contextual semantic disambiguation is performed on ambiguous semantic words (such as "suitable" and "easy") to obtain a structured representation, and finally intent recognition is performed.

[0024] For example, if a user's voice message includes "insufficient data," the user's intent may be to request a data plan. Or, if a user's voice message includes "student," the user's intent may be to request a campus card.

[0025] S106: Based on the complementarity of products, the similarity between profile features and product attributes, and the overlap between products and users' historical behaviors, construct multiple product combinations and calculate the combined generation score for each product combination.

[0026] The complementarity of goods refers to the fact that two goods provide different services and can complement each other. The complementarity of goods can be reflected by the co-occurrence rate or the number of times they co-occur. For example, if goods A and B co-occur 5 times within a certain period, then they are complementary.

[0027] The similarity between user profile features and product attributes can be understood as the correlation between users and products. The more similar the user profile features and product attributes are, the stronger the correlation, indicating a better match between the product and the user; in other words, a high degree of matching. Conversely, the less similar the user profile features and product attributes are, the weaker the correlation, indicating a poorer match between the product and the user; in other words, a low degree of matching.

[0028] Among these, the overlap between a product and a user's historical behavior reflects whether the user repeats the same behavior over a period of time. For example, if a user subscribes to a data package twice within three months, then the product "Data Security Package" overlaps with that user's behavior.

[0029] S108: Generate a score based on the user's intent information and the combination of various product combinations, and calculate a personalized score for each product combination.

[0030] S110: Calculate the composite recommendation score for each product combination based on the product recommendation algorithm model.

[0031] The product recommendation algorithm model can be one or more. For example, multiple product recommendation algorithm models can be used to calculate the composite recommendation score of each product combination. The product recommendation algorithm model can be an existing product recommendation algorithm model, which will not be elaborated here.

[0032] S112: Based on the combined generated score, personalized score, and composite recommendation score, recommend corresponding product combinations to the user.

[0033] In this embodiment, the recommended product combinations can be displayed on the auxiliary view, i.e., the page view where customer service personnel provide auxiliary tools. For example, marketing campaigns can be launched in the operational position of the auxiliary view, specifically displaying recommended marketing campaigns and product combination information. Here, the operational position is a channel touchpoint, i.e., the point where the marketing campaign is launched.

[0034] The method provided in this application, by connecting a user's call, extracting the user's profile features and identifying the user's intent information, constructs multiple product combinations based on the complementarity of products, the similarity between profile features and product attributes, and the overlap between products and the user's historical behavior, and calculates a combination generation score for each product combination. Based on the user's intent information and the combination generation scores of each product combination, a personalized score for each product combination is calculated. Based on a product recommendation algorithm model, a composite recommendation score for each product combination is calculated. Based on the combination generation score, personalized score, and composite recommendation score, the corresponding product combination is recommended to the user. This method can achieve multi-mode collaborative recommendation, solve the problem of recommendation limitations, and, by referencing the user's profile features and intent information, the recommendation results can better adapt to the user's needs, improving the accuracy of recommendations and user satisfaction.

[0035] See Figure 2 In this embodiment of the application, step S106, based on the complementarity of goods, the similarity between profile features and goods attributes, and the overlap between goods and users' historical behaviors, constructs multiple goods combinations and calculates the combination generation score for each goods combination. This step may include the following steps: S202: Construct multiple product combinations.

[0036] This allows for the creation of multiple product combinations based on a pre-defined set of products. Each product combination can include two or more products. For example, product combination A might include a "data package + video membership," while product combination B might include a "video data-free package + family account service," and so on.

[0037] S204: Determine the complementarity score of each product combination based on the correlation between products.

[0038] Among them, CP (Combination Complementarity) can be used to represent the combination complementarity score, which reflects the degree of complementarity between products. A high combination complementarity score means that the multiple products in the product combination are highly complementary, and the product combination is more worthy of recommendation.

[0039] In this embodiment of the application, CP can be calculated based on the product association graph using at least one of the methods of association probability and positive point mutual information.

[0040] Among them, it is possible to statistically analyze the bundled purchase or transaction records of different products (such as) in historical marketing data, thereby establishing the relationship between products and constructing a product relationship graph.

[0041] In one implementation, a method for calculating association probability is used alone. Specifically, this can be based on the product association graph to statistically analyze the historical co-occurrence frequency or historical co-occurrence purchase rate (CR) of product combinations, and use CR as the CP value. For example, if the historical marketing data of products such as data packages, video memberships, package upgrades, and value-added services are analyzed, and the historical co-occurrence frequency or historical co-occurrence purchase rate of the product combination "data package + video membership" is 68%, then the CP of the product combination "data package + video membership" is 0.68.

[0042] In another implementation, the point mutual information calculation method is used alone. Specifically, the point mutual information PMI can be calculated based on the ratio of the true probability to the expected probability of two products appearing in a product combination. Then, the point mutual information PPMI is taken and used as the value of CP. The formula for point mutual information is not elaborated here.

[0043] In another implementation, the correlation probability and positive point mutual information calculation methods can be used together. Specifically, CR and PPMI can be calculated separately, and then weighted summation can be performed to obtain CP, and the result can be mapped to the 0~1 interval.

[0044] S206: Determine the demand matching degree of each product combination based on the similarity between profile features and product attributes.

[0045] Among them, DM (Demand Match Degree) can be used to represent the degree of demand matching, which reflects the degree of similarity between the user's profile characteristics and the product attributes. A high degree of similarity indicates a high degree of demand matching, which means that the products in the product combination are well adapted to the user's needs, and the product combination is worth recommending.

[0046] In this embodiment, user profile tags can first be obtained from the user's historical data, including consumption level (high / medium / low value), demand type (data usage sensitivity / package sensitivity / value-added service sensitivity), and usage scenario (family sharing / personal commuting / business office, etc.). Furthermore, attribute tags are labeled for each product within the product bundle, such as labeling the product "10GB data package" with the attribute tags "data usage sensitivity" and "personal commuting," and labeling the product "family sharing package" with the attribute tags "family scenario" and "package sensitivity," thus establishing a product attribute tag library. Then, the profile tags and product attribute tags are converted into vectors, and the cosine similarity is calculated by dividing the vector dot product by the product of the vector magnitudes, serving as the DM. The result is mapped to the 0~1 interval.

[0047] For example, user profile tags are converted into vector [1,0,1], and product attribute tags in product combinations are converted into vector [1,1,1]. Calculating the cosine similarity yields DM=0.816.

[0048] S208: Determine the historical behavior matching degree of each product combination based on the overlap between the user's historical behavior and the products.

[0049] The History Behavior Match Degree (HM) can be used to represent the degree of overlap between a user's historical behavior and the products they buy. A high HM indicates that the user has chosen that product combination a high number of times in the past, thus making it more likely to be accepted by the user again and thus worthy of recommendation.

[0050] In this embodiment, a user's historical interaction records can be filtered to obtain the user's historical behavior, such as extracting the user's participation records in marketing activities within the past 6 months, including the types of data packages subscribed, purchased membership services, and package adjustment records. Alternatively, a user's browsing records can be filtered to obtain the user's historical behavior, such as records of marketing pages within the app that the user clicked on without any clicks. Alternatively, a user's consultation records can be filtered to obtain the user's historical behavior, such as the keywords mentioned by the user during a phone call.

[0051] In this embodiment, tag-based matching, embedding-based matching, or rule-based matching methods can be used to calculate the overlap between each item in the product bundle and the user's historical behavior. For example, if a user has purchased a "10GB data package," a "20GB data package," and another "20GB data package" three times in the past, and a product bundle contains a "15GB data package" (15GB falls between 10-20GB), then the overlap between this product and the user's historical behavior is 1. If a product bundle contains a "5GB data package" (5GB falls outside the 10-20GB range), then the overlap between this product and the user's historical behavior is 0.3.

[0052] In this embodiment, the weighted average of the overlap between all items in a product combination can be used to obtain HM, which is then mapped to the range of 0 to 1. The weighting can be adjusted based on the time elapsed since the historical behavior occurred, with more recent historical behaviors having higher weights and more distant historical behaviors having lower weights. For example, if a user has repeatedly purchased a combination of data and video memberships, the weighted average of the overlap between all items in a product combination like "10GB data package + video membership" would yield HM = 0.80.

[0053] S210: Calculate the combination generation score of each product combination according to the combination complementarity score, demand matching degree and historical behavior matching degree of each product combination according to the preset formula.

[0054] In this embodiment of the application, CGScore (Combination Generation Score) can be used to represent the combination generation score, which is specifically used to quantify the overall suitability of the recommended combination.

[0055] The above preset formula can be specifically expressed as follows: ; in, For any combination of products The combination generates a score with a value range of 0 to 1. Score the combinatorial complementarity of product combination c. For the demand matching degree of product combination c, The degree of historical behavior matching for product combination c. , and As the weight, satisfying .

[0056] The above formula can also be called a product combination generation algorithm or CGSA model. It is used to generate combination generation scores for product combinations. It can integrate complementarity, demand matching and historical behavior to form a three-dimensional score. The resulting combination generation score is used for product combination recommendation, and the recommendation results are more accurate.

[0057] It should be noted that the above-mentioned complementary weights Demand matching weight and historical behavior matching weight The system can be dynamically adjusted based on business strategies and feedback metrics from the CGSA model, or it can be updated manually. For example, technical personnel can directly modify the values ​​of α, β, and γ on the server based on real-time operational data, such as a decrease in conversion rate under a certain strategy, and the changes will be immediately synchronized to the CGSA model.

[0058] In this embodiment of the application, the weights can be initialized first based on the business strategy. , and Then, the weights are dynamically adjusted based on the feedback metrics of the CGSA model. , and One or more of the following. Business strategies include, but are not limited to, business priorities or strategy activation rules. Feedback metrics for the CGSA model can be pre-set, including but not limited to, "recommended combination click-through rate," "conversion rate," or "user complaint rate."

[0059] For example, if the business priority of "increasing cross-selling rate" is high, the complementarity weight α can be increased; for instance, the complementarity weight α can be adjusted from 0.3 to 0.5. If the business priority of "meeting users' immediate needs" is high, the demand matching weight β can be increased; for instance, the matching weight β can be adjusted from 0.4 to 0.6. If the business priority of "repurchasing from existing users" is high, the historical behavior matching weight γ can be increased; for instance, the historical behavior matching γ can be adjusted from 0.3 to 0.4.

[0060] For example, policy activation rules include, but are not limited to, the policy activation period or the target user group. For instance, emphasizing complementarity during holiday promotions can increase the complementarity weight α. Emphasizing demand matching during periods of low traffic at the end of the month can increase the demand matching weight β. For instance, emphasizing historical behavior for high-value user groups can increase the historical behavior matching weight γ. Emphasizing demand matching for new user groups can increase the demand matching weight β.

[0061] For example, if a product combination with a high CP value obtained from the current weight has a conversion rate of 25%, then the complementary weight α can be adjusted in a positive direction, that is, the value of the complementary weight α can be adjusted proportionally to the conversion rate of the product combination; if a product combination with a low DM value obtained from the current weight has a complaint rate of 8%, then the demand matching weight β can be adjusted in a negative direction, that is, the value of the demand matching weight β can be adjusted inversely to the complaint rate of the product combination.

[0062] In addition, a feedback-weight mapping model can be established, which uses machine learning (such as linear regression, reinforcement learning, etc.) to train the relationship between feedback indicators and weights. For example, "a 5% increase in conversion rate" corresponds to "an increase of α by 0.05", and "a 3% decrease in complaint rate" corresponds to "an increase of β by 0.03".

[0063] In this embodiment, the feedback indicators of the CGSA model can be statistically analyzed periodically (e.g., every 24 hours), input into the aforementioned feedback-weight mapping model to calculate the optimal weight values, and automatically update α, β, and γ to achieve automatic weight updates and optimization. Furthermore, a weight adjustment report can be generated for viewing, analysis, and summarization; the report can include the reasons for the adjustment and the expected results.

[0064] The following example illustrates the calculation process of the above formula. For example, product combinations. Data package A, video membership B Calculated based on historical co-occurrence purchase rate The result is calculated based on the degree of matching between user age, interest tags, and product attributes. It is calculated based on the user's history of purchasing similar combinations multiple times. , , , Then, using the above formula, the combined generation score of product combination c is calculated as follows: .

[0065] In this embodiment of the invention, by constructing multiple product combinations, the complementarity score of each product combination is determined based on the correlation between products, the demand matching degree of each product combination is determined based on the similarity between user profile features and product attributes, the historical behavior matching degree of each product combination is determined based on the overlap between user's historical behavior and products, and the combination generation score of each product combination is calculated according to a preset formula based on the complementarity score, demand matching degree, and historical behavior matching degree of each product combination. By comprehensively considering multiple dimensions such as product complementarity, demand matching degree, and historical behavior matching degree, a more reasonable product combination can be generated.

[0066] See Figure 3 In this embodiment of the application, step S108, which generates a score based on the user's intent information and the combination of various product combinations, and calculates a personalized score for each product combination, may include the following steps: S302: Generate a score based on the user's intent information and the combination of various product combinations, and rank the multiple product combinations.

[0067] The order of multiple product combinations can represent the recommendation priority. The order of the sorting results from high to low represents the order of priority from high to low. The higher the priority, the more worthwhile it is to recommend to the user.

[0068] S304: Calculate a personalized score for each product combination based on the ranking of multiple product combinations.

[0069] See Figure 4 In this embodiment of the application, step S302, which generates a score based on the user's intent information and the combination of various product combinations, and sorts multiple product combinations, may include the following steps: S402: Determine the intent label based on the user's intent information.

[0070] Among them, intent tags can reflect the user's intent information. For example, the intent tag "video buffering" can reflect the user's intent to purchase video products, and the intent tag "insufficient data" can reflect the user's intent to purchase data products.

[0071] S404: Determine the scene label based on the current scene.

[0072] The scenario tag can reflect the scenario in which the current call takes place. For example, the scenario tag "night" means that the user initiated the current call at night, the scenario tag "holiday" means that the user initiated the current call on a holiday, and the scenario tag "high-value user exclusive time period" means that the time period in which the user initiated the call is the time period in which most high-value users initiate calls.

[0073] S406: Generate a score based on the combination of intent tags, scene tags, and various product combinations, calculate the intent-driven ranking of each product combination according to a preset formula, and sort multiple product combinations.

[0074] In this embodiment of the application, it can be used Intent-Driven Rank can be used to represent intent-driven ranking, which can be used to optimize the priority of product combinations based on user intent. It can be used as follows: (Intent TagMatch) represents the degree of match between the product combination and the intent tag. It can be used... (Context Tag Match) indicates the degree of matching between product combinations and context tags.

[0075] The above preset formula can be specifically expressed as follows: ; in, Intention-driven ranking for any product combination c. For product mix The combination of these factors generates a score. The degree of match between product combination c and intent tag i. The degree of matching between product combination c and scene tag t. As the intention weighting factor, .

[0076] The above formula can also be called the intent-driven ranking algorithm or the IDCRA model. It is used to generate intent-driven rankings. Because it uses a dual matching mechanism of intent labels and scene labels, the ranking results are more accurate and intelligent.

[0077] In this embodiment of the application, the matching degree between product combination c and intent tag i is... The following method can be used to calculate it.

[0078] First, construct an intent-product tag mapping library, associate common intent tags, such as "insufficient data", "video buffering", and "package expiration", with product attribute tags, and establish matching rules. For example, the intent tag "insufficient data" corresponds to the attribute tag "data package" or "data security package", and the intent tag "video buffering" corresponds to the attribute tag "high-speed data package" or "5G upgrade package".

[0079] Then, the basic overlap can be calculated by counting the percentage of products associated with intent tag i in product combination c. This basic overlap can then be refined based on the importance of each product to the intent to obtain product importance weights. For example, product combination c is "10GB data package + video membership," with a total of 2 products. Intent tag i is "insufficient data," and the number of products associated with intent tag i is 1. The associated product is "10GB data package," so the percentage of associated products is 0.5. Therefore, the basic overlap for the associated product "10GB data package" is 0.5, and the basic overlap for the unassociated product "video membership" is also 0.5. After refining the basic overlap based on the importance of each product to the intent, the product importance weight for "10GB data package" is 0.8, and the product importance weight for "video membership" is 0.2.

[0080] Finally, the ITM is calculated based on product relevance and product importance weight, mapped to the 0-1 range. Product relevance can be set to 1 to represent relevance, and 0 to represent no relevance. Specifically, the ITM is calculated by weighting the product relevance of each product within the product bundle according to its importance weight. For example, the product importance weight of "10GB data package" is 0.8, and its relevance with "insufficient data" is 1; the product importance weight of "video membership" is 0.2, and its relevance with "insufficient data" is 0. Therefore, ITM = Σ(product relevance × product importance weight) = 1 × 0.8 + 0 × 0.2 = 0.8.

[0081] In this embodiment of the application, the matching degree between product combination c and scene tag t is... The following method can be used for calculation. First, define scene tags and adaptation rules. For example, scene tags can include "Night," "Holidays," "Exclusive Time Slots for High-Value Users," and "Back-to-School Season." Clearly define the products adapted to each scene tag. For example, the scene tag "Night" adapts to products like "Nighttime Data Package" or "Video Data-Free Package," and the scene tag "Back-to-School Season" adapts to products like "Campus Package" or "Family Account Service." Then, calculate the scene adaptation coefficient. Specifically, you can use historical data to calculate the conversion rate of each product in scene t, and use that conversion rate as the scene adaptation coefficient. For example, if the conversion rate of the product "Nighttime Data Package" is 30% in the "Nighttime" scene and 10% in the "Non-Nighttime" scene, then the scene adaptation coefficient for this product is 0.3 in the "Non-Nighttime" scene and 0.1 in the "Non-Nighttime" scene. Finally, take the average of the scene adaptation coefficients for all products in product combination c, or sum them weighted according to their sales volume percentage, to obtain the CTM value, and map the result to the 0-1 range. For example, if the product combination is "Nighttime Data Package + Family Number", and the scenario tag is "Nighttime", then the scenario adaptation coefficient of the product "Nighttime Data Package" is 0.3, and the scenario adaptation coefficient of the product "Family Number" is 0.1. Then CTM = (0.3 + 0.1) / 2 = 0.2.

[0082] It should be noted that, in the embodiments of this application, the intent weighting factor Adjustments can be made dynamically based on operational strategies or feedback metrics from the IDCRA model.

[0083] In one implementation, the intent weighting factor can be adjusted based on the operational strategy. First, prioritize the operational strategies and adjust the intent weighting factor accordingly. For example, if the operational strategy of "precisely responding to users' immediate needs" has a higher priority, then if users explicitly mention "insufficient traffic," the intent weighting factor can be increased. The value, such as The weighting factor was adjusted from 0.6 to 0.8 to prioritize matching intent tags. For example, if the operational strategy of "scenario-based marketing" has a higher priority, the intent weighting factor can be reduced during exclusive holiday events. The value, such as The value was adjusted from 0.6 to 0.4 to focus on scene tags. Then, the policy trigger conditions were configured; specifically, the scenarios in which the policy takes effect can be set, such as when a user explicitly states their needs. =0.8, when the user expresses a vague expression =0.5, or, set the policy's effective period, such as holidays. =0.4, working days =0.6 etc.

[0084] In another implementation, the intent weight factor can be adjusted based on feedback metrics from the IDCRA model. First, define the feedback metrics for the IDCRA model, such as "intent match accuracy" and "scenario-appropriate conversion rate." Then, establish adjustment rules; for example, if the "intent match accuracy" feedback metric is above 80% for three consecutive days, then... Automatically increase by 0.05; if the feedback metric "Scene Adaptation Conversion Rate" is below 20% for three consecutive days, it indicates inaccurate scene tag matching and should be increased. weight This can reduce like Automatically reduced by 0.03.

[0085] In this embodiment, the feedback metrics of the IDCRA model can also be statistically analyzed periodically (e.g., every 12 hours) and automatically updated according to the established adjustment rules. The value is recorded, and adjustments are logged for later analysis.

[0086] The following example illustrates the calculation process of the above formula. For instance, based on the user's intent information, the intent tag $i = "video buffering" is determined; based on the current scenario, the scenario tag $t = "high-value user exclusive time slot" is determined; and the product combination... Data packages, monthly video memberships , Calculated separately , , The intent-driven ranking is then calculated according to the above formula. .

[0087] In this embodiment, by determining intent tags based on user intent information, determining scene tags based on the current scene, generating a score based on the combination of intent tags, scene tags, and various product combinations, calculating the intent-driven ranking of each product combination according to a preset formula, and ranking multiple product combinations, it is possible to achieve product combination ranking based on the matching of intent tags and scene tags, with the product combinations ranked first being more in line with the user's current expressed intent.

[0088] See Figure 5 In this embodiment of the application, step S304, which calculates a personalized score for each product combination based on the order of multiple product combinations, may include the following steps: S502: Determine the similarity between the user's historical behavior and the historical behavior of each product combination based on the user's historical behavior.

[0089] S504: Determine the time decay score of the user and each product combination based on the number of historical behaviors of the user related to each product combination within the time window.

[0090] S506: Calculate the personalized score for each product combination according to a preset formula based on the similarity of the user's historical behavior with each product combination and the time decay score.

[0091] The above-mentioned preset formula is as follows: .

[0092] In this embodiment of the application, it can be used The Personalized Recommendation Optimization Score represents a personalized rating used to improve the accuracy of product combinations in adapting to users. It can be used... (History Behavior Similarity) represents the similarity of historical behaviors. It can be used as... The Time Decay Score measures a user’s recent changes in interest in a product mix.

[0093] Where u represents the user. For any combination of products Personalized rating, Intent-driven sorting The similarity of user u's historical behavior to product combination c. For user u and product combination c, a time decay score is assigned. Historical behavior weighting factor .

[0094] The above formula can also be called the Personalized Recommendation Optimization Algorithm PROA, which is used to generate personalized scores. It combines historical behavior and time decay scores to make personalized recommendations, making the recommendation results more personalized.

[0095] In this embodiment of the application, the historical behavior similarity between user u and product combination c is... The following method can be used to calculate it.

[0096] First, extract the user's historical behavior and convert it into a vector form to construct a historical behavior feature vector. For example, you can obtain the user's marketing activity interaction records for the past 12 months, including "type of goods purchased," "frequency of purchase," and "duration of purchase." For instance, "purchased 10GB data packages twice and video memberships once in the past 3 months, and purchased a 5GB data package once in the past 6-12 months." The dimensions of the historical behavior feature vector can include product type weight, time decay weight, and frequency weight. Product type weights include, for example, 0.6 for data packages, 0.3 for video memberships, and 0.1 for service packages. Time decay weights can be set according to the principle of higher weight for recent behavior, such as 0.7 for the past 3 months and 0.3 for the past 6-12 months. Frequency weights can be set according to the principle of higher weight for higher purchase frequency, such as 0.6 for 2 purchases and 0.3 for 1 purchase.

[0097] Then, a product feature vector is generated based on the preset feature dimensions. The preset feature dimensions can be set as needed. For example, the feature dimensions include [data usage value, video preference, music preference]. Based on this feature dimension, the product feature vector of the product combination "10GB data package + video membership" is [0.6, 0.3, 0].

[0098] Finally, the HBS is calculated based on the product feature vector and historical behavior feature vector of product combination c. For example, the product feature vector of product combination "10GB data package + video membership" is [0.6, 0.3, 0]. The cosine similarity is calculated between this vector and the historical behavior feature vector of user u to obtain the HBS. The result is mapped to the interval 0~1. The cosine similarity is 0.88, so HBS=0.88.

[0099] In this embodiment of the application, the time decay score of user u and product combination c is... The following method can be used for calculation. First, determine the baseline time window. Specifically, the "current user interaction time" can be used as the baseline, and a time decay period can be set, such as the last month, 1-3 months, 3-6 months, or 6-12 months. Then, set the decay coefficient. Specifically, the decay coefficient can be set to decrease according to the time distance, such as 1.0 for the last month, 0.7 for 1-3 months, 0.4 for 3-6 months, and 0.1 for 6-12 months. Finally, count the number of historical behaviors of user u related to product combination c in each time window. For example, no related behavior in the last month, 1 related behavior in 1-3 months, and 2 related behaviors in 3-6 months. Calculate TDS using "TDS = Σ (number of behaviors × corresponding decay coefficient) / maximum possible number of behaviors", and map the result to the interval 0-1. For example, the calculated TDS = 0.95.

[0100] It should be noted that, in the embodiments of this application, the historical behavior weighting factor Automatic optimization can be performed based on feedback from the PROA model. First, optimization metrics can be defined, including "personalized recommendation conversion rate" or "user retention rate." For example, is the conversion rate at θ=0.5 higher than that at θ=0.4, or is there a 3-month user retention rate based on historical behavior recommendations? Then, a reinforcement learning model is built. Specifically, θ can be used as the action variable (range 0-1) of the reinforcement learning model, and the "conversion rate improvement" can be used as the reward value. The model is trained to learn the relationship between θ and the reward value. For example, if the conversion rate at θ=0.5 is 8% higher than that at θ=0.4, the model tends to adjust θ to 0.5.

[0101] In addition, in this embodiment of the application, the data of the optimization index can be collected periodically (e.g., every 7 days), input into the reinforcement learning model, and the optimal θ value can be output. For example, if the original θ=0.5 and the new data shows that the conversion rate is higher when θ=0.6, then θ=0.6 will be automatically updated and an optimization report (including a comparison of the index before and after θ adjustment) will be generated.

[0102] The following example illustrates the calculation process of the above formula. For instance, a user... I've been frequently browsing information about combinations of traffic and video, particularly product combinations. Data package, iQiyi monthly membership card , Intent-driven ranking was calculated separately. , , Then, according to the above formula, we can calculate: .

[0103] In this embodiment, by determining the similarity between a user's historical behavior and that of each product combination based on the user's historical behavior, determining the time decay score between the user and each product combination based on the number of historical behaviors related to the user and each product combination within a time window, and calculating the personalized score of each product combination based on the similarity between the user's historical behavior and that of each product combination and the time decay score, it is possible to achieve personalized recommendation optimization based on the user's historical behavior and improve the suitability of the recommended product combinations for the user.

[0104] In the embodiments of this application, the above formulas / algorithms have obvious advantages, which can be specifically represented by Table 1 below.

[0105]

[0106] In this embodiment of the application, step S110, which calculates the composite recommendation score for each product combination based on the product recommendation algorithm model, may include the following steps: Based on the product recommendation algorithm model, the composite recommendation score for each product combination is calculated according to the following formula: ; ; Where c represents any combination of goods, and p represents any single item in combination c. Let p be the product recommendation rating, M be the set of product recommendation algorithm models, and m be any product recommendation algorithm model in set M. The weights of the product recommendation algorithm model m are... Let m be the product recommendation algorithm model's rating for product p, T(p) be the time period adaptation coefficient, and C(p) be the channel adaptation coefficient. (c) represents the composite recommendation score for product combination c, with a value range of [0,1]. Let N(c) be the normalization function, and let N(c) be the total number of items in the product combination c.

[0107] The product combination *c* can include multiple products, with no limit on the specific number of products; for example, *c* = {product1, product2, ..., product8}. The weight *Wm* can range from 0 to 1 and can be updated at fixed intervals, such as every 15 minutes based on feedback data. *T(p)* can be preset; for example, a time slot for high-value users can be set to *T(p)* = 1.2, and a "Do Not Disturb" time slot can be set to *T(p)* = 0.3. *C(p)* can also be preset; for example, a hotline scenario can be set to *C(p)* = 1.0, an APP scenario can be set to *C(p)* = 1.1, and an SMS scenario can be set to *C(p)* = 0.8.

[0108] It should be noted that the aforementioned weight Wm can be updated based on feedback data, as follows: First, define the dimensions of the feedback data. Specifically, feedback data can include four categories: "recommended combination click-through rate," "product processing conversion rate," "human adjustment rate," and "user satisfaction rating." For example, a product might have a click-through rate of 20%, a conversion rate of 15%, a human adjustment rate of 10%, and a satisfaction rating of 4.5 / 5. Next, set the weight calculation rules, assigning weights to each type of feedback data, such as a click-through rate weight of 0.3, a conversion rate weight of 0.4, a human adjustment rate weight of 0.2 (a negative indicator; the higher the adjustment rate, the lower the weight), and a satisfaction rating weight of 0.1. Finally, calculate the dynamic weight Wm using the formula "Wm = Σ (standardized feedback data value × feedback weight)."

[0109] In this embodiment of the application, feedback data can also be collected periodically (e.g., every 15 minutes), the feedback data is standardized and mapped to the 0~1 range, and then the latest Wm value is calculated by substituting it into the formula and automatically updated. For example, if the original Wm=0.6 and the new calculated result is 0.72, then Wm=0.72 is updated.

[0110] See Figure 6 In this embodiment of the application, step S112, which recommends corresponding product combinations to the user based on the combined generated score, personalized score, and composite recommendation score, may include the following steps: S602: Based on the combined generated score, personalized score, and composite recommendation score, calculate the multimodal collaborative recommendation score for each product combination according to a preset formula.

[0111] The specific formulas mentioned above are as follows: ; Where c represents any combination of goods. For multi-modal collaborative recommendation scoring of product combination c Generate scores for product combination c. Personalized scoring for product combination c The composite recommendation score for product combination c. , and As the weight, satisfying .

[0112] in, weight You can set the default value to 0.3, or you can set it to other values; there are no specific restrictions. weight You can set the default value to 0.3, or you can set it to other values; there are no specific restrictions. weight You can set the default value to 0.4, or you can set it to other values; there are no specific restrictions.

[0113] S604: Based on the multi-mode collaborative recommendation scores of each product combination, recommend corresponding product combinations to the user.

[0114] In this embodiment, after recommending product combinations, if a user successfully selects a recommended combination, the relevant data can be saved to provide data support for subsequent model training. Specifically, the following data can be structured and stored to form a standardized training dataset: user voiceprint features, call recording transcripts (including timestamps), intent recognition tag sequences (e.g., "Insufficient data → Video membership requirement"), recommended product combination IDs and selection results, agent adjustment records, interaction durations, etc. This dataset is connected to the model training library through an incremental learning interface, achieving a closed-loop optimization of "data accumulation - model iteration - performance improvement".

[0115] In this embodiment, by sorting multiple product combinations, a personalized score for each product combination is calculated. Based on the combined score and personalized score of each product combination, a comprehensive recommendation score for each product combination is calculated. Based on the comprehensive recommendation score of each product combination, the corresponding product combination is recommended to the user, which can realize recommendation based on the comprehensive recommendation score and achieve high recommendation accuracy.

[0116] See Figure 7 The intelligent recommendation method provided in this application embodiment can realize intelligent recommendation based on the above-mentioned multiple algorithms. Compared with the existing recommendation technology, it has obvious advantages. A specific comparative analysis is shown in Table 2.

[0117]

[0118] See Figure 8 The following is a specific example illustrating the application scenario of the intelligent recommendation method provided in this application. In this scenario, a customer initiates a call, a server agent receives the call, the server collects customer information, determines a customer profile, performs intelligent marketing recommendations, updates the algorithm model, and finally, the marketing data is stored and the call ends.

[0119] In this embodiment, the aforementioned models can also be updated. Specifically, user response messages regarding their selection of recommended product combinations can be received, and the vertical domain intent recognition model and user profile can be updated based on these response messages. The vertical domain intent recognition model can be updated primarily via weekly offline updates, supplemented by near-real-time streaming updates. User profile updates can be achieved using a combination of T+1 offline batch processing and near-real-time streaming processing. The update is based on user selection response messages and associated data, requiring prior data collection and structured processing to ensure the data meets the input requirements for model and profile updates. This process can include: core response data collection and data preprocessing. Core response data collection refers to extracting key feedback data from user interactions during marketing recommendations, which can specifically include: 1) Selection Results: The user finally confirms the marketing activity combination they signed up for (e.g., "10GB data package + iQiyi monthly membership card"), the recommended combination they rejected, and the reason for rejection (e.g., "no video membership needed" in the voice-to-text).

[0120] 2) Interaction process data: User follow-up questions about the recommended package (e.g., "How long is the data package valid?"), agent manual adjustment records (e.g., the user accepts the agent replacing "5GB data package" with "15GB data package"), and interaction duration (e.g., the time from the recommendation being displayed to the user's confirmation).

[0121] 3) Basic related data: user voiceprint corresponding to identity identifier, intent tag sequence of this call (such as "insufficient data → video request → confirm processing"), and scene tag (such as "high-value user time period at night").

[0122] Data preprocessing refers to data cleaning and feature extraction performed through the "speech-to-text module" and "context management module" mentioned in the document, and can specifically include: 1) Text standardization: Transform unstructured text transcribed from speech (including vague expressions and colloquial content) into structured semantic features (e.g., label “This data package is quite suitable” as “data package demand + positive feedback”). 2) Abnormal data filtering: Remove invalid feedback (such as user misoperation selections, or erroneous transcription content caused by noise); 3) Feature alignment: Align feedback data with model input features (such as intent tags and product attribute tags) (e.g., associating "refuse video membership" with the negative feature of "video preference").

[0123] On one hand, the vertical domain intent recognition model is the core input module of MAFRM, and its update goal is to improve the recognition accuracy of telecommunications industry-specific semantics (such as "package upgrade" and "data security package") and ambiguous expressions (such as "are there any good deals?"). Updating the vertical domain intent recognition model can include: labeling a new intent sample library and incrementally training to optimize model parameters.

[0124] The new intent sample library can construct newly labeled samples based on preprocessed user feedback data, specifically including: 1) Correctly identify samples: If a user selects "10GB data package" and the original intent is identified as "insufficient data", then it is marked as an "intent-feedback matching sample" (e.g., <"insufficient data", "insufficient data", positive matching>). 2) Error Correction Samples: If the original intent is identified as "package consultation", but the user ultimately selects "data package", it will be marked as "intent-feedback mismatch sample", and the intent label will be corrected (e.g., <"Want to see package related information", "Insufficient data", correct label>). 3) Add new intent samples: If new intents that are not covered appear in user feedback (such as "5G plan switch"), add intent tags and label the corresponding text (such as <"Want to switch to 5G plan", "5G plan requirement", add label>).

[0125] Incremental training to optimize model parameters can be achieved by updating the model using an "incremental learning" approach (avoiding the resource waste caused by full training), specifically including: 1) Fixed base network layers: Retain the mature general semantic recognition layers (such as entity extraction modules) in the model, and only update the feature extraction layers for semantics specific to the communications industry; 2) Input new samples: Input the labeled new samples (accounting for 10%-15% of the existing sample library) into the model and optimize the semantic disambiguation rules (such as adjusting the association weight between vague words such as "suitable" and "cost-effective" and intent labels). 3) Loss function optimization: With "intent recognition accuracy" and "feedback matching rate" as optimization objectives, adjust the softmax classification layer parameters of the model (such as reducing the confusion probability between "traffic demand" and "package demand").

[0126] The effectiveness verification and iteration of the above vertical domain intent recognition model are as follows: 1) Offline validation: Use a test set (accounting for 20% of the new samples) to validate the model update effect, ensuring that the accuracy of intent recognition is improved (e.g., from 86% to 92%) and the error recognition rate is reduced (e.g., the error rate of ambiguous expression is reduced from 18% to 8%) after the update. 2) Online gray-scale testing: Deploy the updated model to a portion of the hotline (e.g., 10% of the agents) and compare the feedback matching rate between the gray-scale group and the control group (original model). If the matching rate of the gray-scale group is more than 5% higher than that of the control group, then the updated model will be fully deployed online.

[0127] On the other hand, user profiles are the core basis for preference weight prediction models and personalized recommendation optimization algorithms (PROScore). The goal of updating user profiles is to dynamically match recent changes in user interests (such as a shift from "video preferences" to "game preferences"). Updating user profiles can include: dynamic adjustment of the weights of core profile tags and iterative supplementation of profile dimensions.

[0128] The dynamic adjustment of the core user profile tag weights can be based on user feedback data, adjusting the weights of the profile dimension tags mentioned in the document (such as "traffic-sensitive / package-sensitive", "video preference / gaming preference", "high-value user / ordinary user"), with the specific rules as follows: 1) Positive reinforcement: If a user selects a certain combination of marketing activities, the corresponding tag weight will increase (e.g., if the user selects "10GB data package + free game data service", the weight of "data sensitive" will increase by 0.2, the weight of "game preference" will increase by 0.3, and the weight value range is 0-1). 2) Negative decay: If a user explicitly refuses a certain combination, the corresponding tag weight is reduced (e.g., if the user refuses "video membership", the weight of "video preference" is reduced by 0.2, and can drop to a minimum of 0). 3) Time decay correction: Combine the "Time Decay Score (TDS)" in PROScore to decay the weight of labels that are not recent feedback (such as positive feedback of "video preference" from 3 months ago, the weight decays by 10% per month).

[0129] Additionally, if user feedback reveals new needs not covered by existing profiles, new profile tags can be added and the dimension system improved, as detailed below: 1) Add new tags: If a user inquires about "family shared data" multiple times, add a "family scenario needs" tag and assign weight based on feedback data (e.g., if "family shared package" is selected, the weight will be +0.3). 2) Dimensional segmentation: Subdivide existing tags (e.g., subdivide "traffic sensitive" into "daily traffic sensitive" and "holiday traffic sensitive", and adjust the weight based on the effective period of the data package selected by the user).

[0130] The correlation between the above user profiles and recommendation results can be verified using PROScore to check the fit, as detailed below: 1) Calculate profile - recommendation matching degree: If the weight of "game preference" in the profile increases to 0.8 after the update, the PROScore score of the recommended "game free data package" should be higher than before the update (e.g., from 0.608 to 0.752). 2) Correcting deviations: If the weight of the "high-value user" tag in the profile increases after the update, but the conversion rate of recommending "premium packages" does not increase, then further adjust the matching rules between profile tags and product attributes (such as adding the "business service" attribute to "premium packages" to enhance the matching degree with high-value users).

[0131] After the model and profile are updated, they need to be integrated into the closed loop of "data accumulation - model iteration - effect improvement" mentioned in the document to ensure that the update effect feeds back into the recommendation process. Specifically, real-time application feedback and effect data monitoring can be carried out.

[0132] Among these features, real-time application feedback can integrate the updated intent recognition model and user profile into the MAFRM recommendation process for subsequent user call recommendations, specifically including: 1) Intent recognition module: Improves the semantic recognition accuracy of new callers, reducing "recommendation bias caused by misjudgment of intent"; 2) User profiling: More accurate preference weight prediction makes the output of CGScore (combined generated score) and IDRank (intent-driven ranking) more in line with user needs.

[0133] Among them, performance data monitoring can statistically analyze the core metrics after the update (compared to those before the update), for example: Intent recognition accuracy: Target improvement of 5%-10%; Recommended combination click-through rate: Target increase of 8%-15%; The target for manual seat adjustment rate is a 10%-20% decrease. If the metrics do not meet expectations, the update process (such as the accuracy of sample labeling and the magnitude of model parameter adjustments) should be backtracked for secondary optimization.

[0134] Through the above update mechanism, a closed loop of "user feedback → data accumulation → model / profile optimization → recommendation accuracy improvement" is achieved, ensuring that the Multi-Modal Collaborative Recommendation Model (MMCRM) continuously adapts to changes in user needs in the marketing scenarios of the communications industry, and improves the accuracy and timeliness of recommendations.

[0135] The multi-modal collaborative recommendation model MMCRM in this embodiment performs comprehensive ranking and recommendation by weighted fusion of the outputs of multiple recommendation algorithms. Specifically, it triggers preset product combination rules through keyword combinations, semantic tag combinations, and user profile feature combinations, involving the following five dimensions: 1) Complementarity analysis: Calculate the cross-selling probability of products within the combination through product association graph (e.g., the historical bundled purchase rate of traffic packages and video memberships is 68%). 2) Demand matching degree: based on the cosine similarity score between user profile tags (such as age, consumption level, and historical orders) and product attributes; 3) Dynamic priority adjustment: The product ranking within the package is dynamically updated based on real-time inventory, promotional budget, and channel conversion data (e.g., when a traffic package is out of stock, the display weight of similar alternative products is automatically increased). 4) Adaptive scene switching: When a change in user intent is detected (such as from "checking data plan" to "complaining about network quality"), the recommendation strategy is automatically switched (from marketing-oriented to service-oriented). 5) Personalized solution generation: Combine user's historical interaction sequence (such as the data usage trend in the past 3 months) to generate personalized solutions (such as recommending the "data security package + smart reminder" service to users with large data usage fluctuations).

[0136] The aforementioned multi-modal collaborative recommendation model, MMCRM, enables dynamic generation and precise prioritization of marketing campaign combinations. It updates the model in real-time based on user selection and response data, forming a closed-loop optimization cycle of "identification-recommendation-feedback" to improve the adaptability to various marketing scenarios. By dynamically adjusting recommendation weights based on business strategies or model feedback, and combining this with contextual semantic disambiguation capabilities, it enhances the accuracy of marketing campaign recommendations, adapts to diverse marketing scenarios, and addresses the issue of static algorithm weights. By using real-time user selection data to feed back into the model iteration, it solves the problems of strategy lag and high duplicate recommendation rates in traditional recommendations, offering greater advantages in adaptability and timeliness in complex marketing scenarios. Furthermore, by integrating intent recognition-based product combination generation mechanisms, preference weight prediction, and intent tag matching and ranking technologies, it overcomes the limitations of single-algorithm recommendations, achieving weighted fusion of multi-dimensional features and improving the comprehensiveness of recommendation combinations. Through personalized recommendation optimization technology based on user historical interactions, it enables recommendation combinations to accurately match individual users' behavioral habits and recent interest changes, improving personalization and addressing the issue of insufficient personalized recommendation depth.

[0137] The above describes the intelligent recommendation method provided in the embodiments of this application. Based on the same idea, the embodiments of this application also provide an intelligent recommendation device, such as... Figure 9 As shown, it may include: The connection module 901 is used to connect a user's call.

[0138] The extraction module 902 is used to extract user profile features and identify user intent information.

[0139] The combined rating generation module 903 is used to construct multiple product combinations and calculate the combined rating of each product combination based on the complementarity of products, the similarity between profile features and product attributes, and the overlap between products and users' historical behaviors.

[0140] The personalized rating module 904 is used to generate ratings based on the user's intent information and the combination of various product combinations, and to calculate the personalized ratings for each product combination.

[0141] The composite recommendation scoring module 905 is used to calculate the composite recommendation score for each product combination based on the product recommendation algorithm model.

[0142] The recommendation module 906 is used to generate ratings based on combinations, personalized ratings, and composite recommendation ratings, and recommend corresponding product combinations to users.

[0143] In this embodiment of the application, the combined scoring module 903 is specifically used for: Construct multiple product combinations; Based on the correlation between products, determine the combination complementarity score of each product combination; Based on the similarity between profile features and product attributes, determine the demand matching degree of each product combination; Based on the overlap between users' historical behavior and products, determine the historical behavior matching degree of each product combination; Based on the combination complementarity score, demand matching degree, and historical behavior matching degree of each product combination, the combination generation score of each product combination is calculated according to the following formula: ; in, For any combination of products The combination of these factors generates a score. Score the combinatorial complementarity of product combination c. For the demand matching degree of product combination c, The degree of historical behavior matching for product combination c. , and As the weight, satisfying .

[0144] In this embodiment of the application, the personalized scoring module 904 is specifically used for: A score is generated based on the user's intent information and the combination of various product combinations, and the multiple product combinations are ranked. Calculate personalized scores for each product combination based on the ranking of multiple product combinations.

[0145] In this embodiment of the application, the personalized scoring module 904 generates a score based on the user's intent information and the combination of various product combinations, and sorts the multiple product combinations, including: Determine intent tags based on user intent information; Determine scene labels based on the current scene; A score is generated based on the combination of intent tags, scenario tags, and various product combinations. The intent-driven ranking of each product combination is calculated according to the following formula, and multiple product combinations are ranked: ; in, Intention-driven ranking for any product combination c. For product mix The combination of these factors generates a score. The degree of match between product combination c and intent tag i. The degree of matching between product combination c and scene tag t. As the intention weighting factor, .

[0146] In this embodiment of the application, the personalized scoring module 904 calculates a personalized score for each product combination based on the order of multiple product combinations, including: Based on the user's historical behavior, determine the similarity between the user's historical behavior and the historical behavior of each product combination; The time decay score of the user and each product combination is determined based on the number of historical behaviors of the user related to each product combination within the time window. Based on the similarity of the user's historical behavior with each product combination and the time decay score, the personalized score for each product combination is calculated according to the following formula: ; Where u represents the user. For any combination of products Personalized rating, Intent-driven sorting The similarity of user u's historical behavior to product combination c. For user u and product combination c, a time decay score is assigned. Historical behavior weighting factor .

[0147] In this embodiment of the application, the composite recommendation scoring module 905 is specifically used for: Based on the product recommendation algorithm model, the composite recommendation score for each product combination is calculated according to the following formula: ; ; Where c represents any combination of goods, and p represents any single item in combination c. Let p be the product recommendation rating, M be the set of product recommendation algorithm models, and m be any product recommendation algorithm model in set M. The weights of the product recommendation algorithm model m are... Let m be the product recommendation algorithm model's rating for product p, T(p) be the time period adaptation coefficient, and C(p) be the channel adaptation coefficient. (c) represents the composite recommendation score for product combination c. Let N(c) be the normalization function, and let N(c) be the total number of items in the product combination c.

[0148] In this embodiment of the application, the recommendation module 906 is specifically used for: Based on the combined generated score, personalized score, and composite recommendation score, the multimodal collaborative recommendation score for each product combination is calculated using the following formula: ; Based on the multi-mode collaborative recommendation scores of each product combination, the corresponding product combinations are recommended to the user. Where c represents any combination of goods. For multi-modal collaborative recommendation scoring of product combination c Generate scores for product combination c. Personalized scoring for product combination c The composite recommendation score for product combination c. , and As the weight, satisfying .

[0149] The apparatus provided in the application embodiments can execute the methods provided in any of the above method embodiments. For detailed processes, please refer to the description in the method embodiments, which will not be repeated here.

[0150] The apparatus provided in this application, by connecting a user's call, extracts the user's profile features and identifies the user's intent information. Based on the complementarity of goods, the similarity between profile features and product attributes, and the overlap between goods and the user's historical behavior, it constructs multiple product combinations and calculates a combination generation score for each product combination. Based on the user's intent information and the combination generation scores of each product combination, it calculates a personalized score for each product combination. Based on a product recommendation algorithm model, it calculates a composite recommendation score for each product combination. Based on the combination generation score, personalized score, and composite recommendation score, it recommends the corresponding product combination to the user. This enables multi-mode collaborative recommendation, solves the problem of recommendation limitations, and, by referencing the user's profile features and intent information, the recommendation results are more adaptable to the user's needs, improving the accuracy of recommendations and user satisfaction.

[0151] Figure 10 This is a schematic diagram of the hardware structure of an electronic device to implement the various embodiments of this application. The electronic device 1000 includes, but is not limited to: a radio frequency unit 1001, a network module 1002, an audio output unit 1003, an input unit 1004, a sensor 1005, a display unit 1006, a user input unit 1007, an interface unit 1008, a memory 1009, a processor 1010, and a power supply 1011, etc. Those skilled in the art will understand that... Figure 10 The electronic device structures shown are not intended to limit the electronic device. An electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. In the embodiments of this application, the electronic device includes, but is not limited to, mobile phones, tablets, laptops, PDAs, in-vehicle terminals, wearable devices, and pedometers.

[0152] The processor 1010 is configured to: connect a user's call; extract the user's profile features and identify the user's intent information; construct multiple product combinations and generate combination generation scores for each product combination based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior; rank the multiple product combinations based on the user's intent information and the combination generation scores of each product combination; and recommend corresponding product combinations to the user according to the ranking of the multiple product combinations.

[0153] This application provides an electronic device that, upon receiving a user's call, extracts the user's profile features and identifies the user's intent information. Based on the complementarity of products, the similarity between profile features and product attributes, and the overlap between products and the user's historical behavior, multiple product combinations are constructed, and a combination generation score for each product combination is calculated. Based on the user's intent information and the combination generation scores of each product combination, a personalized score for each product combination is calculated. A composite recommendation score for each product combination is calculated based on a product recommendation algorithm model. Based on the combination generation score, personalized score, and composite recommendation score, the corresponding product combination is recommended to the user. This enables multi-mode collaborative recommendation, solves the problem of recommendation limitations, and, by referencing the user's profile features and intent information, the recommendation results are more tailored to the user's needs, improving recommendation accuracy and user satisfaction.

[0154] It should be understood that, in this embodiment, the radio frequency unit 1001 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink data from the base station and processes it with the processor 1010; additionally, it transmits uplink data to the base station. Typically, the radio frequency unit 1001 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier, a duplexer, etc. Furthermore, the radio frequency unit 1001 can also communicate with networks and other electronic devices via a wireless communication system.

[0155] The electronic device provides users with wireless broadband internet access through the network module 1002, such as helping users send and receive emails, browse web pages, and access streaming media.

[0156] The audio output unit 1003 can convert audio data received by the radio frequency unit 1001 or the network module 1002 or stored in the memory 1009 into audio signals and output them as sound. Furthermore, the audio output unit 1003 can also provide audio output related to specific functions performed by the electronic device 1000 (e.g., call signal reception sound, message reception sound, etc.). The audio output unit 1003 includes a speaker, a buzzer, and a receiver, etc.

[0157] Input unit 1004 is used to receive audio or video signals. Input unit 1004 may include a graphics processing unit (GPU) 10041 and a microphone 10042. GPU 10041 processes image data of still images or videos acquired by an image capture device (such as a camera) in video capture mode or image capture mode. The processed image frames can be displayed on display unit 1006. The image frames processed by GPU 10041 can be stored in memory 1009 (or other storage medium) or transmitted via radio frequency unit 1001 or network module 1002. Microphone 10042 can receive sound and process such sound into audio data. The processed audio data can be converted into a format that can be transmitted to a mobile communication base station via radio frequency unit 1001 in telephone call mode.

[0158] The electronic device 1000 also includes at least one sensor 1005, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor includes an ambient light sensor and a proximity sensor. The ambient light sensor can adjust the brightness of the display panel 10061 according to the ambient light level, and the proximity sensor can turn off the display panel 10061 and / or backlight when the electronic device 1000 is moved to the ear. As a type of motion sensor, an accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes). When stationary, it can detect the magnitude and direction of gravity and can be used to identify the posture of the electronic device (such as landscape / portrait switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc. The sensor 1005 may also include a fingerprint sensor, pressure sensor, iris sensor, molecular sensor, gyroscope, barometer, hygrometer, thermometer, infrared sensor, etc., which will not be described in detail here.

[0159] The display unit 1006 is used to display information input by the user or information provided to the user. The display unit 1006 may include a display panel 10061, which may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0160] User input unit 1007 can be used to receive input numerical or character information, and generate key signal inputs related to user settings and function control of electronic devices. Specifically, user input unit 1007 includes touch panel 10071 and other input devices 10072. Touch panel 10071, also known as a touch screen, can collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near touch panel 10071). Touch panel 10071 may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and the signal generated by the touch operation, and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to processor 1010, which receives and executes commands from processor 1010. In addition, touch panel 10071 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 10071, the user input unit 1007 may also include other input devices 10072. Specifically, other input devices 10072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, and joysticks, which will not be described in detail here.

[0161] Furthermore, the touch panel 10071 can cover the display panel 10061. When the touch panel 10071 detects a touch operation on or near it, it transmits the information to the processor 1010 to determine the type of touch event. Subsequently, the processor 1010 provides corresponding visual output on the display panel 10061 based on the type of touch event. Although in Figure 10 In this embodiment, the touch panel 10071 and the display panel 10061 are two independent components to realize the input and output functions of the electronic device. However, in some embodiments, the touch panel 10071 and the display panel 10061 can be integrated to realize the input and output functions of the electronic device. The specific implementation is not limited here.

[0162] Interface unit 1008 serves as an interface for connecting external devices to electronic device 1000. For example, external devices may include a wired or wireless headphone port, an external power supply (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a device with an identification module, an audio input / output (I / O) port, a video I / O port, a headphone port, and so on. Interface unit 1008 can be used to receive input from external devices (e.g., data, power, etc.) and transmit the received input to one or more components within electronic device 1000, or it can be used to transmit data between electronic device 1000 and external devices.

[0163] The memory 1009 can be used to store software programs and various data. The memory 1009 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback, image playback, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory 1009 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0164] The processor 1010 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 1009, and by calling data stored in the memory 1009, it performs various functions and processes data, thereby providing overall monitoring of the electronic device. The processor 1010 may include one or more processing units; preferably, the processor 1010 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1010.

[0165] The electronic device 1000 may also include a power supply 1011 (such as a battery) for supplying power to various components. Preferably, the power supply 1011 can be logically connected to the processor 1010 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system.

[0166] Preferably, this application embodiment also provides an electronic device, including a processor 1010, a memory 1009, and a computer program stored in the memory 1009 and executable on the processor 1010. When the computer program is executed by the processor 1010, it implements the various processes of the above method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0167] This application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the various processes of the above method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0168] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0169] The computer-readable storage medium provided in this application embodiment, by connecting a user's call, extracts the user's profile features and identifies the user's intent information. Based on the complementarity of goods, the similarity between profile features and product attributes, and the overlap between goods and the user's historical behavior, it constructs multiple product combinations and calculates the combination generation score for each product combination. Based on the user's intent information and the combination generation score of each product combination, it calculates the personalized score for each product combination. Based on the product recommendation algorithm model, it calculates the composite recommendation score for each product combination. Based on the combination generation score, personalized score, and composite recommendation score, it recommends the corresponding product combination to the user. This enables multi-mode collaborative recommendation, solves the problem of recommendation limitations, and, by referencing the user's profile features and intent information, the recommendation results are more adaptable to the user's needs, improving the accuracy of recommendations and user satisfaction.

[0170] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0171] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0172] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0173] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0174] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0175] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0176] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0177] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0178] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0179] It should be noted that the various models involved in the embodiments of this application can be AI models. The training and prediction processes of each AI model adhere to multiple legal and compliant principles, including legal data sources, compliant data content, compliant data governance, compliant training objectives and schemes, compliant training processes, compliant training environments and tools, and compliant ethical verification of training results, thus meeting the requirements of Article 5 of the Patent Law. Specifically: 1. Legality of data source: All datasets used for AI model training were obtained through legal means, covering three categories: publicly authorized data, data authorized by partners, and self-collected compliant data. Publicly authorized data comes from compliant data sources following open-source licenses such as Apache 2.0, with complete copyright attribution and authorization scope clearly marked, and no unauthorized open-source code or data reuse. Data authorized by partners has been subject to formal data usage agreements, clearly defining the scope, duration, and confidentiality obligations, and possessing a complete authorization chain. For self-collected data involving personal information, strict informed consent procedures have been followed, and anonymization processes (including but not limited to field masking, feature anonymization, and differential privacy technology applications) have been implemented to remove personally identifiable information, fully complying with the requirements of relevant laws and regulations such as the "Interim Measures for the Administration of Generative Artificial Intelligence Services" and the "Personal Information Protection Law."

[0180] 2. Data content compliance: The AI ​​model's dataset undergoes multiple screenings and cleaning processes to remove all content that may violate social morality or harm public interests. It contains no obscene, pornographic, violent, discriminatory, or information that endangers national or public safety, nor does it involve the illegal acquisition or use of genetic resources. For data in sensitive fields (such as healthcare and finance), an additional privacy-preserving computation module (including federated learning and secure multi-party computation technologies) ensures that the data is "usable but not visible," avoiding compliance risks during the original data transmission process and ensuring that the data application scenarios and uses comply with public order and good morals and industry regulatory requirements.

[0181] 3. Data governance standards: A complete data traceability system is established during the AI ​​model training process to automatically record the source, collection time, annotation process, cleaning rules, and permission allocation of training data, generating traceable compliance reports to ensure that the data is verifiable throughout its entire lifecycle. The dataset annotation process for AI models is completed by a professional human R&D team, clearly defining the proportion of human creative contributions and avoiding reliance on AI-generated data that has not undergone substantial human modification, thus meeting the examination requirements for "human main contributions" in AI patent applications.

[0182] 4. Training objectives and plans are compliant: The AI ​​model training objective focuses on intelligent recommendation. The training scheme and the final output results do not violate any mandatory provisions of laws and administrative regulations, do not harm the public interest or the legitimate rights and interests of others, and do not pose any potential risks of being used for illegal activities, infringing on privacy, or undermining public safety. The model strictly adheres to the ethical principle of "intelligent for good".

[0183] 5. Training process compliance: A closed-loop training framework is adopted to ensure compliance and controllability of the training process. The specific process is as follows: First, training samples are obtained through compliant data sources. After the aforementioned data cleaning and desensitization, they are input into the neural network model to generate preliminary training results. Second, an expert system is introduced to verify the preliminary results. Based on preset rules and human expert experience, the feasibility of the results is evaluated, and outputs that may pose ethical risks or compliance hazards are corrected (such as removing decision-making logic that violates public order and good morals, and adjusting model parameters that do not comply with safety regulations). Finally, the loss function weights are dynamically optimized based on expert system feedback to strengthen the model's learning of compliant results, avoid overfitting errors or non-compliant labels, and form a closed-loop control of "data input - model training - expert verification - parameter optimization - result feedback" to ensure that the entire training process complies with A5 ethical review requirements.

[0184] 6. Compliance of training environment and tools: AI model training is implemented using nationally licensed chips and a compliant training platform. All open-source frameworks and components used in the training process have obtained their corresponding licenses, and copyright statements and patent citation information are fully retained, with no instances of infringement or reuse. The training environment is built using virtual devices (containers / virtual machines) with fixed random seeds and initial parameter configurations to ensure the reproducibility of the training process. Furthermore, through access control and operation log recording, risks such as data leakage and parameter tampering during training are prevented, ensuring the security and compliance of the training process.

[0185] 7. Training results ethical verification complies with regulations: After the model is trained, it undergoes additional third-party ethical compliance assessment and algorithm filing review to verify that the model output does not violate social morality or harm public interests. For potentially sensitive scenarios (such as public services and intelligent decision-making), a special result verification mechanism is established to ensure that the model always complies with Article 5 of the Patent Law and relevant laws and regulations in practical applications.

[0186] In summary, the data and training process used in the AI ​​model of this application strictly comply with the relevant provisions of Article 5 of the Patent Law and the Patent Examination Guidelines (2023 Edition), and there is no violation of laws, social ethics, public interests, or illegal use of genetic resources. It fully meets the compliance requirements for patent authorization.

Claims

1. An intelligent recommendation method, characterized in that, The method includes: Connect the user's call; Extract the user's profile features and identify the user's intent information; Based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior, multiple product combinations are constructed and a combined score for each product combination is calculated. Based on the user's intent information and the combination of each product combination, a score is generated, and a personalized score for each product combination is calculated. The composite recommendation score for each product combination is calculated based on the product recommendation algorithm model; Based on the combined rating, personalized rating, and composite recommendation rating, corresponding product combinations are recommended to the user.

2. The method according to claim 1, characterized in that, Based on the complementarity of products, the similarity between the profile features and product attributes, and the overlap between the products and the user's historical behavior, multiple product combinations are constructed, and a combined score for each product combination is calculated, including: Construct multiple product combinations; Based on the correlation between the products, determine the combination complementarity score of each product combination; Based on the similarity between the profile features and the product attributes, the demand matching degree of each product combination is determined; Based on the overlap between the user's historical behavior and the products, the historical behavior matching degree of each product combination is determined; Based on the combination complementarity score, demand matching degree, and historical behavior matching degree of each product combination, the combination generation score of each product combination is calculated according to the following formula: ； in, For any combination of products The combination of these factors generates a score. Score the combinatorial complementarity of the product combination c. The demand matching degree of the product combination c. The historical behavior matching degree of the product combination c. , and As the weight, satisfying .

3. The method according to claim 1, characterized in that, The process of generating a score based on the user's intent information and the various product combinations, and calculating a personalized score for each product combination, includes: Based on the user's intent information and the combination of each product combination, a score is generated, and the multiple product combinations are ranked. Based on the ranking of the multiple product combinations, a personalized score is calculated for each product combination.

4. The method according to claim 3, characterized in that, The process of generating a score based on the user's intent information and the various product combinations, and ranking the multiple product combinations, includes: Determine the intent tag based on the user's intent information; Determine scene labels based on the current scene; A score is generated based on the combination of the intent tags, scene tags, and each product combination. The intent-driven ranking of each product combination is calculated according to the following formula, and the multiple product combinations are then ranked: ； in, Intention-driven ranking for any product combination c. For the product combination The combination of these factors generates a score. The degree of match between the product combination c and the intent tag i. Let be the degree of matching between the product combination c and the scene tag t. As the intention weighting factor, .

5. The method according to claim 3, characterized in that, The step of calculating a personalized score for each of the multiple product combinations based on their ranking includes: Based on the user's historical behavior, determine the similarity between the user's historical behavior and the historical behavior of each product combination; Based on the number of historical behaviors of the user and each product combination within the time window, determine the time decay score of the user and each product combination; Based on the similarity of the user's historical behavior with each product combination and the time decay score, the personalized score for each product combination is calculated according to the following formula: ； Where u is the user, For any combination of products Personalized rating, Intent-driven sorting The similarity between the historical behavior of user u and the product combination c. Assign a time decay score to the user u and the product combination c. Historical behavior weighting factor .

6. The method according to claim 1, characterized in that, The calculation of the composite recommendation score for each product combination based on the product recommendation algorithm model includes: Based on the product recommendation algorithm model, the composite recommendation score for each product combination is calculated according to the following formula: ； ； Where c represents any combination of goods, and p represents any one of the goods in combination c. Let m be the product recommendation rating for product p, M be the set of product recommendation algorithm models, and m be any product recommendation algorithm model in the set M. The weights of the product recommendation algorithm model m are... Let T(p) be the rating of the product recommendation algorithm model m for the product p, and C(p) be the time period adaptation coefficient and C(p) be the channel adaptation coefficient. (c) is the composite recommendation score for the product combination c. Let N(c) be the normalization function, and let N(c) be the total number of goods in the product combination c.

7. The method according to claim 1, characterized in that, The step of generating a score, a personalized score, and a composite recommendation score based on the combination, and recommending corresponding products to the user, includes: Based on the combined generated score, personalized score, and composite recommendation score, the multi-modal collaborative recommendation score for each product combination is calculated according to the following formula: ； Based on the multi-mode collaborative recommendation scores of each product combination, the corresponding product combination is recommended to the user; Where c represents any combination of goods. The multi-modal collaborative recommendation score for the product combination c. Generate a score for the combination of the goods c. Personalized ratings for the product combination c. The composite recommendation score for the product combination c. , and As the weight, satisfying .

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the intelligent recommendation method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the intelligent recommendation method as described in any one of claims 1-7.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the intelligent recommendation method according to any one of claims 1-7.

Images