E-commerce commodity recommendation optimization management system

Abstract

The invention belongs to the technical field of commodity recommendation, and discloses an e-commerce commodity recommendation optimization management system, which comprises the steps of collecting e-commerce comprehensive data, and preprocessing the collected e-commerce comprehensive data to generate a comprehensive data set; weight values are distributed for the user behavior characteristics through a comprehensive data set; in combination with the weights of the user behavior characteristics, scores of the users for all the commodities are calculated, and a user-commodity interaction matrix is constructed according to the scores; the festivals and the user positions are comprehensively considered, the preference degree of the users for the commodities is calculated through the user-commodity interaction matrix, and the optimal commodity category recommendation list for attracting the users is generated, so that the problems of improper regional difference processing and difficult recommendation system performance optimization are solved, the commodity recommendation accuracy is improved, the user experience is enhanced, and the user experience is improved. And the user experience and the operation efficiency of the e-commerce platform are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of commodity recommendation, and more specifically, to a commodity recommendation optimization management system for e-commerce. Background Art

[0002] With the rapid development of e-commerce, consumers' requirements for shopping experience are constantly increasing. As a key link for e-commerce platforms to improve user experience and increase sales conversion rate, product recommendation is becoming increasingly important. However, traditional product recommendation systems are often based on historical purchase records or browsing behaviors. Although such recommendation methods can reflect the explicit needs of users, they are insufficient in terms of personalization, novelty of recommendations, and diversity of recommendations.

[0003] The recommendation algorithms widely used in existing e-commerce product recommendation systems include collaborative filtering, content-based recommendation, and deep learning models (such as deep neural networks, convolutional neural networks, etc.). However, these recommendation algorithms have certain limitations when dealing with data sparsity problems, especially when new users or new products appear. It is difficult for the recommendation system to adapt to changes quickly, resulting in poor recommendation effects. In addition, current recommendation systems often fail to fully consider the impact of time factors (such as festivals) and spatial factors (such as geographic location) on user purchasing behavior, which limits the relevance and real-time nature of the recommendation results.

[0004] In view of this, the present invention proposes an e-commerce product recommendation optimization management system to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: an e-commerce commodity recommendation optimization management system, comprising:

[0006] Data processing module: used to collect comprehensive e-commerce data and pre-process it to obtain a comprehensive data set;

[0007] Weight assignment module: used to assign behavioral feature weights to user behavioral features based on the comprehensive data set; Matrix construction module: used to calculate the user's comprehensive score for each product based on the behavioral feature weights and construct the user-product interaction matrix;

[0008] Recommendation optimization module: used to calculate the user's preference for products based on the user-product interaction matrix and generate the optimal product category recommendation list.

[0009] Furthermore, the e-commerce comprehensive data includes user data, product attribute data and calendar data;

[0010] User data includes user’s personal information and behavior data;

[0011] Personal information data includes the user’s age, gender, geographic location, and occupation;

[0012] Behavioral data includes users’ click data, browsing history, favorites, search history, and purchase history;

[0013] Product attribute data includes the product’s name, category, description, brand, price, sales volume, and positive review rate.

[0014] Furthermore, the preprocessing method of the e-commerce comprehensive data includes:

[0015] Encoding conversion is performed on text data in e-commerce comprehensive data, including:

[0016] Step S31: remove punctuation marks, convert to lowercase, and remove stop words from all text data in the e-commerce comprehensive data, then use a word segmentation tool to perform word segmentation, and integrate the word segmentation results into a corpus;

[0017] Among them, the corpus is a two-dimensional table structure, with columns representing texts and rows representing vocabulary;

[0018] Step S32: using the corpus as input data, training the Word2vec model so that each word is mapped into a vector space of fixed length, and obtaining the Word2vec vector of each word in the corpus through the output of the model training;

[0019] Step S33: multiply the Word2vec vector of each word in the corpus by its corresponding TD value to obtain a weighted word vector. For any text, calculate the mean of the weighted word vectors of all words in the text to obtain the text vector.

[0020] The calculation method of the TD value includes:

[0021] The TD value is equal to the product of the word frequency and the inverse text frequency. The word frequency is the ratio of the number of times a word appears in the text to the total number of words in the text.

[0022] Where Nw represents the total number of texts in the corpus, and df(ct) represents the number of texts containing the vocabulary ct;

[0023] Extract numerical data from the comprehensive e-commerce data, perform data cleaning, missing value filling and data standardization together with text vectors to obtain a comprehensive data set.

[0024] Furthermore, the method of allocating behavioral feature weights to the user's behavioral features according to the comprehensive data set includes:

[0025] For the comprehensive data set of all users within the cycle time, extract the behavior data of each user, convert each behavior data type into a behavior feature, and for each user, generate the association rule "if SF, then GL" by mining frequent item sets, filter and calculate and assign the behavior feature weight of each user based on the association rules left by the filter;

[0026] The method of mining frequent itemsets includes:

[0027] Step S41: extracting basic behaviors of all users from the comprehensive data set, and treating each basic behavior as an independent first-order candidate item set;

[0028] Among them, basic behavior is defined as the behavior data generated by users during a single interaction;

[0029] Step S42: Set a minimum support threshold, calculate the support of each first-order candidate item set, and retain the item sets greater than or equal to the minimum support threshold as first-order frequent item sets;

[0030] The support is calculated as the ratio of the number of sessions containing the item set to the total number of sessions;

[0031] A session refers to the behavioral data of a user from login to logout.

[0032] Step S43: Pair all first-order frequent item sets in pairs to generate second-order candidate item sets;

[0033] Step S44: iteratively execute step S42 and step S43 until the support of all newly generated candidate item sets is less than the minimum support threshold, then terminate the iteration process and obtain frequent item sets of all orders.

[0034] Furthermore, the method of generating the association rule "if SF, then GL" is:

[0035] Among them, SF and GL both refer to behavioral data;

[0036] For frequent item sets larger than the first-order frequent item sets, generate all possible association rules;

[0037] The method of screening association rules includes:

[0038] Set the minimum confidence threshold, calculate the confidence and lift of each association rule, and retain the association rules with a confidence greater than or equal to the minimum confidence threshold and a lift greater than 1;

[0039] Among them, the confidence of the association rule is defined as the ratio of the support of the frequent item set {SF, GL} to the support of the frequent item set {SF};

[0040] The calculation method of association rule lift is:

[0041] The lift of an association rule is equal to the ratio of the confidence of the association rule to the support of the frequent itemset {GL}.

[0042] Furthermore, the method of calculating and allocating the behavioral feature weight of each user includes:

[0043] For all retained association rules, the product of their support and confidence is calculated to obtain the weight value of the association rule, and then decomposed into the behavioral feature weight through the weight distribution formula;

[0044] Weight distribution formula

[0045] Among them, SL is the association rule (SF→GL), a i is the i-th behavior feature in SF, Behavioral characteristics i The behavioral feature weight, GZ SL is the weight value of the association rule SL, Lf(a i →GL) is a i The improvement degree of is, i is the index of the behavior feature.

[0046] Furthermore, the method of obtaining the user's comprehensive interaction score for each product based on the behavioral feature weight calculation and constructing the user-product interaction matrix includes:

[0047] According to the behavioral feature weight, for each product, the user's single interaction score pf is calculated, and the user's comprehensive interaction score for the product is accumulated;

[0048] Single interaction rating

[0049] Among them, t represents the time interval from the occurrence of the behavior to the present, λ is the decay rate constant, and e is the natural constant;

[0050] Based on the user's comprehensive interaction score for each product, create a two-dimensional array, where rows represent users and columns represent products. Each element in the array represents the user's comprehensive interaction score for the product. If the user has no behavior data with the product, the element value is set to zero.

[0051] The sparse matrix storage format is used to convert the two-dimensional array into a sparse matrix, and missing value filling and matrix standardization are performed to obtain the user-product interaction matrix.

[0052] Furthermore, the method of calculating the user's preference for a product based on the user-product interaction matrix and generating an optimal product category recommendation list includes:

[0053] Set the holiday factor and geographic location factor, adjust them and integrate them into the user-product interaction matrix, set the total number of recommended categories N, calculate the preference of user-product interaction pairs, select the top N product categories with the highest preference, and form a product category recommendation list;

[0054] Among them, the preference is equal to the product of the festival factor and the festival factor weight plus the comprehensive interaction score and the geographical location factor.

[0055] Furthermore, the method of setting the holiday factor includes:

[0056] Analyze calendar data to get holiday times, extract behavior data during the holiday duration period from behavior data, count the number of times users generate behavior data for product categories during the holiday duration period and the time periods before and after the holiday, and calculate the holiday factor for each product category;

[0057] Festival Factor

[0058] Among them, SS is the number of behaviors, Th is the duration of the festival, Tn1 is the time period before the festival, Tn2 is the time period after the festival, Th = Tn1 = Tn2;

[0059] Methods for setting the geographic location factor include:

[0060] Crawl posts on social media that contain regions and product categories, perform text segmentation and remove stop words, and use sentiment analysis technology to build a sentiment dictionary that contains positive and negative sentiment words and corresponding sentiment scores;

[0061] Based on the sentiment dictionary, the sentiment scores of all positive sentiment words in the post are accumulated, and the difference is calculated with the sentiment scores of all negative sentiment words. The difference is calculated by the ratio of the total number of words in the post to obtain the sentiment tendency score of the post.

[0062] Count and accumulate the number of likes, comments, and reposts of the post to get the popularity score of the post;

[0063] Multiply the sentiment score of the post by the popularity score to get the comprehensive score of the post. Use the mean calculation method to process the comprehensive score of each post to get the average score of the product category in the region.

[0064] The average score of each commodity category in the region is accumulated and the ratio is calculated with the total number of commodity categories to obtain the average comprehensive score of all commodity categories in the region;

[0065] The average score is calculated by ratioing the average comprehensive score to obtain the geographical location factor of any product category in any region.

[0066] Furthermore, the method of integrating into the user-product interaction matrix after adjustment includes:

[0067] Based on the user-product interaction matrix, add festival factors and geographic location factors to each user-product interaction pair, and assign a festival factor weight;

[0068] The method for adding the geographical location factor is:

[0069] Determine the user's location through the user's geographic location, and add a corresponding geographic location factor to each user-product interaction pair;

[0070] The method to add festival factor weight is:

[0071] Items are classified according to holiday relevance. For each user-item interaction pair, if the current date is within the holiday duration period and the item is related to the holiday, the holiday weight is set to 1, otherwise it is set to 0, where 1 indicates the existence of the holiday and 0 indicates the non-existence of the holiday.

[0072] The technical effects and advantages of the e-commerce commodity recommendation optimization management system of the present invention are as follows:

[0073] The present invention not only covers traditional user behavior data, but also incorporates user personal information and detailed attributes of goods, thereby constructing a data set that can comprehensively depict user portraits and product characteristics, laying a solid foundation for achieving accurate recommendations; in addition, the present invention performs a series of operations such as encoding, cleaning, missing value processing and standardization on data, and uses natural language processing technology (such as the Word2vec model) to perform semantic analysis on text data, thereby improving the system's understanding of user needs and product information, thereby improving the accuracy and personalization of recommendations; at the same time, the present invention also introduces a dynamic weight allocation mechanism based on user behavior characteristics, and uses data mining technology to automatically identify user behavior mode, and assign weight values ​​to different features accordingly, so that the recommendation system can be adaptively adjusted according to the characteristics of different user groups, thereby enhancing the relevance of recommendations; and, considering the impact of specific holidays and user geographic locations on shopping preferences, the present invention integrates time and space factors, introduces holiday factors and geographic location factors, and combines user historical behavior data to dynamically adjust the recommendation strategy, so that the recommendation results are closer to the user's immediate needs and regional preferences; finally, the present invention proposes a user-product matrix construction method, which reasonably quantifies the user's interest in the product by calculating the impact of user behavior and the current time interval, thereby effectively solving the data sparsity problem and improving the overall performance of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a schematic diagram of an e-commerce product recommendation optimization management system of the present invention;

[0075] Figure 2 A schematic diagram of an e-commerce commodity recommendation optimization management method according to the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] Example 1

[0078] See also Figure 1 As shown, the e-commerce commodity recommendation optimization management system described in this embodiment includes:

[0079] Data processing module: used to collect comprehensive e-commerce data and pre-process it to obtain a comprehensive data set;

[0080] Weight assignment module: used to assign behavioral feature weights to user behavioral features based on the comprehensive data set;

[0081] Matrix construction module: used to calculate the comprehensive score of each product by the user based on the behavioral feature weights, and construct the user-product interaction matrix;

[0082] Recommendation optimization module: used to calculate the user's preference for products based on the user-product interaction matrix and generate the optimal product category recommendation list;

[0083] Furthermore, the e-commerce comprehensive data includes user data, product attribute data and calendar data;

[0084] User data includes user’s personal information and behavior data;

[0085] Personal information data includes the user’s age, gender, geographic location, and occupation;

[0086] Behavioral data includes users’ click data, browsing history, favorites, search history, and purchase history, which are collected through the website or app’s log system;

[0087] Product attribute data includes product name, category, description, brand, price, sales volume and favorable comment rate, which are collected through database or e-commerce platform API interface;

[0088] Furthermore, the preprocessing method of the e-commerce comprehensive data includes:

[0089] Encoding conversion is performed on text data in e-commerce comprehensive data, including:

[0090] Step S31: remove punctuation marks, convert to lowercase, and remove stop words from all text data in the e-commerce comprehensive data, then use a word segmentation tool to perform word segmentation, and integrate the word segmentation results into a corpus;

[0091] Among them, the corpus is a two-dimensional table structure, with columns representing texts and rows representing vocabulary;

[0092] Step S32: using the corpus as input data, training the Word2vec model so that each word is mapped into a vector space of fixed length, and obtaining the Word2vec vector of each word in the corpus through the output of the model training;

[0093] Step S33: multiply the Word2vec vector of each word in the corpus by its corresponding TD value to obtain a weighted word vector. For any text, calculate the mean of the weighted word vectors of all words in the text to obtain the text vector.

[0094] The calculation method of the TD value includes:

[0095] The TD value is equal to the product of the word frequency and the inverse text frequency. The word frequency is the ratio of the number of times a word appears in the text to the total number of words in the text.

[0096] Where Nw represents the total number of texts in the corpus, and df(ct) represents the number of texts containing the vocabulary ct;

[0097] Extract numerical data from e-commerce comprehensive data, perform data cleaning, missing value filling and data standardization together with text vectors to obtain a comprehensive data set;

[0098] Furthermore, the method of allocating behavioral feature weights to the user's behavioral features according to the comprehensive data set includes:

[0099] For the comprehensive data set of all users within the cycle time, extract the behavior data of each user, convert each behavior data type into a behavior feature, and for each user, generate the association rule "if SF, then GL" by mining frequent item sets, filter and calculate and assign the behavior feature weight of each user based on the association rules left by the filter;

[0100] The method of mining frequent itemsets includes:

[0101] Step S41: extracting basic behaviors of all users from the comprehensive data set, and treating each basic behavior as an independent first-order candidate item set;

[0102] It should be noted that basic behavior is defined as the behavior data generated by users in a single interaction process. These basic behaviors are the basic units that constitute the comprehensive data set. For example, in one session, a user may have clicked on ads, browsed products, searched for keywords, and purchased products. Clicking on ads is considered a basic behavior, and the same applies to browsing products, searching for keywords, and purchasing products.

[0103] Step S42: Set a minimum support threshold, calculate the support of each first-order candidate item set, and retain the item sets greater than or equal to the minimum support threshold as first-order frequent item sets (for example, {click on advertisement}, {browse product}, {search keyword}, ​​{purchase product});

[0104] The support is calculated as the ratio of the number of sessions containing the item set to the total number of sessions (for example, if there are 10 sessions in total and {click on the ad} appears in 6 sessions, then the support of {click on the ad} is 0.6);

[0105] A session refers to the behavior data of a user from login to logout (for example, if a user clicks on an ad, browses products, searches for keywords, and buys products during the entire process from login to logout, then this session includes the behavior data of clicking on an ad, browsing products, searching for keywords, and buying products).

[0106] Step S43: All first-order frequent item sets are combined in pairs to form second-order candidate item sets (for example, {click on the ad, browse the product}, {click on the ad, search for keywords}, {click on the ad, buy the product} {browse the product, search for keywords}, {browse the product, buy the product}, {search for keywords, buy the product});

[0107] Step S44: iteratively execute step S42 and step S43 until the support of all newly generated candidate item sets is less than the minimum support threshold, terminate the iteration process, and obtain frequent item sets of all orders (for example, 3-item set {search keyword, browse product, purchase product}, 4-item set {search keyword, click on advertisement, browse product, purchase product});

[0108] Furthermore, the method of generating the association rule "if SF, then GL" includes:

[0109] Among them, SF and GL both refer to behavioral data;

[0110] For frequent item sets larger than the first-order frequent item sets, generate all possible association rules;

[0111] It should be noted that generating all possible association rules generates candidate association rules. For example, there are first-order frequent item sets: {click on the ad}, {browse the product}, {search for keywords}, second-order frequent item sets: {click on the ad, browse the product}, {browse the product, search for keywords} and third-order frequent item sets: {click on the ad, browse the product, search for keywords}. For each second-order frequent item set, all possible association rules are generated, including: from {click on the ad, browse the product}: Rule 1: If you click on the ad, browse the product, Rule 2: If you browse the product, click on the ad; from {browse the product, search for keywords}: Rule 3: If browsing products, search for keywords, Rule 4: If searching for keywords, browse products; For each third-order frequent item set, generate all possible association rules, including, from {click on ads, browse products, search for keywords}: Rule 5: If clicking on ads, browse products and search for keywords, Rule 6: If browsing products, click on ads and search for keywords, Rule 7: If searching for keywords, click on ads and browse products, Rule 8: If clicking on ads and browsing products, search for keywords, Rule 9: If clicking on ads and searching for keywords, browse products, Rule 10: If browsing products and searching for keywords, click on ads;

[0112] Set a minimum confidence threshold, generate association rules from frequent item sets (for example, if you click on an ad, then browse products), calculate the confidence of each association rule (the probability of GL occurring under a given SF), retain reliable association rules greater than or equal to the threshold, calculate their lift, and retain association rules with a lift greater than 1;

[0113] The calculation method of the confidence of the association rule includes:

[0114] Confidence

[0115] Among them, (SF∪GL) is the frequent item set {SF, GL} containing SF and GL, that is, the situation where SF and GL occur at the same time, Sp(SF∪GL) is the support of the frequent item set {SF, GL} containing both SF and GL in all sessions (for example, the support of "the user clicked on product A and bought product B"), and Sp(SF) is the support of the frequent item set {SF} containing SF in all sessions (for example, the support of "the user clicked on product A");

[0116] The calculation method of the association rule lifting degree includes:

[0117] Lift

[0118] Where Lf(SF→GL) is the degree of improvement of the product of the probabilities of SF and GL occurring simultaneously, relative to their independent occurrence, which is used to measure whether the probability of SF and GL occurring simultaneously is higher than the expected probability of their independent occurrence. Sp(GL) is the support of the frequent item set {GL} containing GL. If the lift is greater than 1, it means that there is a positive correlation between SF and GL (for example, assuming that SF is a user clicking on product A and GL is a user purchasing product B, if the lift is 2.5, it means that the possibility of a user purchasing product B after clicking on product A is 2.5 times the possibility under random conditions). If it is equal to 1, the two are independent. If it is less than 1, there is a negative correlation.

[0119] It should be noted that all the thresholds mentioned above are set by those skilled in the art according to actual conditions;

[0120] Furthermore, the method of calculating the weight of the user behavior feature according to the association rules left by the screening includes:

[0121] For all association rules with a lift greater than 1, calculate the product of their support and confidence to obtain the weight of the association rule, and decompose the association rule weight into each single behavioral feature weight through the weight distribution formula;

[0122] It should be noted that the weight obtained through association rules is usually for the association rules as a whole, rather than the weight of a single feature (i.e., user behavior). This overall weight needs to be decomposed into the weight of each single feature to better reflect the importance of each behavior.

[0123] The method of decomposing the weight of the association rule includes:

[0124] For each association rule SL: SF→GL, where SF is composed of i behavioral features a;

[0125] According to the differences of SF in each association rule, a corresponding weight distribution formula is configured for each association rule;

[0126] For SF=a1∧a2∧...∧a i , the weight distribution formula is defined as:

[0127]

[0128] in, Behavioral characteristics i The behavioral feature weight, GZ SL is the weight of the association rule SL, L f(a i →GL) is a i The improvement degree of , i is the index of the behavior feature;

[0129] Furthermore, the method of obtaining the user's comprehensive interaction score for each product based on the behavioral feature weight calculation and constructing the user-product interaction matrix includes:

[0130] According to the behavioral feature weight, for each product, the user's single interaction score pf is calculated, and the user's comprehensive interaction score for the product is accumulated;

[0131] Single interaction rating

[0132] Among them, t represents the time interval from the occurrence of the behavior to the present, λ is the decay rate constant, and e is the natural constant;

[0133] It should be noted that λ is a preset value that controls the decay rate of the score over time. A larger λ value will cause the score to decay faster, while a smaller λ value will cause the score to decay slower. In practical applications, the value range of λ can range from very small positive numbers to large positive numbers, depending on the application background and data characteristics. For example, 0<λ≤0.01, in this case, the decay rate of the score is very slow, which is suitable for long-term memory of user behavior. When the impact of user behavior on subsequent behavior lasts for a long time, a smaller λ value can be selected; 0.01<λ≤0.5, in this case, the decay rate of the score is moderate, which is suitable for most application scenarios. When the impact of user behavior on subsequent behavior gradually weakens in a short period of time, a medium-sized λ value can be selected; 0.5<λ, in this case, the decay rate of the score is fast, which is suitable for rapidly changing environments or short-term memory of user behavior. When the impact of user behavior on subsequent behavior will soon disappear, a larger λ value can be selected;

[0134] Based on the user's comprehensive interaction score for each product, create a two-dimensional array, where rows represent users and columns represent products. Each element in the array represents the user's comprehensive interaction score for the product. If the user does not interact with the product, the element value is set to zero.

[0135] It should be noted that since not all users have behavioral records for all products, the matrix is ​​usually sparse, and sparse matrix storage technology can be used to reduce the storage space requirements;

[0136] Use a sparse matrix storage format (such as CSR or CSC) to convert the two-dimensional array into a sparse matrix, perform missing value filling and matrix standardization on the sparse matrix to obtain a user-item interaction matrix;

[0137] Among them, missing value filling is to use a filling method (such as global average, matrix decomposition, collaborative filtering) to fill missing values ​​for user-item pairs without behavior records; matrix standardization is to use a standardization method (such as Z-score standardization) to convert each feature into a standard normal distribution, that is, the mean is 0 and the variance is 1;

[0138] Furthermore, the method of calculating the user's preference for a product based on the user-product interaction matrix and generating an optimal product category recommendation list includes:

[0139] Set the holiday factor and geographic location factor, adjust them and integrate them into the user-product interaction matrix, set the total number of recommended categories N, calculate the preference of user-product interaction pairs, select the top N product categories with the highest preference, and form a product category recommendation list;

[0140] Among them, the preference is equal to the product of the festival factor and the festival factor weight plus the comprehensive interaction score and the geographical location factor;

[0141] Furthermore, the methods for setting the holiday factor include:

[0142] Analyze calendar data to get holiday times, extract behavioral data (such as click data, search records, and purchase records) during the holiday duration period from behavioral data, count the number of times users generate behavioral data for product categories during the holiday duration period and the time periods before and after the holiday, and calculate the holiday factor for each product category;

[0143] Festival Factor

[0144] Wherein, SS is the number of behaviors, Th is the duration of the festival, Tn1 is the time period before the duration of the festival, Tn2 is the time period after the duration of the festival, Th=Tn1=Tn2, SS(Th) is the total number of behaviors (such as search, click or purchase) generated by users on a certain commodity category during the duration of the festival (the festival period is the festival day and the period before and after the festival day. For example, the festival day and the week before and after the festival are combined into the festival period, and the festival period is 15 days), SS(Tn1) is the total number of behaviors generated by users on a specific commodity category during the period before the festival (for example, if the festival period is 15 days, the period before the festival is the 15 days before the first day of the festival), and SS(Tn2) is the total number of behaviors generated by users on a specific commodity category during the period after the festival (for example, if the festival period is 15 days, the period after the festival is the 15 days after the 15th day of the festival);

[0145] Methods for setting the geographic location factor include:

[0146] Use API or crawler technology to crawl posts containing keywords of regions and product categories from social media platforms;

[0147] Segment the text in the post and remove stop words to obtain the retained vocabulary;

[0148] Among them, word segmentation is to use word segmentation tools (such as Jieba) to split the post content into words (for example, "This down jacket is really warm, a must-have in the northern winter!" will be split into "this", "down jacket", "real", "warm", "north", "winter", "must-have"); stop word removal is to use natural language processing (NLP) technology to remove meaningless words (for example, remove "this", "real", "north", and keep "down jacket", "warm", "winter", "must-have");

[0149] Based on the reserved words in each post, sentiment analysis techniques (such as TextBlob, SnowNLP, VADER) are used to build a sentiment dictionary containing positive sentiment words and negative sentiment words, as well as corresponding sentiment scores;

[0150] Based on the sentiment dictionary, the sentiment scores of all positive sentiment words in the post are accumulated, and the difference is calculated with the sentiment scores of all negative sentiment words. The difference is compared with the total number of words in the post to obtain the sentiment tendency score of each post.

[0151] Count and accumulate the number of likes, comments, and reposts of the post to get the popularity score of the post;

[0152] Multiply the sentiment score of the post by the popularity score to get the comprehensive score of the post. Use the mean calculation method to process the comprehensive score of each post to get the average score of the product category in the region.

[0153] The average score of each commodity category in the region is accumulated and the ratio is calculated with the total number of commodity categories to obtain the average comprehensive score of all commodity categories in the region;

[0154] The geographical location factor of any product category in any region is calculated by comparing the average score with the average comprehensive score;

[0155] It should be noted that the geographical location factor is the geographical location factor of a product category in a certain region. For example, the geographical location factor of the warm clothing category in a certain region in the north may be 1.38, and the geographical location factor in a certain region in the south may be 0.9. The geographical location factor of each product category in different regions may be different.

[0156] Furthermore, the method of integrating into the user-product interaction matrix after adjustment includes:

[0157] Based on the user-product interaction matrix, add festival factors and geographic location factors to each user-product interaction pair, and assign a festival factor weight;

[0158] The method for adding the geographical location factor is:

[0159] Determine the user's location through the user's geographic location, and add a corresponding geographic location factor to each user-product interaction pair based on the user's location;

[0160] The method for adding festival factor weight is:

[0161] Assign one or more category labels to each product, mark which products are related to festivals, and classify products by festival relevance;

[0162] For each user-item interaction pair, check whether the current date is during a holiday. If the current date is during a holiday and the item is related to the holiday, set the holiday weight to 1, otherwise it is 0, where 1 indicates the existence of a holiday and 0 indicates the non-existence of a holiday.

[0163] In this embodiment, the e-commerce comprehensive data is collected through the data processing module, and the non-numeric data is encoded and pre-processed together with the numeric data, and the pre-processed data is integrated into a comprehensive data set. Then, the comprehensive data set is used to extract the user's behavior characteristics in the weight allocation module, and by generating association rules, the user behavior characteristics are evaluated and weights are allocated according to the importance of the rules. Then, in the matrix construction module, for each user's behavior on a specific product, the behavior score is calculated according to the user's behavior feature weight and the interval between the behavior occurrence time and the current time, and all the user's behavior scores on a certain product are accumulated to obtain the comprehensive score of the user-product pair. Based on the comprehensive score of the user-product pair, the user-product pair is ranked according to the weight of the user's behavior feature and the interval between the behavior occurrence time and the current time. The user-product matrix is ​​constructed. Next, the specific holiday and user geographic location information is integrated into the recommendation optimization module, and the holiday factor and geographic location factor are set. After comprehensive adjustment, they are integrated into the user-product matrix. The user's preference for the product is calculated through the user-product matrix, and the total number of recommended categories N is set according to the preference. The top N product categories with the highest preference are selected to generate a product category recommendation list. This solves the problems of improper handling of regional and time differences and difficulty in optimizing the performance of the recommendation system, and realizes the improvement of the traditional e-commerce recommendation system, especially in improving the accuracy of recommendations and enhancing user experience, thereby improving the user experience and operational efficiency of the e-commerce platform.

[0164] Example 2

[0165] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an e-commerce commodity recommendation optimization management method, including:

[0166] S1, collect e-commerce comprehensive data and preprocess it to obtain a comprehensive data set;

[0167] S2, assigning behavioral feature weights to user behavioral features based on the comprehensive data set;

[0168] S3, calculate the comprehensive score of each product by the user based on the behavioral feature weight, and construct a user-product interaction matrix;

[0169] S4. Based on the user-product interaction matrix, calculate the user's product preference for the product and generate an optimal product category recommendation list.

[0170] Example 3

[0171] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned e-commerce product recommendation optimization management method is implemented.

[0172] Since the electronic device introduced in this embodiment is an electronic device used to implement an e-commerce commodity recommendation optimization management method in the embodiment of this application, based on the e-commerce commodity recommendation optimization management method introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not introduced in detail here. As long as the technical personnel of this field implement the electronic device used in the e-commerce commodity recommendation optimization management method in the embodiment of this application, it belongs to the scope of protection of this application.

[0173] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0174] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An e-commerce product recommendation optimization management system, characterized in that: The e-commerce commodity recommendation optimization management system includes: Data processing module: used to collect comprehensive e-commerce data and pre-process it to obtain a comprehensive data set; Weight assignment module: used to assign behavioral feature weights to user behavioral features based on the comprehensive data set; Matrix construction module: used to calculate the comprehensive score of each product by the user based on the behavioral feature weights, and construct the user-product interaction matrix; Recommendation optimization module: used to calculate the user's preference for products based on the user-product interaction matrix and generate the optimal product category recommendation list.

2. The e-commerce commodity recommendation optimization management system according to claim 1, characterized in that: The e-commerce comprehensive data includes user data, product attribute data and calendar data; User data includes user’s personal information and behavior data; Personal information data includes the user’s age, gender, geographic location, and occupation; Behavioral data includes users’ click data, browsing history, favorites, search history, and purchase history; Product attribute data includes the product’s name, category, description, brand, price, sales volume, and positive review rate.

3. The e-commerce product recommendation optimization management system according to claim 1, characterized in that: The preprocessing method of the e-commerce comprehensive data includes: Encoding conversion is performed on text data in e-commerce comprehensive data, including: Step S31: remove punctuation marks, convert to lowercase, and remove stop words from all text data in the e-commerce comprehensive data, then use a word segmentation tool to perform word segmentation, and integrate the word segmentation results into a corpus; Among them, the corpus is a two-dimensional table structure, with columns representing texts and rows representing vocabulary; Step S32: using the corpus as input data, training the Word2vec model so that each word is mapped into a vector space of fixed length, and obtaining the Word2vec vector of each word in the corpus through the output of the model training; Step S33: multiply the Word2vec vector of each word in the corpus by its corresponding TD value to obtain a weighted word vector. For any text, calculate the mean of the weighted word vectors of all words in the text to obtain the text vector. The calculation method of the TD value includes: The TD value is equal to the product of the word frequency and the inverse text frequency. The word frequency is the ratio of the number of times a word appears in the text to the total number of words in the text. Where Nw represents the total number of texts in the corpus, and df(ct) represents the number of texts containing the vocabulary ct; Extract numerical data from the comprehensive e-commerce data, perform data cleaning, missing value filling and data standardization together with text vectors to obtain a comprehensive data set.

4. The e-commerce commodity recommendation optimization management system according to claim 1, characterized in that: The method of assigning behavioral feature weights to the user's behavioral features according to the comprehensive data set includes: For the comprehensive data set of all users within the cycle time, extract the behavior data of each user, convert each behavior data type into a behavior feature, and for each user, generate the association rule "if SF, then GL" by mining the frequent item sets, filter and calculate and assign the behavior feature weight of each user based on the association rules left by the filter; The method of mining frequent itemsets includes: Step S41: extracting basic behaviors of all users from the comprehensive data set, and treating each basic behavior as an independent first-order candidate item set; Among them, basic behavior is defined as the behavior data generated by users during a single interaction; Step S42: Set a minimum support threshold, calculate the support of each first-order candidate item set, and retain the item sets greater than or equal to the minimum support threshold as first-order frequent item sets; The support is calculated as the ratio of the number of sessions containing the item set to the total number of sessions; A session refers to the behavioral data of a user from login to logout. Step S43: Pair all first-order frequent item sets in pairs to generate second-order candidate item sets; Step S44: iteratively execute step S42 and step S43 until the support of all newly generated candidate item sets is less than the minimum support threshold, then terminate the iteration process and obtain frequent item sets of all orders.

5. The e-commerce commodity recommendation optimization management system according to claim 4, characterized in that: The method of generating the association rule "if SF, then GL" is: Among them, SF and GL both refer to behavioral data; For frequent item sets larger than the first-order frequent item sets, generate all possible association rules; The method of screening association rules includes: Set the minimum confidence threshold, calculate the confidence and lift of each association rule, and retain the association rules with a confidence greater than or equal to the minimum confidence threshold and a lift greater than 1; Among them, the confidence of the association rule is defined as the ratio of the support of the frequent item set {SF, GL} containing both SF and GL to the support of the frequent item set {SF} containing SF; The calculation method of association rule lift is: The lift of an association rule is equal to the ratio of the confidence of the association rule to the support of the frequent itemset {GL} containing GL.

6. The e-commerce commodity recommendation optimization management system according to claim 4, characterized in that: The method of calculating and allocating the behavioral feature weight of each user includes: For all retained association rules, the product of their support and confidence is calculated to obtain the weight value of the association rule, and then decomposed into the behavioral feature weight through the weight distribution formula; Weight distribution formula Among them, SL is the association rule (SF→GL), a i is the i-th behavior feature in SF, Behavioral characteristics i The behavioral feature weight, GZ SL is the weight value of the association rule SL, Lf(a i →GL) is a i The improvement degree of is, i is the index of the behavior feature.

7. The e-commerce commodity recommendation optimization management system according to claim 1, characterized in that: The method of calculating the comprehensive interaction score of the user for each product based on the behavior feature weight and constructing the user-product interaction matrix includes: According to the behavioral feature weight, for each product, the user's single interaction score pf is calculated, and the user's comprehensive interaction score for the product is accumulated; Single interaction rating Among them, t represents the time interval from the occurrence of the behavior to the present, λ is the decay rate constant, and e is the natural constant; Based on the user's comprehensive interaction score for each product, create a two-dimensional array, where rows represent users and columns represent products. Each element in the array represents the user's comprehensive interaction score for the product. If the user has no behavior data with the product, the element value is set to zero. The sparse matrix storage format is used to convert the two-dimensional array into a sparse matrix, and missing value filling and matrix standardization are performed to obtain the user-product interaction matrix.

8. The e-commerce commodity recommendation optimization management system according to claim 1, characterized in that: The method of calculating the user's preference for a product based on the user-product interaction matrix and generating an optimal product category recommendation list includes: Set the holiday factor and geographic location factor, adjust them and integrate them into the user-product interaction matrix, set the total number of recommended categories N, calculate the preference of user-product interaction pairs, select the top N product categories with the highest preference, and form a product category recommendation list; Among them, the preference is equal to the product of the festival factor and the festival factor weight plus the comprehensive interaction score and the geographical location factor.

9. The e-commerce commodity recommendation optimization management system according to claim 8, characterized in that: The method of setting the festival factor includes: Analyze calendar data to obtain holiday times, extract behavior data within the holiday duration period from behavior data, count the number of times users generate behavior data for product categories within the holiday duration period and the time periods before and after the holiday, and calculate the holiday factor for each product category; Festival Factor Among them, SS is the number of behaviors, Th is the duration of the festival, Tn1 is the time period before the festival, Tn2 is the time period after the festival, Th = Tn1 = Tn2; Methods for setting the geographic location factor include: Crawl posts on social media that contain regions and product categories, perform text segmentation and remove stop words, and use sentiment analysis technology to build a sentiment dictionary that contains positive and negative sentiment words and corresponding sentiment scores; Based on the sentiment dictionary, the sentiment scores of all positive sentiment words in the post are accumulated, and the difference is calculated with the sentiment scores of all negative sentiment words. The difference is calculated by the ratio of the total number of words in the post to obtain the sentiment tendency score of the post. Count and accumulate the number of likes, comments, and reposts of the post to get the popularity score of the post; Multiply the sentiment score of the post by the popularity score to get the comprehensive score of the post. Use the mean calculation method to process the comprehensive score of each post to get the average score of the product category in the region. The average score of each commodity category in the region is accumulated and the ratio is calculated with the total number of commodity categories to obtain the average comprehensive score of all commodity categories in the region; The average score is calculated by ratioing the average comprehensive score to obtain the geographical location factor of any product category in any region.

10. The e-commerce commodity recommendation optimization management system according to claim 8, characterized in that: The method of integrating the adjustment into the user-product interaction matrix includes: Based on the user-product interaction matrix, add festival factors and geographic location factors to each user-product interaction pair, and assign a festival factor weight; The method for adding the geographical location factor is: Determine the user's location through the user's geographic location, and add a corresponding geographic location factor to each user-product interaction pair; The method to add festival factor weight is: Items are classified according to holiday relevance. For each user-item interaction pair, if the current date is within the holiday duration period and the item is related to the holiday, the holiday weight is set to 1, otherwise it is set to 0, where 1 indicates the existence of the holiday and 0 indicates the non-existence of the holiday.

Images