Marketing content recommendation method based on data analysis

Abstract

The invention discloses a marketing content recommendation method based on data analysis. The method comprises the following steps: acquiring consumption data of a user; the consumption data of each user form an independent text document, and the text document is standardized and subjected to word segmentation; forming a product marketing keyword list, and combining keywords in the list into a query text; obtaining the relevance between the text document and the query text as the interestingness of the user on the product; executing first push to the user exceeding the interestingness threshold value at the first push time to obtain first push data; obtaining consumption characteristics according to the consumption data and the first push data, establishing a network neural model, and evaluating the purchase intention of the user failing in conversion in the first push to obtain a purchase intention index of the user; and adjusting the purchase intention index of the user according to the first push data, and executing second push to the user whose purchase intention index reaches an index threshold. According to the method, the recommendation accuracy is improved, and the user experience is optimized.

Description

Technical Field

[0001] This application relates to the field of marketing recommendation technology, and more specifically, to a marketing content recommendation method based on data analysis. Background Technology

[0002] In existing technologies, marketing content messages can trigger highly relevant notifications based on user location, behavior, and time, directly displaying on the user's phone lock screen or notification center. Users can directly click to jump to specific pages within the app for action, and push permissions for specific apps can be easily enabled or disabled in system settings. Therefore, marketing push notifications have become the most common way to promote platform products. The core objective of recommendation is to push the most relevant and compelling marketing content, such as advertisements, products, articles, videos, and offers, to target users at the right time and in the right context to improve user engagement, conversion rates, and commercial value. Its technological foundation is mainly built on modern recommendation systems, combined with specific needs and data from the marketing field. Currently, the technology generally involves analyzing the characteristics of the items themselves and user profiles to recommend items that match the user profile. However, since marketing objectives are often multi-dimensional, using linear, standalone models or methods becomes problematic when detailed user behavior data, preference data, purchase history, and lifecycle stages are lacking. This prevents push notifications from automatically adjusting their selection of users, leading to broad or even incorrect tags. Consequently, content may become disconnected from users' actual needs, potentially triggering persistent push notifications based on simple behaviors, such as a single link click, or sending the same content to active and churning users. This results in excessively high push frequency, failing to consider users' information processing capacity and the suitability for different times and scenarios, causing annoyance and even lowering conversion rates. Therefore, a more optimized marketing push method is needed, capable of more flexibly adjusting the target audience, avoiding ineffective pushes, and regulating conversion rates. Summary of the Invention

[0003] This application provides a data analysis-based marketing content recommendation method, including the following steps:

[0004] S1, acquire user consumption data, which includes user information, consumption records, and behavior records; form a separate text document for each user's consumption data, and standardize and segment the text document;

[0005] S2, Based on the product description and product functions, generate a list of keywords for product marketing, and combine the keywords in the list into query text;

[0006] S3, obtain the correlation between the text document and the query text to get the user's interest in the product;

[0007] S4, set the interest threshold, and perform the first push to users who exceed the interest threshold at the first push time to obtain the first push data;

[0008] S5: Based on consumption data and first push data, obtain consumption characteristics, establish a network neural model, evaluate the purchase intention of users who failed to convert in the first push, and obtain the user's purchase intention index.

[0009] S6: Adjust the user's purchase intention index based on the first push data, and execute the second push for users whose purchase intention index has reached the index threshold.

[0010] In step S1, the method for obtaining behavior records is as follows: embed gtag.js or GA4 measurement protocol tracking code globally in the website / APP, and send a page_view event containing URL, timestamp, and device information when each page is loaded.

[0011] The user behaviors include browsing search pages, browsing history, in-site conversations, click paths, search keywords, and adding items to the shopping cart.

[0012] In step S3, a score is used to measure the relevance between the query text Q and the text document D containing user A's consumption data.

[0013] Among them, IDF(q) i ) represents the keyword q i Inverse document frequency, TF(q) i D) represents the keyword q i In the term frequency of the text document (D), TF(q) i ,Q) represents the keyword q i The term frequency of the query text (Q);

[0014] Obtain the inverse document frequency

[0015] Where n represents the keyword q i The number of documents; N represents the total number of documents; This means it does not contain q. i The number of documents and the number of documents containing q i The ratio of the number of documents;

[0016] q i In document D, the word frequency

[0017] Where, f(q) i D) represents q iThe number of times a term appears in user A's text document D, |D| represents the length of text document D, avgdl represents the average length of the document set, k1 represents the rate of increase of the term frequency result in term frequency saturation, and b represents the role of document length normalization.

[0018] q i Querying the term frequency of text Q

[0019] Where f(q) i ,Q) represents q i k2 represents the frequency importance of a term in the query text Q, where k2 is the number of times it appears in the query text Q.

[0020] In step S5, a neural network model is established, consumption features are input into the neural network model, weights are assigned to each feature according to its importance, and the user's purchase intention index for the product is output.

[0021] The specific calculation method is as follows:

[0022]

[0023] Where W is the feature vector, F is the feature weight, and T is the feature data vector after weighted average;

[0024] Establish a neural network model, initialize the weights to a non-zero random number less than 1, and normalize the weighted average data vector T to obtain the data vector X; it is known that the feedforward multilayer network has n layers, and each network device attribute corresponds to a neuron in the layer;

[0025] Set the output of the i-th neuron in the k-th layer of the neural network to... 1≤k≤n, where,

[0026]

[0027] Where f(·) is the activation function, The variable for the activation function;

[0028]

[0029] W i,n+1 =-θ;

[0030] Among them W i,j Let θ be the weights from neuron j to neuron i, and θ be the threshold.

[0031] Using the sigmoid activation function, we get:

[0032]

[0033] Wi,j The formulas for updating weights and thresholds θ are as follows:

[0034] Through backpropagation, let m be the output layer (last layer) of the neural network, and let the error of the i-th neuron in the m-th layer be...

[0035] in, To calculate the output of the i-th neuron in the m-th layer using a neural network, Y i It is the expected value;

[0036] The output layer error term is:

[0037] Where η is the learning rate, η∈(0,1); a is the impulse coefficient, which is used to prevent the addition of an impulse term to local minima, a∈(0,1); p is the number of iterations; and ΔW i,j (p) represents the change in weight between the j-th neuron in the (m-1)-th layer and the i-th neuron in the m-th layer during the p-th iteration;

[0038] Error of the i-th neuron in hidden layer k

[0039]

[0040] Where k represents the layers other than the output layer, and ΔW i,j (p) represents the change in weight between the j-th neuron in layer k-1 and the ith neuron in layer k during the p-th iteration, where l belongs to layer k+1. l,i The weights are the values ​​of the i-th neuron in the k-th layer to the l-th neuron in the (k+1)-th layer. The output is the j-th neuron in the (k-1)-th layer;

[0041] The weight W in the Pth iteration i,j (p)=W i,j (p-1)+ΔW i,j (p).

[0042] The input consumption characteristics include: age, gender, region, number of purchases of similar products, number of times the user clicks in, number of times the user clicks out, scrolling depth of the recommended page, browsing duration, number of times the user adds items to their shopping cart, and number of times the user shares the product.

[0043] In step S6, the method for sending the second push is as follows:

[0044] Retrieve the user's session window from the first push data. Set a valid session duration threshold of h0, and select session windows with durations exceeding the threshold h0 as valid windows M, where M = [M1, M2, M3, ... M]. d ], where the z-th valid window is M zM z The window time is [ZF0, ZF1], and the window duration is TM;

[0045] Set the core conversion time period of the push notification to △T1, and the sending time of the first push notification is T1∈△T1;

[0046] Among them, when At the second sending time T2, a second push is sent again to users whose purchase intention index exceeds the index threshold, where T2∈△T1, T2=T1+λ;

[0047] when At the same time as the first push notification is sent, a second push notification is sent to users whose purchase intention index exceeds the index threshold.

[0048] when At that time, a second push notification is sent to users whose purchase intention index exceeds the index threshold at times when their conversation frequency is high.

[0049] In step S6, the method for adjusting the user's purchase intention index includes:

[0050] Set to session window M z The initial associated extension time period is △ZF, which includes r sub-extensions, △ZF = [△ZF1, △ZF2, △ZF3, ... △ZF]. r ], where △ZF1=M z +△ZF0, △ZF0≥0; Set the p-th sub-extension segment as △ZF0. p , △ZF p For M z The feedback weight is μ p ;

[0051] Within the extended time period △ZF, the system triggered a total of s related applications, including k positive feedback related applications FZP = [FZP1, FZP2, FZP3, ... FZP]. k ], and h negative feedback associated applications NZP = [NZP1, NZP2, NZP3, ... NZP h ], where the g-th positive feedback correlation application is set to FZP g FZP g Occurring in the g-th extension segment of △ZF g The u-th negative feedback correlation application is NZP. u NZP u Occurring in the u-th extension segment △ZF u ;

[0052] Obtain the z-th valid window M z Push feedback rate

[0053] Obtain the total feedback score of the user's d valid windows

[0054] Given that the user's purchase intention index is W B At this point, the user's purchase intention index is adjusted to W. B '=δW B .

[0055] The beneficial effects achieved by this application are as follows:

[0056] By integrating multi-dimensional data to form complete user profiles, the data value density is significantly improved. The structured storage method facilitates subsequent model iteration and training, supports long-term user behavior tracking and analysis, accurately identifies product selling points, establishes a mapping relationship between marketing language and user needs, and can adjust keyword weights in real time based on market feedback to maintain the flexibility of marketing strategies.

[0057] By combining user activity time period analysis, such as delaying push notifications to active users at night, bounce rates can be improved. Model parameters are adjusted in real-time based on initial push feedback, resulting in faster response times compared to traditional methods. An exponential decay algorithm avoids excessive intrusion, reducing user aversion and improving conversion rates. Precise targeting lowers marketing costs, reduces irrelevant pushes, and improves NPS scores. This solution uses data-driven decision-making to maximize marketing effectiveness while building a user-friendly push ecosystem. It allows for more flexible adjustment of push targets, avoiding ineffective pushes, and adjusting conversion rates, effectively maintaining brand reputation while enhancing business value. Attached Figure Description

[0058] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart illustrating the marketing content recommendation method based on data analysis proposed in this application.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0062] This application provides a data analysis-based marketing content recommendation method. First, it acquires user consumption data, including user information, consumption records, and behavioral records. The user information and consumption records are obtained through the e-commerce backend's member database system information, including user age, gender, region, historical orders, payment methods, return / exchange records, and other consumption-related information. Behavioral records are collected using the company's proprietary statistical tools for event tracking and session recording to track user behavior. Tracking code using gtag.js or the GA4 measurement protocol is embedded globally in the website / app. Each page loads a page_view event containing the URL, timestamp, and device information, collecting user page browsing history, in-site conversations, click paths, search keywords, and browsing / search behaviors such as adding items to the cart / favorites. Each user's consumption data is then generated into a separate text document.

[0063] The acquired text document is standardized, unifying its format and capitalization. The Python instantiation statement is: `cleaned_text = text.lower()`. Then, regular expressions are used to filter it, excluding non-alphanumeric and non-Chinese characters. The Python instantiation statement is: `cleaned_text = re.sub(r'[^\w\s\u4e00-\u9fff]',",text`. The standardized text document is then segmented using jieba. The `lcut()` method is used in the instantiation, and setting the `cut_all` parameter to `True` will return all possible segmentation results. Additionally, a marketing language dictionary can be added to the jieba library, increasing the weight of words with relevant brand meaning. After segmentation, the text is cleaned to remove stop words, punctuation, and special characters, and then stemmed or lemmatized.

[0064] Based on the product description and functions of the new product, obtain a list of keywords for product promotion, and combine the product keywords in the list into query text, such as: "portable long-lasting smart Bluetooth bone conduction headphones".

[0065] Using a product-consumption probability correlation model, the correlation between text documents and query text is described by inverse document frequency (IDF) and term frequency (TF). The scores are adjusted based on document length and average length, and the scores of all keywords in the query text are summed to represent the user's interest in the product.

[0066] To prevent a term from appearing too many times in a document and thus having an excessive weight, a saturation function is introduced. To better adapt to documents of different lengths and ensure that the impact of document length on weight is not linear, a document length factor is introduced.

[0067] Specifically, a score is obtained to measure the relevance between the query text (Q) and the text document (D) of user A's session log. Where IDF(q) i ) represents the keyword q i Inverse document frequency, TF(q) i D) represents the keyword q i In the term frequency of the text document (D), TF(q) i ,Q) represents the keyword q i The term frequency of the query text (Q);

[0068] Obtain the inverse document frequency

[0069] Where n represents the keyword q i The number of documents; N represents the total number of documents; This means it does not contain q. i The number of documents and the number of documents containing q i The ratio of the number of documents; 0.5 is a smoothing value to ensure that there is no division by zero; 1 is used to ensure that the argument of log is not 0;

[0070] q i In document D, the word frequency

[0071] Where, f(q) i D) represents q i The number of times a term appears in user A's text document D, |D| represents the length of text document D, avgdl represents the average length of the document set, k1 represents the rate of increase of the term frequency result in term frequency saturation, the default value is 1.2, b represents the role of document length normalization, normalization is disabled when it is 0.0, and fully normalized when it is 1.0, the default value is 0.75.

[0072] Where TF(q) i When D is very large, it tends to a fixed value so as not to weaken the effect of IDF, so TF tends to 1.0.

[0073] q i Querying the term frequency of text Q

[0074] Where f(q) i ,Q) represents q i The number of times a term appears in the query text Q. k2 represents an adjustment factor. The larger the value of k2, the more important the term frequency in the query text Q. If it is set to 0, the term frequency will be disabled.

[0075] In one specific implementation, it is known that user A's consumption text document D contains 1000 documents (N=1000), and the keyword "Bluetooth headset" appears in 50 documents [n(q i =50], "Bluetooth headset" appears 5 times in the current document D [f(q i [,D)=5], document length |D|=800, average length avgdl=500, parameters k1=1.2, k2=1, b=0.75;

[0076]

[0077] The correlation score between user A's text document and the keyword "Bluetooth headset" is Ascore = 1.3·1.63·1≈2.12.

[0078] Using the same method, sum the scores of all keywords in the query statement to obtain user A's level of interest in the product.

[0079] Obtain all users' interest in the product, set an interest threshold G0, and execute the first push to users who exceed the interest threshold at the first push time T1. After sending the push, record the first push data of user interaction with the push.

[0080] Consumer characteristics obtained from the first push data include: age, gender, region, number of purchases of similar products, number of times the user clicks in and out, scrolling depth of the recommended page, browsing time, number of times the user adds items to the shopping cart, and number of times the user shares the product. A neural network model is built based on these consumer characteristics. The model is then fed these consumer characteristics to predict the user's purchase intention for the product. Each input characteristic is assigned a weight based on its importance.

[0081] The specific calculation method is as follows:

[0082]

[0083] Where W is the feature vector, F is the feature weight, and T is the feature data vector after weighted average;

[0084] Establish a neural network model, initialize the weights to a non-zero random number less than 1, and normalize the weighted average data vector T to obtain the data vector X; it is known that the feedforward multilayer network has n layers, and each network device attribute corresponds to a neuron in the layer;

[0085] Set the output of the i-th neuron in the k-th layer of the neural network to... 1≤k≤n, where,

[0086]

[0087] Where f(·) is the activation function, The variable for the activation function;

[0088]

[0089] W i,n+1 =-θ;

[0090] Among them W i,j Let θ be the weights from neuron j to neuron i, and θ be the threshold.

[0091] Using the sigmoid activation function, we get:

[0092]

[0093] W i,j The formulas for updating weights and thresholds θ are as follows:

[0094] Through backpropagation, let m be the output layer (last layer) of the neural network, and let the error of the i-th neuron in the m-th layer be...

[0095] in, To calculate the output of the i-th neuron in the m-th layer using a neural network, Y i It is the expected value;

[0096] The output layer error term is:

[0097] Where η is the learning rate, η∈(0,1); a is the impulse coefficient, which is used to prevent the addition of an impulse term to local minima, a∈(0,1); p is the number of iterations; and ΔW i,j (p) represents the change in weight between the j-th neuron in the (m-1)-th layer and the i-th neuron in the m-th layer during the p-th iteration;

[0098] Error of the i-th neuron in hidden layer k

[0099]

[0100] Where k represents the layers other than the output layer, and ΔW i,j (p) represents the change in weight between the j-th neuron in layer k-1 and the ith neuron in layer k during the p-th iteration, where l belongs to layer k+1. l,i The weights are the values ​​of the i-th neuron in the k-th layer to the l-th neuron in the (k+1)-th layer. The output is the j-th neuron in the (k-1)-th layer;

[0101] The weight of the Pth iteration

[0102] W i,j (p)=Wi,j (p-1)+ΔW i,j (p).

[0103] Based on the aforementioned neural network model, it is possible to predict users' purchase intentions for marketing products. In this embodiment, the purchase intention levels are 1, 2, 3, 4, and 5. When the propensity index is 1, the purchase intention is the lowest, and when the propensity index is 5, the purchase intention is the highest.

[0104] Retrieve the interaction session window between user B and the first push notification. Set the minimum duration of a valid session to h0, and define session windows whose duration exceeds the threshold h0 as valid windows M, where M = [M1, M2, M3, ... M]. d ], where the z-th valid window is M z M z The window time is [ZF0, ZF1], and the window duration is TM;

[0105] Set the core conversion time period of the push notification to △T1, and the sending time of the first push notification is T1∈△T1;

[0106] Among them, when When the user has almost no interaction with the push notification, the reason is likely that the user's idle time does not match the push notification time. At the second sending time T2, a second push notification is sent to users whose purchase intention index exceeds the threshold W0, where T2∈△T1, T2=T1+λ; where the value of λ is preferably 6h-16h.

[0107] when If the timing of the first push (T1) is appropriate for the user, but no conversion occurs, consider adjusting the push content and sending a push again to users whose purchase intention index exceeds the threshold W0 at the same time as the first push.

[0108] when If a user's interaction with the push notification occurs outside of the core conversion time, we can examine the distribution of their conversation frequency and select times when their conversation frequency is higher to send push notifications again to users whose purchase intention index exceeds the threshold W0.

[0109] For example, after analyzing user purchase data, the system sets the core conversion period as Friday 19:00-Saturday 22:00. Therefore, the first push notification for the product is sent on April 14th (Friday) at 19:00. Based on the feedback data after the push, the system obtains user data for those who did not convert successfully in the first push. Then, based on the purchase intention prediction model, the purchase intention reach index W is obtained. B User B, whose score is 3, receives a second push notification;

[0110] If user B does not have a valid conversation window with the first push, the message will be resent to user B 15 hours later, which is 10:00 AM on Saturday.

[0111] If all valid conversation windows between user B and the first push notification are between 19:00 on Friday and 22:00 on Saturday, it means the timing of the push notification is appropriate for the user. Adjust the push notification content (e.g., emphasize features and offers) and send the second push notification at 19:00 on Saturday, April 15th.

[0112] When the effective conversation window between user B and the push notification occurs outside the core conversion period, such as at 15:00 on Saturday, April 15th and 10:00 on Sunday, it means that part of the user's interaction with the push notification is outside the core conversion period. In this case, check the distribution of the frequency of the conversation and send the push notification again when the frequency of the conversation is higher.

[0113] Set to session window M z The initial associated extension time period is △ZF, which includes r sub-extensions, △ZF = [△ZF1, △ZF2, △ZF3, ... △ZF]. r ], where △ZF1=M z +△ZF0, △ZF0≥0; Set the p-th sub-extension segment as △ZF0. p , △ZF p For M z The feedback weight is μ p ;

[0114] Within the extended time period △ZF, the system triggered a total of s related applications, including k positive feedback related applications FZP = [FZP1, FZP2, FZP3, ... FZP]. k ], and h negative feedback associated applications NZP = [NZP1, NZP2, NZP3, ... NZP h ], where the g-th positive feedback correlation application is set to FZP g FZP g Occurring in the g-th extension segment of △ZF g The u-th negative feedback correlation application is NZP. u NZP u Occurring in the u-th extension segment △ZF u ;

[0115] Obtain the z-th valid window M z Push feedback rate

[0116] Obtain the total feedback score of user B's d valid windows.

[0117] Obtain user B's purchase intention index W B '=δW B .

[0118] For example, in one embodiment, the product push notification arrives at 19:00 on April 14 (Friday), and the effective interaction window between the user and the push notification's session page is April 14 [19:00, 19:10], [19:40, 19:45], [20:10, 20:15], and April 15 [10:00, 10:20]; wherein, within the extended time period [19:00, 19:30] starting from session window 1 [19:00, 19:10], there are 2 triggers. There are positive feedback platform applications [Application 1] and [Application 2], and one negative feedback platform application [Application 3]. The session window 1 is divided into a first sub-extension segment [19:00, 19:15], a second sub-extension segment [19:15, 19:25], and a third sub-extension segment [19:25, 19:30], with corresponding feedback weights of 1, 0.8, and 0.6, respectively. [Application 1] is triggered in the first sub-extension segment, [Application 2] is triggered in the second sub-extension segment, and [Application 3] is triggered in the third sub-extension segment.

[0119] At this point, the push feedback score for session window 1 is obtained:

[0120] Positive feedback platforms typically include those that showcase, introduce, and review products, especially professional review sites and authoritative media outlets, which are beneficial for improving conversion rates. Negative feedback platforms typically include price comparison / discount platforms and e-commerce platforms that sell competing products. These platforms tend to display lower prices for the same product or offer more favorable discounts / coupons, making it easier for users to switch to third-party platforms with lower prices.

[0121] Furthermore, the present invention also provides a computer-readable storage medium comprising a stored application, wherein the application executes the method described in the above-described method embodiments when it runs.

[0122] Furthermore, the present invention also provides an electronic device including a memory and a processor, wherein the memory stores a computer application, and the processor is configured to execute the method described in the above-described method embodiments through the computer application.

[0123] Furthermore, it should be understood that since the various modules are only provided to illustrate the functional units of the device of the present invention, the physical devices corresponding to these modules may be the processor itself, or a part of the processor's software, a part of its hardware, or a combination of software and hardware. Therefore, the number of modules shown in the figures is merely illustrative.

[0124] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer application instructing related hardware. The computer application can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer application includes computer application code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include any entity or device capable of carrying the computer application code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0125] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer applications. The internal memory provides the environment for the operation of the operating system and computer applications stored in the non-volatile storage media. The communication interface of the computer device is implemented using a network, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0126] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A marketing content recommendation method based on data analysis, characterized in that, Includes the following steps: S1, acquire user consumption data, which includes user information, consumption records, and behavior records; form a separate text document for each user's consumption data, and standardize and segment the text document; S2, Based on the product description and product functions, generate a list of keywords for product marketing, and combine the keywords in the list into query text; S3, obtain the correlation between the text document and the query text to get the user's interest in the product; S4, set the interest threshold, and perform the first push to users who exceed the interest threshold at the first push time to obtain the first push data; S5: Based on consumption data and first push data, obtain consumption characteristics, establish a network neural model, evaluate the purchase intention of users who failed to convert in the first push, and obtain the user's purchase intention index. S6: Adjust the user's purchase intention index based on the first push data, and execute the second push for users whose purchase intention index has reached the index threshold.

2. The marketing content recommendation method based on data analysis as described in claim 1, characterized in that, In step S1, the method for obtaining behavior records is as follows: embed gtag.js or GA4 measurement protocol tracking code globally in the website / APP, and send a page_view event containing URL, timestamp, and device information when each page is loaded.

3. The marketing content recommendation method based on data analysis as described in claim 2, characterized in that, The user behaviors include browsing search pages, browsing history, in-site conversations, click paths, search keywords, and adding items to the shopping cart.

4. The marketing content recommendation method based on data analysis as described in claim 1, characterized in that, In step S3, a score is used to measure the relevance between the query text Q and the text document D containing user A's consumption data. Among them, IDF(q) i ) represents the keyword q i Inverse document frequency, TF(q) i D) represents the keyword q i In the term frequency of the text document (D), TF(q) i ,Q) represents the keyword q i The term frequency of the query text (Q); Obtain the inverse document frequency Where n represents the keyword q i The number of documents; N represents the total number of documents; This means it does not contain q. i The number of documents and the number of documents containing q i The ratio of the number of documents; q i In document D, the word frequency Where, f(q) i D) represents q i The number of times a term appears in user A's text document D, |D| represents the length of text document D, avgdl represents the average length of the document set, k1 represents the rate of increase of the term frequency result in term frequency saturation, and b represents the role of document length normalization. q i Query the term frequency of text Q Where f(q) i ,Q) represents q i k2 represents the frequency importance of a term in the query text Q, where k2 is the number of times it appears in the query text Q.

5. The marketing content recommendation method based on data analysis as described in claim 4, characterized in that, In step S5, a neural network model is established, consumption features are input into the neural network model, weights are assigned to each feature according to its importance, and the user's purchase intention index for the product is output. The specific calculation method is as follows: Where W is the feature vector, F is the feature weight, and T is the feature data vector after weighted average; Establish a neural network model, initialize the weights to a non-zero random number less than 1, and normalize the weighted average data vector T to obtain the data vector X; it is known that the feedforward multilayer network has n layers, and each network device attribute corresponds to a neuron in the layer; Set the output of the i-th neuron in the k-th layer of the neural network to... 1≤k≤n, where, Where f(·) is the activation function, The variable for the activation function; W i,n+1 =-θ; Among them W i,j Let θ be the weights from neuron j to neuron i, and θ be the threshold. Using the sigmoid activation function, we get: W i,j The formulas for updating weights and thresholds θ are as follows: Through backpropagation, let m be the output layer (last layer) of the neural network, and let the error of the i-th neuron in the m-th layer be... in, To calculate the output of the i-th neuron in the m-th layer using a neural network, Y i It is the expected value; The output layer error term is: Where η is the learning rate, η∈(0,1); a is the impulse coefficient, which is used to prevent the addition of an impulse term to local minima, a∈(0,1); p is the number of iterations; and ΔW i,j (p) represents the change in weight between the j-th neuron in the (m-1)-th layer and the i-th neuron in the m-th layer during the p-th iteration; Error of the i-th neuron in hidden layer k Where k represents the layers other than the output layer, and ΔW i,j (p) represents the change in weight between the j-th neuron in layer k-1 and the ith neuron in layer k during the p-th iteration, where l belongs to layer k+1. l,i The weights are the values ​​of the i-th neuron in the k-th layer to the l-th neuron in the (k+1)-th layer. The output is the j-th neuron in the (k-1)-th layer; The weight W in the Pth iteration i,j (p)=W i,j (p-1)+ΔW i,j (p).

6. The marketing content recommendation method based on data analysis as described in claim 5, characterized in that, The input consumption characteristics include: age, gender, region, number of purchases of similar products, number of times you clicked in, number of times you clicked out, scroll depth of the recommended page, browsing time, number of times you added items to your cart, and number of times you shared the product.

7. The marketing content recommendation method based on data analysis as described in claim 5, characterized in that, In step S6, the method for sending the second push is as follows: Retrieve the user's session window from the first push data. Set a valid session duration threshold of h0, and select session windows with durations exceeding the threshold h0 as valid windows M, where M = [M1, M2, M3, ... M]. d ], where the z-th valid window is M z M z The window time is [ZF0, ZF1], and the window duration is TM; Set the core conversion time period of the push notification to △T1, and the sending time of the first push notification is T1∈△T1; Among them, when At the second sending time T2, a second push is sent again to users whose purchase intention index exceeds the index threshold, where T2∈△T1, T2=T1+λ; when At the same time as the first push notification is sent, a second push notification is sent to users whose purchase intention index exceeds the index threshold. when At that time, a second push notification is sent to users whose purchase intention index exceeds the index threshold at times when their conversation frequency is high.

8. The marketing content recommendation method based on data analysis as described in claim 7, characterized in that, In step S6, the method for adjusting the user's purchase intention index includes: Set to session window M z The initial associated extension time period is △ZF, which includes r sub-extensions, △ZF = [△ZF1, △ZF2, △ZF3, ... △ZF]. r ], where △ZF1=M z +△ZF0, △ZF0≥0; Set the p-th sub-extension segment as △ZF0. p , △ZF p For M z The feedback weight is μ p ; Within the extended time period △ZF, the system triggered a total of s related applications, including k positive feedback related applications FZP = [FZP1, FZP2, FZP3, ... FZP]. k ], and h negative feedback associated applications NZP = [NZP1, NZP2, NZP3, ... NZP h ], where the g-th positive feedback correlation application is set to FZP g FZP g Occurring in the g-th extension segment of △ZF g The u-th negative feedback correlation application is NZP. u NZP u Occurring in the u-th extension segment △ZF u ; Obtain the z-th valid window M z Push feedback Obtain the total feedback score of the user's d valid windows Given that the user's purchase intention index is W B At this point, the user's purchase intention index is adjusted to W. B '=δW B .

Images