Intelligent sales link ai assistance method based on deep learning

Abstract

The application discloses an intelligent sales link AI auxiliary method based on deep learning, and relates to the technical field of deep learning. Target sales data of a sales whole link is acquired by using interface real-time synchronization, database log monitoring, front-end burying, etc.; then, a business feature set is extracted through feature engineering, a training set, a verification set and a test set are divided according to a time period, and a deep learning network architecture with multiple inputs and multiple outputs is designed; then, three sub-models of customer demand prediction, sales process vulnerability identification and marketing effect evaluation are trained respectively and training results are output; based on the output results of the sub-models, dynamic decision rules are formulated to generate optimization instructions; finally, new data after decision execution is captured through a real-time data monitoring architecture, the new data is taken as an incremental training set, a transfer learning strategy is adopted to retrain the model in a small scale, and model parameters are continuously optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent sales link AI assistance, more specifically, the present application relates to an intelligent sales link AI assistance method based on deep learning. BACKGROUND

[0002] With the rapid development of global market economy and the in-depth promotion of digital transformation, the enterprise sales mode is undergoing a profound change from traditional manual driving to data driving. As the core link of connecting customers and realizing value conversion for enterprises, the efficiency and accuracy of the sales link directly determine the market competitiveness of enterprises. However, the current sales link management still faces multiple technical bottlenecks, and the limitations of traditional sales mode and existing intelligent sales technology are increasingly prominent.

[0003] The traditional sales mode highly depends on the experience accumulation of sales personnel, and there are problems such as fuzzy customer positioning, subjective demand judgment, and blind resource allocation. In order to break through the bottleneck of the traditional mode, intelligent sales link systems have appeared in the industry, realizing partial process automation, but there are still some shortcomings in actual use: on the one hand, the existing technology only relies on basic customer purchase records and transaction amounts, and cannot integrate advertisement click volume and customer page dwell time, resulting in a large demand prediction deviation rate of customers, ignoring the deep browsing behavior of customers on non-purchased goods, and missing potential business opportunities.

[0004] On the other hand, the existing technology has a long adaptation period when facing market sudden changes, and cannot meet the real-time adjustment needs of business. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides an intelligent sales link AI assistance method based on deep learning, which solves the problems in the background art by the following scheme.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: an intelligent sales link AI assistance method based on deep learning, comprising the following steps:

[0007] S1: obtaining sales data of a sales full link, denoted as target sales data;

[0008] S2: based on the target sales data, extracting a business feature set, dividing a training set, a validation set and a test set, and obtaining a deep learning network architecture;

[0009] S3: inputting the target sales data into the deep learning network architecture, training a customer demand prediction sub-model, a sales process vulnerability identification sub-model and a marketing effect evaluation sub-model, and outputting the probability of a customer purchasing a specific product, positioning an inefficient node in the sales link, and outputting a marketing input-output ratio;

[0010] The training customer demand prediction sub-model, the specific analysis method of which comprises:

[0011] analyzing the purchase probability P of the customer to the product, wherein P is the cubic root of the sum of the ratio of the customer purchase time decay coefficient to its maximum value, the ratio of the customer access depth index to its maximum value, and the ratio of the customer conversion efficiency to its maximum value;

[0012] The time decay coefficient, the specific analysis method of which comprises:

[0013] The time decay coefficient is equal to the product of the average repurchase rate of the category of the natural constant e and the number of days from the last purchase time to the current time raised to a power; wherein the reciprocal of the average repurchase rate of the category is greater than 0;

[0014] The customer access depth index, the specific analysis method of which comprises:

[0015] The product of the page level of the i-th customer access to the product and its corresponding stay duration is summed for all access times, and the sum is divided by the total access times, and the quotient is the customer access depth index;

[0016] The customer conversion efficiency, the specific analysis method of which comprises:

[0017] The quotient of the sum of the payment amount of the customer purchasing a product and the smoothing coefficient, and the sum of the total amount of the customer adding and the smoothing coefficient, is the customer conversion efficiency;

[0018] The sales process vulnerability identification sub-model, the specific analysis method of which comprises:

[0019] Analyzing the inefficient node, the determination rule of the inefficient node is: when the customer abandonment rate of a certain link is greater than or equal to 30%, or the actual average time consumption of the link exceeds the historical average time consumption of the link, the link is determined as an inefficient node;

[0020] The customer abandonment rate of the certain link, the specific analysis method of which comprises:

[0021] The quotient of the number of abandoned customers of the link divided by the total number of customers entering the link is multiplied by 100%, and the final result is the customer abandonment rate of the link;

[0022] The actual average time consumption of the certain link, the specific analysis method of which comprises:

[0023] The sum of the time consumption of all customers entering the link in the link is divided by the total number of customers entering the link, and the final result is the actual average time consumption of the link;

[0024] The training marketing effect evaluation sub-model, the specific analysis method of which comprises:

[0025] The product net profit brought by the marketing strategy is divided by the total input of the marketing strategy, and the quotient is the input-output ratio;

[0026] S4: Based on the training results of the customer demand prediction sub-model, the sales process vulnerability identification sub-model and the marketing effect evaluation sub-model, a dynamic decision instruction is generated.

[0027] The dynamic decision instruction includes a sales process optimization instruction, a marketing strategy adjustment instruction and a customer demand response instruction.

[0028] The sales process optimization instruction trigger condition is based on the value range of the inefficient node output by the sales process vulnerability identification sub-model; the marketing strategy adjustment instruction trigger condition is based on the marketing input-output ratio output by the marketing effect evaluation sub-model; and the customer demand response instruction trigger condition is based on the purchase probability P value output by the customer demand prediction sub-model.

[0029] S5: Monitor the newly generated sales data after the execution of the decision instruction, use the new data as an incremental training set, input it into each sub-model, and based on the new data, retrain the model on a small scale to optimize its internal parameters.

[0030] Preferably, the target sales data includes customer product purchase records, purchase frequency, purchase amount, purchase time, advertisement click volume and customer activity page stay duration.

[0031] Preferably, the business feature set includes the price, category of the product, the consumption amount of the customer on a specific product, the purchase frequency, the average purchase time interval and the input cost of each marketing strategy.

[0032] Preferably, the training set, the validation set and the test set are divided into 6-month cycles, the target sales data of the first 4 months is the training set, the 5th month is the validation set, and the 6th month is the test set.

[0033] Preferably, the deep learning network architecture generates customer demand response, process optimization and strategy adjustment instructions by inputting target sales data to identify inefficient nodes in the sales process.

[0034] The technical effects and advantages of the present application are:

[0035] 1. The present application breaks through the limitations of the existing technology by integrating multi-source data and driving deep learning, realizes a complete closed loop from sales data collection, feature analysis, model prediction to decision execution and feedback optimization, and improves sales efficiency and decision accuracy.

[0036] 2. The three sub-models provide quantitative analysis from three core dimensions: customer needs, process loopholes, and marketing effectiveness. Customer needs prediction combines dynamic factors such as time decay and access depth, process loophole identification locates inefficient nodes in real time, and marketing effectiveness evaluation quantifies ROI, providing comprehensive support for sales decisions.

[0037] 3. This invention continuously iterates model parameters through incremental training and combines transfer learning capabilities to quickly respond to market changes, industry trends, and business adjustments, maintaining a high level of efficiency even in new sales scenarios.

[0038] 4. By extracting sales data and business feature sets, this invention can improve data utilization efficiency, provide high-quality input for model training, improve the analytical accuracy of each sub-model, and help enterprises accurately grasp key nodes in the sales chain. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0040] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0041] Figure 2 This is a schematic diagram of the business feature extraction and architecture construction process of the present invention.

[0042] Figure 3 This is a schematic diagram of the sub-model training process structure of the present invention.

[0043] Figure 4 This is a schematic diagram of the dynamic decision-making and incremental training process structure of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] like Figures 1-4 The AI-assisted intelligent sales link method based on deep learning, as shown, includes the following steps:

[0046] S1: Obtain sales data from the entire sales chain and record it as target sales data;

[0047] The acquisition of the sales data of the whole sales chain includes customer basic transaction data, customer behavior data and marketing activity related data. The customer basic transaction data is extracted from the enterprise ERP system and the e-commerce platform transaction log interface in a structured transaction data storage format of CSV, and the fields include user_id, product_id, purchase_time, purchase_amount and purchase_quantity.

[0048] It should be particularly noted that for old systems that do not support real-time interfaces, a database log Binlog and a listening tool Canal are deployed to real-time parse the change log of the transaction database. When a new purchase record table data is added in the database, the Canal tool captures the record in real time and transmits it to the data buffer layer through the TCP protocol. The transaction data delay is less than or equal to 1 minute, and after each new transaction is generated, it enters the monitoring system within 5 seconds and completes format verification within 30 seconds.

[0049] The customer behavior data is embedded with a lightweight JavaScript script monitoring code in all customer contact front-end pages. When the user triggers a specific behavior, such as clicking on an advertisement or staying for more than 3 seconds, the code automatically collects behavior data, including behavior type, occurrence time, page identifier and device information, and transmits it to the API service based on SpringBoot in real time through the HTTPS protocol.

[0050] It should be particularly noted that the front-end page includes a product detail page, an advertisement landing page and a settlement page. The behavior data reported by the front-end is first entered into a Kafka message queue to set the behavior data theme, and the queue is configured with 3 partitions to ensure concurrent processing capability. Each partition corresponds to one consumer instance, which consumes data in real time and writes it into the time series database InfluxDB. The delay of a single data from generation to writing into the database is less than or equal to 5 seconds.

[0051] The marketing activity related data uses an enterprise marketing automation platform to obtain new marketing strategy execution data through a timed API call every 30 seconds. When the marketing strategy adjustment instruction is executed, the marketing system sends a "strategy execution complete" notification to the monitoring module of the present application immediately after completing the adjustment operation, with the adjusted real-time data.

[0052] It should be particularly noted that the marketing strategy execution data includes the current activity input cost, the number of new customers and the new transaction amount. The adjusted real-time data includes the adjusted amount of investment, the remaining budget, etc.

[0053] It needs to be specifically pointed out that for the records with missing key fields such as purchase time and amount, the adjacent transaction mean is used for filling; the missing stay duration in the customer behavior data is filled with 30% of the average stay duration of all users on the page. Through the IQR rule, extreme high-value transactions, extreme low-value transactions, high-frequency high-value transactions and low-frequency ultra-low-value transactions are marked as abnormal and stored separately, and are not included in the initial training set; the time format is unified as yyyy-MM-dd HH:mm:ss, the amount unit is unified as RMB yuan, and the duration unit is unified as seconds.

[0054] S2: Based on the target sales data, a business feature set is extracted, a training set, a validation set and a test set are divided, and a deep learning network architecture is obtained.

[0055] The extracted business feature set includes the price, category of the product, the consumption amount of the customer on the specific product, the purchase frequency, the average purchase time interval and the input cost of each marketing strategy.

[0056] It needs to be specifically pointed out that the product price is directly extracted from the product basic information field in the target sales data; the product category is mapped to a structured label by associating the product ID with the enterprise product classification table, and is encoded using enumeration values.

[0057] The consumption amount of the customer on the specific product, the total number of times that the customer purchases the specific product and the average purchase time interval of the customer are grouped according to user_id and product_id, and the cumulative consumption total of each customer on the specific product is obtained by summing the purchase_amount field in the target sales data. The number of purchase_record in the target sales data is counted, that is, the total number of times that the customer purchases the specific product. The time difference between the adjacent two purchases is calculated after sorting the purchase_time field by time, and then the average value of all time differences is calculated. If the customer has only purchased once, the average purchase interval of the product category to which the product belongs is taken as a supplement, that is, the average purchase time interval of the customer.

[0058] The input cost of each marketing strategy is obtained by associating the marketing activity ID with the enterprise marketing expense account, summarizing the total input of the specific marketing strategy, and then classifying and counting the feature values according to the strategy type.

[0059] The training set, the validation set and the test set are divided, 6 months are taken as a data cycle, and the data are divided in chronological order. The target sales data of the first 4 months is used for model parameter learning of the training set; the data of the 5th month is used for adjusting the model parameters of the validation set; and the data of the 6th month is used for evaluating the generalization ability of the model.

[0060] The deep learning network architecture adopts a multi-input multi-output deep learning architecture, where the input layer receives a business feature matrix with dimensions of n_samples×n_features, where n_features includes three types of features: product, customer, and marketing; the processing layer adopts a 3-layer fully connected network, with 128 neurons in the first layer, 64 neurons in the second layer, and 32 neurons in the third layer; the three branches of the output layer correspond to three sub-models that output customer purchase probability, inefficient node index, and ROI, respectively.

[0061] S3: Input the target sales data into the deep learning network architecture to train the customer demand prediction sub-model, the sales process loophole identification sub-model and the marketing effect evaluation sub-model, and output the probability of customers purchasing specific products, locate inefficient nodes in the sales chain, and output the impact of different marketing strategies on the final sales conversion rate and transaction amount.

[0062] The trained customer demand prediction sub-model is based on the probability of a customer purchasing a product, specifically expressed as follows: α is the customer purchase time decay coefficient, D is the customer visit depth index, and δ is the customer conversion efficiency.

[0063] It is important to note the customer purchase time decay factor. ∆t is the number of days since the last purchase, and λ is the reciprocal of the category's average repurchase rate, where λ is greater than 0. Customer Visit Depth Index C represents the total number of visits, d i For the i-th customer visit to the product page level, t i The dwell time corresponding to the page level of the i-th customer visit to the product; customer conversion efficiency. N is the amount paid by the customer for a certain product. t The total amount added to the customer's purchase is calculated using b as a smoothing coefficient to avoid a denominator of 0; b = 10.

[0064] The training process uses P as the target value and employs the Adam optimizer, iterating for 50 rounds. Training stops when the accuracy on the validation set reaches 90% or higher in each round.

[0065] The training sales process vulnerability identification sub-model is based on locating inefficient nodes, specifically as follows:

[0066] R1 represents the customer abandonment rate at a certain stage, and T represents the actual average time spent at that stage. a This represents the historical average time for this step.

[0067] It is important to note the customer abandonment rate at a particular stage. C a C TC is the total number of customers entering this link; the actual average time consumption of a certain link , C O C is the total time consumption of all customers entering this link in this link T C is the total number of customers entering this link.

[0068] The training marketing effect evaluation sub-model is based on the marketing input-output ratio, which is specifically expressed as:

[0069] NP is the net profit of a specific product brought by the marketing strategy, T P is the total input of the marketing strategy. The marketing effect evaluation sub-model takes ROI as the output, uses a linear regression model to fit the relationship between the marketing strategy input and the net profit, R 2 coefficient reaches 0.8 or more.

[0070] S4: Based on the training results of the customer demand prediction sub-model, the sales process vulnerability identification sub-model and the marketing effect evaluation sub-model, generate dynamic decision instructions;

[0071] The dynamic decision instruction is based on the output results of the three sub-models, and generates customer demand response instructions, sales process optimization instructions and marketing strategy adjustment instructions.

[0072] It should be particularly noted that the customer demand response instruction is based on the purchase probability P value output by the customer demand prediction sub-model, and through the Drools rule engine, the threshold is set to customer purchase probability P=0.7, automatically dividing customers into different groups, and marking high-intention customers and high-medium-intention customers, Real-time synchronization to the customer relationship management system forms a dynamically updated customer portrait library, supports quick retrieval of target customer list when the instruction is triggered, generates a send targeted coupon instruction for high-intention customers with a customer purchase probability P≥0.8; for high-medium-intention customers with 0.7≤P<0.8, generate a product detail push instruction.

[0073] The sales process optimization instruction is based on the value range of the inefficient node output by the sales process vulnerability identification sub-model, and the trigger condition is that the inefficient node A is 1, if R1>30% generate a simplified purchase process instruction; if T>T a *1.5, generate an optimized payment page loading speed instruction.

[0074] The marketing strategy adjustment instruction is based on the marketing input-output ratio output by the marketing effect evaluation sub-model, if the marketing input-output ratio ROI<0.8, generate a stop-loss instruction, suspend the advertisement of the channel with low input-output ratio, and transfer the budget to the channel with ROI=1.2; if 0.8≤ROI<1.2, generate an optimization instruction, increase the discount strength of the product.

[0075] The generated dynamic decision instruction is automatically sent to the corresponding person in charge, and the execution state of the instruction is marked in the CRM system, and a reminder is set for not completing the task beyond the deadline.

[0076] S5: Monitor the newly generated sales data after the execution of the decision instruction, input the new data as an incremental training set into each sub-model, and retrain the model on a small scale based on the new data to optimize its internal parameters.

[0077] The new data is used as an incremental training set by mixing the new data with 10% of the historical training set to form an incremental training set; the initial learning rate is 20%, each incremental training iteration is 10 rounds, and when the validation set loss decreases by 5%, the focus is on updating the repeat purchase rate coefficient Lambda in the customer demand prediction sub-model and the weight coefficient of the return on investment ROI in the marketing effect evaluation sub-model. When a new product line is added to the business scenario, transfer learning is used.

[0078] It should be particularly noted that the small-scale retraining uses transfer learning to freeze the first two layers of the model weights, only trains the output layer, and fine-tunes the output layer with 3 months of new scenario data until the model accuracy is greater than or equal to 85% under the new scenario.

[0079] Firstly, the present application uses interface real-time synchronization, database log monitoring, front-end point embedding and other technologies to obtain target sales data of the entire sales chain; secondly, through feature engineering, a business feature set is extracted, and the training set, validation set and test set are divided according to the time period, and a multi-input and multi-output deep learning network architecture is designed; then, three sub-models of customer demand prediction, sales process vulnerability identification and marketing effect evaluation are trained respectively, and three key indicators of customer purchase probability, inefficient node index and return on investment are output through quantitative calculation; based on the output results of the sub-models, dynamic decision rules are formulated to generate optimization instructions; finally, through the real-time data monitoring architecture, new data after decision execution is captured, which is used as an incremental training set, and transfer learning strategy is used to retrain the model on a small scale, continuously optimize the model parameters, and form a closed-loop system of data collection-model training-decision generation-effect feedback-model optimization.

[0080] Secondly: the present application discloses an embodiment in the drawings, only the structures involved in the present disclosure are involved, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0081] Finally: the above-mentioned only for the preferred embodiment of the present application, and does not limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based AI-assisted intelligent sales link method, characterized in that, include: S1: Obtain sales data from the entire sales chain and record it as target sales data; S2: Based on the target sales data, extract the business feature set, divide it into training set, validation set and test set, and obtain the deep learning network architecture; S3: Input the target sales data into the deep learning network architecture to train the customer demand prediction sub-model, the sales process loophole identification sub-model and the marketing effect evaluation sub-model, and output the probability of customers purchasing specific products, locate inefficient nodes in the sales link and output the marketing input-output ratio. The specific analysis methods for training the customer demand prediction sub-model include: Analyze the probability P of a customer purchasing a product, where P is the cube root of the sum of the following three terms: the ratio of the customer purchase time decay coefficient to its maximum value, the ratio of the customer visit depth index to its maximum value, and the ratio of the customer conversion efficiency to its maximum value. The specific analysis methods for the time decay coefficient include: The time decay coefficient is equal to the product of the reciprocal of the category's average repurchase rate (e) and the number of days since the last purchase, raised to the power of 0; where the reciprocal of the category's average repurchase rate is greater than 0. The customer visit depth index is analyzed using the following methods: The sum of the product of the page level of the i-th customer visit and its corresponding dwell time is calculated, and then the sum is divided by the total number of visits. The quotient is the customer visit depth index. The specific analysis methods for customer conversion efficiency include: The customer conversion efficiency is the sum of the amount paid by a customer for a product and the smoothing factor, divided by the sum of the total amount added by the customer and the smoothing factor. The specific analysis methods of the sales process vulnerability identification sub-model include: The analysis of inefficient nodes is based on the following criteria: when the customer abandonment rate of a certain stage is greater than or equal to 30%, or when the actual average time spent in a certain stage exceeds the historical average time spent in that stage, that stage is considered an inefficient node. The specific analysis methods for customer abandonment rate at a certain stage include: Divide the number of customers who abandon the process at this stage by the total number of customers who enter the process at this stage, and then multiply the result by 100%. The final result is the customer abandonment rate at this stage. The specific analysis methods for the actual average time spent in a certain step include: The total time spent by all customers in this stage is divided by the total number of customers in this stage, and the final result is the actual average time spent in this stage. The specific analysis methods of the training marketing effectiveness evaluation sub-model include: The return on investment (ROI) is the ratio of the net profit generated by a marketing strategy to the total investment in that marketing strategy. S4: Based on the training results of the customer demand prediction sub-model, the sales process loophole identification sub-model, and the marketing effectiveness evaluation sub-model, generate dynamic decision instructions; The dynamic decision-making instructions include sales process optimization instructions, marketing strategy adjustment instructions, and customer demand response instructions. The trigger conditions for the sales process optimization instruction are based on the value range of inefficient nodes output by the sales process vulnerability identification sub-model; the trigger conditions for the marketing strategy adjustment instruction are based on the marketing input-output ratio output by the marketing effectiveness evaluation sub-model; and the trigger conditions for the customer demand response instruction are based on the purchase probability P-value output by the customer demand prediction sub-model. S5: Monitor the new sales data generated after the execution of the decision-making instructions, use the new data as an incremental training set, input it into each sub-model, and retrain the model on a small scale based on the new data to optimize its internal parameters.

2. The AI-assisted intelligent sales link method based on deep learning according to claim 1, characterized in that: The target sales data mentioned in S1 includes customer product purchase records, purchase frequency, purchase amount, purchase time, ad clicks, and customer activity page dwell time.

3. The AI-assisted intelligent sales link method based on deep learning according to claim 1, characterized in that: The business feature set includes product price, category, customer spending on a specific product, number of purchases, average purchase interval, and cost of marketing strategies.

4. The AI-assisted intelligent sales link method based on deep learning according to claim 1, characterized in that: As described in S2, the training set, validation set, and test set are divided into three sets with a 6-month cycle. The target sales data for the first 4 months is used as the training set, the data for the 5th month is used as the validation set, and the data for the 6th month is used as the test set.

5. The AI-assisted intelligent sales link method based on deep learning according to claim 1, characterized in that: The deep learning network architecture described in S2 identifies inefficient nodes in the sales process by inputting target sales data, generating customer demand responses, process optimization, and strategy adjustment instructions.

Images