Data report generation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of natural language data processing, in particular to a data report generation method and system based on artificial intelligence, and the method comprises the following steps: collecting interaction data of e-commerce platform users, and generating cleaned data through data cleaning and preprocessing; and based on the cleaned data, using a clustering algorithm to classify user behavior patterns, and determining a user group. By collecting and cleaning the user interaction data of the e-commerce platform, high quality and applicability of the data are guaranteed, and a foundation is laid for analysis. And the user behaviors are classified carefully by using a clustering algorithm, so that the understanding of the user group behaviors is improved, and the individuation and the accuracy are enhanced. And particularly, key user behavior characteristics are extracted from user behavior classification, so that the relevance and effect of recommendation are optimized. And the influence of different features is calculated through statistical analysis, so that the recommendation process is further refined, and the recommendation model can more flexibly adapt to user requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language data processing, and in particular to a data report generation method and system based on artificial intelligence. Background Art

[0002] Natural language data processing is an important branch of artificial intelligence and computational linguistics, which involves enabling computers to understand, interpret, generate and respond to forms of human language. This technical field focuses on developing algorithms and systems that can extract information from text and voice data. Data report generation usually includes data summarization, feature extraction, trend analysis and prediction.

[0003] However, existing technologies usually rely on fixed algorithm models in the generation of personalized data reports and lack dynamic adjustment mechanisms, which limits their adaptability when user behavior changes rapidly. For example, if they fail to update and reflect changes in market trends and user interests in a timely manner, the generated content may not meet user needs, thereby reducing user experience and purchase conversion rates. Therefore, improvements are needed. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a data report generation method and system based on artificial intelligence.

[0005] In order to achieve the above object, the present invention adopts the following technical solution, a data report generation method based on artificial intelligence, comprising the following steps: Collect interaction data of users of the e-commerce platform, and generate cleaned data through data cleaning and preprocessing; based on the cleaned data, apply clustering algorithms to classify user behavior patterns, determine user groups, and generate user behavior classification results; Extracting user behavior features from the user behavior classification results, including the most frequently visited product categories and average consumption levels, and generating a user feature summary; applying statistical analysis to the user feature summary to calculate the influence of each feature and generating a feature impact analysis result; Based on the feature impact analysis results, user features and product information are input, and a recommendation model is obtained through collaborative filtering algorithm processing; the recommendation model is applied to the e-commerce platform data to make real-time recommendations, and the recommendation effects and user feedback are recorded to obtain the recommendation effects; Collect user feedback on the recommendation effect, analyze the feedback, extract satisfaction indicators, and generate user feedback analysis results; adjust the parameters of the collaborative filtering algorithm according to the user feedback analysis results to generate an optimized recommendation model.

[0006] Preferably, the steps of obtaining the cleaned data are: Collect interaction data of e-commerce platform users, including click-through rate, purchase history, and search history, to obtain a complete original interaction data set; Based on the complete original interaction data set, incomplete and abnormal data are identified and removed, outlier detection and missing data processing are performed to obtain a screened cleaned data set; Based on the screened cleaned data set, missing value filling and data format conversion are applied to generate cleaned data.

[0007] Preferably, the steps of obtaining the user behavior classification result are: Based on the cleaned data, a K-means algorithm is used and the number of clusters and the distance metric are set to obtain an algorithm configuration result; Based on the algorithm configuration result, clustering is performed on the cleaned data, and it is iterated until the clustering result is stable to obtain a preliminary user behavior classification result; Based on the preliminary user behavior classification results, abnormal clusters are eliminated, the clustering results are rationally processed, and similar cluster groups are merged to obtain user behavior classification results.

[0008] Preferably, the steps of obtaining the user feature summary are: Load the user behavior classification results, identify the access frequency, product category preference and average consumption amount, combine the user group characteristics, perform attribute aggregation, analyze and integrate the behavior data, and obtain the user behavior attribute set; Based on the user behavior attribute set, the user feature score is calculated using the following formula: ; in, represents the user's preference for the i-th category of products, is the weight factor of the i-th category product, is the total number of product categories, is the maximum preference in the product category, and F is the user feature score; Based on the user feature scores, the user's most frequently visited product categories and average consumption levels are analyzed, user consumption features are extracted, and a user feature summary is formed.

[0009] Preferably, the steps for obtaining the feature impact analysis result are: Based on the user feature summary, the influence of each feature is calculated using the following formula: ; in, It is The regression coefficient of the feature, It is User to The value of the feature, It is The mean of the features, is the influence score, J is the total number of features, is the intercept of the regression model; Based on the influence scores, the features that have the greatest impact on the purchase decision are identified, and feature impact analysis results are generated.

[0010] Preferably, the steps of obtaining the recommendation model are: Based on the feature impact analysis results, a collaborative filtering algorithm is used, user features and product information are used as input, the number of neighbors in the collaborative filtering algorithm is set, and a preliminary design of a recommendation algorithm is obtained; Based on the preliminary design of the recommendation algorithm, the similarity threshold and the weight of the score prediction are adjusted, and multiple rounds of tests are conducted to evaluate the prediction accuracy and response speed of the model to obtain an optimized and adjusted recommendation algorithm; Based on the optimized and adjusted recommendation algorithm, model training is performed, and user data and product database are used for training to obtain a recommendation model.

[0011] Preferably, the steps of obtaining the recommendation effect are: Deploy the recommendation model to the server, monitor the browsing and transaction behaviors of users, and generate a personalized product recommendation list through real-time processing by the recommendation model; Based on the personalized product recommendation list, the comprehensive score of the recommendation effect is calculated using the formula: ; in, Indicates The user's rating of the recommendation satisfaction. Indicates The number of interactions per user, is the total number of evaluation users, R is the comprehensive score of the recommendation effect; Based on the comprehensive score of the recommendation effect, click-through rate, purchase conversion rate and user feedback data are collected to form a recommendation effect report.

[0012] Preferably, the steps of obtaining the user feedback analysis results are: Based on the recommendation effect, the user's comprehensive satisfaction score is calculated using the formula: ; in, For the The user's rating of the recommendation satisfaction. Indicates The time interval between users providing feedback, is the total number of users, MF is the comprehensive satisfaction score, and R is the comprehensive score of the recommendation effect; Based on the comprehensive satisfaction score, collaborative filtering algorithm parameters and user experience that need to be optimized are identified to form a user feedback analysis result.

[0013] Preferably, the steps for obtaining the optimized recommendation model are: Analyze the user feedback analysis results, determine the collaborative filtering algorithm parameters that need to be adjusted according to the recommendation accuracy and personalization level, and obtain an adjustment requirement list; Based on the adjustment requirement list, adjust the parameters of the collaborative filtering algorithm, adjust the similarity threshold and weight distribution, and obtain a modified algorithm configuration; Based on the modified algorithm configuration, an offline test is performed to verify the recommendation model after deployment optimization.

[0014] The present invention provides a data report generating system, comprising: The data preprocessing module collects user interaction data from the e-commerce platform, filters the data and standardizes its format to generate cleaned data; The user behavior classification module, based on the cleaned data, applies partition aggregation and distance calculation to group the user's interactive behaviors and obtain the user behavior classification results; The feature summary module counts the user's access frequency and consumption amount based on the user behavior classification results, and synthesizes the user feature summary; The influence analysis module calculates the weight of each feature in the user's selection through difference comparison and influence measurement for the user feature summary, and generates feature influence analysis results; The recommendation optimization module uses the results of feature impact analysis and combines product information to adjust the weight parameters in the collaborative filtering algorithm, optimize the recommendation model, record real-time recommendation effects and user feedback, extract satisfaction indicators from them, and obtain the optimized recommendation model.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention ensures the high quality and applicability of data by collecting and cleaning user interaction data on e-commerce platforms, laying the foundation for analysis. The clustering algorithm is used to classify user behaviors in detail, which improves the understanding of user group behaviors, thereby enhancing personalization and accuracy. In particular, key user behavior features are extracted from user behavior classification to optimize the relevance and effect of recommendations. The recommendation process is further refined by calculating the influence of different features through statistical analysis, making the recommendation model more flexible to adapt to user needs. The integration of these optimization measures not only enhances the real-time response capability of the model, but also continuously optimizes the algorithm through user feedback, achieving self-learning and continuous improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the steps of the present invention; Figure 2 It is a schematic diagram of the steps of obtaining cleaned data in the present invention; Figure 3 Schematic diagram of the steps for obtaining the user behavior classification results in the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] See also Figure 1-3 The present invention provides a technical solution, a data report generation method based on artificial intelligence, comprising the following steps: Collect the interaction data of e-commerce platform users, generate cleaned data through data cleaning and preprocessing; based on the cleaned data, apply clustering algorithms to classify user behavior patterns, determine user groups, and generate user behavior classification results; Extract user behavior features from the user behavior classification results, including the most frequently visited product categories and average consumption levels, and generate a user feature summary; for the user feature summary, use statistical analysis to calculate the influence of each feature and generate feature impact analysis results; Based on the results of feature impact analysis, user features and product information are input and processed through collaborative filtering algorithm to obtain a recommendation model; the recommendation model is applied to e-commerce platform data to make real-time recommendations, and the recommendation effects and user feedback are recorded to obtain the recommendation results; Collect user feedback on recommendation effects, analyze the feedback, extract satisfaction indicators, and generate user feedback analysis results; based on the user feedback analysis results, adjust the parameters of the collaborative filtering algorithm to generate an optimized recommendation model.

[0019] The steps to obtain the cleaned data are: Collect interaction data of e-commerce platform users, including click-through rate, purchase history, and search history, to obtain a complete original interaction data set; Based on the complete original interaction data set, incomplete and abnormal data are identified and removed, outlier detection and missing data processing are performed to obtain a screened cleaned data set; Based on the screened cleaned data set, missing value filling and data format conversion are applied to generate cleaned data.

[0020] Specifically, when aggregating click-through rate data, purchase record data, and search record data from multiple sources, it is necessary to first compare them based on the user ID and timestamp of the e-commerce platform, merge the interactive behaviors of the same user on the same day into one record, and then conduct a preliminary verification of various types of interactive information. For example, compare the daily click-through rate with a pre-established click range. The click range can be set to 0 to 2,000 times. This range is based on the average level obtained by statistical analysis of past operating data plus an appropriate reserved value. If records exceeding this range are found, the corresponding marks are retained for subsequent verification. At the same time, the single purchase quantity and price in the purchase record are verified. For comparison, the single purchase quantity and price can be delimited in the range of 0 to 100 pieces and 0 to 5,000 yuan. This range is determined by the distribution statistics of the platform's sales records in the past 12 months. For search records, the reasonable range of keyword length is compared. A range of 2 to 50 characters can be selected. This character range is determined by the empirical value of the length of common phrases in users' daily searches. The identified abnormal records are temporarily stored, and all remaining records that meet basic completeness are summarized to form a preliminary data set that can cover the user's main interaction behaviors. Finally, these preliminary data sets are combined and serialized in chronological order to obtain a complete original interaction data set.

[0021] After obtaining the complete original interactive data set, in the process of identifying and removing incomplete and abnormal data, it is necessary to first compare and judge the click-through rate, purchase records and search records obtained previously. The click-through rate can be compared with the previously agreed range of 0 to 2000 times one by one. If the click-through rate is less than 0 times or exceeds 2000 times, the record is regarded as abnormal and listed separately. For purchase records, the number of single transactions can be set between 0 and 100, and the amount of single transactions can be set between 0 yuan and 5000 yuan, and check whether the amount and quantity meet the previously agreed logical relationship. Then check that the keyword length in the search record is between 2 and 50 characters. If missing fields are found, this record is included in the missing information classification, and then a simple loop traversal is used to verify all items marked as abnormal or missing one by one. If it is determined that it is indeed abnormal or cannot obtain effective information through compensation, the record is uniformly eliminated. Finally, the overall data after eliminating the abnormal records is marked and the filtered information is recorded to obtain a filtered cleaned data set.

[0022] Based on the previously obtained cleaned data set, it is necessary to fill the missing values ​​appropriately and unify the data format. First, for the positions marked with missing values, mean interpolation, neighboring interpolation or median interpolation in a specific time interval can be used. These interpolation methods will be comprehensively compared based on the contextual timing of the click-through rate or the specific peak segment of the purchase record. For example, if a missing value is detected in a non-peak time period, it will be supplemented according to the mean of the adjacent time period. If a missing value is detected in a peak time period, it will be calculated according to the average value of the historical highest transaction period. Subsequently, all supplemented data will be reorganized according to a unified character encoding and a unified numerical precision to ensure consistency with the parsing requirements of the next analysis step. If some fields belong to classified information, they will be converted into index values ​​or enumeration values ​​that can be used for subsequent calculations. Finally, after completing the integration of all data formats, the cleaned data is generated.

[0023] The steps to obtain the user behavior classification results are as follows: Based on the cleaned data, the K-means algorithm is used and the number of clusters and distance metrics are set to obtain the algorithm configuration results; Based on the algorithm configuration results, cluster the cleaned data and iterate until the clustering results are stable to obtain preliminary user behavior classification results; Based on the preliminary user behavior classification results, abnormal clusters are eliminated, the clustering results are rationalized, and similar cluster groups are merged to obtain the user behavior classification results.

[0024] Specifically, when reading the cleaned data, you need to first determine the number of clusters and distance measurement method used for K-means. The number of clusters can be set by statistically analyzing the distribution of past samples and observing the separation of data under different numbers of clusters. For example, you can first set And conduct multiple rounds of tests. If it is found that there are large differences within some clusters, the number of clusters can be increased or decreased. The distance measurement method can refer to the Euclidean distance or other known measurement paradigms, but the dimensions and dimensions involved in the calculation process must be clearly defined in advance, and normalized comparisons must be performed before preparing clustering. If it is found that the data exceeds a preset mean interval range in one or more dimensions, it can be recorded and numerically smoothed using a preset offset correction method. The preset mean interval is generally generated by the average and standard deviation calculated when comparing thousands of records. For example, the average value is determined in When it is between 40 and 60, the standard deviation can be doubled to define the maximum and minimum thresholds. If the actual value exceeds this interval, the deviation is marked. This makes it easier to maintain dimensional consistency when calculating the distance. Subsequently, K-means is initialized, and each data point is represented by coordinates based on multiple dimensional values. Distance is calculated according to dimensions related to user behavior until the first round of distance results are obtained between all data points and the assigned initial center and the initial center of gravity of each cluster can be calculated. Finally, the number of clusters and the corresponding distance metrics are summarized as the algorithm configuration results.

[0025] After obtaining the algorithm configuration results defined above, it is necessary to perform K-means clustering on the cleaned data based on the number of clusters and distance measurement methods specified therein. During the execution process, the cluster center is initialized first, and then the data is scanned one by one and the distance between each cluster center is calculated respectively. The center corresponding to the minimum distance is selected for attribution marking. After the current attribution is completed, the center of gravity is updated by recalculating the coordinates of each cluster center and continuing to the next round of iteration. If it is found that the internal distribution of a single cluster deviates, a monitoring threshold can be set to determine whether additional correction is required. The threshold can be selected based on the total number and distribution of clusters. For example, it is determined by multiplying the mean deviation of the internal distance of the cluster by a proportional coefficient. If the mean deviation exceeds the critical value obtained by this proportional coefficient, further judgment is made and it is recorded whether it is necessary to add a center or eliminate abnormal data. The iterative process continues until the single center of gravity offset is lower than a convergence standard or the maximum number of cycles is reached. Finally, the attribution of all data points is recorded to obtain preliminary user behavior classification results.

[0026] After obtaining the preliminary user behavior classification results, it is necessary to screen for possible abnormal clusters. First, check whether the proportion of each cluster is far below the pre-judged lower limit or obviously exceeds the upper limit. The upper and lower limits can be obtained by statistically analyzing the range of previous clustering situations in different usage scenarios. For example, the proportion is set between 2% and 40% as an acceptable range. If the proportion of any cluster is lower than 2% or exceeds 40%, the cluster number is recorded and it is determined whether it is an abnormal distribution. It is then reviewed in combination with other indicators such as the average distance or variance of data points. If it is confirmed that some clusters have abnormal characteristics, they are eliminated, and the data points in the abnormal clusters are re-merged or split according to the previously calculated centroid coordinates. Then, the overlap and separation between the remaining clusters are observed. If it is found that the center distances of multiple clusters are close and the separation is insufficient, they can be merged into the same cluster group based on the analysis results of the behavioral characteristics of this part of the population, and finally the user behavior classification results are formed.

[0027] The steps to obtain the user feature summary are: Load the user behavior classification results, identify the access frequency, product category preference and average consumption amount, combine the user group characteristics, perform attribute aggregation, analyze and integrate the behavior data, and obtain the user behavior attribute set; Based on the user behavior attribute set, the user feature score is calculated using the following formula: ; in, represents the user's preference for the i-th category of products, is the weight factor of the i-th category product, is the total number of product categories, is the maximum preference in the product category, and F is the user feature score; Based on the user feature scores, analyze the user's most frequently visited product categories and average consumption levels, extract user consumption features, and form a user feature summary.

[0028] Specifically, the user behavior classification results are loaded, the access frequency, product category preference and average consumption amount are identified, and the field information of the user group characteristics is first aggregated, and the cumulative number of visits of each user is compared with the pre-established access frequency range. For example, the number of visits is determined to be between 0 and 3,000 times as a reference interval with reference to the usage records of the recent three months, and the visit behaviors exceeding 3,000 times are marked. Then, the product category preferences are disassembled, and the purchase proportions of different product categories are matched one by one with the previously obtained number of visits. Then, the user's single consumption amount is extracted from the sales end records and multiple consumptions are checked. For example, the consumption amount can be compared with the range of 0 to 20,000 yuan one by one and the outliers can be marked for further verification. The common attribute characteristics of the user group can be counted as the basis for the merged label. The layers are merged according to the fields such as the access frequency and product category preference. During the merging process, the product categories with higher correlation are merged and packaged, the consumption mean is recorded, and the median and variance are calculated. These statistical values ​​are summarized and compared with the user group characteristics to see if there is any obvious deviation and reconfirm the deviation items. Finally, all the merged statistical results are summarized to obtain the user behavior attribute set.

[0029] The benefit of the formula is that by incorporating the preferences and weight factors of different product categories and normalizing them with the maximum preference, it is able to combine the influence of multiple categories of preferences into a single feature score, forming comparable numerical results among different users.

[0030] The acquisition steps are as follows: count the user's purchase records and browsing records by item. The number of clicks and purchases of products in the same category is calculated by multiplying the number of purchases by a certain consumption amount and then weighting the result with the number of clicks to get a preference ratio between 0 and 1. For example, in a scenario with five categories of products, when a user has 30 clicks and 3 purchases for products in category 1, the comprehensive preference can be statistically obtained based on the corresponding amount distribution. .

[0031] The steps to obtain are as follows: The importance of this type of product in the actual operation process can be numerically quantified. It can be evaluated by observing the contribution of this type of product to the total sales and user return rate and other indicators. In the same dimension, if it is found that this type of product accounts for 20% of the overall revenue, its initial weight can be set to around 1.2, and then fine-tuned up and down according to secondary indicators such as return rate, and finally determined. Specific value.

[0032] The steps to obtain are to count the actual number of product categories on the platform. If the current platform has 5 categories of products, then , this value comes from the actual configuration of the platform database and is updated every time a new product category is launched.

[0033] The steps to obtain are: Select the one with the largest preference value, assuming that the previously obtained ,but .

[0034] Calculation process: Bring in specific parameters, here , , , , , , , , , , , ; calculate For each : , , , , ; Square these values ​​and sum them: ; Take the square root: ; The result shows that the comprehensive feature score of the current user is about 1.98. The larger the value, the higher the match between the user's overall preference and important categories. If the score is greater than 2.5, it means that the user is highly involved in the main product categories. If it is lower than 1.0, it means that the user's activity in each category is relatively low. In the current step, this value shows that the user's purchase and click behaviors are relatively concentrated in the five categories and have a more obvious preference for certain categories.

[0035] When analyzing the user's most frequently visited product categories and average consumption level based on user feature scores, it is necessary to compare the number of visits and consumption amounts of users on each product category in parallel based on the preference information and consumption amount distribution obtained previously. If a product category has more than 200 visits and the cumulative consumption amount is more than 5,000 yuan, this value can be obtained by continuously tracking platform orders and click traffic statistics. The statistical method can be split by time period, for example, the consumption amount is summarized once every week, and the number of visits is cross-validated to determine the user's stickiness to the product category. If it is multiple If the number of visits in the comparison of each period is more than 30% more than that of other categories, this category can be regarded as the most frequently visited product category. At the same time, when calculating the average consumption level, it is necessary to smooth the amounts of multiple transactions and eliminate abnormal transactions that are too large at one time. If it is found that the amounts of multiple single transactions exceed three times the overall consumption average of the previous year, they will be marked separately after further verification of the actual payment information. Then, the total amount of all remaining normal transactions is divided by the number of transactions to obtain the average consumption level of the user. After completing these parallel comparisons, the user's consumption characteristics can be extracted to form a user feature summary.

[0036] The steps to obtain the feature impact analysis results are: Based on the user feature summary, the influence of each feature is calculated using the following formula: ; in, It is The regression coefficient of the feature, It is User to The value of the feature, It is The mean of the features, is the influence score, J is the total number of features, is the intercept of the regression model; Based on the influence scores, identify the features that have the greatest impact on purchase decisions and organize and generate feature impact analysis results.

[0037] Specifically, the formula is useful in introducing This item is used to measure the deviation of users from the mean of each feature value. It combines the combined influence of the regression coefficient and the intercept term to aggregate the change ranges of multiple features into a comparable influence score, and conducts quantitative analysis on the differences between different users under multi-dimensional features.

[0038] The steps for obtaining are as follows: when monitoring the initial regression samples of multiple users, first collect user interaction information for at least one month and record the mean difference range of each user in the main features, and then use the least squares method or other regression fitting methods to iterate the intercept term. This intercept remains relatively stable in the same regression model. For example, when sampling 1,000 users, the intercept term is obtained through repeated regression fitting. .

[0039] The steps to obtain are: The corresponding coefficients of each feature are extracted according to the results of multiple regressions. The size of the coefficients is usually related to the importance of the feature in the overall purchase decision. Based on the statistical data of the past three months or longer periods, by checking the residual distribution after each round of regression and performing multiple rounds of verification, a relatively stable value can be gradually converged. For example, in the data analysis of the above 1,000 users, wait.

[0040] The steps for obtaining are: Users collect their The actual detection values ​​on the feature dimensions are as follows. For example, the user has recorded 8 orders under the feature of "Monthly Order Frequency", the detection value under the feature of "Coupon Usage Ratio" is 0.4 (indicating that 40% of the purchase records use coupons), and the detection value under the feature of "Interaction Duration" is an average of 1.5 hours per week. These specific values ​​are normalized or converted into continuous data that can be directly processed to form .

[0041] The steps to obtain are: The average value of each feature among all users is recorded and counted. This requires monitoring a sufficient number of users and eliminating extreme outliers in the data. For example, the average of "monthly order frequency" is 5 orders, the statistical mean of "coupon usage ratio" is 0.3, and the statistical mean of "interaction time" is 1.2 hours.

[0042] The step of obtaining J is to count the total number of features that have been included in the modeling analysis. For example, if 10 features are selected that best represent the purchase decision, then , this value can be dynamically updated with the subsequent expansion of the feature library.

[0043] Calculation process: Bring in the required parameters. Here, select Assume that the three main characteristic coefficients of the user are ,intercept , and the detection values ​​of these three features of the user are , and its characteristic means in the overall user sample are ; calculate For each j: , , ; Multiply the regression coefficient of each feature by the above distance: ; Add them together and add the intercept : ; The results show that for the first user, the deviations of these three features from the mean accumulate into an influence score of 4.851 in the regression model. If the same model is applied to other users, the influence scores obtained for each user can be compared horizontally. When the score is higher than 6.0, it means that the deviation of these features is more significant. If the score is lower than 2.0, it means that the user is closer to the overall mean level, which can provide a numerical reference for the subsequent identification of the features that have the greatest impact on purchase decisions.

[0044] Based on the influence score, when identifying the features that have the greatest impact on purchase decisions, we first need to summarize the influence scores obtained previously, then break down the contribution values ​​of each feature under this regression model, and compare the contribution value of each feature with the pre-established threshold. For example, we can combine the feature contribution mean previously summarized from at least tens of thousands of historical purchase data, and set it to be twice higher than the contribution mean, which is considered to be extremely high and marked. After marking, we need to review the parameters corresponding to these high-contribution features. If it is found that a certain feature accounts for more than 30% of the influence score, then record the feature number for comparison with other features in subsequent analysis, and then list the remaining features one by one based on the contribution value sorting and draw the change curve in different user groups. For features with high similarity, we will merge and list them. In this process, if it is detected that some user groups show significantly different deviations from a certain feature, we can manage the user group in layers and continue to track their subsequent performance. When the contribution comparison of all features is completed, the feature number and name with the highest contribution can be associated with the user group identifier and output, and then the feature impact analysis result is formed.

[0045] The steps to obtain the recommendation model are: Based on the results of feature impact analysis, a collaborative filtering algorithm is used, with user features and product information as input, and the number of neighbors in the collaborative filtering algorithm is set to obtain a preliminary design of the recommendation algorithm; Based on the preliminary design of the recommendation algorithm, the similarity threshold and the weight of the score prediction were adjusted, and multiple rounds of tests were conducted to evaluate the prediction accuracy and response speed of the model to obtain an optimized and adjusted recommendation algorithm; Based on the optimized and adjusted recommendation algorithm, model training is carried out, and user data and product database are used for training to obtain the recommendation model.

[0046] Specifically, based on the results of feature impact analysis, when using collaborative filtering algorithm to take user features and product information as input, it is necessary to first read the user feature information obtained previously and the corresponding product data, mark the product categories preferred by the user and match them one by one with the specific attributes contained in the product itself, and then set the number of neighbors in the collaborative filtering algorithm. When setting, it is necessary to refer to the overall scale of users on the platform. For example, the number of neighbors can be selected between 5 and 50. This range is evaluated based on the similarity distribution collected from tens of thousands of user samples in the early stage. In the actual implementation process, it is necessary to first extract several typical The user groups are compared with their product browsing paths and purchase records one by one. The browsing paths and purchase records are serialized and then the similarity is calculated. If the category browsing tendencies of a user and another user overlap by more than 80%, they can be regarded as high-similarity users. Then, the most comparable neighbor set is screened out based on the set number of neighbors. Next, the product scores between each user and the neighbors are compared, and the overlapping score distributions and difference items are registered and counted to obtain the potential products recommended to each other by both parties. After completing a large-scale scan, the matching results of all users and neighbors are comprehensively calculated and merged to obtain the preliminary design of the recommendation algorithm.

[0047] Based on the preliminary design of the recommendation algorithm, when adjusting the similarity threshold and the weight of the rating prediction, an initial threshold can be set to determine whether there is enough similarity between users. The initial threshold needs to be extracted from real user scenarios. For example, a test is conducted on 100,000 interaction data of different users, and the recall rate and precision in the similarity intervals of 0.5, 0.6, and 0.7 are selected. If it is detected that a better balance can be achieved when the similarity is divided at 0.6, this value is determined as the preset threshold and combined with the distribution of user ratings on products, the weight of the rating prediction is dynamically adjusted. For example, high-frequency purchases are considered as the weight. The scoring weight of the group is set slightly higher than that of the group that only browses but does not purchase, and the weight is managed in segments. Users who purchase 1 to 10 times are defined as low frequency, users who purchase 10 to 30 times are defined as medium frequency, and users who purchase more than 30 times are defined as high frequency. The scoring deviations of these groups are then integrated, and the prediction accuracy and response speed are recorded through multiple rounds of tests. If it is found that the accuracy is stable at more than 85% after multiple cycles and the response speed is maintained within milliseconds, it is confirmed that there is no major deviation in the adjustment. Finally, after multiple rounds of testing, an optimized and adjusted recommendation algorithm is obtained.

[0048] Based on the optimized and adjusted recommendation algorithm, when training the model, the user data and product database obtained previously need to be included in the training process at the same time. First, read the number of purchases and click behaviors from the user's interaction history in chronological order, and then match these behavior data with the optimized similarity threshold and scoring weight obtained previously, and establish mappings between users and products in batches. If the proportion of certain product types in the overall product library is less than 1%, a small amount of weighted interpolation can be performed separately for such products to ensure that they are not ignored in training. Then, all matching sequences of users and products are input for batch calculation and iterative. After each iteration, the intermediate results are verified. If it is observed that the deviation between the prediction and the actual purchase behavior does not exceed the 5% reference interval obtained by the previous statistics, the next round of training will continue. Otherwise, the weights or thresholds will be fine-tuned accordingly. This process is repeated until the system meets the comprehensive requirements of accuracy and speed on the sample data and presents a basic convergence state. Finally, a stable output model parameter combination is selected from all training results to obtain a recommendation model.

[0049] The steps to obtain the recommended effect are: Deploy the recommendation model to the server, monitor the user's browsing and transaction behavior, and generate a personalized product recommendation list through real-time processing through the recommendation model; Based on the personalized product recommendation list, the comprehensive score of the recommendation effect is calculated using the formula: ; in, Indicates The user's rating of the recommendation satisfaction. Indicates The number of interactions per user, is the total number of evaluation users, R is the comprehensive score of the recommendation effect; Based on the comprehensive score of recommendation effect, click-through rate, purchase conversion rate and user feedback data are collected to form a recommendation effect report.

[0050] Specifically, when deploying the recommendation model to the server and monitoring the user's browsing and transaction behaviors, it is necessary to record the page link visited by the user each time on the server side and extract the browsing time, click action and order submission timestamp. In the recording process, these browsing data must be checked against the pre-established range. For example, the page dwell time is compared with the range of 0 seconds to 600 seconds. If the dwell time exceeds 600 seconds, the behavior is marked to further confirm its validity. At the same time, the order information is matched with the range of 0 to 100,000 yuan, and the user's potential preference category is judged in combination with the product types contained in the order. After recording the time and quantity of these transaction behaviors, they are combined with the product data obtained previously. Product preference weights are calculated to integrate user characteristics into real-time operations. During this process, if it is found that the number of clicks on a specific product by a user suddenly soars to more than 100 times within an hour, it is necessary to verify again whether the data is caused by false orders or abnormal visits. After confirming that there is no abnormality, it will be included in a unified interaction sequence, and then the interaction sequence will be associated with the user feature set and calculated online. Based on these records, the user's shopping intention trend is determined and cross-compared with all user groups with similar behavioral characteristics, and then the same or similar product information in the product database is retrieved, the recommendation results are arranged centrally and generated within the specified response time limit, and when all conditions are interacted, all data matching situations are combined to generate a personalized product recommendation list.

[0051] The formula is useful because it introduces By taking the index, it can exponentially amplify the contribution value of users with high satisfaction and frequent interactions while comprehensively considering user satisfaction and the number of interactions, thereby obtaining a more differentiated comprehensive score in the overall calculation of the recommendation effect.

[0052] The steps for obtaining it are: after each product browsing or ordering, push the rating interface to the user and record the user's evaluation score of the recommendation satisfaction. The score is usually set between 1 and 5. For a larger user group, it can be converted into a continuous scale after statistics. If more detailed information is needed, the specific products corresponding to the rating can be collected simultaneously during each rating, and then a personal rating curve can be generated based on the rating history.

[0053] The steps to obtain are: The number of interactions of a user in a period of time, including the total number of page clicks and details viewed, can be combined with the number of order submissions to be regarded as the actual amount of interaction of the user with the product or page. In the actual environment, the number of interactions can be associated with the user's device fingerprint or login account to ensure a one-to-one correspondence and reduce duplicate counting. After the statistics are completed, this value can be compared with the average interaction volume of the group to identify whether there are extreme values ​​that are too high or too low.

[0054] The acquisition step is to record the number of users who participated in the evaluation and gave ratings. When conducting large-scale online tests, there are often hundreds to thousands of users generating evaluation data at the same time. After the statistical cycle is over, the total number of actual evaluation users can be confirmed. The statistical result is defined as .

[0055] Calculation process: Bring in example parameters, such as selecting , the three users’ ratings and interaction times for the current recommendation list are , , ; Calculate item by item : ; ; ; Sum the three terms and divide by : ; ; The result shows that based on the ratings and interaction times of the three users in the current example, the final comprehensive rating of the recommendation effect is approximately 54435.54. When the value is higher than 30,000, it indicates that the interaction frequency and satisfaction of most users are relatively active during this time period and they have a positive attitude towards the recommendation results. If the value is lower than 10,000, it indicates that both the user ratings and interaction needs are at a relatively low level.

[0056] Based on the comprehensive score of recommendation effect, when collecting click-through rate, purchase conversion rate and user feedback data, it is necessary to first extract the number of clicks and order data of different users during the browsing period from the comprehensive score sequence obtained above, and compare the number of clicks with the interval of 0 to 2000 times one by one. The records exceeding this interval are marked as abnormal high-frequency behaviors. Then compare the order transaction amount with 0 to 50,000 yuan. If a user has a single transaction amount that is too high, a second verification is required. Then, the feedback data is summarized and the difference between each user in the access information and the subsequent purchase process is summarized. If it is found that a group of people click If the click-through rate is at a high level but the purchase conversion rate is extremely low, further inspection is conducted to determine whether there are malicious clicks or abnormal wait-and-see operations. This phenomenon can also be compared vertically with the conversion rates of other groups and divided into segments. If the average order amount of some groups is generally more than three times higher than that of other groups, the aforementioned evaluation indicators can be referred to again to determine whether it should be regarded as a special sales behavior. Finally, the click data and conversion data of all users in each link are interactively summarized, and the feedback scores of individual users and groups are combined to integrate the full performance. After the process is completed, a recommendation effect report can be formed.

[0057] The steps to obtain user feedback analysis results are as follows: Based on the recommendation effect, the user's comprehensive satisfaction score is calculated using the formula: ; in, For the The user's rating of the recommendation satisfaction. Indicates The time interval between users providing feedback, is the total number of users, MF is the comprehensive satisfaction score, and R is the comprehensive score of the recommendation effect; Based on the comprehensive satisfaction score, identify the collaborative filtering algorithm parameters and user experience that need to be optimized, and form user feedback analysis results.

[0058] Specifically, the formula is beneficial in that by With time interval Combine and incorporate and The weighted form of the recommendation effect is combined with the comprehensive score obtained above. By performing calculations, the overall user satisfaction performance can be quantified in multi-user and multi-time scenarios, providing comparable quantitative indicators for subsequent judgment of collaborative filtering algorithm parameters that need to be optimized and user experience details.

[0059] The acquisition steps are as follows: after each recommendation is pushed to the user, the system will collect satisfaction scores after the user completes browsing or purchasing operations. The scores can be set to 1 to 5, and can be subdivided into continuous quantitative intervals of 1 to 10 or even 1 to 100 on a larger scale. The system compares the score distribution and historical scoring records of different users and records the scores as .

[0060] The acquisition step is to monitor the time interval between each user completing the recommended content browsing and submitting feedback. The interval is measured in hours or days, and can also be accurate to seconds for short-term and high-frequency application scenarios. When comparing multiple groups of people, the shortest and longest feedback times are used to determine the time interval. If it is detected that some users submit a large number of feedbacks in a very short period of time, it will be confirmed again whether it is an invalid repeated behavior. After eliminating the anomaly, the final time interval will be recorded as .

[0061] The steps to obtain are as follows, which have been obtained through the comprehensive scoring formula of the recommendation effect. The value of is based on the interaction data, satisfaction and product preferences of all users. If or other values, which can be used directly in this formula .

[0062] The acquisition steps are to count the number of users with ratings or interaction records in the actual environment. After each round of recommendation cycle, the total number of users with feedback records is regarded as If it is detected that some users’ ratings are invalid or the evaluation process is incomplete, they need to be eliminated and only valid records are retained to determine .

[0063] Calculation process: Bring in the example parameters, for example, select Indicates that 3 users provide ratings and time intervals, and assumes the comprehensive rating of the recommendation effect obtained previously ; For each user enter their and , example values ​​are: First user , User 2 , User 3 ; Calculate separately and and multiply by : To the first user: ; ; Multiply these together and then multiply : ; To the second user: ; ; ; To the third user: ; ; ; Add the values ​​obtained by the three users: ; The result shows that the comprehensive satisfaction score of the three users in the current recommendation scenario is 111266.483. If more user ratings and time intervals are included in the calculation, the same formula can be used to batch sum and obtain a new ,when When the value is higher than 100,000, it usually indicates that the overall user satisfaction is relatively active. If it is lower than 20,000, it means that the overall user rating and interaction frequency are low. From this, we can preliminarily judge the collaborative filtering algorithm parameters or related user experience processes that need to be optimized to form user feedback analysis results.

[0064] Based on the comprehensive satisfaction score, when identifying the collaborative filtering algorithm parameters and user experience that need to be optimized, it is necessary to first split and compare the score deviation items in the previously recorded user rating data and time interval data. The comparison results can divide the users with obviously low or high scores into specific groups for observation. If some of them have only submitted a rating once within a week or have repeated the ratings multiple times with short intervals, their feedback stability will be checked according to the previously determined time intervals. For example, the behavior with a time interval of 0 to 1 hour is regarded as a fast feedback group, the group with a time interval of 1 to 8 hours is regarded as a medium period, and the group with a time interval of more than 1 day is regarded as a long period. The satisfaction trends and recommendations of these groups are analyzed. The recommendation effect comprehensive score is compared with the average order amount or total browsing time of users in the same group, and then the value is compared with the range of 0 yuan to 2000 yuan or 0 minutes to 480 minutes for multiple rounds. If it is found that a large number of users only give low scores when the average browsing time reaches more than 300 minutes, there may be a mismatch in recommendation or incorrect algorithm parameters. At this time, the collaborative filtering algorithm is further disassembled in the calculation method of the similarity threshold or the scoring weight, or the system interface interaction is improved. After executing the above screening steps, a clear improvement suggestion can be output to cover the required parameter adjustment, interaction delay or content matching, etc. After completing all parameters and experience improvement items, they can be integrated into the user feedback analysis results.

[0065] The steps to obtain the optimized recommendation model are: Analyze the results of user feedback analysis, determine the collaborative filtering algorithm parameters that need to be adjusted based on the recommendation accuracy and personalization level, and obtain a list of adjustment requirements; Based on the adjustment requirement list, adjust the parameters of the collaborative filtering algorithm, adjust the similarity threshold and weight distribution, and obtain the modified algorithm configuration; Based on the modified algorithm configuration, offline testing is performed to verify the deployment of the optimized recommendation model.

[0066] Specifically, when analyzing the user feedback analysis results, based on the recommendation accuracy and personalization level, it is necessary to first disassemble the various rating records and conversion rate comparison information in the feedback data obtained earlier, and check item by item whether the users with high click-through rates but no purchases generally have a score lower than 3 points or show other characteristics. Then, combined with the personalization dimensions obtained in the previous step, for example, the recommendation hit rate and error value are compared among users of different age groups and users in different geographical locations. The features of the marked low-accuracy groups or user groups with obvious personalization contrast are summarized. During the feature summary process, if it is found that the recommendation accuracy of some groups is lower than the benchmark value set in advance, the benchmark value is usually the average of the actual evaluation data in previous years. Add 2 to 3 times the standard deviation. If the current accuracy is lower than this benchmark value, it means that extra attention should be paid to the parameters of this group, and then their personalized weight distribution and collaborative filtering similarity threshold should be observed. When comparing, the score fluctuation and personalized discrimination should be carried out simultaneously. If the discrimination of multiple groups of people is not enough or the accuracy has not reached the specified upper limit for several consecutive weeks, the specified upper limit can be determined by adding 5 percentage points to the average of the highest accuracy in all recommended scenarios in the past year. Once a large-scale deviation is confirmed, the parameters corresponding to these groups will be sorted into suspicious items. Subsequently, the main impact range and usage time and other information will be screened out in these deviations and situations where the personalization level is insufficient, and finally a list of required adjustment requirements will be generated.

[0067] Based on the adjustment requirements list, when adjusting the parameters of the collaborative filtering algorithm, first list the similarity threshold and weight distribution to be corrected. The similarity threshold can be split by interval to observe the relationship between the click and the actual purchase given in the previous step. For example, the user group with a similarity between 0.3 and 0.5 is defined as weak correlation, and 0.5 to 0.7 is defined as moderate correlation, and more than 0.7 is high correlation. If a large number of low-scoring users are found in the weak correlation interval, the similarity threshold may be increased to exclude users who are not similar enough. If the high correlation interval is too narrow and leads to insufficient samples, the threshold is appropriately lowered. Similarly, the weight distribution needs to be based on user browsing behavior, purchase times, and The product preference depth is redistributed, and the numerical situation of each dimension is compared with the pre-established measurement interval. For example, the number of purchases is compared with the interval of 0 to 30 times, and the browsing time is compared with the interval of 0 minutes to 600 minutes. If it is found that the distribution in a certain interval is significantly higher than that in other intervals but does not bring effective clicks or score improvements, the weight of the corresponding dimension will be reduced, and the actual visits and ratings of the user group will be tracked in multiple iterations. If the score changes in a more reasonable direction after each adjustment and the similarity range can cover more users, the adjustment idea will be retained and updated to the record. Finally, all new thresholds and weight data will be merged to obtain the modified algorithm configuration.

[0068] Based on the modified algorithm configuration, when conducting offline testing, it is necessary to extract user interaction data within a certain period from the historical database for playback, and calculate the recommendation results according to the collaborative filtering process with completed parameter adjustment. During the playback process, the real purchase history will be compared one by one to compare the accuracy of the recommendation, and the response speed in different time periods will be recorded. If it is found that the data processing speed exceeds the pre-set time limit when the test volume increases to a certain multiple, the time limit is usually defined based on the average value obtained by evaluating the previous system load and adding a part of the safety margin on this basis. Once exceeded, detailed performance data will be collected during the test phase and the algorithm configuration will be reviewed again. If there are no significant abnormalities in the accuracy and response speed, the playback scale can be increased in batches or different types of user samples can be replaced to compare the final matching results. If it is observed that the overall accuracy is not much different from the target value set previously, and the personalized recommendation of the core user group remains stable, it can be determined that the model has more suitable conditions for real-world deployment. Finally, all offline test records are classified to verify the deployment of the optimized recommendation model.

[0069] The present invention provides a data report generating system, comprising: The data preprocessing module collects user interaction data from the e-commerce platform, filters the data and standardizes its format to generate cleaned data; The user behavior classification module, based on the cleaned data, applies partition aggregation and distance calculation to group the user's interactive behaviors and obtain the user behavior classification results; The feature summary module counts the user's access frequency and consumption amount based on the user behavior classification results, and synthesizes the user feature summary; The influence analysis module calculates the weight of each feature in the user's selection through difference comparison and influence measurement for the user feature summary, and generates feature influence analysis results; The recommendation optimization module uses the results of feature impact analysis and combines product information to adjust the weight parameters in the collaborative filtering algorithm, optimize the recommendation model, record real-time recommendation effects and user feedback, extract satisfaction indicators from them, and obtain the optimized recommendation model.

[0070] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A data report generation method based on artificial intelligence, characterized in that: The following steps are involved: Collect interaction data from users of e-commerce platforms, and generate cleaned data through data cleaning and preprocessing; Based on the cleaned data, a clustering algorithm is applied to classify user behavior patterns, determine user groups, and generate user behavior classification results; Extracting user behavior features from the user behavior classification results, including the most frequently visited product categories and average consumption levels, and generating a user feature summary; applying statistical analysis to the user feature summary to calculate the influence of each feature and generating a feature impact analysis result; Based on the feature impact analysis results, user features and product information are input, and a recommendation model is obtained through collaborative filtering algorithm processing; the recommendation model is applied to the e-commerce platform data to make real-time recommendations, and the recommendation effects and user feedback are recorded to obtain the recommendation effects; Collect user feedback on the recommendation effect, analyze the feedback, extract satisfaction indicators, and generate user feedback analysis results; adjust the parameters of the collaborative filtering algorithm according to the user feedback analysis results to generate an optimized recommendation model.

2. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the cleaned data are as follows: Collect interaction data of e-commerce platform users, including click-through rate, purchase history, and search history, to obtain a complete original interaction data set; Based on the complete original interaction data set, incomplete and abnormal data are identified and removed, outlier detection and missing data processing are performed to obtain a screened cleaned data set; Based on the screened cleaned data set, missing value filling and data format conversion are applied to generate cleaned data.

3. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the user behavior classification result are: Based on the cleaned data, a K-means algorithm is used and the number of clusters and the distance metric are set to obtain an algorithm configuration result; Based on the algorithm configuration result, clustering is performed on the cleaned data, and it is iterated until the clustering result is stable to obtain a preliminary user behavior classification result; Based on the preliminary user behavior classification results, abnormal clusters are eliminated, the clustering results are rationally processed, and similar cluster groups are merged to obtain user behavior classification results.

4. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps of obtaining the user feature summary are: Load the user behavior classification results, identify the access frequency, product category preference and average consumption amount, combine the user group characteristics, perform attribute aggregation, analyze and integrate the behavior data, and obtain the user behavior attribute set; Based on the user behavior attribute set, the user feature score is calculated using the following formula: ； in, represents the user's preference for the i-th category of products, is the weight factor of the i-th category product, is the total number of product categories, is the maximum preference in the product category, and F is the user feature score; Based on the user feature scores, the user's most frequently visited product categories and average consumption levels are analyzed, user consumption features are extracted, and a user feature summary is formed.

5. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the feature impact analysis results are: Based on the user feature summary, the influence of each feature is calculated using the following formula: ； in, It is The regression coefficient of the feature, It is User to The value of the feature, It is The mean of the features, is the influence score, J is the total number of features, is the intercept of the regression model; Based on the influence scores, the features that have the greatest impact on the purchase decision are identified, and feature impact analysis results are generated.

6. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the recommendation model are: Based on the feature impact analysis results, a collaborative filtering algorithm is used, user features and product information are used as input, the number of neighbors in the collaborative filtering algorithm is set, and a preliminary design of a recommendation algorithm is obtained; Based on the preliminary design of the recommendation algorithm, the similarity threshold and the weight of the score prediction are adjusted, and multiple rounds of tests are conducted to evaluate the prediction accuracy and response speed of the model to obtain an optimized and adjusted recommendation algorithm; Based on the optimized and adjusted recommendation algorithm, model training is performed, and user data and product database are used for training to obtain a recommendation model.

7. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the recommendation effect are as follows: Deploy the recommendation model to the server, monitor the browsing and transaction behaviors of users, and generate a personalized product recommendation list through real-time processing by the recommendation model; Based on the personalized product recommendation list, the comprehensive score of the recommendation effect is calculated using the formula: ； in, Indicates The user's rating of the recommendation satisfaction. Indicates The number of user interactions, is the total number of evaluation users, R is the comprehensive score of the recommendation effect; Based on the comprehensive score of the recommendation effect, click-through rate, purchase conversion rate and user feedback data are collected to form a recommendation effect report.

8. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the user feedback analysis result are: Based on the recommendation effect, the user's comprehensive satisfaction score is calculated using the formula: ； in, For the The user's rating of the recommendation satisfaction. Indicates The time interval between users providing feedback, is the total number of users, MF is the comprehensive satisfaction score, and R is the comprehensive score of the recommendation effect; Based on the comprehensive satisfaction score, collaborative filtering algorithm parameters and user experience that need to be optimized are identified to form a user feedback analysis result.

9. The method for generating data reports based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the optimized recommendation model are: Analyze the user feedback analysis results, determine the collaborative filtering algorithm parameters that need to be adjusted according to the recommendation accuracy and personalization level, and obtain an adjustment requirement list; Based on the adjustment requirement list, adjust the parameters of the collaborative filtering algorithm, adjust the similarity threshold and weight distribution, and obtain a modified algorithm configuration; Based on the modified algorithm configuration, offline testing is performed to verify the deployment of the optimized recommendation model.

10. A data report generation system according to the artificial intelligence-based data report generation method according to any one of claims 1 to 9, characterized in that: include: The data preprocessing module collects user interaction data from the e-commerce platform, filters the data and standardizes its format to generate cleaned data; The user behavior classification module, based on the cleaned data, applies partition aggregation and distance calculation to group the user's interactive behaviors and obtain the user behavior classification results; The feature summary module counts the user's access frequency and consumption amount based on the user behavior classification results, and synthesizes the user feature summary; The influence analysis module calculates the weight of each feature in the user's selection through difference comparison and influence measurement for the user feature summary, and generates feature influence analysis results; The recommendation optimization module uses the results of feature impact analysis and combines product information to adjust the weight parameters in the collaborative filtering algorithm, optimize the recommendation model, record real-time recommendation effects and user feedback, extract satisfaction indicators from them, and obtain the optimized recommendation model.

Images