Cross-device association method based on mobile intelligent terminal sensor data fusion analysis

Abstract

The present invention discloses a cross-device association method based on mobile smart terminal sensor data fusion analysis, which includes the following steps: collecting multiple sensor data of two types of devices as model training data; performing pre-processing operations such as splitting, equalization, and normalization on the data; designing a neural network model. Since the present invention needs to compare the similarity of sensor data collected by the two types of devices, a twin neural network is used as a discriminant model, and the cosine distance is used to calculate the similarity of the output vectors of the two convolutional neural networks, and the loss function is further calculated based on the similarity; the designed twin neural network is used to train the corresponding model for each scenario; a similarity threshold is set, and if the threshold is exceeded, the data matching is successful, otherwise it fails, thereby realizing cross-device association. The method of the present invention is the first method for associating multiple mobile smart terminal devices of a user, and the amount of user data collected is small, the association speed is fast, and the association accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-device association, and in particular relates to a cross-device association method based on mobile intelligent terminal sensor data fusion analysis. Background Art

[0002] Cross-device association is a natural extension and development of terminal device identification technology, and it is to determine whether multiple devices belong to the same user. For example, online advertisers need to collect the browsing history of a single user on different devices to complete a more complete and comprehensive portrayal of the shopping preferences of online users in order to achieve precision marketing and advertising push. According to the different association methods used, devices and device association methods can be divided into two categories: deterministic association and probabilistic association. Deterministic association mainly involves users using the same account to log in to different devices, and then associating different devices based on the same account name or strong identifiers such as generated cookies. This method is generally applicable to major service providers. Some existing applications do not require users to log in to the system to access services, and strong identifiers such as cookies are easily cleared by users, resulting in their failure. Therefore, the current focus of academic research is on probabilistic cross-device association. The feasibility of probabilistic cross-device association lies in the fact that the user attributes, preferences, and user behaviors reflected by the same user when using different devices have certain similarities.

[0003] Relatively speaking, the research on probabilistic cross-device association technology is still in its infancy. The industry is still using evercookies, cookie synchronization, cookie sharing and single-terminal device fingerprints to track and identify user devices. In the cross-device association competition organized by the ICDM2015 and CIKM2016 international conferences, the focus is also on how to design algorithms based on existing data sets, such as pairwise ranking, to achieve device association analysis. However, existing work has also pointed out a feasible way to probabilistic cross-device association, that is, in different application scenarios, we should study how to obtain multi-terminal related attributes, user related attributes and environment related attributes from network traffic and browser data. In addition, how to make full use of various sensors carried by mobile smart terminals is also an important topic that needs to be studied in depth. Summary of the invention

[0004] The purpose of the present invention is to propose a cross-device association method based on mobile intelligent terminal sensor data fusion analysis, so as to realize the analysis and association of sensor data in different types of devices of the same user.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme:

[0006] The cross-device association method based on mobile intelligent terminal sensor data fusion analysis includes the following steps:

[0007] Step 1. Collect multiple sensor data of two types of devices in different scenarios as model training data;

[0008] Step 2: Split, equalize, and normalize the sensor data collected in step 1, and perform labeling preprocessing operations on the data of different types of scenes;

[0009] Step 3. Build a twin neural network model to compare the sensor data collected by the two types of devices. The twin neural network includes two sub-networks. The two sub-networks use convolutional neural networks with the same structure and share weights.

[0010] Step 4. Input the sensor data preprocessed in step 2 into the twin neural network model for training, and use the cosine distance to calculate the similarity and loss function of the output vectors of the two convolutional neural networks;

[0011] Step 5. Based on the above step 4, the twin neural network model is trained using the data preprocessed in step 2, and finally multiple different network models for different types of scenarios are obtained;

[0012] Step 6. Input multiple sets of sensor data of the two types of devices to be associated into the corresponding twin neural network model for matching according to different application scenarios to obtain a similarity set;

[0013] A similarity threshold is preset. If the similarity between the two sets of data to be matched in the two types of devices exceeds the similarity threshold, the two sets of data are considered to be matched successfully and marked as 1; if the similarity is less than the similarity threshold, the match is considered to fail and marked as 0;

[0014] The data to be matched is summed up. If more than half of the data is matched successfully, the two devices are considered to be matched and related.

[0015] Among them, the two types of devices specifically refer to different types of mobile smart terminals.

[0016] The present invention has the following advantages:

[0017] As described above, the present invention describes a cross-device association method based on mobile smart terminal sensor data fusion analysis. This method is the first method for associating multiple mobile smart terminal devices of a user. Through the method steps proposed in the present invention, the association method of the present invention collects less user data, has a faster association speed, and has a higher association accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1It is an overall block diagram of the cross-device association method based on mobile intelligent terminal sensor data fusion analysis in the present invention;

[0019] Figure 2 It is the overall structure diagram of the twin neural network model in the present invention;

[0020] Figure 3 It is a structural diagram of the convolutional neural network in the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0022] Example 1

[0023] This embodiment 1 describes a cross-device association method based on mobile smart terminal sensor data fusion analysis. The association method aims to obtain the training data required for the model by collecting certain specific actions performed by the user. After a series of data preprocessing, the twin neural network model training data is built, and the appropriate similarity calculation and loss function are selected to train the corresponding model for each specific action scenario; according to different user operations, the corresponding model for each specific action scenario can be selected, and the model can be used to realize cross-device association based on mobile smart terminal sensor data fusion analysis.

[0024] like Figure 1 As shown, the cross-device association method based on mobile intelligent terminal sensor data fusion analysis includes the following steps:

[0025] Step 1. Collect multiple sensor data of two types of devices in different scenarios as model training data. The two types of devices specifically refer to different types of mobile smart terminals, such as Pad devices and Phone devices.

[0026] The sensor data to be collected in this embodiment include six types, namely, accelerometer data, light sensor data, gyroscope data, magnetic field sensor data, gravity sensor data, and linear accelerator data.

[0027] Taking Android devices as an example, Android software is written to use the sensor API provided by Android to collect data, collect user behavior, operation and environmental information, and the API functions used are shown in Table 1 below.

[0028]

[0029] Data collection requires users to complete specified operations, so that the data is more targeted and processable.

[0030] Data collection involves various types of scenarios. In this embodiment, for example, data is collected in the following six scenarios:

[0031] a. Walking with the equipment;

[0032] b. Enter a specified text in the device software;

[0033] c. Browse news on the device for 30 seconds;

[0034] d. Use the device to open the short video software and watch the 30-second short video;

[0035] e. Use the device to open the shopping app and browse for 30 seconds;

[0036] f. Use your device to open the photo, zoom in and out, and take a photo.

[0037] The above six device usage scenarios are designed mainly considering that the operations in these six scenarios are common user operations on mobile smart terminals with relatively high frequency. The file format of the collected data is csv file.

[0038] Of course, the above does not constitute a limitation to the present invention. The present invention is not limited to collecting data in the above six scenarios, and operations in other application scenarios can also be collected according to actual needs, which will not be described in detail here.

[0039] Step 2. Split, equalize, and normalize the various sensor data collected in step 1, and perform preprocessing operations such as labeling for data of different types of scenes.

[0040] First, the collected data is split. Since step 1 is for data collection in different application scenarios, after collecting the data, the data generated by different operations must be classified.

[0041] Therefore, when starting an operation during the collection process, click Start Experiment, and when ending the operation, click the End button. This will insert a separator into the data, and different operations will be separated according to the separator in the subsequent data processing.

[0042] Secondly, according to the different sampling frequencies, the data is equalized, that is:

[0043] For data with high sampling rate, the operation of dividing the interval into intervals and taking the average is performed; for data with low sampling rate and insufficient values ​​in the interval, interpolation processing is performed, and the final sampling value is the actual sampling value closest to the midpoint of the interval.

[0044] Specifically, the equalization process is as follows:

[0045] Since the sampling frequency of each sensor is different, the sensor data can be divided into two categories: fine-grained sampling with a high sampling rate and coarse-grained sampling with a low sampling rate.

[0046] Fine-grained sampling data includes accelerometer data, gyroscope data, magnetic field sensor data, gravity sensor data, and linear accelerator data. These data are characterized by rapid changes and more data collected per second. Coarse-grained sampling data includes light sensor data. Since the lighting conditions around the device do not change too drastically, the sampling frequency is low.

[0047] For fine-grained sampling data, a threshold of 10 groups of single sensor data is set per second. The overall sampling time interval is relatively uniform. Assume that there are N groups of samples, where N is a natural number greater than 10.

[0048] The present invention utilizes a uniform downsampling method to sample 10 groups of data within 1 second.

[0049] Perform interpolation operations on coarse-grained sampling data.

[0050] Assume that there are n sampling groups, where n is a natural number and n is less than 10; 1 second is divided into 10 adjacent intervals of 0.1 seconds, and the final sampling value of each interval is the real sampling value closest to the interval.

[0051] The obtained data is split, balanced, and finally normalized.

[0052] Since the data distribution of different sensors varies greatly, the same sensor needs to be normalized.

[0053] Considering that the data obtained by the fine-grained sampling sensor data part is relatively precise, the 6 decimal places are first taken for calculation, and then the average value μ(s) of the data collected by each triaxial sensor is calculated. i ) and standard deviation σ(s i ).

[0054]

[0055] Where N is a single sensor s i The total number of data, x n ,y n ,z n is the three-axis data of the sensor, n=1,...,N.

[0056] Since the present invention involves multiple scenarios of user operation of devices, corresponding labels can be added to each scenario. For example, the six scenarios mentioned above are represented by a to f to identify the data collected in different scenarios. Specifically:

[0057] a. Carry the device and walk; b. Enter a paragraph of text in the device software; c. Use the device to browse news; d. Use the device to open the short video software to watch short videos; e. Use the device to open the shopping software to browse; f. Use the device to open the photo, zoom in and view the photo and take a photo.

[0058] The present invention performs multiple steps of preprocessing operations on the data to obtain data that meets the model training requirements.

[0059] Step 3. Build a twin neural network model to compare the sensor data collected by the two types of devices.

[0060] Since the purpose of the present invention is to analyze and compare the similarity between sensor data, the twin neural network model is used as the discrimination model of the present invention in this embodiment, and its network structure is as follows: Figure 2 shown.

[0061] The inputs S1 and S2 of the model are data that can be processed by the neural network. The present invention designs the input data of the model as a 72×72 two-dimensional array. First, the sensor data structure at a certain time t0 is as follows:

[0062] S t0 =[x1,y1,z1,x2,y2,z2,…,x5,y5,z5,L,L,L].

[0063] The one-dimensional array is 18 bits long, of which the first 15 bits are the data values ​​of the five three-axis sensors, and the last three bits are the value L of the light intensity collected at the current moment. In order to facilitate subsequent convolution calculations, three identical values ​​are copied here.

[0064] The present invention combines the arrays of 4 moments into a long one-dimensional array, and the length at this time is 18×4=72 bits, and S4=[S t0 ,S t1 ,S t2 ,S t3 ], similar 72 groups of S4 constitute a 72×72 input matrix.

[0065] The twin neural network consists of two sub-networks. The two sub-networks use convolutional neural networks with the same structure and share the weight W.

[0066] Take one of the convolutional neural network structures as an example, Figure 3 Considering that the data correlation between sensors is weak, while the data collected by the same sensor is highly correlated, the convolutional neural network adopts a seven-layer structure.

[0067] Among them, after the input data, the first layer of convolution uses a 3×3 convolution kernel, a step size of 3, a total of 18 convolution kernels, and the output data is 24×24×18. Considering that the sensor data has positive and negative values, the second layer activation function selects the Tanh function; the third layer of convolution uses a 4×4 convolution kernel, a total of 36, and the corresponding step size is 4, the output data is 6×6×36, and the fourth layer activation function is also the Tanh function; the fifth and sixth layers are fully connected layers, each with 100 nodes. In order to reduce the dimension of the output data, the seventh layer of the fully connected layer is 18 nodes. Therefore, after the convolutional neural network, the output data is an 18-dimensional vector, such as the vector Net(W,S1) and vector Net(W,S2) shown in 3.

[0068] Step 4. Input the sensor data preprocessed in step 2 into the twin neural network model for training, and use the cosine distance to calculate the similarity and loss function of the output vectors of the two convolutional neural networks.

[0069] For the output vectors Net(W,S i ) and Net(W,S j ), similarity Sim(S i ,S j )for:

[0070]

[0071] Among them, COS(Net(W,S i ),Net(W,S j )) represents the cosine distance between two sets of vectors output by the twin neural network.

[0072] The similarity range is [-1, +1]. When Sim(S i ,S j )=1, the two output vectors are in the same direction, and the similarity is maximum. i ,S j )=-1, the direction is opposite and the similarity is minimum.

[0073] In the loss function, the distance metric of the common vectors is used, and the cosine distance between the two vectors needs to be calculated as follows:

[0074] SimDist(S i ,S j )=1-Sim(S i ,S j ).

[0075] Among them, SimDsit(S i ,S j) represents the cosine distance between two vectors; according to the formula of cosine distance, the value range of cosine distance is [0,2]. The larger the distance, the smaller the similarity, and vice versa.

[0076] Since the present invention needs to determine whether two sets of sample data are generated by the same user, it can be regarded as a binary classification problem. For the binary classification problem, the loss function of the twin neural network is:

[0077]

[0078] Among them, Loss(Y,D) represents the loss function. Y represents whether it is the label of the same user, and D is the abbreviation of distance, indicating that the loss function uses the distance function for calculation.

[0079] When the two sets of data are generated by the same user, Y = 0, then The goal is to reduce the loss function, that is, the smaller the distance between the two, the better. When the two sets of data are not generated by the same user, Y = 1, then The goal is also to reduce the loss function.

[0080] Since the cosine distance has a value range of [0,2], the boundary value 1 is set here based on experience.

[0081] When SimDist(S i ,S j )≥1, the loss value is 0, and when it is less than 1, the loss value is a positive number. The significance of this boundary value is that when the two arrays are generated by different users, the optimization stops when the distance is greater than 1 to avoid overfitting.

[0082] Step 5. Based on the above step 4, the data preprocessed in step 2 is input into the twin neural network model for training, and finally different network models for different types of scenarios are obtained.

[0083] During network training, sensor data from two types of devices (mainly mobile phones and tablets) are collected and trained using the designed twin neural network model. Multiple models are trained in various scenarios for subsequent device association.

[0084] Specifically, each person and each group of experiments will take about ten minutes to collect data.

[0085] The present invention uses a sliding window to capture data every 10 seconds. Each experiment can collect about 60 groups of data. A total of 20 users are invited to participate in the experiment. There are 60×20×6×2 groups of data for six experiments and two types of equipment.

[0086] Then, the data of the two types of devices are extracted respectively. The data of the two types of devices belonging to the same user are labeled as 0, which is a positive sample pair, and the data of different users are labeled as 1, which is a negative sample pair.

[0087] Finally, the twin neural network was used for training, and six models were trained for the six scenarios for subsequent association.

[0088] Step 6. Input the sensor data of the devices to be associated into the corresponding twin neural network model for pairwise matching according to different application scenarios to obtain a similarity set.

[0089] A similarity threshold is preset. If the similarity between the two sets of data to be matched in the two types of devices exceeds the similarity threshold, the two sets of data are considered to be matched successfully and marked as 1; if it is less than the similarity threshold, the match is considered to fail and marked as 0.

[0090] The data to be matched is summed up. If more than half of the data is matched successfully, the two devices are considered to be matched and related.

[0091] When associating devices, first select a model corresponding to the scenario from multiple trained twin neural network models based on different application scenarios.

[0092] For example, in a scenario where the user is walking, you can select a model trained using data collected while the user is walking with the device.

[0093] The cross-device association problem is described as collecting data from two types of devices to form a Phone device set and a Pad device set.

[0094] For a single device D in the Phone device collection i , match D in the Pad device set j ; Assume that device D i There are n sets of data, namely Equipment D j There are m groups of data, namely Among them, n and m are both odd numbers.

[0095] If the data collected by the device is an even number, the last set of data can be removed to facilitate subsequent voting operations.

[0096] For D i Each piece of data of the device, from the data Start by combining it with D j The m sets of data of the equipment are input into the twin neural network model one by one for matching, and a similarity set CosSim{CS1,CS2,…,CS m}.

[0097] Preset a similarity threshold T D , when the similarity value in the similarity set exceeds the similarity threshold T D , the group of data is considered to be matched successfully, success is 1, and failure is 0, then the similarity set is transformed into the matching set Match.

[0098] Sum the number of successful matches of data in the matching set Match.

[0099] When the sum result sum1 is greater than When more than half of the data in the Pad device data set matches successfully, it is considered that device D i Data in and device D j If the match is successful, it is recorded as 1, and if the match is unsuccessful, it is recorded as 0.

[0100] D i The remaining data in Each data in the The processing method is similar to that of device D j The m groups of data are input one by one into the twin neural network model for matching.

[0101] Equipment D i After all the data in are matched, the device D is formed. i With device D j The matching sequence.

[0102] The matching sequence consists of 0 and 1, with a total of n numbers. The sum of the n numbers is calculated as sum2. When sum2 is greater than n / 2, the device D is determined. i and device D j If the match is successful, the cross-device association task is completed.

[0103] The method of the present invention is the first method for associating multiple mobile intelligent terminal devices of a user. Through the steps of the method of the present invention, the amount of user data collected is small, the association speed is fast, and the association accuracy is high.

[0104] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A cross-device association method based on mobile intelligent terminal sensor data fusion analysis, characterized in that: The steps include: Step 1. Collect multiple sensor data of two types of devices in different scenarios as model training data; Step 2: Split, equalize, and normalize the sensor data collected in step 1, and perform labeling preprocessing operations on the data of different types of scenes; Step 3. Build a twin neural network model to compare the sensor data collected by the two types of devices. The twin neural network includes two sub-networks. The two sub-networks use convolutional neural networks with the same structure and share weights. Step 4. Input the sensor data preprocessed in step 2 into the twin neural network model for training, and use the cosine distance to calculate the similarity and loss function of the output vectors of the two convolutional neural networks; Step 5. Based on the above step 4, the twin neural network model is trained using the data preprocessed in step 2, and finally multiple different network models for different types of scenarios are obtained; Step 6. Input multiple sets of sensor data of the two types of devices to be associated into the corresponding twin neural network model for matching according to different application scenarios to obtain a similarity set; A similarity threshold is preset. If the similarity between the two sets of data to be matched in the two types of devices exceeds the similarity threshold, the two sets of data are considered to be matched successfully and marked as 1; if the similarity is less than the similarity threshold, the match is considered to fail and marked as 0; The data to be matched is summed up. If more than half of the data is matched successfully, the two devices are considered to be matched and related. Among them, the two types of devices specifically refer to different types of mobile smart terminals; In step 1, the sensor data includes six types, namely, accelerometer data, light sensor data, gyroscope data, magnetic field sensor data, gravity sensor data and linear accelerator data; In step 1, the device is used to collect data in the following scenarios: Carry the device for walking; input a paragraph of text in the device software; use the device to browse news; use the device to open short video software to watch short videos; use the device to open shopping software to browse; use the device to open photos, zoom in and out to view photos and take photos.

2. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: In step 2, the equalization process is as follows: First, according to the different sampling frequencies of sensors, the sensor data is divided into fine-grained sampling data and coarse-grained sampling data; The fine-grained sampling data includes accelerometer data, gyroscope data, magnetic field sensor data, gravity sensor data, and linear accelerator data, and the coarse-grained sampling data includes light sensor data; For fine-grained sampling data, a threshold of 10 groups of single sensor data is set per second; Assume that there are N groups of sampled data within one second, N is a natural number and is greater than 10. Use the uniform downsampling method to sample 10 groups of data within one second. Interpolation operation is performed on coarse-grained sampling data; assuming there are n sampling groups, n is a natural number, and n is less than 10; 1 second is divided into 10 adjacent intervals of 0.1 seconds, and the final sampling value of each interval is the actual sampling value closest to the interval.

3. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: In step 2, the normalization process is as follows: Perform batch normalization on data from the same sensor; Considering that the data obtained by the fine-grained sampling sensor data part is relatively precise, the 6 decimal places are first taken for calculation, and then the average value μ(s) of the data collected by each triaxial sensor is calculated. i ) and standard deviation σ(s i ); Where N is a single sensor s i The total number of data, x n ,y n , x n is the sensor three-axis data, n=1,...,N.

4. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: The input data of the twin neural network model is designed to be a 72×72 two-dimensional array; First, the sensor data structure at a certain time t0 is as follows: <h2 style=";text-align:left;direction:ltr">S<h2 style=";text-align:left;direction:ltr"> t0 <h2 style=";text-align:left;direction:ltr"> ([x1, y1, z1, x2, y2, z2,…, x5, y5, z5, L, L, L]) The length of the above one-dimensional array is 18 bits, of which the first 15 bits are the data values ​​of the five three-axis sensors, and the last three bits are the value L of the light intensity collected at the current moment. To facilitate subsequent convolution calculations, here are three identical values; The arrays of the four moments t0, t1, t2, and t3 are combined into a long one-dimensional array. The length of this array is 18×4=72 bits, and S4=[S t0 , S t1 , S t2 , S t3 ], there are 72 groups of S4 in total, forming a 72×72 input matrix.

5. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: In step 3, the convolutional neural network adopts a seven-layer structure; After inputting the data, the first layer of convolution uses a 3×3 convolution kernel with a step size of 3, a total of 18 convolution kernels, and the output data is 24×24×18; the second layer activation function selects the Tanh function; the third layer of convolution uses a 4×4 convolution kernel, a total of 36, and the corresponding step size is 4, the output data is 6×6×36, and the fourth layer activation function is also the Tanh function; The fifth and sixth layers are fully connected layers, each with 100 nodes. To reduce the dimension of the output data, the seventh fully connected layer has 18 nodes. Therefore, after passing through the convolutional neural network, the output data is an 18-dimensional vector.

6. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: In step 5, the calculation process of the similarity and loss function is: For the output vectors Net(W, S i ) and Net(W, S j ), similarity Sim(S i , S j )for: Among them, COS(Net(W,S i ), Net(W, S j )) represents the cosine distance between two sets of vectors output by the twin neural network; The similarity range is [-1, +1]. When Sim(S i , S j )=1, the two output vectors are in the same direction, and the similarity is maximum. i , S j )=-1, the direction is opposite, and the similarity is minimum; In the loss function, the distance metric of the common vectors is used, and the cosine distance between the two vectors needs to be calculated as follows: SimDist(S i ,S j )=1-Sim(S i ,S j ); Among them, SimDsit(S i , S j ) represents the cosine distance between two vectors; according to the formula of cosine distance, the value range of cosine distance is [0, 2]. The larger the distance, the smaller the similarity, and vice versa. Judging whether two sets of data are generated by the same user is a binary classification problem, so the loss function of the twin neural network is: Among them, Loss(Y, D) represents the loss function; Y represents whether it is the label of the same user, and D represents the abbreviation of distance, indicating that the loss function uses the distance function for calculation; When the two sets of data are generated by the same user, Y = 0, then When the two sets of data are not generated by the same user, Y = 1, then Set the boundary value to 1, when SimDist(S i , S j )≥1, the loss value is 0, and when it is less than 1, the loss value is a positive number. This boundary value makes the optimization stop when the distance between the two arrays is greater than 1 when they are generated by different users, thus avoiding overfitting.

7. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 1 is characterized in that: In step 5, during network training, sensor data of two types of devices are collected and trained using the designed twin neural network model. Multiple models are trained in multiple scenarios for subsequent device association.

8. The cross-device association method based on mobile intelligent terminal sensor data fusion analysis according to claim 7 is characterized in that: In step 6, when associating devices, first select a corresponding model from multiple trained twin neural network models according to different application scenarios; The cross-device association problem is described as collecting data from two types of devices to form a Phone device set and a Pad device set; For a single device D in the Phone device collection i , match D in the Pad device set j ; Assume that device D i There are n sets of data, namely Equipment D j There are m groups of data, namely Where n and m are both odd numbers; For D i Each piece of data of the device, from data Start by combining it with D j The m sets of data of the equipment are input into the twin neural network model one by one for matching, and a similarity set CosSim{CS1, CS2, …, CS m }; Preset a similarity threshold T D , when the similarity value in the similarity set exceeds the similarity threshold T D When , the group of data is considered to be matched successfully, success is 1, and failure is 0, then the similarity set is transformed into the matching set Match; Sum the number of successful matches of data in the matching set Match; When the sum result suml is greater than When more than half of the data in the Pad device data set matches successfully, it is considered that device D i Data in and device D j If the match is successful, it is recorded as 1, and if the match is unsuccessful, it is recorded as 0; D i The remaining data in Each data in the The processing method is similar to that of device D j The m groups of data are input into the twin neural network model one by one for matching; Equipment D i After all the data in are matched, the device D is formed. i With device D j The matching sequence of The matching sequence consists of 0 and 1, with a total of n numbers. The sum of the n numbers is calculated as sum2. When sum2 is greater than n / 2, the device D is determined. i and device D j If the match is successful, the cross-device association task is completed.

Images