Single-user cross-domain behavior modeling method based on machine learning algorithm
Through a single user cross-domain behavior modeling method based on machine learning algorithms, the problems of cross-domain behavior integration and personalized service in the existing technology are solved, and accurate prediction and personalized recommendation of user behavior patterns are realized, providing a more comprehensive Service experience.
Patent Information
- Application Number
- CN202411080374.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2024-12-06
AI Technical Summary
Existing Internet applications lack cross-domain integration in user behavior modeling and prediction, cannot automatically build personalized scenarios, and rely on manual definition of user types, making it difficult to achieve truly personalized services.
Using a single user cross-domain behavior modeling method based on machine learning algorithms, by collecting and preprocessing user cross-domain behavior data, building features that reflect user behavior patterns and environmental characteristics, selecting appropriate machine learning algorithm training models, and making real-time predictions and personalized recommendations, regularly updated models to adapt to changes in user behavior.
It realizes accurate prediction of single user behavior patterns and preferences, provides personalized services, integrates behavioral data from multiple applications, captures user behavioral connections and transformation patterns, provides more comprehensive services, and truly realizes a personalized experience.
Smart Images

Figure CN119089037A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of behavior prediction technology, and more specifically, to a single user cross-domain behavior modeling method based on a machine learning algorithm. Background Art
[0002] At present, all Internet applications will model, predict and recommend user behaviors for their own users. Typical examples include content applications such as Douyin and bilibili, which record and model user behaviors such as likes, comments, forwarding, collections, and rewards to predict user preferences for a certain type of content, thereby recommending more user-favorite content and improving user stickiness. Another example is e-commerce applications such as Taobao and JD.com, which record and model user browsing, adding to cart, and ordering behaviors to predict user demand for a certain type of product, thereby recommending more products that users need and promoting sales.
[0003] In the above application directions, various Internet applications have achieved good results. However, since each application only records and predicts the behavior of its own users in its own single field, it does not conduct cross-domain modeling and prediction of the user's behavior on devices that integrate multiple applications. In addition, due to the limitations of business scope and access rights to hardware, these applications do not incorporate environmental information as a condition in the modeling process, and cannot automatically build personalized scenarios. In essence, modeling and prediction are based on the big data of massive users. For a single user, his or her behavior is matched to a certain user type after big data modeling, and then the behavior prediction for this type of user is applied to the single user. This type of solution relies on manual definition of user types, so the number of types is very limited, and it is difficult to achieve true personalization and "one thousand faces for one thousand people" for a single user. Summary of the invention
[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a single user cross-domain behavior modeling method based on a machine learning algorithm.
[0005] To achieve the above purpose, the present invention provides the following technical solution: a single user cross-domain behavior modeling method based on a machine learning algorithm, the specific method steps are as follows:
[0006] Step 1: Data collection and preprocessing: collect cross-domain behavior data of a single user, and perform preprocessing operations such as cleaning, deduplication, and formatting on the collected data;
[0007] Step 2: Feature construction: Based on the user's behavior records, construct features that reflect the user's behavior patterns and behavior preferences. Combined with the collected environmental information, construct features that reflect the conditions under which the user's behavior occurs, and introduce sentiment analysis features into the model.
[0008] Step 3: Model selection and training: According to the specific application scenarios and requirements, select the appropriate machine learning algorithm to build the user behavior model, and use the preprocessed data and constructed features to train the machine learning model;
[0009] Step 4: Model evaluation and optimization: evaluate the performance of the model through cross-validation or error analysis methods, and optimize the model based on the evaluation results;
[0010] Step 5: Real-time behavior prediction and personalized recommendation: When a user takes action in a new field or scenario, the trained model is used to predict the user's behavior in real time. Based on the user's behavior prediction results, personalized recommendation content is generated for the user.
[0011] Step 6: Model update and maintenance: Update the model regularly based on the passage of time and changes in user behavior, and continuously monitor and maintain the model.
[0012] As a further improvement of the technical solution of the present invention, the data collected in step one comes from the user's behavior records and all operations in various Internet applications, including those actively triggered by the user, as well as the user's confirmation actions after the traditional scenario mode actively recommends, and at the same time collects environmental information related to the user's behavior, including time, location and device status.
[0013] As a further improvement of the technical solution of the present invention, the user's behavior record in step 2 includes the user's browsing preferences, purchase frequency and interaction habits, the collected environmental information includes time characteristics and location characteristics, and the introduced sentiment analysis features conduct a comprehensive analysis of user comments, sharing, facial expression detection results, voice command emotion prediction and other data to understand the user's emotional changes and demand changes, and input this information into the model as important features.
[0014] As a further improvement of the technical solution of the present invention, the machine learning algorithm in step three constructs the user behavior model, including using a sequence model to process the user's time series data, or using a decision tree or neural network algorithm for classification or regression prediction.
[0015] As a further improvement of the technical solution of the present invention, the personalized recommendation content in step five includes products, services and information that the user may be interested in, or activities or suggestions that are consistent with the user's behavior patterns and preferences.
[0016] As a further improvement of the technical solution of the present invention, the model update and maintenance in step six includes regularly retraining the model or updating the model by online learning, and the model maintenance includes checking the performance indicators of the model, processing outliers and noise data, and optimizing the computational efficiency of the model.
[0017] As a further improvement of the technical solution of the present invention, the sequence model is used to process the continuous behavior record of the user by time series data, and a recurrent neural network (sequence model (including but not limited to RNN, transformer)) sequence model is used for processing, and the formula design of the sequence model (including but not limited to RNN, transformer) includes an input layer, a hidden layer and an output layer;
[0018] Input layer: input time series data (x_1, x_2, …, x_t) into the network;
[0019] Hidden layer: At each time step (t), the hidden state (h_t) is calculated from the current input (x_t) and the hidden state of the previous time step (h_{t-1}), using the formula: (h_t=f(W_x x_t+W_h h_{t-1}+b)), where (W_x) and (W_h) are weight matrices, (b) is the bias term, and (f) is the activation function (such as tanh or ReLU);
[0020] Output layer: Generates output (y_t) based on hidden state (h_t). For classification tasks, use softmax function to convert hidden state into probability distribution. For regression tasks, you can directly use linear output layer.
[0021] As a further improvement of the technical solution of the present invention, the neural network algorithm performs classification or regression prediction. In the neural network, the output of each neuron is the weighted sum of the input plus the bias term, which is then converted through an activation function. For modeling of a single user's cross-domain behavior, the input is the user's historical behavior characteristics and environmental characteristics, and the output is the predicted user behavior or related classification;
[0022] The formula is: (z=\sum_{i=1}^{n}w_i x_i+b), where (z) is the weighted input, (w_i) is the weight of the (i)th input feature (x_i), (b) is the bias term, and after passing through the activation function (f), the output of the neuron is (a=f(z)).
[0023] Beneficial effects of the present invention:
[0024] 1. By collecting cross-domain behavioral data of a single user and combining it with environmental characteristics, and conducting comprehensive analysis of the user's comments, sharing, facial expression detection results, voice command emotion prediction and other data, we can understand the user's emotional changes and demand changes. This information can be input into the model as an important feature to improve the model's predictive ability. Using machine learning algorithms, we can build a model that reflects the behavior patterns and preferences of a single user. The model can more accurately predict the behavior of a single user in different scenarios or fields in the future, thereby providing more personalized services.
[0025] 2. The cross-domain method designed can integrate behavioral data from multiple sources, capture the behavioral connections and transformation patterns of users across various domains, and reflect all cross-domain behaviors of users on a single device, which helps to gain insight into the overall behavioral dynamics of users and provide more comprehensive services;
[0026] 3. The present invention is based on the data of a single user rather than big data, truly realizing personalization and “one thousand faces for one thousand people”, making full use of the rich sensors of smart devices, integrating environmental information as a condition, and reflecting the interaction between users and the environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION
[0028] 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.
[0029] As attached Figure 1 The single user cross-domain behavior modeling method based on machine learning algorithm is shown, and the specific steps are as follows:
[0030] Step 1: Data collection and preprocessing: collect cross-domain behavior data of a single user, and perform preprocessing operations such as cleaning, deduplication, and formatting on the collected data;
[0031] Step 2: Feature construction: Based on the user's behavior records, construct features that reflect the user's behavior patterns and behavior preferences. Combined with the collected environmental information, construct features that reflect the conditions under which the user's behavior occurs, and introduce sentiment analysis features into the model.
[0032] Step 3: Model selection and training: According to the specific application scenarios and requirements, select the appropriate machine learning algorithm to build the user behavior model, and use the preprocessed data and constructed features to train the machine learning model;
[0033] Step 4: Model evaluation and optimization: evaluate the performance of the model through cross-validation or error analysis methods, and optimize the model based on the evaluation results;
[0034] Step 5: Real-time behavior prediction and personalized recommendation: When a user takes action in a new field or scenario, the trained model is used to predict the user's behavior in real time. Based on the user's behavior prediction results, personalized recommendation content is generated for the user.
[0035] Step 6: Model update and maintenance: Update the model regularly based on the passage of time and changes in user behavior, and continuously monitor and maintain the model.
[0036] The data collected in step one comes from the user's behavior records and all operations in various Internet applications, including those actively triggered by the user and the user's confirmation actions after the traditional scenario mode actively recommends. At the same time, environmental information related to user behavior is collected, including time, location and device status.
[0037] Among them, the user's behavior records in step 2 include the user's browsing preferences, purchase frequency and interaction habits. The collected environmental information includes time characteristics and location characteristics. The introduced sentiment analysis features conduct a comprehensive analysis of user comments, sharing, facial expression detection results, voice command emotion prediction and other data to understand the user's emotional changes and demand changes, and input this information into the model as important features.
[0038] Among them, in step three, the machine learning algorithm builds the user behavior model, including using a sequence model to process the user's time series data, or using a decision tree or neural network algorithm for classification or regression prediction.
[0039] Among them, the personalized recommendation content in step five includes products, services and information that the user may be interested in, or activities or suggestions that are in line with the user's behavior patterns and preferences.
[0040] Among them, step six, model update and maintenance, includes regular retraining of the model or updating the model through online learning. Model maintenance includes checking the performance indicators of the model, processing outliers and noise data, and optimizing the computational efficiency of the model.
[0041] Among them, the sequence model is used to process the continuous behavior records of users in time series data, and a recurrent neural network (sequence model (including but not limited to RNN, transformer)) sequence model is used for processing. The formula design of the sequence model (including but not limited to RNN, transformer) includes an input layer, a hidden layer and an output layer;
[0042] Input layer: input time series data (x_1, x_2, …, x_t) into the network;
[0043] Hidden layer: At each time step (t), the hidden state (h_t) is calculated from the current input (x_t) and the hidden state of the previous time step (h_{t-1}), using the formula: (h_t=f(W_x x_t+W_h h_{t-1}+b)), where (W_x) and (W_h) are weight matrices, (b) is the bias term, and (f) is the activation function (such as tanh or ReLU);
[0044] Output layer: Generates output (y_t) based on hidden state (h_t). For classification tasks, use softmax function to convert hidden state into probability distribution. For regression tasks, you can directly use linear output layer.
[0045] Among them, the neural network algorithm performs classification or regression prediction. In the neural network, the output of each neuron is the weighted sum of the input plus the bias term, and then converted through the activation function. For the cross-domain behavior modeling of a single user, the input is the user's historical behavior characteristics and environmental characteristics, and the output is the predicted user behavior or related classification;
[0046] The formula is: (z=\sum_{i=1}^{n}w_ix_i+b), where (z) is the weighted input, (w_i) is the weight of the (i)th input feature (x_i), (b) is the bias term, and after passing through the activation function (f), the output of the neuron is (a=f(z)).
[0047] Smart devices such as smart cars and smart phones have rich sensors to obtain environmental information. Their manufacturers' overall control over the software and hardware systems enables them to simultaneously obtain rich environmental information (such as time, geographic location, temperature and humidity, vehicle speed, power, etc.) and cross-domain user behavior information (such as the behavior of users using different applications in a certain period of time) while meeting current regulatory requirements. This enables them (smart device manufacturers) to have the basis for automatically building personalized scenarios for a single user. The present invention automatically implements fully personalized scenario construction and operation recommendations for the cross-domain behavior of a single user on a single device. At the same time, based on a unified account system, the present invention can also be expanded to the use of a single user on multiple types of devices or multiple devices of the same type. The basic process remains consistent. The following takes the in-vehicle scenario as an example to illustrate the specific solution:
[0048] 1. Recording (data collection and preprocessing). The substantive content of the records can be divided into two categories: behavior and conditions. Both types of substantive content have a common key value - time;
[0049] Behavior: refers to all operations that users actively perform on the vehicle, including but not limited to direct and active operation of the vehicle computer by users (steering wheel, accelerator / brake, various buttons / switches, touch, voice control, etc.), and operation of the vehicle computer by users actively setting scene modes (for example, setting a nap mode to achieve one-touch operation of the air conditioner, windows, seats, etc.). After these behaviors are recorded, they will form a collection of several behaviors according to their time concentration;
[0050] Conditions: refers to all vehicle states and environmental information collected by sensors, including but not limited to vehicle position, speed, direction, degree of inclination of each axis, power / fuel level, vibration, road conditions collected and identified by ADAS sensors, cabin temperature, humidity, number of passengers and their location distribution, passenger identity, air conditioning status, audio status, etc. After these conditions are recorded, several sets of conditions will be formed according to the time segments corresponding to the above behavior sets;
[0051] Record format: Depending on business needs, the substantive content can be in JSON or XML format. The time needs to be broken down to year, month, day, hour, minute, second, and millisecond, and the day of the week, whether it is a working day, etc. should be recorded;
[0052] 2. Modeling (feature construction, model selection and training, model evaluation and optimization): The above records are used as data and provided to the machine learning algorithm for training to obtain the corresponding model. From a simple principle level, the model is the mapping relationship between the above condition set and the behavior set. We call this mapping relationship a habit.
[0053] 3. Prediction (real-time behavior prediction and personalized recommendation): When users actually use the vehicle, the real-time recorded conditions and behaviors are input into the model, and the model predicts the user's next behavior based on the user's habits;
[0054] 4. Interaction (real-time behavior prediction and personalized recommendation): After the model predicts user behavior, it is divided into two categories according to the triggering method of the interaction:
[0055] (Model) Active triggering: When the current real-time recorded conditions match the condition set of a user's habit, and the user actively performs an action in the behavior set of the habit, the model initiates an interaction, asking the user whether to perform the remaining actions in the behavior set of the habit. After the user responds, the corresponding actions are executed and the data is recorded;
[0056] (Model) Passive triggering: When the current real-time recorded conditions match the condition set of a user's habit, and the user actively performs the triggering action (for example, the user wakes up the voice assistant and says "Let's go") after getting in the car, the model directly executes all the actions in the behavior set of the habit and records the data;
[0057] 5. Correction (model update and maintenance): During multiple model iterations, newly generated data will correct past models, continuously improving model prediction performance while reflecting changes in user habits.
[0058] In summary, the present invention is based on the data of a single user rather than big data, truly realizing personalization and "one thousand faces for one thousand people", making full use of the rich sensors of smart devices, integrating environmental information as a condition, reflecting the interaction between users and the environment, and collecting cross-domain behavior data of a single user and combining it with environmental characteristics, and through comprehensive analysis of user comments, sharing, facial expression detection results, voice command emotion prediction and other data, to understand the user's emotional changes and demand changes. This information can be input into the model as an important feature to improve the model's predictive ability. The machine learning algorithm can be used to construct a model that reflects the behavior patterns and preferences of a single user. The model can more accurately predict the behavior of a single user in different scenarios or fields in the future, thereby providing more personalized services. The designed cross-domain method can integrate behavioral data from multiple sources, capture the behavioral connections and transformation patterns of users in various fields, help to gain insight into the overall behavioral dynamics of users, and provide more comprehensive services.
[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. Single user cross-domain behavior modeling method based on machine learning algorithm, It is characterized in that The specific steps are as follows: Step 1: Data collection and preprocessing: collect cross-domain behavior data of a single user, and perform preprocessing operations such as cleaning, deduplication, and formatting on the collected data; Step 2: Feature construction: Based on the user's behavior records, construct features that reflect the user's behavior patterns and behavior preferences. Combined with the collected environmental information, construct features that reflect the conditions under which the user's behavior occurs, and introduce sentiment analysis features into the model. Step 3: Model selection and training: According to the specific application scenarios and requirements, select the appropriate machine learning algorithm to build the user behavior model, and use the preprocessed data and constructed features to train the machine learning model; Step 4: Model evaluation and optimization: evaluate the performance of the model through cross-validation or error analysis methods, and optimize the model based on the evaluation results; Step 5: Real-time behavior prediction and personalized recommendation: When a user takes action in a new field or scenario, the trained model is used to predict the user's behavior in real time. Based on the user's behavior prediction results, personalized recommendation content is generated for the user. Step 6: Model update and maintenance: Update the model regularly based on the passage of time and changes in user behavior, and continuously monitor and maintain the model.
2. According to claim 1, the single user cross-domain behavior modeling method based on machine learning algorithm, Features: The data collected in step 1 comes from the user's behavior records and all operations in various Internet applications, including those actively triggered by the user and the user's confirmation actions after the traditional scenario mode actively recommends. At the same time, environmental information related to the user's behavior is collected, including time, location and device status.
3. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 1, Features: The user's behavior record in step 2 includes the user's browsing preferences, purchase frequency and interaction habits, and the collected environmental information includes time characteristics and location characteristics. The introduced sentiment analysis features conduct a comprehensive analysis of the user's comments, sharing, facial expression detection results, and voice command emotion prediction data to understand the user's emotional changes and demand changes, and input this information into the model as important features.
4. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 1, Features: In the step three, the machine learning algorithm constructs the user behavior model, including using a sequence model to process the user's time series data, or using a decision tree or neural network algorithm for classification or regression prediction.
5. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 1, Features: The personalized recommendation content in step five includes products, services and information that the user may be interested in, or activities or suggestions that are consistent with the user's behavior patterns and preferences.
6. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 1, Features: The step six of model updating and maintenance includes regularly retraining the model or updating the model by online learning. Model maintenance includes checking the performance indicators of the model, processing outliers and noise data, and optimizing the computational efficiency of the model.
7. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 4, Features: The sequence model processes the continuous behavior records of the user using the time series data, and uses a recurrent neural network (sequence model (including but not limited to RNN, transformer)) sequence model for processing. The formula design of the sequence model (including but not limited to RNN, transformer) includes an input layer, a hidden layer and an output layer; Input layer: input time series data (x_1, x_2, …, x_t) into the network; Hidden layer: At each time step (t), the hidden state (h_t) is calculated from the current input (x_t) and the hidden state of the previous time step (h_{t-1}), using the formula: (h_t=f(W_x x_t+W_h h_{t-1}+b)), where (W_x) and (W_h) are weight matrices, (b) is the bias term, and (f) is the activation function (such as tanh or ReLU); Output layer: Generates output (y_t) based on hidden state (h_t). For classification tasks, use softmax function to convert hidden state into probability distribution. For regression tasks, you can directly use linear output layer.
8. The single user cross-domain behavior modeling method based on machine learning algorithm according to claim 4, Features: The neural network algorithm performs classification or regression prediction. In the neural network, the output of each neuron is the weighted sum of the input plus the bias term, which is then transformed through an activation function. For modeling the cross-domain behavior of a single user, the input is the user's historical behavior characteristics and environmental characteristics, and the output is the predicted user behavior or related classification; The formula is: (z=\sum_{i=1}^{n}w_i x_i+b), where (z) is the weighted input, (w_i) is the weight of the (i)th input feature (x_i), (b) is the bias term, and after passing through the activation function (f), the output of the neuron is (a=f(z)).
Citation Information
Patent Citations
Recommendation method and related device
CN115905680A
Cross-domain recommendation method and system based on user preference diversity
CN116910375A
Content recommendation system based on artificial intelligence, electronic equipment and storage medium
CN117972206A
Intelligent service system and method for automobile consumption
CN118297756A
System and method for multi-task lifelong learning on personal device with improved user experience
US20230052903A1
