Multi-terminal message reminding method based on user habit analysis

Abstract

The application discloses a multi-terminal message reminding method based on user habit analysis, which can improve user experience, increase user effective operation, provide guidance for potential intention and has high identification accuracy. The method comprises the following steps: step 1) obtaining first user habit data; step 2) comparing the first user habit data with pre-stored standard user habit data; step 3) controlling the general processor to extract normal user habit data from the first user habit data when the similarity value is less than a first preset threshold; step 4) comparing the first user current operation data with the normal user habit data, and obtaining user browsing operation of the next step according to the comparison result; and step 5) pushing the next step user operation to one or more terminals of the first user in a message reminding mode.

Description

Technical Field

[0001] This invention relates to the field of big data intelligent analysis and application technology, and in particular to a multi-terminal message notification method based on user habit analysis. Background Technology

[0002] Currently, the vast majority of content publishing adopts a web broadcasting approach, meaning the published content is identical to all audiences. Some content publishing systems also employ a simple matching and filtering approach based on user characteristics (age, gender, region), which we'll call narrowcasting. Narrowcasting can filter information more relevant to users based on their features, fostering a stronger sense of belonging and loyalty among users compared to broadcasting, and increasing business traffic. Existing technology, such as Chinese invention patent CN107357833B, discloses a data processing method and related products. In this method, a mobile terminal includes a general-purpose processor that performs the following steps: acquiring first user habit data; comparing the first user habit data with pre-stored standard user habit data to obtain a similarity value; when the similarity value is less than a first preset threshold, extracting normal user habit data from the first user habit data; and updating the standard user habit data based on the normal user habit data. The embodiments of the present invention can acquire user habit data and update the user habit data when the user habit data is outside the prediction range. In this way, the user habit data can be updated in a timely manner, thereby improving the intelligence of the mobile terminal. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-terminal message notification method based on user habit analysis that improves user experience, increases effective user operations, guides potential intentions, and has high recognition accuracy, in contrast to existing technologies. This method includes the following steps:

[0004] Step 1) Obtain the first user habit data;

[0005] Step 2) Compare the first user habit data with the pre-stored standard user habit data;

[0006] Step 3) Control the general processor to extract normal user habit data from the first user habit data when the similarity value is less than the first preset threshold;

[0007] Step 4) Compare the current operation data of the first user with the normal user habit data, and obtain the next browsing operation of the user whose current operation data is matched based on the comparison result;

[0008] Step 5) Push the next user action to one or more terminals of the first user via message notification. By acquiring user actions and extracting normal user habit data, and comparing it with standard user habit data, the system predicts the user's next action. When the user habit data is outside the prediction range, the standard user habit data is updated. In this way, user habit data can be updated in a timely manner, improving the level of intelligence.

[0009] To further optimize the technical solution, the following optimization measures were also taken:

[0010] Normal user habit data and primary user habit data contain operation sequence information. Normal user data is extracted from the primary user habit data (i.e., the current user's operation habits) for prediction and comparison.

[0011] The operation sequence information includes: starting a session, searching for products, viewing details, logging in, starting registration, checking in, adding to an order, adding to favorites, and the timing of these actions. This operation sequence information is a library used to predict user actions, exhaustively enumerating them through preset methods. Since app and website operations can be pre-defined and developed by developers, this enumeration is effective.

[0012] Pre-stored standard user habit data is obtained through the following training:

[0013] A prediction network is constructed based on all samples contained in the training set, where node i in the network corresponds to sample x in the training set. i A directed edge e from node i to node j will be established if the nodes correspond to each other and have the same behavioral information type; and if the nodes have the same behavioral information as their neighbors. ij :

[0014] 1) Node j is one of the k neighbors of node i determined by the KNN algorithm;

[0015] 2) Node i and node j have the same behavioral information.

[0016] Using existing user habit data as the training set, and after establishing directed edges between all nodes in the prediction network according to the above two conditions, the entire prediction network Q will be composed of subnetwork q. c Composed of, and 1≤c≤C, and subnetwork q c Each node in the network has the same type of behavioral information, and there is no connection between any two subnetworks. Clearly, the prediction network Q built based on these two conditions can powerfully represent the topological relationships of all training samples, and further, the similarity features represented by nodes can be mathematically expressed in the constructed prediction network.

[0017] The notification methods include: a response area appearing near the edge of the terminal's display device; the response area displays a scaled-down thumbnail of the push notification operation page. The page thumbnail also includes links, allowing users to easily navigate to the corresponding webpage when clicking the thumbnail, improving ease of use and capturing potential new customers or potential business from existing users. For example, if an unregistered user opens a product details page, they will typically look for the registration page. In this case, directly pushing the registration page to the browser and / or mobile device can increase the number of new user registrations.

[0018] Standard user habit data is generated by the terminal based on the user's usage records for at least one application over a period of time and / or updated based on normal user habit data. This method can send push notifications to one or more apps as needed.

[0019] The present invention also discloses: a computer program for implementing the above-described multi-terminal message notification method based on user habit analysis, and a storage medium storing the above-described computer program.

[0020] This invention employs the following steps: 1) acquiring first user habit data; 2) comparing the first user habit data with pre-stored standard user habit data; 3) controlling a general-purpose processor to extract normal user habit data from the first user habit data when the similarity value is less than a first preset threshold; 4) comparing the first user's current operation data with normal user habit data, and obtaining the next user browsing operation that matches the current operation data based on the comparison result; and 5) pushing the next user operation to one or more terminals of the first user as a message reminder. By acquiring user operations, extracting normal user habit data, comparing it with standard user habit data, predicting the user's next operation, and updating the standard user habit data when the user habit data is outside the prediction range, this invention can update user habit data in a timely manner, improving the level of intelligence. Therefore, this invention has the advantages of improving user experience, increasing effective user operations, providing guidance for potential intentions, and high recognition accuracy. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the sequence of steps in an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the predictive recommendation effect of an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram comparing the recognition effects of an embodiment of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying embodiments.

[0025] Example:

[0026] Reference Figures 1 to 3 A multi-terminal message notification method based on user habit analysis includes the following steps:

[0027] Step 1) Obtain the first user habit data;

[0028] Step 2) Compare the first user habit data with the pre-stored standard user habit data;

[0029] Step 3) Control the general processor to extract normal user habit data from the first user habit data when the similarity value is less than the first preset threshold;

[0030] Step 4) Compare the current operation data of the first user with the normal user habit data, and obtain the next browsing operation of the user whose current operation data is matched based on the comparison result;

[0031] Step 5) Push the next user action to one or more terminals of the first user via message notification. By acquiring user actions and extracting normal user habit data, and comparing it with standard user habit data, the system predicts the user's next action. When the user habit data is outside the prediction range, the standard user habit data is updated. In this way, user habit data can be updated in a timely manner, improving the level of intelligence.

[0032] To further optimize the technical solution, the following optimization measures were also taken:

[0033] Normal user habit data and primary user habit data contain operation sequence information. Normal user data is extracted from the primary user habit data (i.e., the current user's operation habits) for prediction and comparison.

[0034] The operation sequence information includes: starting a session, searching for products, viewing details, logging in, starting registration, checking in, adding to an order, adding to favorites, and the timing of these actions. This operation sequence information is a library used to predict user actions, exhaustively enumerating them through preset methods. Since app and website operations can be pre-defined and developed by developers, this enumeration is effective.

[0035] Pre-stored standard user habit data is obtained through the following training:

[0036] A prediction network is constructed based on all samples contained in the training set, where node i in the network corresponds to sample x in the training set. i A directed edge e from node i to node j will be established if the nodes correspond to each other and have the same behavioral information type; and if the nodes have the same behavioral information as their neighbors.ij :

[0037] 1) Node j is one of the k neighbors of node i determined by the KNN algorithm;

[0038] 2) Node i and node j have the same behavioral information.

[0039] Using existing user habit data as the training set, and after establishing directed edges between all nodes in the prediction network according to the above two conditions, the entire prediction network Q will be composed of subnetwork q. c Composed of, and 1≤c≤C, and subnetwork q c Each node in the network has the same type of behavioral information, and there is no connection between any two subnetworks. Obviously, the prediction network Q built based on the above two conditions can effectively show the topological relationship of all training samples, and further, the similarity features represented by the nodes can be mathematically expressed in the constructed prediction network. The KNN algorithm is referenced from Keller JM, Gray MR, Givens J AA fuzzy K-nearest neighbor algorithm[J].IEEE Trans. Syst., Man, Cybern., 1985, SMC-15(4):580-585. Due to space limitations, it will not be elaborated further.

[0040] The notification methods include: a response area appearing near the edge of the terminal's display device; the response area displays a scaled-down thumbnail of the push notification operation page. The page thumbnail also includes links, allowing users to easily navigate to the corresponding webpage when clicking the thumbnail, improving ease of use and capturing potential new customers or potential business from existing users. For example, if an unregistered user opens a product details page, they will typically look for the registration page. In this case, directly pushing the registration page to the browser and / or mobile device can increase the number of new user registrations.

[0041] Standard user habit data is generated by the terminal based on the user's usage records for at least one application over a period of time and / or updated based on normal user habit data. This method can send push notifications to one or more apps as needed.

[0042] The pseudocode is as follows:

[0043] Input: Given a dataset (i.e., a collection of standard user habit data) X = [x1, x2, ..., x N ] T , where x∈R d d is the feature dimension, and Y = [y1, y2, ..., y] is the set of behavioral information corresponding to a given dataset X. N ]T The test set (i.e., the set of normal user habit data) T = [t1, t2, ..., t M ] T , where t M ∈R d d is the dimension of the sample features, y i ∈{1,2,,C}; Parameters k1, truncation distance dc, width parameter σ, parameter k2, damping coefficient λ, maximum number of iterations H, and threshold θ when the iteration termination condition is met during the process of constructing the prediction network using the KNN algorithm.

[0044] Output: A set of predicted behavior information for the test set.

[0045] Training phase

[0046] Step 1: Use the KNN algorithm to construct a prediction network Q based on the given training set X and the corresponding behavioral information set Y.

[0047] Step 2: Calculate the concentration ρ of each node in the prediction network Q using equation (7). j .

[0048] Step 3: Calculate the weight ε of the i-th node in the subnetwork q using equations (6) and (5) respectively. c i and the weights ε of the subnetwork q c .

[0049] Step 4: Set h = 0;

[0050] Step 5: Repeat until the iteration termination condition is met. Or it satisfies h>H.

[0051] Step 5.1: h = h + 1;

[0052] Step 5.2: Repeat until j = N;

[0053] Step 5.2.1: Update the node influence factor using equation (1)

[0054] Prediction phase

[0055] Step 6: Set m = 1 and loop the program until m > M.

[0056] Step 6.1: Using equation (2), implement the similarity hypothesis through the KNN algorithm to determine the test sample t in the test set T. m The nearest neighbor set v i ;

[0057] Step 6.2: Calculate the maximum probability c using equations (5) and (4). * Then based on c * The test sample t will be determined. m Behavioral information type and t m It should be classified into class c;

[0058] Step 6.3: m = m + 1

[0059] Step 7: Output the predicted behavior information set [y1, y2, ..., y] for the test set T. M ]

[0060] Step 8: Convert the predicted behavior information set into push messages and then push them to users.

[0061] Finish

[0062] Predictive behavior information sets generally consist of various operations within the sequence of actions, which are then manually preset for corresponding push notifications. For example, if the predicted behavior information set recommends pushing "Add to Order," then the manually preset "Add to Order" message should correspond to displaying a thumbnail and prompt of the "Add to Order" page near the edge of the screen, and mapping the link information to the thumbnail. When the user clicks the thumbnail, the corresponding linked page opens. Simultaneously, for other devices of the same user, such as mobile devices, the link is also synchronously pushed via SMS or thumbnail display. Figure 3 Compared with the accuracy of push prediction and recognition in existing background technologies, this embodiment shows that the recognition effect is better than that of existing technologies due to the KNN adjustment algorithm used in this embodiment, which can effectively improve the user's application experience.

[0063] In i h The influence factor of node i during the h-th iteration is represented.

[0064]

[0065] v t =KNN(t), Equation (2)

[0066] In equation (1), when h = 0, the node concentration is set as the node's influence factor, i.e. The influence factor of node i during the h-th iteration is represented by de. i This represents the out-degree of node i, i.e., the directed edge e. ij The number.

[0067] In equation (2), v t v represents the set of k neighbor nodes of t. tNodes in the dataset may come from different subnetworks qc, where 1 ≤ c ≤ C. To implement the style similarity hypothesis from a data perspective, we will apply the maximum probability c of t for the c-th category (i.e., subnetwork qc). * =arg max c Ψ c Equation (3);

[0068] Ψ c This represents the probability that t belongs to the c-th category from a style perspective.

[0069] In the prediction network Q, Ψ is defined by the weights of the subnetworks and the influence factors of the nodes. c ,

[0070]

[0071] In equation (4), Represents set v t The i-th neighbor node of node t, and Its impact factor.

[0072] Define subnetwork q c The weights are as follows:

[0073]

[0074] Subnetwork q c The weight of a subnetwork q is the average of the weights of all its constituent nodes. c The weight of the i-th node in the array is:

[0075]

[0076] e ij Representative subnetwork q c The directed edges from node i to node j, N i Represents a directed edge e ij The number of nodes. The local concentrations of nodes are not equal, and calculating the influence factor of a node using 1 / N iterations does not reflect the actual distribution of nodes. Here, the concentration of node j (corresponding to the j-th training sample in the training set) is defined as follows:

[0077]

[0078] N represents the total number of nodes in the prediction network Q, d jl Let dc represent the Euclidean distance between node j and node l, dc represent the cutoff distance, and χ(...) represent the distance function. When x(...) = 1, otherwise x(...) = 0.

[0079] Although the invention has been described in conjunction with preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make various changes, substitutions and modifications to the subject matter listed herein without departing from the spirit and scope of the invention. Therefore, the scope of protection of the invention shall be determined by the scope defined in the claims.

Claims

1. A multi-terminal message notification method based on user habit analysis, characterized by: Includes the following steps, Step 1) Obtain the first user habit data; Step 2) Compare the first user habit data with the pre-stored standard user habit data; Step 3) Control the general processor to extract normal user habit data from the first user habit data when the similarity value is less than the first preset threshold; Step 4) Compare the current operation data of the first user with the normal user habit data, and obtain the next browsing operation of the user that hits the current operation data based on the comparison result. Step 5) Push the next user operation to one or more terminals of the first user as a message notification; The standard user habit data is trained by the terminal based on the user's usage records for at least one application over a period of time and / or updated based on normal user habit data, specifically including: Input: Given a dataset, i.e., a set of standard user habit data, X = [x1, x2, ..., x...]. N ] T Where x1, x2, ..., x N ∈R d d represents the feature dimension of the sample, and Y = [y1, y2, ..., y] represents the set of behavioral information corresponding to the given dataset X. N ]T, the test set is the set of normal user habit data T = [t1, t2, ..., t M ] T , where t1, t2, ..., t M ∈R d d is the dimension of the sample features, y i ∈{1,2,…,C}; Parameters k1 and cutoff distance d are used in the process of constructing the prediction network using the KNN algorithm. c Width parameter σ, parameter k2 in the KNN algorithm when implementing the physical similarity assumption, damping coefficient λ, maximum number of iterations H, and threshold θ when the iteration termination condition is met; Output: A set of predicted behavior information for the test set; Training phase: Step 1: Use the KNN algorithm to construct a prediction network Q based on the given training set X and the corresponding behavioral information set Y; Step 2: Equation (7), where N represents the total number of nodes in the prediction network Q; This represents the Euclidean distance between node j and node l; Represents the cutoff distance; Represents the distance function, when When x(...) = 1, otherwise x(...) = 0; use equation (7) to calculate the concentration ρ of each node in the prediction network Q. j ; Step 3: Equation (5), Equation (6), where, Subnetwork The weights; Subnetwork The weight of the i-th node in the equation; e ij Representative subnetwork q c The directed edges from node i to node j; calculate the subnetwork q using equations (6) and (5) respectively. c The weight of the i-th node and subnetwork q c weight ε c ; Step 4: Set h = 0; Step 5: Repeat until the iteration termination condition is met. Or satisfying h>H; Step 5.1: h = h + 1; Step 5.2: Repeat until j = N; Step 5.2.1: Equation (1), where, The influence factor of node j in the (h+1)th iteration; The influence factor of node i in the h-th iteration process is represented; when h = 0, the concentration of the node is set as the influence factor of the node, that is, the influence factor of node i in the h-th iteration process. This represents the out-degree of node i, i.e., the directed edge. The number of nodes; update the node influence factor using equation (1); Prediction phase: Step 6: Set m = 1, and loop the program until m > M; Step 6.1: Equation (2), where, Represents the set of k neighbor nodes of t. The nodes in the network may come from different subnetworks q c , 1≤c≤C; Using equation (2), the similarity assumption is implemented through the KNN algorithm to determine the test sample t in the test set T. m The nearest neighbor set v i ; Step 6.2: Equation (3), Equation (4), where, This represents the probability that t belongs to the c-th category from a style perspective; Representative set The i-th neighbor node of node t, and Its influence factor; calculate the maximum probability c using equations (3), (4), and (5). * Then based on c * The test sample t will be determined. m Behavioral information type and t m It should be classified into class c; Step 6.3: m = m + 1; Step 7: Output the predicted behavior information set [y1, y2, ..., y] for the test set T. M Step 8: Convert the predicted behavior information set into push messages and then push them to the user; The pre-stored standard user habit data was obtained through the following training: A prediction network is constructed based on all samples contained in the training set, where node i in the network corresponds to sample x in the training set. i A directed edge e from node i to node j will be established if the nodes correspond to each other and have the same behavioral information type; and if the nodes have the same behavioral information as their neighbors. ij : 1) Node j is one of the k neighbors of node i determined by the KNN algorithm; 2) Node i and node j have the same behavioral information; Using existing user habit data as the training set, and after establishing directed edges between all nodes in the prediction network according to the above two conditions, the entire prediction network Q will be composed of subnetwork q. c Composed of, and 1≤c≤C, and subnetwork q c Each node in the network has the same type of behavioral information, and there is no connection between any two subnetworks.

2. The multi-terminal message notification method based on user habit analysis as described in claim 1, characterized in that: The normal user habit data and the first user habit data contain operation sequence information.

3. The multi-terminal message notification method based on user habit analysis as described in claim 2, characterized in that: The operation sequence information includes: starting a session, searching for products, viewing details, logging in, starting registration, checking in, adding to an order, adding to favorites, and the timing information of the above actions.

4. The multi-terminal message notification method based on user habit analysis as described in claim 1, characterized in that: The message notification method includes: a response area appearing near the edge of the terminal's display device; the response area displays a scaled-down thumbnail of the push operation page.

5. The multi-terminal message notification method based on user habit analysis as described in claim 1, characterized in that: The standard user habit data is trained by the terminal based on the user's usage records for at least one application over a period of time and / or updated based on the normal user habit data.

6. A storage medium storing a computer program that implements the multi-terminal message notification method based on user habit analysis as described in claim 1.

Images