A real-time feature-based accurate recommendation method driven by multimodal cues

Abstract

The present invention discloses a real-time feature-based precise recommendation method driven by multimodal clues, which belongs to the technical field of automobile sales market prediction. The method includes obtaining request samples entered by users, building a clue database, building dynamic offset features, calculating intent scores and sorting them, and realizing real-time precise recommendations in automobile sales scenarios. The method solves the technical problem of improving recommendation accuracy by modeling user dynamic behavior offsets and combining multimodal data fusion. The present invention realizes the deep fusion of multimodal clues, enhances the causality of recommendations, realizes precise recommendations under low-latency constraints, adapts to the real-time decision-making needs of the sales site, designs field standardization processing for automobile sales scenarios, and improves business matching.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile sales market prediction, and in particular relates to a multimodal clue-driven real-time feature precision recommendation method. Background Art

[0002] Currently, in the automotive sales process, traditional sales lead recommendation systems mainly rely on technologies such as collaborative filtering recommendation, content filtering recommendation, temporal behavior analysis, and rule engine recommendation. The specific technologies are as follows:

[0003] Collaborative filtering recommendations generally mine similar user preferences through the user-car model interaction matrix.

[0004] Content filtering recommendations are generally based on static matching of vehicle attributes and user profiles, such as price, brand, etc.

[0005] The principle of time series behavior analysis is to predict user intentions using time series models, such as the LSTM model.

[0006] The general practice of rule engine recommendations is to set static rules based on human experience and generate recommendations based on the static rules.

[0007] Deficiencies in the existing technology:

[0008] 1. Insufficient utilization of multimodal data: Existing sales lead recommendation systems fail to effectively integrate user structured data (budget, city) and unstructured behavior (browsing sequence, test drive appointments), resulting in one-sided feature representation.

[0009] 2. Weak ability to capture dynamic intent: Traditional methods often ignore the deviation between users' real-time behavior and historical transaction data (such as budget fluctuations and compressed car purchase cycles), making it difficult to quantify the probability of upgraded purchase intent.

[0010] 3. Poor scenario adaptability: Existing technologies often fail to model the high decision-making costs and long conversion cycles unique to automobile sales scenarios, resulting in recommendations that deviate from actual needs. Summary of the Invention

[0011] The purpose of the present invention is to provide a real-time feature-based accurate recommendation method driven by multimodal clues, which solves the technical problem of improving recommendation accuracy by modeling user dynamic behavior deviations and combining multimodal data fusion.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] A multimodal cue-driven real-time feature-based accurate recommendation method includes the following steps:

[0014] 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;

[0015] The fields of the requested sample include city, budget, car purchase time, down payment ratio, browsed model sequence, consultation frequency, and test drive appointment;

[0016] Step 2: The user feature extraction module retrieves the request sample, encodes or normalizes the fields in the request sample, and obtains a static feature vector. ;

[0017] Step 3: Create a lead database and enter multiple sales lead data as historical data;

[0018] The fields of the sales lead data include vehicle model, customer profile, whether the transaction was completed, transaction cycle, number of follow-up times, salesperson ID and timestamp;

[0019] The lead building module retrieves the sales lead data of the same model from the lead database based on 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;

[0020] Step 4: Intermediate feature generation module generates feature vectors and eigenvectors Combine them to generate historical shift features. Based on the historical shift features, use the Sigmoid function to calculate the score of the user's purchase intention upgrade ;

[0021] The feature vector , eigenvector and purchase intent upgrade scores Splice and generate the intermediate fusion feature vector corresponding to each sales lead data ;

[0022] Step 5: The recommendation score calculation module calculates the intermediate fusion feature vector Perform scoring calculations and use the Sigmoid function to calculate the matching confidence score between each sales lead data and the current customer;

[0023] 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 results, forms a recommendation result set, and pushes the recommendation result set to the sales front-end system.

[0024] Preferably, when executing step 2, the following steps are specifically included:

[0025] Step 2-1: The user feature extraction module retrieves the request sample and performs feature processing on the fields in the request sample, including:

[0026] One-hot encode the city field; standardize the budget field; normalize the down payment ratio field; convert the car purchase time field to "days from now" and normalize it; extract the ID set of recently viewed models based on the browsed model sequence field; count the number of inquiries in the past T days based on the consultation frequency field and normalize it; and binarize the test drive reservation field.

[0027] Step 2-2: Merge the processing results of step 2-1 to generate a static feature vector :

[0028] ;

[0029] Where R represents the field of the requested sample; Represents a static feature processing function, specifically the feature processing method in step 2-1; i represents the field number.

[0030] Preferably, when executing step 3, the following steps are specifically included:

[0031] Step 3-1: Establish a lead database and enter sales lead data;

[0032] Step 3-2: Retrieve the request sample and retrieve all sales lead data from the lead database. Filter the sales lead data based on the browsed vehicle model sequence in the request sample, retain the sales lead data with the same vehicle model ID, and build a preliminary candidate set.

[0033] Step 3-3: Standardize the sales lead data in the preliminary candidate set, including:

[0034] One-hot encoding is performed on the vehicle model field; the customer profile field is categorized by age; the transaction completion field is binary-labeled; the transaction cycle field is standardized; the follow-up number field is normalized; and the salesperson identification field is one-hot encoded.

[0035] The feature vector of each sales lead data Expressed as:

[0036] ;

[0037] 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;

[0038] Step 3-4: All eigenvectors Gather and build a pool of candidate leads.

[0039] Preferably, when executing step 4, the following steps are specifically included:

[0040] Step 4-1: Intermediate feature generation module generates feature vectors and eigenvectors Perform combined splicing;

[0041] Step 4-2: Obtain the splicing results from step 4-1 and extract historical migration features, including:

[0042] Step 4-2-1: Set the time window T and search for sales lead data with transaction records within the time window T;

[0043] Step 4-2-2: Calculate the offset latitude, that is, the historical offset characteristics include:

[0044] Calculate the budget offset based on the difference between the user budget and the average transaction budget ;

[0045] Calculate the purchase time offset based on the current purchase intention days and the average transaction cycle ;

[0046] Calculate frequency deviation based on current consultation frequency and average follow-up times ;

[0047] Step 4-3: Use the Sigmoid function to calculate the score of the user's purchase intention upgrade :

[0048] ;

[0049] in, is the Sigmoid function, are weights, b is bias;

[0050] Step 4-4: Scoring Purchase Intent Upgrades , eigenvector and eigenvectors Splice and get the intermediate fusion feature vector .

[0051] Preferably, when executing step 5, the following steps are specifically included:

[0052] Step 5-1: Obtain the intermediate fusion feature vector ;

[0053] Step 5-2: Use the Sigmoid function to calculate the matching score of each sales lead data in the candidate pool :

[0054] ;

[0055] in, is the weight, and T represents the transpose operation.

[0056] Preferably, when executing step 6, the following steps are specifically included:

[0057] Step 6-1: Retrieve the settlement result of step 5 and obtain the matching score set of all sales lead data in the candidate pool ;

[0058] Step 6-2: Set matching scores Sort from high to low;

[0059] Step 6-3: Set the upper limit of recommendation quantity to K, select the top K sales leads in the ranking to form the final recommendation set;

[0060] Step 6-4: Push the recommended result set to the sales front end.

[0061] The multimodal clue-driven real-time feature-based precise recommendation method described in the present invention solves the technical problem of improving recommendation accuracy by modeling user dynamic behavior offsets and combining multimodal data fusion. The present invention combines user static attributes (city, budget), dynamic behavior (browsing sequence, test drive appointment) and historical transaction data to construct a global feature representation and achieve deep fusion of multimodal clues. The present invention quantifies the probability of user intention upgrade through three-dimensional offset features of budget, time and frequency, enhances recommendation causality, adopts Sigmoid function design, realizes precise recommendation under low latency constraints, adapts to real-time decision-making needs at the sales site, and designs field standardization processing for automobile sales scenarios to improve business matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is the main flow chart of the present invention;

[0063] Figure 2 is a flow chart of step 4 of the present invention;

[0064] Figure 3 is a system architecture diagram in this embodiment;

[0065] Figure 4 It is a data flow diagram of the system architecture of the present invention. DETAILED DESCRIPTION

[0066] Depend on Figure 1-Figure 4The multimodal cue-driven real-time feature-based accurate recommendation method shown in FIG1 includes the following steps:

[0067] 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;

[0068] In this embodiment, the sales front end is a client or APP, and the request sample is entered into the sales front end in the form of text after the form. When the sales front end obtains the request sample entered by the user or salesperson, the request sample is structured and packaged to generate a standard field set. :

[0069] ;

[0070] in Corresponding to the fields of the request sample, in this embodiment, the fields of the request sample include city, budget, car purchase time, down payment ratio, browsed model sequence, consultation frequency and test drive appointment.

[0071] Step 2: The user feature extraction module retrieves the request sample, encodes or normalizes the fields in the request sample, and obtains a static feature vector. ;

[0072] When executing step 2, the specific steps include:

[0073] Step 2-1: The user feature extraction module retrieves the request sample and performs feature processing on the fields in the request sample, including:

[0074] One-hot encoding is performed on the city field. The city field is a categorical feature. This embodiment uses one-hot encoding to convert each city into an independent binary feature. For example, "Beijing", "Shanghai", and "Guangzhou" will generate three new features, respectively indicating whether the sample belongs to the city.

[0075] The budget field is normalized. The budget field is a numerical feature. In this embodiment, a Min-Max normalization process is performed on it, and the budget field is scaled to a specified interval [0, 1] through a linear transformation.

[0076] The down payment ratio field is normalized. In this embodiment, the down payment ratio field is scaled to a unit interval [0, 1] through normalization.

[0077] The vehicle purchase time field is converted to "number of days from the current day" and normalized. In this embodiment, the current day is used as the benchmark to calculate the difference between the vehicle purchase time and the current day.

[0078] A recently browsed vehicle model ID set is extracted based on the browsed vehicle model sequence field. The browsed vehicle model sequence field is sequence-type behavior data. In this embodiment, the IDs of vehicle models within the past 30 days are extracted to form a vehicle model ID set.

[0079] The number of consultations in the past T days is counted according to the consultation frequency field and normalized. In this embodiment, the consultation frequency field is normalized to a reasonable range by dividing the consultation frequency by the maximum number of consultations in T days.

[0080] The test drive reservation field is binarized and converted into a 0 / 1 value. In this embodiment, 1 indicates a reservation and 0 indicates no reservation.

[0081] Step 2-2: Merge the processing results of step 2-1 to generate a static feature vector :

[0082] ;

[0083] Where R represents the field of the requested sample; Represents a static feature processing function, specifically the feature processing method in step 2-1. Each field corresponds to its corresponding processing method; i represents the field number.

[0084] Step 3: Create a lead database and enter multiple sales lead data as historical data;

[0085] The fields of the sales lead data include vehicle model, customer profile, whether the transaction was completed, transaction cycle, number of follow-up times, salesperson ID and timestamp;

[0086] The lead building module retrieves the sales lead data of the same model from the lead database based on 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;

[0087] When executing step 3, the specific steps include:

[0088] Step 3-1: Establish a lead database and enter sales lead data;

[0089] Step 3-2: Retrieve the request sample and retrieve all sales lead data from the lead database. Filter the sales lead data based on the browsed vehicle model sequence in the request sample, retain the sales lead data with the same vehicle model ID, and build a preliminary candidate set.

[0090] Step 3-3: Standardize the sales lead data in the preliminary candidate set, including:

[0091] Perform one-hot encoding on the vehicle model field.

[0092] 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.

[0093] The "Transaction" field is binary marked. In this embodiment, "yes" is marked as 1 and "no" is marked as 0.

[0094] 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.

[0095] Normalize the follow-up times field using Min-Max normalization.

[0096] Perform one-hot encoding 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.

[0097] The feature vector of each sales lead data Expressed as:

[0098] ;

[0099] 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;

[0100] Step 3-4: All eigenvectors Gather and build a pool of candidate leads.

[0101] Step 4: Intermediate feature generation module generates feature vectors and eigenvectors Combine them to generate historical shift features. Based on the historical shift features, use the Sigmoid function to calculate the score of the user's purchase intention upgrade ;

[0102] The feature vector , eigenvector and purchase intent upgrade scores Splice and generate the intermediate fusion feature vector corresponding to each sales lead data ;

[0103] When executing step 4, the specific steps include:

[0104] Step 4-1: Intermediate feature generation module generates feature vectors and eigenvectors Perform combined splicing;

[0105] Step 4-2: Obtain the splicing results from step 4-1 and extract historical migration features, including:

[0106] Step 4-2-1: Set the time window T and search for sales lead data with transaction records within the time window T;

[0107] Step 4-2-2: Calculate the offset latitude, that is, the historical offset characteristics include:

[0108] Calculate the budget offset based on the difference between the user budget and the average transaction budget , the specific formula is as follows:

[0109] ;

[0110] in, For the set time window, Represents the set of completed leads from the sales lead data extracted from the candidate lead pool. Represents the budget, specifically the user's current desired car purchase budget, The budget is the average value, that is, the average budget of similar users in historical data. If there is no historical data, it is filled with the preset average price of the car model.

[0111] Calculate the purchase time offset based on the current purchase intention days and the average transaction cycle ;

[0112] ;

[0113] in, Indicates the pre-purchase time, specifically the time the user expects to purchase the car. The average car purchase time is the average number of days from browsing to purchasing for similar users in historical data. If there is no historical data, it is filled in using the average purchase cycle of the car model in the preset market.

[0114] Calculate frequency deviation based on current consultation frequency and average follow-up times ;

[0115] ;

[0116] in, Indicates the actual consultation frequency of the user, Indicates the average consultation frequency, that is, the average number of consultations by similar users in historical data. If there is no historical data, it will be filled in with the preset average number of consultations for this model.

[0117] Step 4-3: Use the Sigmoid function to calculate the score of the user's purchase intention upgrade :

[0118] ;

[0119] in, is the Sigmoid function, are weights, b is bias;

[0120] Step 4-4: Scoring Purchase Intent Upgrades , eigenvector and eigenvectors Splice and get the intermediate fusion feature vector .

[0121] In this embodiment, the final spliced as follows:

[0122] .

[0123] Step 5: The recommendation score calculation module calculates the intermediate fusion feature vector Perform scoring calculations and use the Sigmoid function to calculate the matching confidence score between each sales lead data and the current customer;

[0124] When executing step 5, the specific steps include:

[0125] Step 5-1: Obtain the intermediate fusion feature vector ;

[0126] Step 5-2: Use the Sigmoid function to calculate the matching score of each sales lead data in the candidate pool :

[0127] ;

[0128] in, is the weight, T represents the transposition operation, in this embodiment, It can be learned through training samples or initialized heuristically.

[0129] 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 results, forms a recommendation result set, and pushes the recommendation result set to the sales front-end system.

[0130] When executing step 6, the specific steps include:

[0131] Step 6-1: Retrieve the settlement result of step 5 and obtain the matching score set of all sales lead data in the candidate pool ;

[0132] Step 6-2: Set matching scores Sort from high to low;

[0133] Step 6-3: Set the upper limit of recommendation quantity to K, select the top K sales leads in the ranking to form the final recommendation set;

[0134] Step 6-4: Push the recommended result set to the sales front end.

[0135] In this embodiment, in actual application, a layered system architecture may be used to deploy various modules. For example, the following is a specific hardware architecture implementation of the present invention:

[0136] The system includes data acquisition layer, data processing layer, computing layer and application layer.

[0137] Deployed at the data collection layer:

[0138] The sales front-end system includes the sales front-end (client server or APP). The sales front-end system collects user request samples (city, budget, browsing model series, etc.) in real time.

[0139] Deployment at the data processing layer:

[0140] Request parsing module, used to store original request data in a structured manner;

[0141] User feature extraction module, used to encode / standardize user fields and generate static feature vectors;

[0142] Lead database, used to store historical sales lead data (car model, transaction cycle, number of follow-ups, etc.).

[0143] Deployment at the computing layer:

[0144] The lead building module is used to filter historical data based on the browsed models and build a candidate pool;

[0145] The intermediate feature generation module is used to calculate historical offset features (budget / time / frequency offset) and generate a fused feature vector;

[0146] The recommendation score calculation module is used to calculate the matching confidence score using the Sigmoid function.

[0147] Deployment at the application layer:

[0148] The recommendation sorting and distribution module is used to sort and return the top-K recommendation results to the sales front end.

[0149] All of the above modules are server devices and can adopt a distributed server architecture. Each layer communicates with each other through the Internet.

[0150] The multimodal clue-driven real-time feature-based precise recommendation method described in the present invention solves the technical problem of improving recommendation accuracy by modeling user dynamic behavior offsets and combining multimodal data fusion. The present invention combines user static attributes (city, budget), dynamic behavior (browsing sequence, test drive appointment) and historical transaction data to construct a global feature representation and achieve deep fusion of multimodal clues. The present invention quantifies the probability of user intention upgrade through three-dimensional offset features of budget, time and frequency, enhances recommendation causality, adopts Sigmoid function design, realizes precise recommendation under low latency constraints, adapts to real-time decision-making needs at the sales site, and designs field standardization processing for automobile sales scenarios to improve business matching.

Claims

1. A multimodal cue-driven real-time feature-based accurate recommendation method, characterized by: The steps include: 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 requested sample include city, budget, car purchase time, down payment ratio, browsed model sequence, consultation frequency, and test drive appointment; Step 2: The user feature extraction module retrieves the request sample, encodes or normalizes the fields in the request sample, and obtains a static feature vector z user ; Step 3: Create a lead database and enter multiple sales lead data as historical data; The fields of the sales lead data include vehicle model, customer profile, whether the transaction was completed, transaction cycle, number of follow-up times, salesperson ID and timestamp; The lead building module retrieves the sales lead data of the same model from the lead database based on 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; Step 4: The intermediate feature generation module generates the feature vector z user and eigenvectors Combine them to generate historical shift features. Based on the historical shift features, use the Sigmoid function to calculate the score s of the user's purchase intention upgrade intent ; The eigenvector z user , eigenvector and purchase intent upgrade scores intent Splice and generate the intermediate fusion feature vector corresponding to each sales lead data When executing step 4, the specific steps include: Step 4-1: Intermediate feature generation module generates feature vectors and the eigenvector z user Perform combined splicing; Step 4-2: Obtain the splicing results from step 4-1 and extract historical migration features, including: Step 4-2-1: Set the time window T and search for sales lead data with transaction records within the time window T; Step 4-2-2: Calculate the offset latitude, that is, the historical offset characteristics include: Calculate the budget offset Δ1 based on the difference between the user budget and the average transaction budget; Calculate the purchase time offset Δ2 based on the current purchase intention days and the average transaction cycle; Calculate the frequency shift Δ3 based on the current consultation frequency and the average number of follow-up visits; Step 4-3: Use the Sigmoid function to calculate the score s of the user's purchase intention upgrade intent : s intent =σ(w1×Δ1+w2×Δ2+w3×Δ3+b); Where σ(·) is the Sigmoid function, w1, w2, w3 are weights, and b is the bias; Step 4-4: Score the purchase intent upgrade intent , eigenvector z user and eigenvectors Splice and get the intermediate fusion feature vector Step 5: The recommendation score calculation module calculates the intermediate fusion feature vector Perform scoring calculations and use the Sigmoid function to calculate the matching confidence score between each sales lead data and the current customer; When executing step 5, the specific steps include: Step 5-1: Obtain the intermediate fusion feature vector Step 5-2: Use the Sigmoid function to calculate the matching score s for each sales lead data in the candidate pool (i) : Among them, W is the weight and T represents the transposition operation; 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 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-based accurate recommendation method according to claim 1, characterized in that: When executing step 2, the specific steps include: Step 2-1: The user feature extraction module retrieves the request sample and performs feature processing on the fields in the request sample, including: One-hot encode the city field; standardize the budget field; normalize the down payment ratio field; convert the car purchase time field to "days from now" and normalize it; extract the ID set of recently viewed models based on the browsed model sequence field; count the number of inquiries in the past T days based on the consultation frequency field and normalize it; and binarize the test drive reservation field. Step 2-2: Combine the processing results of step 2-1 to generate a static feature vector z user : with user ={f stat (R1),...,f stat (R i )}; Where R represents the field of the requested sample; f stat (·) represents a static feature processing function, specifically the feature processing method in step 2-1; i represents the field number.

3. The multimodal clue-driven real-time feature-based accurate recommendation method according to claim 2, characterized in that: 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 and retrieve all sales lead data from the lead database. Filter the sales lead data based on the browsed vehicle model sequence in the request sample, retain the sales lead data with the same vehicle model ID, and build a preliminary candidate set. Step 3-3: Standardize the sales lead data in the preliminary candidate set, including: One-hot encoding is performed on the vehicle model field; the customer profile field is categorized by age; the transaction completion field is binary-labeled; the transaction cycle field is standardized; the follow-up number field is normalized; and the salesperson identification field is one-hot encoded. The feature vector of each sales lead data Expressed as: Where, j is the number of the sales lead data field; i represents the i-th sales lead data; f line (·) indicates the processing method, corresponding to the standardized structure processing method in step 3-3; Step 3-4: All eigenvectors Gather and build a pool of candidate leads.

4. The multimodal clue-driven real-time feature-based accurate recommendation method according to claim 3, characterized in that: When executing step 6, the specific steps include: Step 6-1: Retrieve the settlement result of step 5 and obtain the matching score set of all sales lead data in the candidate pool {s (1) ,...,s (i) }; Step 6-2: Match the score set {s (1) ,...,s (i) }Sort from high to low; Step 6-3: Set the upper limit of recommendation quantity to K, select the top K sales leads in the ranking to form the final recommendation set; Step 6-4: Push the recommended result set to the sales front end.

Images