Multi-modal clue-driven real-time feature accurate recommendation method

Abstract

The invention discloses a multi-modal clue-driven real-time feature accurate recommendation method, and belongs to the technical field of automobile sales market prediction, and the method comprises the steps: obtaining a request sample inputted by a user, constructing a clue database, constructing dynamic offset features, calculating intention scores, sorting the scores, and achieving the real-time accurate recommendation of an automobile sales scene. The technical problem of improving recommendation accuracy by modeling user dynamic behavior offset and combining with multi-modal data fusion is solved, deep fusion of multi-modal clues is realized, recommendation causality is enhanced, accurate recommendation is realized under low-delay constraint, real-time decision-making requirements of a sales site are adapted, and the recommendation accuracy is improved. Field standardization processing is designed for an automobile sales scene, and the business matching degree is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automotive sales market prediction, and particularly relates to a real-time feature precise recommendation method driven by multi-modal clues. Background Art

[0002] Currently, in the process of automotive sales, traditional sales lead recommendation systems mainly rely on technologies such as collaborative filtering recommendation, content filtering recommendation, time series behavior analysis, and rule engine recommendation. The specific technologies are as follows: Collaborative filtering recommendation generally mines similar user preferences through a user-vehicle interaction matrix.

[0003] Content filtering recommendation is generally based on the static matching of vehicle attributes and user portraits, such as price, brand, etc.

[0004] The principle of time series behavior analysis is to use a time series model to predict user intentions, such as the LSTM model.

[0005] The general practice of rule engine recommendation is to set static rules based on manual experience and generate recommendations according to the static rules.

[0006] Deficiencies in the prior art: 1. Insufficient utilization of multi-modal data: Existing sales lead recommendation systems fail to effectively integrate user structured data (budget, city) and unstructured behaviors (browsing sequence, test drive reservation), resulting in one-sided feature representation.

[0007] 2. Weak ability to capture dynamic intentions: Traditional methods often ignore the deviation of user real-time behaviors from historical transaction data (such as budget fluctuations, compressed car purchase cycles), and it is difficult to quantify the probability of purchase intention upgrade.

[0008] 3. Poor scene adaptability: In the prior art, the unique characteristics of high decision-making cost and long conversion cycle in the automotive sales scene are often not modeled specifically, and the recommendation results deviate from the real needs. Summary of the Invention

[0009] The purpose of the present invention is to provide a real-time feature precise recommendation method driven by multi-modal clues, which solves the technical problem of improving the recommendation accuracy by modeling the deviation of user dynamic behaviors and jointly fusing multi-modal data.

[0010] To achieve the above purpose, the present invention adopts the following technical solutions: A real-time feature precise recommendation method driven by multi-modal clues, comprising the following steps: Step 1: The request parsing module obtains the request samples collected by the sales front end and performs structured storage processing on the request samples; The fields of the request sample include city, budget, car purchase time, down payment ratio, browsing model sequence, consultation frequency, and test drive reservation; Step 2: The user feature extraction module retrieves the request sample and performs encoding or numerical standardization processing on the fields in the request sample to obtain a static feature vector ; Step 3: Establish a lead database and enter multiple sales lead data as historical data; The fields of the sales lead data include model, customer portrait, whether a deal is made, deal cycle, follow-up times, salesperson identifier, and timestamp; The lead construction module retrieves the sales lead data of the same model in the lead database according to the browsing model sequence, performs standardized structure vector processing on the retrieved sales lead data, and obtains the feature vector of each sales lead data , and constructs a candidate pool; Step 4: The intermediate feature generation module combines the feature vector and the feature vector to generate historical offset type features, and calculates the score of the user's purchase intention upgrade using the Sigmoid function according to the historical offset type features ; Concatenate the feature vector , the feature vector and the score of the purchase intention upgrade to generate an intermediate fusion feature vector corresponding to each sales lead data ; Step 5: The recommended score calculation module calculates the score of the intermediate fusion feature vector , and calculates the matching confidence score between each sales lead data and the current customer using the Sigmoid function; Step 6: The recommended sorting and distribution module sorts the sales lead data in the candidate pool according to the matching confidence score obtained in Step 5, selects the top K sales lead data as the recommended results, forms a recommended result set, and pushes the recommended result set to the sales front-end system.

[0011] Preferably, when performing Step 2, it specifically includes the following steps: Step 2-1: The user feature extraction module retrieves the request sample and performs feature processing on the fields in the request sample, specifically including: Perform one-hot encoding on the city field; perform normalization on the budget field; perform normalization on the down payment ratio field; convert the car purchase time field to "number of days from the current day" and perform normalization; extract the set of recently viewed vehicle model IDs according to the viewed vehicle model sequence field; count the number of consultations in the past T days according to the consultation frequency field and perform normalization; perform binary encoding on the test drive reservation field; Step 2-2: Combine the processing results of Step 2-1 to generate a static feature vector : ; Among them, R represents the fields of the request sample; represents the static feature processing function, specifically the feature processing method in Step 2-1; i represents the field number.

[0012] Preferably, when performing Step 3, it specifically includes the following steps: Step 3-1: Establish a lead database and enter sales lead data; Step 3-2: Retrieve the request sample, retrieve all sales lead data from the lead database, and screen the sales lead data according to the viewed vehicle model sequence in the request sample, retaining the sales lead data with the same vehicle model ID to construct a preliminary candidate set; Step 3-3: Perform standardized structure processing on the sales lead data in the preliminary candidate set, specifically including: Perform one-hot encoding on the vehicle model field; perform category encoding on the customer profile field according to age; perform binary marking on the whether closed field; perform normalization on the closing cycle field; perform normalization on the follow-up times field; perform one-hot encoding on the salesperson identification field; The feature vector of each sales lead data is expressed as: ; Among them, j is the field number of the sales lead data; i represents the i-th sales lead data; represents the processing method, corresponding to the standardized structure processing method in Step 3-3; Step 3-4: Aggregate all the feature vectors to construct a candidate lead pool.

[0013] Preferably, when performing Step 4, it specifically includes the following steps: Step 4-1: The intermediate feature generation module combines and splices the feature vector and the feature vector ; Step 4-2: Obtain the splicing result of Step 4-1, and extract historical offset type features, specifically including: Step 4-2-1: Set a time window T, and search for sales lead data with transaction records within the time window T; Step 4-2-2: Calculate the offset dimension, that is, the historical offset type features specifically include: Calculate the budget offset according to the difference between the user budget and the average transaction budget ; Calculate the car purchase time offset according to the current car purchase intention days and the average transaction cycle ; Calculate the frequency offset according to the current consultation frequency and the average follow-up times ; Step 4-3: Use the Sigmoid function to calculate the score of the user's purchase intention upgrade : ; Among them, is the Sigmoid function, are all weights, and b is the bias; Step 4-4: Splice the score of the purchase intention upgrade , the feature vector and the feature vector to obtain the intermediate fusion feature vector .

[0014] Preferably, when performing Step 5, it specifically includes the following steps: Step 5-1: Obtain the intermediate fusion feature vector ; Step 5-2: Use the Sigmoid function to calculate the matching score of each sales lead data in the candidate pool : ; Among them, is the weight, and T represents the transpose operation.

[0015] Preferably, when performing Step 6, it specifically includes the following steps: Step 6-1: Retrieve the settlement result of Step 5 to obtain the set of matching scores of all sales lead data in the candidate pool ; Step 6-2: Sort the set of matching scores from high to low; Step 6-3: Set the recommended upper limit number to K, and select the top K sales lead data in the sorting to form the final recommendation set; Step 6-4: Push the recommended result set to the sales front end.

[0016] A multi-modal clue-driven real-time feature precise recommendation method according to the present invention solves the technical problem of improving the recommendation accuracy by modeling the dynamic behavior deviation of users and combining multi-modal data fusion. The present invention combines user static attributes (city, budget), dynamic behaviors (browsing sequence, test drive reservation) and historical transaction data to construct a global feature representation, realizing the deep fusion of multi-modal clues. The present invention quantifies the probability of user intention upgrade through three-dimensional deviation features of budget, time, and frequency, enhances the causality of recommendation, adopts the Sigmoid function design, realizes precise recommendation under low-latency constraints, adapts to the real-time decision-making needs of the sales site, and designs field standardization processing for the automobile sales scenario, improving the business matching degree. Brief Description of the Drawings

[0017] Figure 1 is the main flowchart of the present invention; Figure 2 is the flowchart of Step 4 of the present invention; Figure 3 is a system architecture diagram in this embodiment; Figure 4 is the schematic diagram of the data flow of the system architecture of the present invention. Detailed Embodiment

[0018] As Figures 1-4 shown, a multi-modal clue-driven real-time feature precise recommendation method includes the following steps: Step 1: The request parsing module obtains the request samples collected by the sales front end and performs structured storage processing on the request samples; In this embodiment, the sales front end is a client or an APP. The request samples are entered into the sales front end in the form of text after a table. When the sales front end obtains the request samples entered by users or salespersons, it performs structured encapsulation on the request samples to generate a standard field set : ; Among them correspond to the fields of the request samples. In this embodiment, the fields of the request samples include city, budget, car purchase time, down payment ratio, browsing model sequence, consultation frequency, and test drive reservation.

[0019] Step 2: The user feature extraction module retrieves the request samples and performs encoding or numerical standardization processing on the fields in the request samples to obtain static feature vectors ; When performing Step 2, it specifically includes the following steps: Step 2-1: The user feature extraction module retrieves the request sample and performs feature processing on the fields in the request sample, specifically including: Perform One-hot encoding on the city field. The city field belongs to categorical features. In this embodiment, One-hot encoding is used to convert each city into an independent binary feature. For example, "Beijing", "Shanghai", and "Guangzhou" will generate 3 new features, respectively indicating whether the sample belongs to that city.

[0020] Perform standardization processing on the budget field. The budget field is a numerical feature. In this embodiment, it is scaled to the specified interval [0,1] through linear transformation by performing Min-Max standardization on it.

[0021] Perform normalization processing on the down payment ratio field. In this embodiment, the down payment ratio field is scaled to the unit interval [0,1] through normalization processing.

[0022] Convert the car purchase time field to "number of days from the current day" and perform standardization processing. In this embodiment, taking the current day as the benchmark, the difference between the car purchase time and the current day is calculated.

[0023] Extract the set of recently viewed vehicle model IDs according to the viewed vehicle model sequence field. The viewed vehicle model sequence field is sequential behavioral data. In this embodiment, the IDs of the vehicle models within the past 30 days are extracted to form the vehicle model ID set.

[0024] Count the number of consultations in the past T days according to the consultation frequency field and perform normalization processing; in this embodiment, by dividing the consultation frequency by the maximum number of consultations within T days, the consultation frequency field is normalized to a reasonable interval.

[0025] Perform binarization processing on the test drive reservation field. After binarization, it is converted into 0 / 1 numerical values. In this embodiment, 1 indicates reservation and 0 indicates non-reservation.

[0026] Step 2-2: Merge the processing results of Step 2-1 to generate a static feature vector : ; where R represents the fields of the request sample; represents the static feature processing function, specifically the feature processing methods in Step 2-1. Each type of field corresponds to its corresponding processing method; i represents the field number.

[0027] Step 3: Establish a lead database and enter multiple sales lead data as historical data; The fields of the sales lead data include vehicle model, customer portrait, whether a deal is made, deal cycle, follow-up times, salesperson identifier, and timestamp; The lead construction module retrieves the sales lead data of the same model from the lead database according to the browsed model sequence, performs standardized structure vector processing on the retrieved sales lead data, and obtains the feature vector of each sales lead data. , build a candidate pool; When executing step 3, the specific steps include: Step 3-1: Establish a lead database and enter sales lead data; Step 3-2: retrieve the request sample, retrieve all sales lead data from the lead database, filter the sales lead data according to the browsed model sequence in the request sample, retain the sales lead data with the same model ID, and build a preliminary candidate set; Step 3-3: Standardize the sales lead data in the preliminary candidate set, including: Perform one-hot encoding on the vehicle model field.

[0028] The customer portrait field is coded by age. In this embodiment, the continuous age data is divided into different categories (such as youth, middle-aged, and elderly), and then these categories are coded.

[0029] The whether transaction is completed field is binary marked. In this embodiment, "yes" is marked as 1 and "no" is marked as 0.

[0030] The transaction cycle field is standardized. In this embodiment, the data is converted into a distribution with a mean of 0 and a standard deviation of 1 through Z-score standardization.

[0031] The follow-up times field is normalized using Min-Max normalization.

[0032] One-hot encoding is performed on the salesperson ID field. For example, if there are salespersons "A", "B", and "C", three new columns "Salesperson ID_A", "Salesperson ID_B", and "Salesperson ID_C" will be added after one-hot encoding.

[0033] The feature vector of each sales lead data It is expressed as: ; Wherein, j is the number of the sales lead data field; i represents the i-th sales lead data; Indicates the processing method, corresponding to the standardized structure processing method in step 3-3; Step 3-4: All feature vectors Gather and build a pool of candidate leads.

[0034] Step 4: The intermediate feature generation module combines the feature vector and the feature vector to generate historical offset class features. Based on the historical offset class features, the Sigmoid function is used to calculate the score of the user's purchase intention upgrade ; The feature vector , the feature vector and the score of the purchase intention upgrade are concatenated to generate the intermediate fusion feature vector corresponding to each sales lead data ; When executing Step 4, it specifically includes the following steps: Step 4-1: The intermediate feature generation module combines and concatenates the feature vector and the feature vector ; Step 4-2: Obtain the concatenation result of Step 4-1 and extract the historical offset class features, specifically including: Step 4-2-1: Set the time window T and search for sales lead data with transaction records in the time window T; Step 4-2-2: Calculate the offset dimension, that is, the historical offset class features specifically include: Calculate the budget offset according to the difference between the user budget and the average transaction budget , and the specific formula is as follows: ; Among them, is the set time window, represents the set of closed leads among the sales lead data extracted from the candidate lead pool, represents the budget, specifically the current expected car purchase budget of the user, is the average value of the budget, that is, the average budget of users of the same type in historical data. If there is no historical data, the average price of this model preset is used for filling.

[0035] Calculate the car purchase time offset according to the current car purchase intention days and the average transaction cycle ; ; Among them, represents the pre-car purchase time, specifically representing the user's expected car purchase time, is the average car purchase time, that is, the average number of days from browsing to purchase for users of the same type in historical data. If there is no historical data, the average purchase cycle of this model in the market preset is used for filling.

[0036] Calculate the frequency offset according to the current consultation frequency and the average follow-up times ; ; Among them, represents the actual consultation frequency of the user, represents the average consultation frequency, that is, the average number of consultations of users of the same type in historical data. If there is no historical data, the average consultation frequency of this vehicle model preset is used for filling.

[0037] Step 4-3: Calculate the score of the user's purchase intention upgrade using the Sigmoid function : ; Among them, is the Sigmoid function, are all weights, and b is the bias; Step 4-4: For the score of the purchase intention upgrade , the feature vector and the feature vector are concatenated to obtain the intermediate fusion feature vector .

[0038] In this embodiment, the finally concatenated is as follows: .

[0039] Step 5: The recommendation score calculation module calculates the score of the intermediate fusion feature vector , and uses the Sigmoid function to calculate the matching confidence score between each sales lead data and the current customer; When performing Step 5, it specifically includes the following steps: Step 5-1: Obtain the intermediate fusion feature vector ; Step 5-2: Use the Sigmoid function to calculate the matching score of each sales lead data in the candidate pool : ; Among them, is the weight, and T represents the transpose operation. In this embodiment, can be learned through training samples or initialized heuristically.

[0040] Step 6: The recommendation sorting and distribution module sorts the sales lead data in the candidate pool according to the matching confidence score obtained in Step 5, selects the top K sales lead data as the recommendation result, forms the recommendation result set, and pushes the recommendation result set to the sales front-end system.

[0041] When performing Step 6, it specifically includes the following steps: Step 6-1: Retrieve the settlement result of Step 5 to obtain the matching score set of the sales lead data in the entire candidate pool ; Step 6-2: Sort the matching score set in descending order; Step 6-3: Set the recommended upper limit quantity as K, and select the top K sales lead data after sorting to form the final recommendation set; Step 6-4: Push the recommendation result set to the sales front end.

[0042] In this embodiment, in practical applications, a hierarchical system architecture can be adopted to deploy each module. For example, the following is a specific hardware architecture implementation of the present invention: The system includes a data acquisition layer, a data processing layer, a computing layer, and an application layer.

[0043] Deploy in the data acquisition layer: The sales front-end system, which includes the sales front end (client server or APP), and the sales front-end system collects user request samples (city, budget, browsing model sequence, etc.) in real time.

[0044] Deploy in the data processing layer: The request parsing module, which is used to store the original request data in a structured manner; The user feature extraction module, which is used to encode / normalize the user fields to generate static feature vectors; The lead database, which is used to store historical sales lead data (model, transaction cycle, follow-up times, etc.).

[0045] Deploy in the computing layer: The lead construction module, which is used to screen historical data according to the browsed models and construct a candidate pool; The intermediate feature generation module, which is used to calculate historical offset features (budget / time / frequency offset) and generate fused feature vectors; The recommendation score calculation module, which is used to calculate the matching confidence score through the Sigmoid function.

[0046] Deploy in the application layer: The recommendation sorting and distribution module, which is used to sort and return the top-K recommendation results to the sales front end.

[0047] Each of the above modules is a server device, and a distributed server architecture can be adopted, and each layer communicates with each other through the Internet.

[0048] A real-time feature precise recommendation method driven by multimodal cues according to the present invention solves the technical problem of improving the recommendation accuracy by modeling the dynamic behavior deviation of users and combining multimodal data fusion. The present invention combines user static attributes (city, budget), dynamic behaviors (browsing sequence, test drive reservation) and historical transaction data to construct a global feature representation, realizing the deep fusion of multimodal cues. The present invention quantifies the probability of user intention upgrade through three-dimensional deviation features of budget, time and frequency, enhances the causality of recommendation, adopts the Sigmoid function design, realizes precise recommendation under the low-latency constraint, adapts to the real-time decision-making requirements of the sales site, designs field standardization processing for the automobile sales scenario, and improves the business matching degree.

Claims

1. A real-time feature precise recommendation method driven by multi-modal cues, characterized in that: It includes the following steps: Step 1: The request parsing module obtains the request samples collected by the sales front-end and performs structured storage processing on the request samples; The fields of the request samples include city, budget, car purchase time, down payment ratio, viewed model sequence, consultation frequency, and test drive reservation; Step 2: The user feature extraction module retrieves the request sample, encodes or normalizes the values of the fields in the request sample to obtain a static feature vector ; Step 3: Establish a lead database and enter multiple sales lead data as historical data; The fields of the sales lead data include model, customer profile, whether a deal is made, deal cycle, follow-up times, salesperson identifier, and timestamp; The clue construction module retrieves the sales clue data of the same model from the clue database according to the browsed model sequence, performs standardized structure vector processing on the retrieved sales clue data, and obtains the feature vector of each sales clue data , and constructs a candidate pool; Step 4: The intermediate feature generation module combines the feature vector and the feature vector to generate historical offset class features, and based on the historical offset class features, uses the Sigmoid function to calculate the score of the user's purchase intention upgrade ; Concatenate the feature vector , the feature vector and the score of the purchase intention upgrade to generate the intermediate fusion feature vector corresponding to each sales lead data; Step 5: The recommended score calculation module calculates the score for the intermediate fusion feature vector and calculates the matching confidence score between each sales lead data and the current customer by using the Sigmoid function; Step 6: The recommendation sorting and distribution module sorts the sales lead data in the candidate pool according to the matching confidence scores obtained in Step 5, selects the top K sales lead data as the recommendation results, forms a recommendation result set, and pushes the recommendation result set to the sales front-end system.

2. The multimodal clue-driven real-time feature precise recommendation method according to claim 1, characterized in that: When performing Step 2, it specifically includes the following steps: Step 2-1: The user feature extraction module retrieves the request samples and performs feature processing on the fields in the request samples, specifically including: Performing One-hot encoding on the city field; performing standardization processing on the budget field; performing normalization processing on the down payment ratio field; converting the car purchase time field to "days from the current date" and performing standardization processing; extracting the set of recently viewed model IDs according to the viewed model sequence field; counting the number of consultations in the past T days according to the consultation frequency field and performing normalization processing; performing binary processing on the test drive reservation field; Step 2-2: Merge the processing results of Step 2-1 to generate a static feature vector : ； Among them, R represents the field of the request sample; represents the static feature processing function, specifically the feature processing method in step 2-1; i represents the number of the field.

3. The multimodal cue-driven real-time feature precise recommendation method according to claim 2, wherein: When performing Step 3, it specifically includes the following steps: Step 3-1: Establish a lead database and enter sales lead data; Step 3-2: Retrieve the request samples, retrieve all sales lead data from the lead database, and filter the sales lead data according to the viewed model sequence in the request samples, retaining the sales lead data with the same model ID to construct a preliminary candidate set; Step 3-3: Perform standardized structure processing on the sales lead data in the preliminary candidate set, specifically including: Performing One-hot encoding processing on the model field; performing category encoding on the customer profile field according to age; performing binary marking processing on the whether a deal is made field; performing standardization processing on the deal cycle field; performing normalization processing on the follow-up times field; performing One-hot encoding processing on the salesperson identifier field; The feature vector of each sales lead data is expressed as: ； Among them, j is the number of the sales lead data field; i represents the i-th sales lead data; represents a processing method, corresponding to the standardization structure processing method in step 3-3; Step 3-4: Aggregate all the feature vectors to construct a candidate clue pool.

4. The multimodal cue-driven real-time feature precise recommendation method according to claim 3, wherein: When performing Step 4, it specifically includes the following steps: Step 4-1: The intermediate feature generation module combines and splices the feature vector and the feature vector for combination and splicing. Step 4-2: Obtain the concatenation result of Step 4-1 and extract historical offset type features, specifically including: Step 4-2-1: Set a time window T and find the sales lead data with deal records in the time window T; Step 4-2-2: Calculate the offset dimension, that is, the historical offset type features specifically include: Calculate the budget deviation based on the difference between the user's budget and the average transaction budget ; Calculate the purchase time offset based on the current number of days of car purchase intention and the average transaction cycle ; Calculate the frequency offset based on the current consultation frequency and the average follow-up times ; Step 4-3: Calculate the score of the user's purchase intention upgrade using the Sigmoid function : ； Among them, is the Sigmoid function, are all weights, and b is the bias; Step 4-4: Score for the upgrade of purchase intention , feature vector and feature vector are concatenated to obtain an intermediate fusion feature vector .

5. The multimodal cue-driven real-time feature precise recommendation method according to claim 4, characterized in that: When performing Step 5, it specifically includes the following steps: Step 5-1: Obtain the intermediate fusion feature vector ; Step 5-2: Calculate the matching score of each sales lead data in the candidate pool using the Sigmoid function : ； Among them, is the weight, and T represents the transpose operation.

6. The multimodal clue-driven real-time feature precise recommendation method according to claim 5, wherein: When performing Step 6, it specifically includes the following steps: Step 6-1: Retrieve the settlement result of Step 5 to obtain the matching score set of the sales lead data in the entire candidate pool ; Step 6-2: Sort the matching score set in descending order; Step 6-3: Set the recommended upper limit number to K, select the top K sales lead data in the sorting to form the final recommended set; Step 6-4: Push the recommendation result set to the sales front-end.

Images