Method and system for generating activity plan based on artificial intelligence
Through feature extraction and vectorization representation, the activity scheme is generated, and the time-limited discount strategy is adjusted in combination with collaborative filtering and reinforcement learning algorithms, the problem of event scheme generation and time-limited discount strategy is solved, and efficient personalization of marketing activities and user satisfaction is achieved. promote.
Patent Information
- Application Number
- CN202410623889.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-20
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-05-20
AI Technical Summary
During the generation of activity plan, the prompt elements entered by the user have vague semantic expression, missing key information, and uneven data structure. How to quickly match and generate personalized activity plans that are close to user needs, and need to be balanced when there is a limited time offer. Discount strength and marketing costs, real-time tracking of user feedback data to optimize generation strategies, has problems.
Through feature extraction and vectorization representation, the object subject, copy summary, style, etc. faced by the activity plan are converted into input feature vectors, and related text is generated using deep neural networks, and combined with collaborative filtering algorithms and reinforcement learning algorithms, the limited-time discount strategy is dynamically adjusted. , collect user feedback data in real time for incremental learning optimization, and generate personalized limited-time discount strategies.
It improves the efficiency and effectiveness of marketing activities, provides a personalized experience, improves user satisfaction and loyalty, and enhances the market competitiveness of merchants.
Smart Images

Figure CN118537044B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information technology, and in particular to an activity plan generation method and system based on artificial intelligence. Background Art
[0002] In the process of generating activity plans, how to efficiently match and associate the prompt elements input by users with the data in the pre-stored activity plan database is a technical problem that needs to be solved urgently. First, the prompt elements input by users may have problems such as ambiguous semantic expression and missing key information, which need to be semantically understood and supplemented through natural language processing technology. Secondly, in the matching process, the degree of structure of different activity plan data varies, and the amount of data is huge. How to quickly find the data with the highest correlation with the prompt elements and generate activity plans that are close to user needs based on this is a complex optimization problem. At the same time, there are large differences in the historical behavior data and preference characteristics of different users. When generating personalized activity plans, it is also necessary to fully explore and utilize user portrait information, which puts higher requirements on the design and implementation of algorithm models. In addition, when generating limited-time offers, how to balance the strength of the offer with the marketing cost, and dynamically adjust the offer generation strategy to adapt to market changes, so as to stimulate users' impulse consumption while taking into account the normal ordering speed of users is also a difficult problem to be solved. It is necessary to balance the length of the offer. If the offer time limit is too long, it may not stimulate users' consumption. If the offer time limit is too short, it may lead to insufficient time for users to understand and place orders, resulting in customer loss. Finally, how to track user feedback data in real time and use the incremental learning mechanism to continuously optimize the activity plan generation model and offer generation strategy is the key to ensuring the long-term and effective operation of the system. Summary of the invention
[0003] The present invention provides an activity plan generation method based on artificial intelligence, which mainly includes:
[0004] By extracting features or vectorizing the prompt elements of the activity plan, including the object subject, copywriting summary, activity style, and marketing products, an input feature vector of the activity plan generation model is obtained, and based on the input feature vector, the activity plan generation model is used to output the current activity plan text;
[0005] Obtain the consumption data of users in historical activities, extract the consumption characteristics of users in various historical activities through data mining and statistical analysis, and use collaborative filtering algorithm to generate personalized limited-time discount strategies through discount strategy generation model according to the current activity plan text and the corresponding user consumption characteristics;
[0006] In the current activity plan text, dynamically adjust the discount strength and validity period of the limited-time discount strategy, monitor the user's response to different limited-time discount items in the limited-time discount strategy in real time, and use reinforcement learning algorithms to continuously optimize the limited-time discount strategy;
[0007] During the current activity, we continuously collect actual user feedback data, including user evaluation of the current activity, user purchase behavior and interaction behavior, and extract key information and sentiment information from user feedback through the natural language processing model.
[0008] Incremental learning and online learning algorithms are used for key information and emotional tendencies in user feedback to update and optimize the activity plan generation model and the limited-time discount strategy generation model in real time to obtain the optimal limited-time discount strength and time limit in different activity plans.
[0009] The present invention provides an activity plan generation system based on artificial intelligence, which mainly includes:
[0010] An activity plan generation module is used to generate an activity plan text according to activity elements;
[0011] User consumption feature extraction module, used to analyze historical data to extract user consumption features;
[0012] A personalized discount strategy generation module is used to create personalized limited-time discounts for users;
[0013] Discount strategy optimization module, which is used to optimize discount intensity and validity period through reinforcement learning;
[0014] User feedback analysis module, used to extract key and emotional information of user feedback through NLP;
[0015] Real-time update optimization module, used to update activity plans and discount strategy models.
[0016] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:
[0017] The present invention discloses an activity plan generation method and system based on artificial intelligence. Aiming at the business scenario problems of marketing activity plan formulation and limited-time discount strategy personalization, the present invention converts prompt elements such as the object subject, copywriting summary, required style, and marketing products of the activity plan into input feature vectors through feature extraction and vectorization representation. Then, these input feature vectors are passed into a pre-trained deep neural network model, and the relevant activity plan text is automatically generated by using multi-layer nonlinear transformation and self-attention mechanism, and semantic optimization and adjustment are performed by natural language processing technology to ensure the fluency and relevance of the text. In solving the problem of personalized limited-time discount strategy, the present invention adopts data mining and statistical analysis technology to analyze the user's historical consumption data and discover consumption preferences. Using this information and combining with collaborative filtering algorithm, a personalized limited-time discount strategy matching the user's consumption characteristics is generated. By dynamically adjusting the discount strength and validity period of the discount, and using reinforcement learning algorithm to monitor and optimize the strategy in real time, the present invention aims to improve the user's order speed and conversion rate.
[0018] During the activity, the present invention continuously collects actual feedback data from users, and uses sentiment analysis and natural language processing technology to extract key information and sentiment tendencies in the feedback. Through incremental learning and online learning algorithms, the present invention can update and optimize the activity plan generation model and limited-time discount generation strategy in real time, and fine-tune the limited-time discount strength and time limit.
[0019] In general, the present invention provides an intelligent activity plan generation and personalized marketing strategy optimization solution by integrating machine learning algorithms, natural language processing technology and data analysis. This not only improves the efficiency and effectiveness of marketing activities, but also strengthens the personalized experience of users, thereby enhancing the market competitiveness of merchants while improving user satisfaction and loyalty. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 The present invention is a flowchart of an activity plan generation method based on artificial intelligence.
[0021] Figure 2 A schematic diagram of an activity plan generation method and system based on artificial intelligence of the present invention.
[0022] Figure 3 This is another schematic diagram of an activity plan generation method and system based on artificial intelligence of the present invention. DETAILED DESCRIPTION
[0023] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.
[0024] like Figure 1-3 In this embodiment, an activity plan generation method and system based on artificial intelligence may specifically include:
[0025] S101. Extract features or vectorize the prompt elements of the activity plan, including the object subject, text summary, activity style, and marketing products, to obtain the input feature vector of the activity plan generation model, and output the current activity plan text through the activity plan generation model based on the input feature vector.
[0026] According to the prompt elements faced by the activity plan, including the object subject, the copy summary, the activity style, and the marketing product, text analysis and semantic understanding are performed through a natural language processing model to extract keywords and subject words to form a structured feature representation; the extracted keywords and subject words are converted into multi-dimensional word vectors, and the semantically similar word vectors are clustered through word vector clustering to obtain a semantic feature vector; the semantic feature vector is used as the input of the activity plan generation model, and the activity plan generation model includes a mapping relationship between the input feature vector and the output activity plan text; before the activity plan text is generated by the trained activity plan generation model, the parameters and weights of the activity plan generation model are dynamically adjusted based on the training set; the generated activity plan text is post-processed and optimized, and the language model and knowledge graph are used to perform grammatical error correction, semantic completion, and logic optimization on the text; according to the generated activity plan text, in combination with the marketing product and activity style, corresponding visual materials and design plans are generated to form an activity plan solution.
[0027] For example, according to the prompt elements faced by the activity plan, including the object subject, copywriting summary, activity style, marketing products and other information, text analysis and semantic understanding are performed through natural language processing technologies such as TF-IDF and Word2Vec, and keywords and subject words are extracted to form a structured feature representation. Then, word embedding models such as Word2Vec or GloVe are used to map keywords and subject words into a low-dimensional semantic space to obtain word vector representation. Next, word vector clustering algorithms such as K-means and DBSCAN are used to aggregate semantically similar word vectors to obtain higher-level semantic feature vectors. The semantic feature vector is used as the input of the activity plan generation model, and the generation rules and patterns of the activity plan are learned through deep learning networks such as Seq2Seq model and Transformer model. Among them, the Seq2Seq model encodes the input feature vector into a hidden state through the encoder-decoder structure, and then generates the output text through the decoder; the Transformer model captures the long-distance dependency in the input sequence through the self-attention mechanism and position encoding to generate high-quality text. When training the activity plan generation model, a large amount of historical activity consumption data or hypothetical activity consumption data with true value annotations is collected as a training set, and the parameters and weights of the activity plan generation model are dynamically adjusted through the difference between the training output and the true value annotation. When generating the activity plan text, attention mechanisms such as BahdanauAttention or LuongAttention are used to dynamically generate a weight matrix by calculating the similarity between the input feature vector and the hidden state, and used to adjust the output of the model so that it can pay more attention to the contextual information related to the current input and improve the coherence and relevance of the generated text. In order to improve the diversity and innovation of the generated text, technologies such as random noise and variational autoencoder (VAE) are introduced. Random noise increases the diversity of generated text by adding certain random perturbations to the input or hidden state; variational autoencoder introduces hidden variables to learn the probability distribution of input data, thereby generating more diverse and innovative text. The generated activity plan text is post-processed and optimized, and language models such as GPT and BERT and knowledge graphs such as ConceptNet are used to perform grammatical error correction, semantic completion, and logic optimization on the text to improve the readability and rationality of the text. Among them, language models can learn language rules and semantic information in large-scale text data through pre-training, which can be used for text generation and optimization; knowledge graphs can provide rich common sense knowledge and conceptual relationships, which can be used to infer logical relationships in texts and supplement related information.According to the optimized activity plan text, combined with information such as marketing products and activity styles, natural language processing technology is used to extract key information in the text, such as themes, emotions, scenes, etc., and then image generation models (such as GAN, VAE) are used to generate corresponding visual materials based on this information. Through style transfer and image synthesis technology, the materials are matched and optimized with the text content to generate supporting posters, promotional pictures, videos, etc., forming a complete activity plan solution to provide support for the execution and promotion of the activity. When performing text analysis on the prompt elements of the activity plan, the TF-IDF algorithm can be used to calculate the importance of each word, and the top-50 words with the highest weight are selected as keywords. At the same time, the Word2Vec model is used to convert these keywords into 200-dimensional word vectors. The K-means algorithm is used to cluster the word vectors, and the number of clusters is set to 10. After 50 iterations, 10 semantic feature clusters are obtained. The Seq2Seq model is used to generate activity plan text. Both the encoder and decoder use 3-layer LSTM, the hidden layer dimension is 512, and the Dropout rate is 0.2. In the generation process, the BahdanauAttention mechanism is used to calculate the attention weight between the current decoding state and the encoder output, and the weighted summed context vector is concatenated with the current decoding state as the input for the next decoding. Gaussian noise with a mean of 0 and a standard deviation of 0.1 is added to the decoder input to improve the diversity of the generated text. After the generation is completed, the BERT model is used to optimize the text, and the MaskedLanguageModel task and NextSentencePrediction task are used to perform grammatical error correction and semantic completion on the text. Concepts and relationships related to the text are retrieved from the ConceptNet knowledge graph to infer and supplement the logical relationships in the text. For example, if the text mentions "red" and "festive", the common sense that "red symbolizes festivity and auspiciousness" can be inferred from the knowledge graph to enrich the text content. When generating supporting visual materials, the YOLOv5 model can be used to extract key information from the text, such as "spring", "outdoors", "young people", etc., and then the StyleGAN2 model can be used to generate corresponding images based on this information, and the CycleGAN model can be used to transfer the style to match the text content. The generated text and visual materials are integrated to form a complete activity plan.
[0028] Based on the historical data in the pre-stored activity plan database, a machine learning algorithm is used to train the activity plan generation model, including passing the historical input feature vectors of the historical data into the pre-trained deep neural network model, generating activity plan text related to the prompt elements through multi-layer nonlinear transformation and self-attention mechanism, and semantically optimizing and adjusting the generated activity plan text.
[0029] The historical input feature vectors of the historical data are passed into a pre-trained deep neural network model, and an activity plan text related to the prompt elements is generated through multi-layer nonlinear transformation and self-attention mechanism; the generated activity plan text is semantically optimized and adjusted, including using a natural language processing model to perform structured analysis on the generated text and identify key information and semantic relationships; according to predefined optimization rules and strategies, errors in the activity plan text are corrected, and the activity plan text is adjusted and supplemented through context information and knowledge base; random noise and variational autoencoders are introduced in the generation process, and random perturbations are added to the input or hidden state to enrich the activity plan text; the enriched activity plan text is output, and corresponding visual materials and design plans are generated according to the marketing products and activity styles to form a complete activity plan solution.
[0030] Exemplarily, based on the historical data in the pre-stored activity plan database, the sequence generation model in deep learning, such as RNN, LSTM, Transformer, etc., is used to train the activity plan generation model. By learning the input feature vectors in the historical data and the corresponding activity plan text, a mapping relationship between the feature vector and the text is established. During the training process, the input historical feature vectors are subjected to multi-layer nonlinear transformation using pre-trained deep neural network models such as BERT and GPT, and high-level semantic features are extracted using multi-layer fully connected neural networks and activation functions such as ReLU and Sigmoid. The self-attention mechanism is introduced to dynamically adjust the weight of the features by calculating the correlation between feature vectors, so that the model can pay more attention to the historical information related to the current input. Based on the activity plan generation model obtained through training, the current prompt elements are converted into corresponding input feature vectors and passed into the model for reasoning and generation. The generated activity plan text may have problems such as grammatical errors and semantic incoherence, so it needs to be further semantically optimized and adjusted. The generated text is structured and parsed using natural language processing technologies such as syntactic analysis and semantic role labeling. Through syntactic analysis, the main components of the sentence are identified to determine whether there are grammatical errors; through semantic role labeling, the predicate and argument in the sentence are identified to determine whether there is semantic incoherence. Then, according to predefined rules and strategies, the identified errors are corrected and adjusted, such as deleting redundant words, adjusting the order of words, and supplementing missing components. In order to further improve the diversity and innovation of the generated text, random noise and variational autoencoders are introduced in the generation process. A certain amount of Gaussian noise is added to the input or hidden state of the generation model, so that the results of each generation are slightly different; the variational autoencoder is combined with the generation model, and the generated text can present richer and more varied characteristics through the sampling and reconstruction of hidden variables. It is also possible to consider introducing technologies such as reinforcement learning and adversarial generative networks, and further improve the quality and diversity of the generated text by designing appropriate reward functions and loss functions. The optimized and adjusted activity plan text can better meet the needs and expectations of users and improve the quality and readability of the activity plan. In order to better achieve the title goal, after the text is generated, the quality of the generated activity plan is evaluated and screened by manual evaluation or automatic evaluation, and the high-quality plan is selected as the final output. Establish a user feedback channel to collect users' opinions and suggestions on the generated plan, and use them to iteratively optimize the model and improve the generation strategy. You can also consider introducing external knowledge sources such as knowledge graphs and domain ontologies to enrich the model's background knowledge and common sense reasoning ability, making the generated activity plan more reasonable and convincing. Output the optimized activity plan text, and generate corresponding supporting materials and design plans as needed, such as posters, promotional pictures, videos, etc., to form a complete activity plan solution.When training the activity plan generation model, the long short-term memory network (LSTM) can be used as the sequence generation model, and the input layer dimension is set to 512, the hidden layer dimension is 1024, the output layer dimension is 512, the learning rate is 0.001, the batch size is 64, and the training epochs is 100. The Transformer model is used as the pre-training model, and the number of encoder and decoder layers is set to 6, the number of attention heads is 8, the feedforward network dimension is 2048, and the dropout rate is 0.1. When extracting semantic features, a 3-layer fully connected neural network is used, and the dimensions of each layer are 768, 1024, and 1280 respectively, and the activation functions are ReLU and Sigmoid. When performing syntactic analysis, Stanford Parser is used to perform dependency syntactic analysis on the generated text, and grammatical components such as subject-verb-object structure and attributive clauses are identified through rule matching, and the height and width of the syntactic tree are calculated to determine whether there are sentences that are too long or too complex. When performing semantic role labeling, the BERT pre-trained model is used to fine-tune the text, identify the predicates and arguments in the sentence, and calculate the semantic similarity between the arguments and the predicates to determine whether there is semantic incoherence. When introducing random noise, the mean of the noise can be set to 0 and the standard deviation to 0.1, and the influence of the noise can be gradually reduced during the training process. When introducing the variational autoencoder, the dimension of the latent variable can be set to 256, and the KL divergence can be used as a regularization term to control the distribution of the latent variable to be close to the standard normal distribution. When performing reinforcement learning, a reward function based on indicators such as ROUGE and BLEU can be designed, and the PPO algorithm can be used for policy optimization, with the learning rate set to 0.0001, the discount factor set to 0.99, and the batch size set to 32. When performing adversarial generation, the conditional generative adversarial network (CGAN) can be used, and the network structure of the generator and discriminator can be set to a 3-layer fully connected neural network, with the dimensions of each layer being 512, 1024, and 2048 respectively, and the WGAN-GP loss function can be used for training, and the penalty coefficient can be set to 10. During manual evaluation, 100 generated activity plans can be randomly selected and scored by three experts. The scoring dimensions include topic relevance, content richness, language fluency, logical rationality, etc. A 5-point scoring system is used, and the average score and standard deviation are calculated to screen out the top 10% with the highest scores as high-quality plans.
[0031] S102. Obtain the consumption data of the user in historical activities, extract the consumption characteristics of the user corresponding to each historical activity through data mining and statistical analysis, and use collaborative filtering algorithm according to the current activity plan text and the corresponding user's consumption characteristics to generate personalized limited-time discount strategies through the discount strategy generation model.
[0032] The consumption data of the user in historical activities is obtained, and the consumption data includes consumption amount, consumption frequency, consumption time, and consumption commodity category; the consumption data is preprocessed, including cleaning, deduplication, and formatting, to obtain a structured consumption data set; the consumption data set is analyzed to obtain the consumption characteristics of the user in different historical activities, and the consumption characteristics include consumption amount range, consumption frequency distribution, and consumption commodity preference; according to the current activity plan text, key information is extracted, including activity theme, target user group, and promotion intensity, and matched with the consumption characteristics to screen out a user subset related to the current activity; for the user subset, a collaborative filtering prediction model based on items is used to calculate the similarity matrix between users, and the participation and consumption potential of each user in the current activity are predicted according to the similarity matrix; through the preferential strategy generation model, the user consumption characteristics and the collaborative filtering prediction results are combined to generate a personalized limited-time preferential strategy; in the process of generating the preferential strategy, a reinforcement learning module is introduced to dynamically optimize the preferential strategy by real-time tracking of user feedback and activity effects; and the generated personalized limited-time preferential strategy is pushed to all target users in the target user subset.
[0033] Exemplarily, the consumption data of users in historical activities are obtained, including multi-dimensional data such as consumption amount, consumption frequency, consumption time, and consumption commodity category, and the data is cleaned, deduplicated, formatted, and other pre-processing operations are performed to form a structured consumption data set. The Apriori algorithm is used to mine frequent item sets and association rules to find the correlation between different commodities; the K-means algorithm is used to cluster users, and the consumption characteristics of different user groups are characterized according to the clustering results to form a user consumption feature matrix. According to the current activity plan text, key information such as activity theme, target user group, and promotion intensity are extracted, and matched with the user consumption feature matrix to filter out user subsets whose correlation with the current activity is higher than the threshold. For the filtered user subset, the item-based collaborative filtering algorithm is used. First, a user-item rating matrix is constructed based on the user's historical consumption record; the similarity between items is calculated using similarity measurement methods such as Euclidean distance and Pearson correlation coefficient; based on item similarity and user historical ratings, the user's interest in unrated items is predicted; the Top-N items with the highest interest are selected as recommendation results to predict each user's participation in the current activity and consumption potential. Combining user consumption characteristics and collaborative filtering results, a preferential strategy generation model is constructed through an RNN-based sequence generation model. The input is user consumption characteristics and activity information, and the output is preferential strategy text. Specifically, the preferential text corresponding to the user consumption characteristics is generated. For example, if the user consumption characteristics are preferential consumption preferences for discounted goods, then "emphasize the discount logo" or "strengthen the original discount" is added to the preferential text; for example, if the activity information is the issuance of coupons, then "coupon targeted issuance" or "coupon can be used in a superimposed manner" is added to the preferential text. During the training process, the cross entropy loss function and Adam optimizer are used, and the attention mechanism and BeamSearch decoding are added to improve the accuracy and diversity of the generation, and automatically generate personalized limited-time preferential strategies, such as giving higher discounts to high-consumption users and first-order discounts to new users, and dynamically adjust the preferential strength and validity period to maximize the activity conversion rate and user satisfaction. In the process of generating preferential strategies, the reinforcement learning algorithm DDPG (Deep Deterministic Policy Gradient) is introduced to define the state space including user characteristics, activity information, historical preferential strategies, etc., and the value range of preferential strength and validity period is used as the action space. The reward function is designed according to indicators such as coupon usage rate and user satisfaction, guiding the intelligent agent to generate more effective preferential strategies, dynamically optimize preferential strategies by real-time tracking of user feedback and activity effects, and continuously improve the effectiveness and accuracy of strategies in multiple rounds of iterations. Finally, the generated personalized limited-time preferential strategies are pushed to target users, and users are reached through multiple channels such as SMS, email, APP push, and public account push, and the best channel combination is selected according to user attributes and preferences.Introduce technologies such as AB testing and multi-armed bandit to dynamically adjust the push strategies and content of different channels, improve the user's response rate and conversion rate, and ultimately achieve precision marketing and maximize the long-term value of users. When obtaining user historical consumption data, big data processing tools such as Hive can be used to perform ETL processing on TB-level raw log data, extract key fields such as user ID, product ID, consumption amount, consumption time, and generate a structured consumption record table through data cleaning and conversion. Use the FP-Growth algorithm in SparkMLlib, set the minimum support to 0.01 and the minimum confidence to 0.5, mine frequent item sets and association rules, and find the correlation between different products. For example, the association support of "milk→bread" is 0.08 and the confidence is 0.7. Use the K-means algorithm to cluster user consumption records, set the number of clusters K to 5, the number of iterations to 10, and get 5 user groups. Calculate the average consumption amount, purchase frequency and other indicators of each group to form a user consumption feature matrix. When performing collaborative filtering, cosine similarity is used to calculate the similarity between items, and the nearest neighbor algorithm is used to select the top-50 similar items. Based on the user's historical rating, the UserCF algorithm is used to predict the user's participation in the current activity. If the predicted score is greater than 4 points, it is considered that the user has a high willingness to participate. When generating preferential strategies, a GRU-based sequence generation model is used, the word vector dimension is set to 128, the hidden layer dimension is set to 256, the learning rate is 0.001, and 100 epochs are trained. Using the DDPG algorithm in reinforcement learning, the Actor network and the Critic network are both set to be 3-layer fully connected networks, with hidden layer dimensions of 64 and 128 respectively, and the reward function is "0.6 coupon usage rate + 0.4 user satisfaction", and the exploration noise is attenuated from 0.9 to 0.1 to achieve adaptive optimization of the strategy. When pushing through multiple channels, the multi-armed bandit algorithm is used to dynamically select the best channel combination, set the exploration rate to 0.2, and update the return rate of each channel based on feedback indicators such as click-through rate and conversion rate to continuously improve the push effect.
[0034] S103. In the current activity plan text, dynamically adjust the discount strength and validity period of the limited-time discount strategy, monitor the user's response to different limited-time discount items in the limited-time discount strategy in real time, and use a reinforcement learning algorithm to continuously optimize the limited-time discount strategy.
[0035] Through real-time monitoring of user behavior data, key behavior indicators, including click-through rate and shopping cart information, are captured to obtain users' initial reactions to the current limited-time discount. If the user's click-through rate for any limited-time discount item is higher than the preset threshold, the current discount parameters are judged to be attractive to the user group corresponding to the user, so as to maintain or increase similar discounts. Based on the analysis of product attributes, promotional products suitable for further discounts are determined, and the discount range is adjusted according to the profit margin, seasonality and inventory of the promotional products. Data analysis tools are used to analyze historical sales data and time factors, predict the best time period for promotional activities, and set the corresponding validity period. By monitoring the market environment, the limited-time discount strategy is adjusted to cope with the promotional activities of competitors and fluctuations in market demand. Platform traffic data is obtained, peak access time is analyzed, and the release time of limited-time discounts is adjusted according to user activity to increase exposure. Feedback information on limited-time discount items is collected through the user feedback mechanism. If the user evaluation is lower than the threshold, the limited-time discount strategy is adjusted in time to improve user satisfaction. Reinforcement learning algorithms are used to adjust the current discount parameters based on real-time feedback data and optimize each limited-time discount item. Combined with cost-benefit analysis, the actual benefits brought by each limited-time discount project should be evaluated regularly, and those limited-time discount projects that do not meet the standards should be adjusted or eliminated to ensure the economic rationality of the entire limited-time discount strategy.
[0036] Exemplarily, according to the text of the current activity plan, the limited-time preferential policy information contained therein, such as discount strength, validity period, applicable commodity range, etc., is extracted to form a structured preferential policy representation. Through methods such as embedding points and log collection, various behaviors of users on the activity page, such as clicking, browsing, and receiving coupons, are recorded in real time, and these data are stored in a log system or data warehouse in a structured form to form a user behavior log data set. The preferential policy representation and the user behavior data set are encoded as vectors respectively, and they are spliced together to form a complete state vector as the input of the reinforcement learning algorithm. In the reinforcement learning algorithm, the preferential policy is used as the action space of the intelligent agent, and the improvement of user response is used as the reward function. Through continuous trial and error and learning, the optimal preferential policy combination is found. The deep Q network (DQN) algorithm is adopted, and a neural network is used to fit the Q value function. The state vector is used as the input of the network, and the corresponding Q value is output. The strategy with the highest Q value is used as the current optimal strategy. The experience replay mechanism is introduced, and a circular queue is used to store the transfer samples obtained by the interaction between the intelligent agent and the environment, including the current state, action, reward, next state and other information. During training, uniform sampling or priority sampling is used to extract small batches of samples from the experience pool and use them for neural network training and parameter update. In order to balance exploration and utilization, the ε-greedy strategy is adopted, and the initial value of ε is set to 0.9. As the training progresses, the value of ε is gradually reduced in an exponential decay manner until a smaller threshold such as 0.1 is reached. When implementing the algorithm, a random number generator is used to decide whether to choose the action with the highest Q value or a random action. The optimal discount strategy obtained by reinforcement learning is applied to the current activity plan, and the discount strength, validity period, and applicable product range of the limited-time discount strategy are dynamically adjusted. The effect of the discount strategy is continuously monitored, new user behavior data is collected, and a closed-loop feedback is formed. The discount strategy is continuously optimized and improved to achieve adaptive adjustment and personalized recommendation of the discount strategy. When extracting the limited-time discount strategy information in the activity plan text, regular expressions can be used to match discount strengths such as "70% off", "8.5% off", validity periods such as "3 days", "7 days", and applicable product ranges such as "universal" and "specified products available". Keywords, and the matching results are stored in the discount strategy object in JSON format. When collecting user behavior logs, you can use tracking tools such as GrowingIO, SensorsData, etc. By inserting JS code into the page, you can record the user's clicks, browsing, coupon collection and other behaviors in real time, and clean and convert the logs in the form of Map / Reduce tasks. Finally, store them in the HDFS distributed file system in Parquet format, and generate a new partition every day.In the reinforcement learning algorithm, the MLP network can be used as an approximate function of the Q function, with an input state vector dimension of 100, including preferential strategy features and user behavior features, three hidden layers, and each layer dimension is 256, 128, and 64, respectively. The output layer dimension is the size of the preferential strategy action space, such as 10. The size of the experience replay pool is set to 10,000, the batch size is set to 128, the learning rate is set to 0.001, and the discount factor is set to 0.95. In the ε-greedy strategy, the initial value of ε can be set to 0.9, and ε can be decayed to 0.99 times the original value every 1,000 iterations until the minimum value is 0.1. According to the optimal strategy obtained by reinforcement learning, the discount strength, validity period and applicable product range of the preferential strategy are dynamically adjusted, such as adjusting the discount strength from 20% to 30%, extending the validity period from 3 days to 5 days, and adjusting the applicable product range from designated products to the entire store, etc., and real-time monitoring of the discount strategy's key indicators such as the write-off rate, number of new users, and repurchase rate. According to the changes in indicators, it is decided whether to further adjust the preferential strategy to form a closed-loop feedback and continuous optimization. For example, it is monitored that a limited-time offer has a click-through rate of 5% within the first hour after release, exceeding the preset threshold of 3%. Therefore, it is judged that the discount is highly attractive to users, and it is decided to maintain the current discount strength. Based on seasonal considerations, an additional discount of 10% is provided to accelerate inventory turnover. Historical data shows that the second week of November each year is the peak period for sales. Therefore, a series of promotional activities can be launched during this period, and a validity period of one week can be set. In response, a flash sale with a 25% discount can be launched for 48 hours to maintain market competitiveness. Data shows that Saturday afternoon from 2 to 5 is the peak time for user visits. Therefore, a limited-time discount strategy can be scheduled to be released at this time to increase exposure and participation. After an event that includes a limited-time discount strategy, 30% of users expressed dissatisfaction with the discount in user feedback. In response, it is planned to increase the discount by at least 5 percentage points in the next event to improve user satisfaction. After 100 rounds of iterations, the algorithm found that when the discount is set between 15% and 20%, the probability of user purchase is the highest. Based on this, the algorithm is adjusted to self-optimize within this discount range. For example, the total cost of a promotion is 10,000 yuan, the additional turnover generated is 15,000 yuan, and the net profit is 5,000 yuan. Considering the profit margin of 15%, the economic rationality of the event is relatively high. For those limited-time discount projects that generate a profit margin of less than 10%, adjustments or cancellations are made. For example, by real-time monitoring of user behavior data, it is found that the click-through rate of a certain product has reached 8%, which is higher than the preset threshold of 5%. It is judged that the current discount strength of the product is more attractive to users, and the discount range can be maintained or appropriately increased. Based on the analysis of product attributes, it is determined that a seasonal product with a profit margin of 30% and a stock volume of 1,000 pieces is suitable for further discounts. The discount rate can be adjusted from the original 20% to 30%.For example, by analyzing historical sales data, it is found that during the National Day, the average daily sales can reach 3 times the usual level. Therefore, the National Day can be set as the best time period for promotional activities, and a 7-day validity period can be set. By monitoring the promotional activities of competitors, it is found that they have launched a discount of 100 yuan for purchases over 500 yuan. In order to cope with the competition, the limited-time discount strategy of our store can be adjusted to 100 yuan for purchases over 400 yuan. By analyzing the platform traffic data, it is found that 20:00-22:00 every day is the peak access period, and the average daily UV can reach 100,000. Therefore, the release time of the limited-time discount can be adjusted to this time period to obtain higher exposure. Through the user feedback mechanism, the average score of a certain limited-time discount project is collected as 5 points (out of 5 points), which is 4 points lower than the preset threshold. Therefore, it is necessary to adjust the discount strength and promotion method of the project in a timely manner to improve user satisfaction. Reinforcement learning algorithms are used to dynamically adjust the discount of products based on real-time feedback data from users, such as click-through rate, purchase conversion rate, etc. If the click-through rate increases from 5% to 8% after the discount is adjusted from 20% to 30%, the next step is to try to adjust the discount to 50%. Through cost-benefit analysis, it is found that the promotion cost of a certain limited-time discount project is 10,000 yuan, but the actual sales revenue is only 8,000 yuan, which does not meet the expected goal. Therefore, the project needs to be adjusted or eliminated to ensure the economic rationality of the overall promotion strategy.
[0037] Based on user behavior data and market dynamics, a discount strategy adjustment model is established to predict the optimal discount level and optimal validity period.
[0038] Collect user behavior data, including user visit frequency, shopping cart addition history, and purchase history. Through data preprocessing and feature engineering, extract user behavior characteristics, including average purchase time and probability of responding to discounts. Collect market dynamics, including competitive product discount strategies, market supply and demand information, and seasonal influences, integrate time factors, and use time series analysis methods to obtain market dynamic trends. Obtain discount product information, including cost price, inventory level, and current price, and analyze the discount range based on market dynamic trends. The discount strategy adjustment model calculates the consumer sensitivity coefficient based on the probability of responding to discounts and the average purchase time. Combined with the consumer discount sensitivity coefficient and inventory level, the discount strategy adjustment model calculates the optimal discount strength and the optimal validity period.
[0039] For example, the consumption behavior data of users in different marketing activities are obtained, including the purchase amount, purchase frequency, and purchase category of users, etc. Through data cleaning and feature engineering, a structured user consumption feature data set is constructed. The preferential promotion information of competitors in the market is collected, including discount strength, validity period, participation conditions, etc., and a structured market dynamic data set is formed through automatic crawling by web crawlers or manual entry. For the manual entry method, it is necessary to formulate standardized entry templates and processes, train and assess the entry personnel, establish a data verification and review mechanism, and conduct regular data quality inspections and evaluations to ensure the accuracy and efficiency of the entry. The discount intensity, validity period and other parameters of historical preferential strategies are analyzed in association with their corresponding sales, profits, user conversion rate and other performance indicators. The Apriori algorithm or FP-growth algorithm is used to mine frequent item sets and association rules to find the association patterns between different preferential parameter combinations and performance indicators. Decision tree algorithms such as ID3, C4.5 or CART are used to generate decision trees with preferential strategy parameters as features and performance indicators as target variables. The performance of decision trees is optimized through pruning, cross-validation and other methods to find the rules between preferential strategy parameters and performance indicators. The sensitivity of different user groups to price discounts is analyzed. Through user questionnaires, A / B tests and other methods, the user's acceptance and preference data for different discount intensity are collected. Using the price-demand elasticity theory, a linear regression or logarithmic linear regression model is established, with the commodity price as the independent variable and the sales volume or purchase probability as the dependent variable to estimate the price elasticity coefficient; or discrete choice models such as Logit and Probit are used to estimate the user's price sensitivity parameters and establish a user price sensitivity model based on the user's choice behavior at different price levels. According to the user price sensitivity model, users are divided into different groups such as high sensitivity, medium sensitivity, and low sensitivity. Then, the purchasing behavior and product preferences of different sensitivity groups under different price discounts are analyzed. The historical sales data, inventory data, profit margin data, etc. of the products are integrated. Through data analysis and mining, the most suitable product categories and SKUs for promotion, as well as the promotion time period and target user groups are identified, and differentiated pricing strategies for different user groups are found. Based on the user consumption feature data set and the market dynamic data set, machine learning algorithms such as logistic regression and random forest are used to establish a preferential strategy adjustment model. With sales, profit, user conversion rate, etc. as optimization targets, the optimal discount intensity and validity period combination for different user groups under different market environments are predicted. Methods such as grid search or random search are used to tune the hyperparameters of the algorithm; feature selection methods such as filtering, wrapping, or embedding are used to select the most relevant and discriminative feature subsets; feature engineering techniques such as standardization, normalization, and binning are used to transform and extract the original features to improve the performance and generalization ability of the model.Compare and evaluate the prediction results of the preferential strategy adjustment model with the overall goals of the marketing activities. By setting confidence thresholds, combined with business rules and constraints, generate executable preferential strategy adjustment plans, and conduct AB testing and effect monitoring. Collect user evaluations, suggestions and satisfaction data on preferential strategies through user behavior tracking, questionnaires, customer feedback and other channels; clean, annotate and structure the feedback data to form a data set that can be used for model training and evaluation. Regularly compare and analyze the model's prediction results and actual effects, identify prediction deviations and anomalies, and retrain, fine-tune or upgrade the model based on feedback data to continuously improve the accuracy and applicability of the model and form a closed-loop optimization of preferential strategy adjustment. When collecting user consumption behavior data, you can use big data processing platforms such as Hive to perform ETL processing on the original user behavior logs, extract key fields such as user ID, product ID, purchase time, purchase amount, etc., and aggregate them on a monthly or quarterly basis to generate a structured user consumption feature data set. When crawling competitors' promotional information, you can use Python's Scrapy framework to write crawler scripts, regularly obtain detailed information on promotional activities from competitors' official websites, apps and other channels, and use the Pandas library to clean and structure the crawled data to form a market dynamic data set. When mining association rules through the Apriori algorithm, you can set the minimum support to 0.05 and the minimum confidence to 0.8 to find out that users who buy product A have an 80% probability of also buying product B. When using the CART decision tree algorithm to establish a preferential strategy effect prediction model, you can set the maximum depth of the tree to 5 and the minimum number of samples of leaf nodes to 100. The accuracy of the model is evaluated through 10-fold cross validation, and the average accuracy rate reaches more than 85%. When conducting user price sensitivity analysis, a logarithmic linear regression model can be used, with the product discount rate as the independent variable and the purchase conversion rate as the dependent variable. When the discount rate is increased from 10% off the original price to 70% off, the purchase conversion rate of highly sensitive users increases from 5% to 20%, while the purchase conversion rate of low-sensitivity users only increases from 3% to 8%. When establishing a discount strategy adjustment model, you can use the XGBoost algorithm, set the number of trees to 100, the learning rate to 0.1, optimize hyperparameters through grid search, and use feature importance sorting to select the top-20 features to improve the prediction accuracy of the model. For different user groups and market environments, predict the optimal discount strength and validity period combination, such as giving a 20% discount to highly sensitive users at a competitive promotion node with a validity period of 3 days, which is expected to increase sales by more than 20%. For example, obtain the user behavior data of all target users who have purchased discounted products. The target users visited the website an average of 3 times a week in the past 6 months, added 6 items to the shopping cart on average per month, and actually purchased 3 items per month on average.The target user added the product to the shopping cart 10 times on average without any discount, and the product was purchased after the discount was paid for 2 times. After feature engineering processing, the average purchase time interval (from adding to the shopping cart to payment) of the user is 5 days, and the probability of responding to the discount is 20%. Through time series analysis, it is found that at the end of summer, the discount strategy of competing products usually increases the discount intensity, and the market demand increases during this period. For example, from July to September, the average discount intensity of competing products increased from 10% to 15%, and the overall market demand increased from 1,000 units per month to 1,500 units. Specifically, for a popular summer dress, the cost price is 100 yuan, the current inventory level is 1,000 pieces, and the current price is 200 yuan. Combined with the trend of market dynamics, the discount intensity range should be 15%-30%. Combined with the probability of user A responding to the discount (20%) and the average purchase time (5 days), the consumer discount sensitivity coefficient can be calculated as follows: Discount sensitivity coefficient = response discount probability * average purchase time = 20 / 5 = 4. Through the calculation of the discount strategy adjustment model, it is found that providing a 20% discount (5% more than the highest discount of competing products) and a validity period of 25 days can bring the greatest benefits to the platform. In summary, through the comprehensive analysis and calculation of user behavior data, market dynamics and product information, a discount strategy that can both attract users and maximize profits can be obtained.
[0040] S104. During the current activity, actual feedback data from users is continuously collected, including users' evaluation of the current activity, users' purchasing behavior and interactive behavior, and key information and sentiment information from user feedback is extracted through a natural language processing model.
[0041] Emotional information includes emotional tendency and emotional intensity. Through the activity information system, the activity name, activity type, activity time range and participation conditions of the current activity are extracted to confirm the relevance of user evaluation to the current activity. Text analysis tools are used to extract evaluation content, evaluation time, evaluation score or star rating and evaluation label from user evaluation, and the emotional tendency is analyzed through the sentiment analysis unit. Specifically, the user feedback content is parsed through the natural language processing model, the user comment text data is extracted, and the sentiment analysis algorithm is used to determine the emotional tendency of the extracted text data to distinguish positive, negative or neutral emotions. Through the user purchase history information system, purchase behavior data is obtained, including purchased products or services, purchase quantity, purchase frequency, purchase time, order amount and purchase channel. Through the user interaction tracking system of the website and APP, the user's interactive behavior such as page browsing history, click path, stay time, and shopping cart addition and deletion behavior is recorded. Analyze the information of collected or liked products and the user's interactive data on social media, including likes, shares and comments, and extract key emotional vocabulary and determine the corresponding emotional intensity through the natural language processing model. Use keyword extraction and topic model analysis tools in the natural language processing model to process user evaluation texts and determine the main topics and semantic relevance of the evaluations. If the semantic relevance exceeds a reasonable threshold, further analyze the text content of the user feedback content on the main topics through part-of-speech tagging and entity recognition, identify and associate important entity information, including names of people, places, and / or organizations. Perform data cleaning on the collected data, remove invalid or erroneous data, perform data integration, integrate structured data (including purchase behavior data) and unstructured data (including evaluation content), and encapsulate them into key information.
[0042] For example, by embedding user feedback forms and evaluation components in the activity page, users' evaluations and opinions on the current activity are collected in real time, including text data such as user satisfaction, suggestions and complaints. Use HTML5 form and Ajax asynchronous submission, realize dynamic rendering and verification of the form through front-end frameworks such as Vue.js, and use third-party evaluation components such as Trustpilot and ReviewTrackers, etc., integrated into the page through JSSDK, realize the collection and display of evaluation data, and transmit these feedback data to the data processing server. Collect users' purchase behavior data during the activity, including users' browsing history, purchase history, order history, payment history, etc., insert JSSDK into the page through third-party tracking platforms such as GrowingIO and SensorsData, automatically collect users' behavior data, and report it to the data warehouse in real time. Collect users' interactive behavior data on the activity page, including users' clicks, scrolling, hovering and other operations. For the APP side, use iOS non-tracking SDK and Android non-tracking SDK, and use AOP aspect programming and Runtime method replacement and other technologies to realize automatic collection and reporting of user interactive behavior data to the data analysis platform. The collected user feedback text data is processed and analyzed using a natural language processing model. A sentiment classification model based on BERT is used to make positive, neutral, negative and other sentiment judgments on user feedback. Keyword extraction algorithms such as TextRank are used to extract keywords and subject terms from the feedback text. The BiLSTM-CRF named entity recognition model is used to identify key entities such as goods, brands, and services mentioned in user feedback. A semantic role labeling model based on dependency parsing is used to analyze semantic information such as opinion subjects, sentiment polarity, and evaluation objects in user feedback to enrich the dimensions and granularity of sentiment analysis. The collected user purchase behavior data and interaction behavior data are processed and mined using a user behavior analysis model. The user behavior conversion and loss in the activity are analyzed through methods such as funnel analysis, retention analysis, and path analysis. Use the Markov chain model to establish the user behavior transition probability matrix, and use the Viterbi algorithm to decode and predict the state sequence; use association rule mining algorithms such as Apriori and FP-Growth to discover frequent patterns and association rules of user behavior; use the user clustering algorithm based on the RFM model to divide users into high-value, medium-value, low-value and other groups according to their recent consumption time, consumption frequency, consumption amount and other dimensions, to achieve precise marketing and personalized services, and discover key behavior nodes and bottlenecks that affect the effectiveness of activities.Combine the sentiment analysis results and keyword extraction results of user feedback with the user's purchasing behavior and interactive behavior, and use BI tools such as Tableau and PowerBI to clean, integrate, model and visualize data from different sources and dimensions, generate interactive data analysis reports and dashboards, and form a comprehensive evaluation report on the effect of the activity, including visualization charts such as activity conversion funnel charts, user retention curve charts, user behavior path charts, as well as numerical tables of key indicators and text descriptions such as trend analysis, to display key indicators such as the overall conversion rate, sales, and average order value of the activity, as well as experience indicators such as user satisfaction and complaint rate. According to the activity effect evaluation report and user feedback analysis results, optimize and adjust the current activity, including adjusting the rules of play, discount strength, page layout, etc., and gradually apply the optimized activity plan through AB testing and gray release, and continuously monitor the optimization effect to form a closed loop of continuous improvement. When embedding a user feedback form in an activity page, you can use the v-form component of Vue.js to implement dynamic validation of form fields through declarative form validation rules, such as required, min, max, etc., and asynchronously submit form data to the backend API interface through Ajax to achieve real-time data transmission. For the collection of user evaluation data, you can use the JSSDK provided by ReviewTrackers to asynchronously load the SDK script after the page is loaded, and automatically render the evaluation component on the right side of the page by specifying parameters such as productId, so as to achieve one-click submission of user evaluation data and real-time synchronization to the ReviewTrackers platform. When collecting user purchase behavior data, you can use the non-embedded SDK provided by the GrowingIO platform to automatically collect user browsing, adding to cart, ordering and other behavior data by introducing a line of JS code in the head of the page, and set the collection granularity to the button level to achieve detailed collection and analysis of key behaviors. For the collection of APP user behavior data, you can use AspectOrientedProgramming on the iOS platform and XposedFramework on the Android platform. Through the AOP aspect, you can hook the AppDelegate and Activity lifecycle methods, automatically collect data when the user enters and leaves the page, and replace the Runtime method to implement parameter parsing and reporting of key behavior buttons. The amount of user behavior data generated every day can reach tens of millions. When performing sentiment analysis on user feedback text, you can use a fine-tuning model based on BERT, with three sentiment classifications (positive, neutral, and negative) as the task, and train it on a data set of 300,000 hotel reviews. The final model has an accuracy rate of more than 95% on the test set.When extracting keywords from user feedback, the TextRank algorithm can be used to construct a co-occurrence matrix between candidate keywords, iteratively calculate the centrality score of each candidate keyword, and finally select the top-10 keywords with the highest scores. The BiLSTM-CRF named entity recognition model is used to identify key entities such as goods and brands mentioned in user feedback with an F1 value of 0.95. The BI-LSTM-based semantic role labeling model is used to analyze the semantic roles such as evaluation subjects, evaluation objects, and evaluation words in user feedback with an F1 value of 0.8. When analyzing user purchase behavior, the FPGrowth algorithm can be used to mine frequent item sets and association rules of user purchases with a minimum support of 0.01 and a minimum confidence of 0.5, such as "users who bought product A have an 85% probability of also buying product B." When performing user grouping, the K-Means algorithm can be used, with the consumption amount, consumption frequency, and average order value in the last month as features, the K value is set to 5, and iterates 100 times to divide users into five groups: high value, medium-high value, medium value, medium-low value, and low value. The RFM model is used to label each user with numerical labels of three dimensions: R, F, and M. For example, the activity name retrieved by the system is "Spring Big Sale", the activity type is "Limited Time Discount", the activity time range is from April 1, 2023 to April 7, 2023, and the participation condition is "Cumulative shopping of more than $100", confirming the user reviews associated with the current activity. The text analysis tool extracts a review from the user review, "I really like this promotion, the product quality is very good", the review time is April 2, 2023, the review score is 5 stars, and the review tags include "satisfied" and "high cost performance". The sentiment analysis unit analyzes the above evaluation content and determines that the sentiment tendency is positive. The sentiment intensities of the sentiment words "like" and "very good" in the evaluation are 8 and 9 respectively (on a sentiment intensity scale of 0 to 1). The purchase behavior data shows that the user purchased a product worth $150 on April 1, 2023, which is his fifth purchase, and he purchases it once every 2 months on average. The order amount is $150 and it is purchased through the official website. The user's browsing history on the website shows that the user stayed on a product page for 5 minutes, and the click path included homepage->promotion area->product details. During this period, 2 items were added to the shopping cart, and only 1 item was settled in the end. The user interacted with a post about the "Spring Big Sale" on social media, which was expressed as 150 likes, 30 shares, and 20 comments. The most frequently appearing sentiment keyword was "super value" with a sentiment intensity of 7. The keyword extraction results show that the main topics in the user's evaluation are "discount strength" and "logistics speed". Through the topic model analysis, it is found that the semantic relevance is 85, which exceeds the set threshold of 7. Part-of-speech tagging and entity recognition techniques analyze user reviews, identify “Spring Sale” as the main topic, and associate it with the entity information “XYZ Mall” (organization name).Among the 1,000 pieces of collected data, 30 pieces of erroneous data (such as duplicate or erroneous format data) were removed through data cleaning, and then 970 pieces of structured purchase behavior data and unstructured evaluation content data were combined and encapsulated into key information.
[0043] S105. Incremental learning and online learning algorithms are used for key information and emotional tendencies fed back by users to update and optimize the activity plan generation model and the limited-time discount strategy generation model in real time, so as to obtain the optimal limited-time discount strength and time limit in different activity plans.
[0044] The sentiment analysis algorithm is used to convert the key information and sentiment tendencies of user feedback, convert the key information into user satisfaction, preferred activity type, and support for the discount strength, and convert the sentiment tendencies into positive and / or negative scores. The incremental learning algorithm is used to analyze the changing trends of user sentiment tendencies and key information, and dynamically adjust the parameters of the activity plan generation model. The online learning algorithm is used to update the limited-time discount strategy generation model in real time. According to the activity plan generation model, multiple alternative activity plans are output, each of which contains different discount strengths and time limits. The multi-objective optimization algorithm is used to evaluate the output multiple alternative activity plans, and the combination of the strength and time limit of the limited-time discount in each alternative plan is optimized. Through simulation tests, different market responses are simulated to evaluate the effects of each optimized alternative activity plan. Combined with the simulation test results and actual market data, the optimal activity plan is determined from multiple alternative activity plans, including the most suitable limited-time discount strength and time limit.
[0045] For example, user feedback is collected through channels such as in-App feedback, customer service records, and social media comments. Feedback data is captured and preprocessed using a distributed crawler system and data cleaning tools. Natural language processing technology is used to perform word segmentation, part-of-speech tagging, named entity recognition, sentiment analysis, and other processing on user feedback text data. Key information such as goods, services, and problems mentioned in the feedback is extracted, and the positive, negative, and neutral sentiment tendencies of the feedback are judged to form a structured feedback feature vector. The extracted user feedback feature vector is used as the input of the incremental learning and online learning algorithms to dynamically update the activity plan generation model and the limited-time discount strategy generation model. Online learning algorithms based on gradient descent such as OnlineGradientDescent are used. Every time a new user feedback sample is received, the loss function value of the current model is calculated based on the feature vector and label of the sample, and the partial derivative of the model parameters is calculated based on the loss function to obtain the gradient vector. The model parameters are updated at a certain learning rate along the direction of gradient descent. At the same time, adaptive learning rate adjustment strategies such as AdaGrad and RMSprop are used to dynamically adjust the learning rate of each parameter based on historical gradient information to accelerate the convergence speed. When updating the activity plan generation model, the Transformer-based Seq2Seq model architecture is adopted. The Encoder part uses recurrent neural networks such as LSTM or GRU to serialize the user feedback feature vector into a context vector of fixed length; the Decoder part also uses LSTM or GRU to predict the probability distribution of the next character based on the context vector and the previously generated characters, and generates the complete activity plan text through decoding algorithms such as BeamSearch. Attention mechanisms such as dot product attention and multi-head attention are added between the Encoder and the Decoder, so that the Decoder can adaptively focus on different parts of the Encoder output sequence when generating each character, thereby improving the accuracy and coherence of the generated text. When updating the limited-time discount strategy generation model, reinforcement learning algorithms such as DeepQ-Network are used, activity effect indicators (such as conversion rate, user satisfaction, etc.) are used as reward values for environmental feedback, and the combination of limited-time discount strength and time limit is used as the action space of the agent. The action-state value function is fitted through Q-value networks such as multi-layer perceptrons and convolutional neural networks, and the experience replay mechanism and TD error are used as loss functions to update the parameters of the Q-value network. At the same time, improved algorithms such as DoubleDQN and DuelingDQN are introduced, and mechanisms such as target network and priority experience replay are added to improve the stability of the algorithm and sample utilization efficiency. When generating limited-time discount strategies for different activity plans, the Monte Carlo tree search algorithm is used to take each activity plan as the root node of the tree and each limited-time discount strategy as a child node. Through random sampling and strategy evaluation, the search tree is continuously expanded and updated.In the expansion phase, according to the current node's visit times, expected benefits and other statistical information, heuristic strategies such as UCB are used to select the optimal child node for expansion; in the evaluation phase, leaf nodes are quickly evaluated using methods such as fast chess moves to obtain the expected benefit estimate of the node, and the statistical information of the ancestor node is updated by backpropagation; in the application phase, according to the statistical information of the child nodes of the root node, the child node with the most visits or the highest expected benefit is selected as the optimal strategy, and deep reinforcement learning technologies such as DeepMCTS and AlphaZero are introduced to continuously optimize the evaluation function and strategy function of the search tree through self-learning. The optimized activity plan and limited-time discount strategy are applied to actual marketing activities, and the effectiveness indicators of the activities, such as user engagement, conversion rate, sales, etc., are continuously monitored. On the one hand, the activity effect indicators are used as new samples for incremental learning and online learning algorithms, and the model parameters are updated regularly or in real time; on the other hand, active learning, transfer learning and other technologies are used to screen and enhance samples according to activity effect indicators, and automated machine learning technologies such as AutoML and Neural Architecture Search are used to automatically search and optimize the model's hyperparameters and architecture according to activity effect indicators, continuously improve the generalization ability and robustness of the model, and form a closed-loop learning and optimization process. When collecting user feedback data, you can use Python crawler frameworks such as Scrapy to write distributed crawler scripts, regularly crawl user comments from app stores such as AppStore and GooglePlay, and use data processing libraries such as Pandas to clean and deduplicate the crawled data. On average, more than 100,000 user feedback data can be collected every day. When preprocessing text, you can use NLP libraries such as Jieba and NLTK to perform operations such as word segmentation, part-of-speech tagging, and named entity recognition on user feedback texts, and use sentiment analysis libraries such as SnowNLP and TextBlob to judge the sentiment tendency of the text, with an accuracy rate of more than 85%. When conducting online learning, you can use online learning tools such as vowpalwabbit, set the learning rate to 0.01, the regularization parameter to 0.1, and perform an incremental update on the model for every 1,000 new user feedback samples received. The entire process takes no more than 10 minutes. When using Transformer to generate activity plan text, you can use pre-trained language models such as OpenAI GPT and BERT, set the hidden layer dimensions of Encoder and Decoder to 768, the number of attention heads to 12, the training batch size to 32, the learning rate to 0.0001, and use the BeamSearch decoding algorithm, set the beam width to 5, and the generated activity plan text is fluent, with an average BLEU score of more than 75 points.When using DeepQ-Network to optimize the limited-time discount strategy, the Q value network can be set to 3-layer MLP, the hidden layer dimension is 256, the activation function is ReLU, the optimizer is Adam, the learning rate is 0.001, the experience replay pool size is 10000, the target network update frequency is 1000 times, and the ε-greedy exploration strategy is used, the initial exploration rate is 0.9, and the decay is 0.99 every 1000 times. The discount strategy after final convergence can increase the activity conversion rate by more than 10%. When using the Monte Carlo tree search algorithm, the UCB exploration coefficient can be set to 1.414, the number of simulations for each search is 1000, the search depth is 10, the evaluation function is random chess, and the multi-armed bandit algorithm is used to dynamically adjust the structure of the search tree. When no better strategy can be found after 5 consecutive searches, the search tree is pruned and reconstructed. The optimal activity plan finally searched out has a click-through rate and conversion rate of more than 15% in the AB test. When fine-tuning and optimizing the model, you can use automated hyperparameter search tools such as Google Vizier, set the search space to: hidden layer dimension [128, 256, 512], learning rate [0.0001, 0.001, 0.01], regularization coefficient [0.01, 0.1, 1.0], and use the Hyperband algorithm for adaptive resource allocation and early stopping. The optimal hyperparameter combination can improve the performance of the model by 5%-10%. For example, from the user feedback data, the text mining unit identified that 60% of users expressed high satisfaction, 30% of users preferred outdoor activities, and 10% of users were dissatisfied with the current discount strength. After being processed by the sentiment analysis algorithm, the user's emotional tendency was quantified as: positive score 8 (satisfied), negative score -2 (dissatisfied). The incremental learning algorithm found that in the past month, the user's positive emotional tendency towards the discount strength has dropped from 7 to 5, indicating that the user's satisfaction with the discount strength has declined. Online learning algorithm updates limited-time discount strategy: For example, the online learning algorithm increases the discount strength parameter from 20% to 25% based on real-time data to quickly adapt to market changes. The model outputs three plans: Plan A offers a 30% discount for 1 week; Plan B offers a 20% discount for 2 weeks; and Plan C offers a 25% discount for 10 days. The multi-objective optimization algorithm predicts that Plan B will attract the most additional customers and has the best combination of discount strength and time limit. Simulation tests show that sales are expected to increase by 15% under simulated market response with Plan B, while Plans A and C can only increase by 10% and 12% respectively. Based on simulation tests and actual market data analysis, Plan B (20% discount for 2 weeks) was finally determined to be the optimal activity plan. After implementing Plan B, the newly collected user feedback data showed that the satisfaction with the discount strategy increased to 8.5, indicating a significant improvement in user satisfaction.
[0046] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for generating activity plans based on artificial intelligence, It is characterized in that The method comprises: By extracting features or vectorizing the prompt elements of the activity plan, including the object subject, copywriting summary, activity style, and marketing products, an input feature vector of the activity plan generation model is obtained, and based on the input feature vector, the activity plan generation model is used to output the current activity plan text; Obtain the consumption data of users in historical activities, extract the consumption characteristics of users in various historical activities through data mining and statistical analysis, and use collaborative filtering algorithm to generate personalized limited-time discount strategies through discount strategy generation model according to the current activity plan text and the corresponding user consumption characteristics; In the current activity plan text, dynamically adjust the discount strength and validity period of the limited-time discount strategy, monitor the user's response to different limited-time discount items in the limited-time discount strategy in real time, and use reinforcement learning algorithms to continuously optimize the limited-time discount strategy; During the current activity, we continuously collect actual user feedback data, including user evaluation of the current activity, user purchase behavior and interaction behavior, and extract key information and sentiment information from user feedback through the natural language processing model. Incremental learning and online learning algorithms are used for key information and emotional tendencies in user feedback to update and optimize the activity plan generation model and the limited-time discount strategy generation model in real time to obtain the optimal limited-time discount strength and time limit in different activity plans.
2. The method according to claim 1, in, The method extracts features or vectorizes the prompt elements of the activity plan, including the object subject, the copy summary, the activity style, and the marketing product, to obtain the input feature vector of the activity plan generation model, and outputs the current activity plan text through the activity plan generation model based on the input feature vector, including: According to the prompt elements of the activity plan, including the subject, copywriting summary, activity style and marketing products, the natural language processing model is used to perform text analysis and semantic understanding, extract keywords and subject terms, and form a structured feature representation; The extracted keywords and subject words are converted into multi-dimensional word vectors, and word vectors with similar semantics are aggregated through word vector clustering to obtain semantic feature vectors; Using the semantic feature vector as input to an activity plan generation model, the activity plan generation model including a mapping relationship between the input feature vector and an output activity plan text; Before generating the activity plan text by using the trained activity plan generation model, dynamically adjusting the parameters and weights of the activity plan generation model based on the training set; Post-processing and optimizing the generated activity plan text, using a language model and knowledge graph to perform grammatical error correction, semantic completion and logic optimization on the text; According to the generated activity plan text, combined with the marketing products and activity style, corresponding visual materials and design plans are generated to form an activity plan solution; It also includes: using a machine learning algorithm to train an activity plan generation model based on historical data in a pre-stored activity plan database, including passing the historical input feature vectors of the historical data into a pre-trained deep neural network model, generating activity plan text related to the prompt elements through multi-layer nonlinear transformations and self-attention mechanisms, and semantically optimizing and adjusting the generated activity plan text.
3. The method according to claim 2, in, The method uses a machine learning algorithm to train an activity plan generation model based on historical data in a pre-stored activity plan database, including passing historical input feature vectors of historical data into a pre-trained deep neural network model, generating activity plan text related to prompt elements through multi-layer nonlinear transformation and self-attention mechanism, and semantically optimizing and adjusting the generated activity plan text, including: The historical input feature vector of the historical data is passed into the pre-trained deep neural network model, and the activity plan text related to the prompt element is generated through multi-layer nonlinear transformation and self-attention mechanism; Performing semantic optimization and adjustment on the generated activity plan text, including using a natural language processing model to perform structural analysis on the generated text and identify key information and semantic relationships; Correct errors in the activity plan text according to predefined optimization rules and strategies, and adjust and supplement the activity plan text through context information and knowledge base; Introducing random noise and variational autoencoders in the generation process to enrich the activity scenario text by adding random perturbations in the input or hidden state; The enriched activity plan text is output, and corresponding visual materials and design plans are generated according to the marketing products and activity styles to form a complete activity plan solution.
4. The method according to claim 1, in, The method of obtaining the consumption data of the user in historical activities, extracting the consumption characteristics of the user in each historical activity through data mining and statistical analysis, and using the collaborative filtering algorithm to generate a personalized limited-time preferential strategy through a preferential strategy generation model according to the current activity plan text and the corresponding user's consumption characteristics, includes: Obtaining the consumption data of the user in historical activities, wherein the consumption data includes consumption amount, consumption frequency, consumption time and consumption commodity category; Preprocessing the consumption data, including cleaning, deduplication and formatting, to obtain a structured consumption data set; Analyze the consumption data set to obtain the consumption characteristics of the user in different historical activities, wherein the consumption characteristics include consumption amount range, consumption frequency distribution and consumption commodity preference; Extract key information from the current activity plan text, including activity theme, target user group, and promotion intensity, and match it with the consumption characteristics to filter out a user subset related to the current activity; For the user subset, an item-based collaborative filtering prediction model is used to calculate a similarity matrix between users, and the participation and consumption potential of each user in the current activity are predicted based on the similarity matrix; Generate a personalized limited-time discount strategy by combining the user consumption characteristics and the collaborative filtering prediction results through the discount strategy generation model; In the process of generating the preferential strategy, a reinforcement learning module is introduced to dynamically optimize the preferential strategy by tracking user feedback and activity effects in real time; The generated personalized limited-time discount strategy is pushed to all target users in the target user subset.
5. The method according to claim 1, in, In the current activity plan text, the discount strength and validity period of the limited-time discount strategy are dynamically adjusted, and the limited-time discount strategy is continuously optimized by using a reinforcement learning algorithm through real-time monitoring of the user's response to different limited-time discount items in the limited-time discount strategy, including: Through real-time monitoring of user behavior data, key behavior indicators, including click-through rate and shopping cart information, are captured to obtain users’ initial reactions to the current limited-time offers; If the user's click rate on any limited-time discount item is higher than a preset threshold, the current discount parameters are judged to be attractive to the user group corresponding to the user, so as to maintain or increase the strength of similar discounts; Based on product attribute analysis, identify promotional products that are suitable for further discounts, and adjust the discount amount based on the profit margin, seasonality and inventory level of the promotional products; Use data analysis tools to analyze historical sales data and time factors, predict the best time period for promotional activities, and set corresponding validity periods; By monitoring the market environment, adjust the limited-time discount strategy to cope with the promotional activities of competitors and fluctuations in market demand; Obtain platform traffic data, analyze peak visit periods, and adjust the release time of limited-time offers according to user activity to increase exposure; Collect feedback information on limited-time discount items through the user feedback mechanism. If the user evaluation is below the threshold, adjust the limited-time discount strategy in time to improve user satisfaction; Adopt reinforcement learning algorithm to adjust current discount parameters according to real-time feedback data and optimize various limited-time discount items; Combined with cost-benefit analysis, regularly evaluate the actual benefits of each limited-time discount project, and adjust or eliminate the limited-time discount projects that do not meet the standards to ensure the economic rationality of the entire limited-time discount strategy; It also includes: establishing a discount strategy adjustment model based on user behavior data and market dynamics, and predicting the optimal discount level and optimal validity period.
6. The method according to claim 5, in, The above-mentioned method of establishing a preferential strategy adjustment model based on user behavior data and market dynamics to predict the optimal discount intensity and optimal validity period includes: Collect user behavior data, including user visit frequency, shopping cart addition history, and purchase history. Through data preprocessing and feature engineering, extract user behavior characteristics, including average purchase time and probability of responding to discounts; Collect market dynamics, including competitive product discount strategies, market supply and demand information, and seasonal impacts, integrate time factors, and use time series analysis methods to obtain market dynamic trends; Obtain information about discounted products, including cost price, inventory level and current price, and analyze the discount range based on market trends; The preferential strategy adjustment model calculates the consumer sensitivity coefficient based on the probability of responding to preferential treatment and the average purchase time; Combining the consumer's discount sensitivity coefficient and inventory level, the discount strategy adjustment model calculates the optimal discount intensity and optimal validity period.
7. The method according to claim 1, in, During the current activity, actual feedback data from users is continuously collected, including the user's evaluation of the current activity, the user's purchase behavior and interaction behavior, and key information and sentiment information from user feedback are extracted through a natural language processing model, including: Emotional information includes emotional tendency and emotional intensity; Through the activity information system, the activity name, activity type, activity time range and participation conditions of the current activity are extracted to confirm the relevance of the user evaluation to the current activity; Use text analysis tools to extract evaluation content, evaluation time, evaluation score or star rating, and evaluation tags from user evaluations, and use sentiment analysis units to analyze sentiment tendencies. Specifically, use natural language processing models to parse user feedback content, extract user comment text data, and use sentiment analysis algorithms to determine sentiment tendencies of the extracted text data, distinguishing between positive and negative or neutral emotions. Obtain purchase behavior data through the user purchase history information system, including purchased products or services, purchase quantity, purchase frequency, purchase time, order amount and purchase channel; Through the user interaction tracking system of the website and APP, record the user's interaction behavior such as page browsing history, click path and dwell time, as well as shopping cart addition and deletion behavior; Analyze the information of favorite or favourite products, as well as the user interaction data on social media, including likes, shares and comments, and extract key sentiment words and determine the corresponding sentiment intensity through the natural language processing model; Use keyword extraction and topic model analysis tools in the natural language processing model to process user evaluation texts and determine the main topics and semantic relevance of the evaluations; If the semantic relevance exceeds a reasonable threshold, the text content of the user feedback content on the main topic is further analyzed through part-of-speech tagging and entity recognition to identify and associate important entity information, including names of people, places and / or organizations; Clean the collected data, remove invalid or erroneous data, perform data integration, integrate structured data and unstructured data, and encapsulate them into key information, where structured data includes purchase behavior data and unstructured data includes evaluation content.
8. The method according to claim 1, in, The key information and sentiment tendency of user feedback are incrementally learned and online learned, and the activity plan generation model and the limited-time discount strategy generation model are updated and optimized in real time to obtain the optimal limited-time discount strength and time limit in different activity plans, including: Use sentiment analysis algorithms to convert key information and sentiment tendencies in user feedback into data, convert key information into user satisfaction, preferred activity types, and support for discounts, and convert sentiment tendencies into positive and / or negative scores; Analyze the changing trends of user sentiment and key information through incremental learning algorithms, and dynamically adjust the parameters of the activity plan generation model; Apply online learning algorithms to update the limited-time discount strategy generation model in real time; Generate a model based on the activity plan and output multiple alternative activity plans, each of which contains different preferential strengths and time limits; Use a multi-objective optimization algorithm to evaluate the output of multiple alternative activity plans and optimize the combination of the strength and time limit of the limited-time discount in each alternative plan; Through simulation testing, different market responses are simulated to evaluate the effectiveness of each optimized backup activity plan; Combine simulation test results with actual market data to determine the optimal activity plan from multiple alternative activity plans, including the most suitable limited-time discount intensity and time limit.
9. An activity plan generation system according to the activity plan generation method based on artificial intelligence as claimed in claim 1, It is characterized in that The system comprises: An activity plan generation module is used to generate an activity plan text according to activity elements; User consumption feature extraction module, used to analyze historical data to extract user consumption features; A personalized discount strategy generation module is used to create personalized limited-time discounts for users; Discount strategy optimization module, which is used to optimize discount intensity and validity period through reinforcement learning; User feedback analysis module, used to extract key and emotional information of user feedback through NLP; Real-time update optimization module, used to update activity plans and discount strategy models.
