CN110991711A - Multi-factor perception terminal switching prediction method based on deep neural network - Google Patents
Multi-factor perception terminal switching prediction method based on deep neural network Download PDFInfo
- Publication number
- CN110991711A CN110991711A CN201911135219.1A CN201911135219A CN110991711A CN 110991711 A CN110991711 A CN 110991711A CN 201911135219 A CN201911135219 A CN 201911135219A CN 110991711 A CN110991711 A CN 110991711A
- Authority
- CN
- China
- Prior art keywords
- user
- terminal
- code
- behavior
- attributes
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 12
- 230000008447 perception Effects 0.000 title claims abstract description 9
- 230000006399 behavior Effects 0.000 claims abstract description 85
- 230000008859 change Effects 0.000 claims abstract description 67
- 238000010606 normalization Methods 0.000 claims abstract description 14
- 238000007781 pre-processing Methods 0.000 claims abstract description 12
- 238000013507 mapping Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 13
- 230000002159 abnormal effect Effects 0.000 claims description 12
- 239000011159 matrix material Substances 0.000 claims description 12
- 238000000605 extraction Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 9
- 230000008030 elimination Effects 0.000 claims description 6
- 238000003379 elimination reaction Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012886 linear function Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 125000004122 cyclic group Chemical group 0.000 claims description 2
- 238000001514 detection method Methods 0.000 claims description 2
- 230000009286 beneficial effect Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/40—Business processes related to the transportation industry
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Marketing (AREA)
- Biophysics (AREA)
- General Business, Economics & Management (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Tourism & Hospitality (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Primary Health Care (AREA)
- Telephonic Communication Services (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a multi-factor perception terminal switching prediction method based on a deep neural network, which comprises the steps of firstly collecting terminal attributes, related user natural attributes, conversation behavior data, flow use behavior data and historical switching records, and preprocessing the terminal attributes; secondly, coding is carried out to obtain a user natural attribute code and a terminal attribute code, a user history change machine information code is constructed according to a user history change machine record and the corresponding terminal attribute code, a user conversation and flow use behavior time sequence is constructed, and coarsening and normalization processing are carried out; extracting time sequence characteristics of user conversation and flow use behaviors; and finally, splicing the natural attribute code of the user, the historical switch-over information code, the conversation behavior time sequence characteristic and the flow use behavior time sequence characteristic, and sending the spliced result into the full-connection network to predict whether the user can change the switch. The prediction method is beneficial to operators to accurately market the mobile terminal and the matched products thereof for users, so that the sales volume of the mobile terminal is increased, and the market scale is enlarged.
Description
Technical Field
The invention relates to the technical field of deep learning, in particular to a multi-factor perception terminal switching prediction method based on a deep neural network.
Background
The sale of the mobile terminal and the matched products becomes the strategic core of telecommunication operators, and the accurate prediction of user change is beneficial to the operators to increase the sales volume of the mobile terminal and expand the market scale. However, user switch behavior is affected by a number of factors, such as: the income condition, the consumption habit, the dependence degree of the mobile phone, the conversation behavior, the traffic use behavior and the like of the user bring great challenges to the terminal switching prediction. In order to realize accurate marketing of an operator terminal and a matched product thereof, the influence of natural attributes and behavior data of a user on the change behavior of the user needs to be deeply mined, and then the change behavior of the user is accurately predicted.
The existing prediction method for terminal switching can be divided into two types: and in the first category, prediction is carried out only based on static attributes of the user (such as natural attributes of the user, terminal use preference, package use data, common positions of the user and the like). Such models have poor performance because of neglected user behavior information and the inability to model dynamic user representations. And in the second category, dynamic behavior data of the user is introduced, and statistical characteristics (such as average call duration, traffic usage and the like) of the behavior data of the user in the last week are defined manually by depending on professional knowledge. The statistical features constructed by relying on the domain knowledge can not model the variation trend of the user behavior sequence in a display mode and can cause information loss to a certain extent.
Disclosure of Invention
The invention aims to overcome the defects and provides a multi-factor perception terminal switching prediction method based on a deep neural network, and the technical problem to be solved is how to fully mine natural attributes of users, call and flow use behavior data and factors influencing user switching in historical switching records to predict the user switching; the method fully excavates factors influencing the user change by utilizing the natural attributes of the user, the call and flow use behavior data and the historical change machine record, and further constructs the terminal change machine prediction model with high generalization capability. The method comprises the steps of firstly, acquiring terminal attributes and related natural user attributes, conversation behavior data, flow use behavior data and historical change machine records, and preprocessing the terminal attributes and the related natural user attributes; then, encoding the preprocessed user natural attributes and terminal attributes to obtain user natural attribute codes and terminal attribute codes, constructing user history change information codes according to user history change records and corresponding terminal attribute codes, constructing a user conversation and flow use behavior time sequence, and performing coarsening and normalization processing; then, using the LSTM network to extract the time sequence characteristics of the user conversation and flow use behaviors; and finally, splicing the natural attribute code of the user, the historical switch-over information code, the conversation behavior time sequence characteristic and the flow use behavior time sequence characteristic, and sending the spliced result into the full-connection network to predict whether the user can change the switch. The method and the system realize multi-factor perception terminal replacement prediction based on the deep neural network, and are beneficial to operators to accurately market the mobile terminal and matched products thereof for users, so that the sales volume of the mobile terminal is increased, and the market scale is enlarged.
The invention achieves the aim through the following technical scheme: a multi-factor perception terminal switching prediction method based on a deep neural network comprises three stages of data acquisition and preprocessing, feature construction and extraction, and feature fusion and prediction, and specifically comprises the following steps:
(1) data acquisition and preprocessing stage: acquiring terminal attributes and related natural user attributes, call behavior data, traffic use behavior data and historical change records, and preprocessing the terminal attributes and the related natural user attributes;
(2) and (3) feature construction and extraction stages: coding the preprocessed user natural attribute and terminal attribute to obtain a user natural attribute code and a terminal attribute code, constructing a user history change information code according to a user history change record and the corresponding terminal attribute code, constructing a user conversation and flow use behavior time sequence, and performing coarsening and normalization processing; and utilizes LSTM network to extract time sequence characteristics of user's talking and traffic using behaviors;
(3) and (3) feature fusion and prediction stage: and splicing the natural attribute code of the user, the historical switch-over information code, the conversation behavior time sequence characteristic and the flow use behavior time sequence characteristic, and then sending the spliced result into the full-connection network to predict whether the user can change the switch.
Preferably, the data acquisition and preprocessing stages are specifically as follows:
(1.1) collecting relevant user natural attributes, including: gender, age, network access duration, client star level, whether the user is an in-person network user, the number of CDMA (code division multiple access) users under the same client and App (application profile) use preference;
(1.2) collecting relevant terminal attributes, including: brand, operating system, price, screen size, whether 4G and rear camera pixels are supported;
(1.3) collecting related user call behavior data and flow use behavior data, and performing missing value completion and abnormal value elimination processing; the user call behavior data comprises daily call times, daily call duration, daily call times and daily call duration, and the flow use behavior data comprises daily flow use times, daily flow use duration, daily uplink flow and daily downlink flow;
and (1.4) collecting user history change machine records.
Preferably, in the step (1.3), missing value and abnormal value detection is performed on the user call and traffic usage behavior data, and the missing value and the abnormal value are filled in by using a linear interpolation method.
Preferably, the user history change machine record is represented in the form of a triple < u, s, t >, where u is a user id, s is a terminal id, and t is a date when the user u uses the terminal s first.
Preferably, the feature construction and extraction stages are specifically as follows:
(2.1) coding the natural attribute of the user to obtain a natural attribute code v of the user;
(2.2) mapping the user natural attribute code v to a feature space to obtain a user natural attribute feature representation fattr(ii) a The user natural attribute code v is mapped to the feature space by using a linear transformation mode, and the formula is as follows:
fattr＝Wa Tv
wherein, WaIs a mapping matrix;
(2.3) coding the terminal attribute to obtain a terminal attribute code l;
(2.4) constructing a user history change machine information code r according to the user history change machine record and the corresponding terminal attribute code;
(2.5) mapping the user history change machine information code r to a feature space to obtain a user history change machine information feature representation fterm(ii) a Wherein, the user change information code r is mapped to the feature space by using a linear transformation mode, and the formula is as follows:
fterm＝Wt Tr
wherein, WtIs a mapping matrix;
(2.6) according to the user conversation and flow use behavior data after the missing value completion and the abnormal value elimination, constructing a conversation behavior time sequence ξ with the time interval of 1 day and the span of T dayscallAnd traffic usage behavior timing ξdataCoarsening and normalizing the granularity g;
(2.7) constructing two multi-layer LSTM networks with the same structure, wherein each layer comprises T/g LSTM units, and carrying out coarsening and normalization on conversation behavior time sequence ξcallAnd traffic usage behavior timing ξdataRespectively input into two LSTM networks, and extracting the time sequence characteristic f of the conversation behaviorcallAnd flow usage behavior timing characteristics fdata。
Preferably, the method for obtaining the user natural attribute code v includes:
(2.1.1) directly performing one-hot encoding on the discrete attributes;
(2.1.2) dividing continuous attributes except age into 5 intervals through equal frequency sub-boxes and then carrying out one-hot coding;
(2.1.3) regarding the age attribute, considering that users in different age groups have different preference of changing machines, dividing the age into 8 intervals, and then carrying out unique hot coding; wherein 8 intervals are under 16 years old, 16 to 21 years old, 22 to 27 years old, 28 to 33 years old, 34 to 39 years old, 40 to 45 years old, 46 to 51 years old, and over 51 years old, respectively;
and (2.1.4) splicing the one-hot codes of all the user natural attributes to obtain a user natural attribute code v.
Preferably, the method for obtaining the terminal attribute code l includes:
(2.3.1) directly one-hot coding discrete attributes other than brand;
(2.3.2) counting the use distribution of the terminal brands, dividing the terminal brands into 10 categories for coding, wherein the brands of 9 bits before the use proportion are divided into 9 categories, and the rest terminal brands are classified into other categories;
(2.3.3) dividing the continuous attribute into 5 intervals through equal frequency sub-boxes and then carrying out one-hot coding;
and (2.3.4) splicing the one-hot codes of all the terminal attributes to obtain a terminal attribute code l.
Preferably, the method for constructing the user history change machine information code r is as follows:
(2.4.1) assume that user u's historical switch machine records are:
<u，s1，t1>，<u，s2，t2>，…，<u，sk，tk>the current date is tnowB, carrying out the following steps of; dividing the terminal corresponding to the user history change machine record into user history use terminals { s1，…，sk-1And currently used terminal sk；
(2.4.2) encoding a plurality of terminal attributes corresponding to the terminals used by the user history
Maximum pooling processing is carried out to obtain historical use terminal codes lhisIndicating a user's historical preference for the terminal;
(2.4.3) encoding the terminal attribute corresponding to the terminal currently used by the user
As the currently used terminal code lnowIndicating the current preference of the user for the terminal;
(2.4.4) calculating the historical average change time interval p of the userhisAnd the current terminal usage duration pnowThe formula is as follows:
pnow＝interval(tk，tnow)
wherein interval (·,) represents the number of days between two dates;
(2.4.5) terminal code l used by user historyhisCurrently used terminal code lnowNormalized historical average change machine time interval phisAnd the normalized current terminal use time pnowAnd splicing to obtain the historical user change information code r.
Preferably, the specific steps of the coarsening and the normalization include:
(2.6.1) performing maximum and minimum normalization processing on each sequence to normalize the processed data to [0, 1%]X is the original value, xmaxIs the maximum value of the sequence in which the value lies, xminFor the minimum value of the sequence in which the value lies, the formula is as follows:
(2.6.2) calculating the average value of the behavior time sequence with the time span of T days every g days, and coarsening the sequence with the original step size of T to the step size of T/g.
Preferably, the step (2.7) adopts a three-layer LSTM network to extract the time sequence characteristics of the user conversation and traffic use behaviors; the LSTM network is a cyclic neural network, and each LSTM unit comprises a memory unit ctAnd three gates: input door itOutput gate otAnd forget door ftRespectively controlling the input, the output and the updating of data; with xtFor input at time t, ht-1And ct-1The calculation formula for the hidden state and the memory unit state at the previous moment is as follows:
it＝sigm(wxixt+Whiht-1+bi)
ft＝sigm(Wxfxt+Whfht-1+bf)
ot＝sigm(Wxoxt+Whoht-1+Wcoct-1+bo)
wherein the operator
The point multiplication operation is represented, W and b represent a weight matrix and a bias vector respectively, and sigm and tanh represent a sigmoid function and a hyperbolic tangent function respectively.
Preferably, the feature fusion and prediction stage is specifically as follows:
(3.1) characterizing the natural attributes of the user by fattrAnd f, representing the information characteristics of the user history change machinetermTime sequence characteristic f of conversation behaviorcallAnd flow usage behavior timing characteristics fdataSplicing to obtain a depth characteristic f;
and (3.2) sending the depth feature f into the full-connection network, and predicting whether the user will change the machine within the future T' days.
Preferably, in the step (3.2), the full-connection network is used for processing the depth feature x, and the probability p of switching the machine within the future T' days of the user is output; the activation function of a hidden layer of the full-connection network is ReLU, the neuron number of an output layer is 1, the activation function is Sigmoid, the range of an output result is [0, 1], and the probability of changing the machine in the future T' day of a user is represented; the nonlinear capability of the model can be enhanced by stacking a plurality of fully-connected layers, and in order to balance the fitting capability and the complexity of the model, a three-layer fully-connected network is adopted; wherein the full connectivity layer maps the input h using a non-linear function ReLU activation according to:
z＝ReLU(WTh)
where W is the full link layer weight matrix and z is the output of the full link layer.
The invention has the beneficial effects that: (1) the invention effectively utilizes the natural attributes of the users, the call behavior data, the traffic use behavior data and the abundant information contained in the historical change machine record, and can fully mine the factors influencing the change machine of the users, so that the model has stronger performance and higher generalization capability; (2) the invention utilizes the LSTM network to extract the time sequence characteristics of the user conversation and flow using behaviors, the depth characteristics are automatically obtained in a data driving mode, the change trend of the user behavior sequence can be modeled in a display mode, and the information loss caused by manually defining the characteristics is avoided.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention;
FIG. 2 is a schematic diagram of user behavior sequence feature extraction according to an embodiment of the present invention;
FIG. 3 is a schematic illustration of feature fusion for an embodiment of the present invention.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto:
example (b): as shown in fig. 1, a multi-factor perception terminal switch prediction method based on a deep neural network is divided into three stages, namely data acquisition and preprocessing, feature construction and extraction, and feature fusion and prediction, and specifically includes the following steps:
(1) data acquisition and processing stage: the specific steps of data acquisition and preprocessing are as follows:
Step 2, collecting relevant terminal attributes, including: brand, operating system, price (dollar), screen size (inches), whether 4G and rear camera pixels are supported.
And 3, collecting related user call behavior data (including daily call out times, daily call out time, daily call in times and daily call in time) and flow use behavior data (including daily flow use times, daily flow use time, daily uplink flow and daily downlink flow), and performing missing value completion and abnormal value elimination processing.
And detecting missing values and abnormal values of the user call and the traffic using behavior data, and filling the missing values and the abnormal values by using a linear interpolation method.
And 4, collecting historical switch change records of the user.
The change machine record is represented in the form of a triplet < u, s, t >, where u is the user id, s is the terminal id, and t is the date on which user u used terminal s the earliest.
(2) A characteristic construction and extraction stage; the specific steps of the feature construction and extraction stage are as follows:
The specific steps of constructing the user natural attribute code comprise:
a) one-hot encoding is done directly for discrete attributes (e.g., gender, customer star rating).
b) And dividing continuous attributes except age (such as the network access time and the number of CDMA under the same client) into 5 intervals through equal frequency sub-boxes, and then carrying out unique hot coding.
c) Regarding the age attribute, considering that users in different age groups prefer different switches, the ages are divided into 8 intervals (16 years old or less, 16 to 21 years old, 22 to 27 years old, 28 to 33 years old, 34 to 39 years old, 40 to 45 years old, 46 to 51 years old, and 51 years old or more), and then are subjected to unique heat coding.
d) And splicing the one-hot codes of all the user natural attributes to obtain a user natural attribute code v.
Step 2, mapping the user natural attribute code v to a feature space to obtain a user natural attribute feature representation fattr。
Mapping the user natural attribute code v to a feature space by using a linear transformation mode, wherein the formula is as follows:
fattr＝Wa Tv (1)
wherein, WaIs a mapping matrix.
And 3, coding the terminal attribute to obtain a terminal attribute code l.
The specific steps of constructing the terminal attribute code comprise:
a) and directly carrying out one-hot coding on discrete attributes (such as an operating system and whether 4G is supported) except brands.
b) Given the large number of terminal brands, directly encoding them one-hot can cause a dimensional disaster. And (4) counting the use distribution of the terminal brands, dividing the terminal brands into 10 categories for coding, wherein the brands using the first 9 bits of the proportion are divided into 9 categories, and the rest terminal brands are classified into other categories.
c) And dividing continuous attributes (such as price and screen size) into 5 intervals through equal frequency sub-boxes and then carrying out one-hot coding.
d) And splicing the one-hot codes of all the terminal attributes to obtain a terminal attribute code l.
And 4, constructing a user history change machine information code r according to the user history change machine record and the corresponding terminal attribute coding mode.
Suppose the historical change machine record for user u is:
<u，s1，t1>，<u，s2，t2>，…，<u，sk，tk>the current date is tnow. The specific steps for constructing the user history change machine information code comprise:
a) dividing the terminal corresponding to the user history change machine record into user history use terminals { s1，…，sk-1And currently used terminal sk。
b) Encoding a plurality of terminal attributes corresponding to the terminal used by the user history
Maximum pooling processing is carried out to obtain historical use terminal codes lhisIndicating the user's historical preference for the terminal.
c) Coding the terminal attribute corresponding to the terminal currently used by the user
As the currently used terminal code lnowIndicating the current preference of the user for the terminal.
d) Calculating historical average change time interval p of userhisAnd the current terminal usage duration pnow，
The formula is as follows:
pnow＝interval(tk，tnow) (3)
where interval (·, ·) represents the number of days between two dates.
e) Terminal code l used by user historyhisCurrently used terminal code lnowNormalized historical average change machine time interval phisAnd the normalized current terminal use time pnowAnd splicing to obtain the historical user change information code r.
Step 5, mapping the user history change machine information code r to a characteristic space to obtain a user history change machine information characteristic representation fterm。
Mapping the user change information code r to a feature space by using a linear transformation mode, wherein the formula is as follows:
fterm＝Wt Tr (4) wherein WtIs a mapping matrix.
Step 6, according to the user conversation and flow use behavior data after missing value completion and abnormal value elimination, constructing a conversation behavior time sequence ξ with the time interval of 1 day and the span of T dayscallAnd traffic usage behavior timing ξdataAnd performing coarsening treatment and normalization treatment with the granularity of g.
The specific steps of the coarsening and the normalization processing comprise:
a) performing maximum and minimum normalization processing on each sequence to normalize the processed data to [0, 1%]X is the original value, xmaxIs the maximum value of the sequence in which the value lies, xminFor the minimum value of the sequence in which the value lies, the formula is as follows:
b) when the model is trained, the sequence with too many steps can significantly increase the complexity of model calculation. And calculating the average value of the behavior time sequence with the time span of T days every g days, and coarsening the sequence with the original step length of T to the step length of T/g.
Step 7, constructing two multi-layer LSTM networks with the same structure, wherein each layer comprises s ═ T/g LSTM units, and performing coarsening and normalization on the conversation behavior time sequence ξcallAnd traffic usage behavior timing ξdataRespectively input into two LSTM networks, and extracting the time sequence characteristic f of the conversation behaviorcallAnd flow usage behavior timing characteristics fdata。
To effectively extract the features of time series data, a multi-layer LSTM network is generally adopted to enhance the nonlinear capability of the model. In order to balance the fitting capability and complexity of the model, three layers of LSTM networks are adopted to extract the time sequence characteristics of the user conversation and traffic use behaviors.
The LSTM network is a kind of recurrent neural network, and each LSTM unit contains a memory unit ctAnd three gates: input door itOutput gate otAnd forget door ftAnd respectively controlling the input, the output and the updating of data. With xtFor input at time t, ht-1And ct-1The calculation formula for the hidden state and the memory unit state at the previous moment is as follows:
it＝sigm(Wxixt+Whiht-1+bi) (6)
ft＝sigm(Wxfxt+Whfht-1+bf) (7)
ot＝sigm(Wxoxt+Whoht-1+Wcoct-1+bo) (9)
wherein the operator
The point multiplication operation is represented, W and b represent a weight matrix and a bias vector respectively, and sigm and tanh represent a sigmoid function and a hyperbolic tangent function respectively.
As shown in FIG. 2, the coarsened and normalized conversation activity timing ξcallAnd traffic usage behavior timing ξdataAnd s-T/g values are contained and input into corresponding LSTM units respectively. In the LSTM network, the state of the last moment is input into the next LSTM unit, the time sequence information of the data is reserved, the vector output by the last layer of LSTM is activated by using a Sigmoid function, and the time sequence characteristic f of the call behavior is extractedcallAnd flow usage behavior timing characteristics fdata。
(3) And (3) feature fusion and prediction stage:
And 2, sending the depth characteristic f into a full-connection network, and predicting whether the user will change the machine within the future T' days.
As shown in fig. 3, the depth feature x is processed by using the full-connection network, and the probability p of changing the machine within the future T' days of the user is output. The activation function of the hidden layer of the full-connection network is ReLU, the neuron number of the output layer is 1, the activation function is Sigmoid, the range of the output result is [0, 1], and the probability of changing the machine in the future T' day of the user is represented. The nonlinear capability of the model can be enhanced by stacking a plurality of fully-connected layers, and in order to balance the fitting capability and complexity of the model, a three-layer fully-connected network is adopted. Wherein the full connectivity layer maps the input h using a non-linear function ReLU activation according to:
z＝ReLU(WTh) (11)
where W is the full link layer weight matrix and z is the output of the full link layer.
While the invention has been described in connection with specific embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (12)
1. A multi-factor perception terminal switching prediction method based on a deep neural network is characterized by comprising three stages of data acquisition and preprocessing, feature construction and extraction, and feature fusion and prediction, and specifically comprises the following steps:
(1) data acquisition and preprocessing stage: acquiring terminal attributes and related natural user attributes, call behavior data, traffic use behavior data and historical change records, and preprocessing the terminal attributes and the related natural user attributes;
(2) and (3) feature construction and extraction stages: coding the preprocessed user natural attribute and terminal attribute to obtain a user natural attribute code and a terminal attribute code, constructing a user history change information code according to a user history change record and the corresponding terminal attribute code, constructing a user conversation and flow use behavior time sequence, and performing coarsening and normalization processing; and utilizes LSTM network to extract time sequence characteristics of user's talking and traffic using behaviors;
(3) and (3) feature fusion and prediction stage: and splicing the natural attribute code of the user, the historical switch-over information code, the conversation behavior time sequence characteristic and the flow use behavior time sequence characteristic, and then sending the spliced result into the full-connection network to predict whether the user can change the switch.
2. The method according to claim 1, wherein the method comprises the following steps: the data acquisition and preprocessing stage is specifically as follows:
(1.1) collecting relevant user natural attributes, including: gender, age, network access duration, client star level, whether the user is an in-person network user, the number of CDMA (code division multiple access) users under the same client and App (application profile) use preference;
(1.2) collecting relevant terminal attributes, including: brand, operating system, price, screen size, whether 4G and rear camera pixels are supported;
(1.3) collecting related user call behavior data and flow use behavior data, and performing missing value completion and abnormal value elimination processing; the user call behavior data comprises daily call times, daily call duration, daily call times and daily call duration, and the flow use behavior data comprises daily flow use times, daily flow use duration, daily uplink flow and daily downlink flow;
and (1.4) collecting user history change machine records.
3. The method according to claim 2, wherein the method comprises the following steps: in the step (1.3), missing value and abnormal value detection is carried out on the user call and the traffic using behavior data, and a linear interpolation method is used for filling the missing value and the abnormal value.
4. The method according to claim 2, wherein the method comprises the following steps: in the step (1.4), the historical user change record is represented in a form that a triple is less than u, s and t, wherein u is a user id, s is a terminal id, and t is a date when the user u uses the terminal s earliest.
5. The method according to claim 1, wherein the method comprises the following steps: the characteristic construction and extraction stage is as follows:
(2.1) coding the natural attribute of the user to obtain a natural attribute code v of the user;
(2.2) mapping the user natural attribute code v to a feature space to obtain a user natural attribute feature representation fattr(ii) a The user natural attribute code v is mapped to the feature space by using a linear transformation mode, and the formula is as follows:
fattr＝WaTv
wherein, WaIs a mapping matrix;
(2.3) coding the terminal attribute to obtain a terminal attribute code l;
(2.4) constructing a user history change machine information code r according to the user history change machine record and the corresponding terminal attribute code;
(2.5) mapping the user history change machine information code r to a feature space to obtain a user history change machine information feature representation fterm(ii) a Wherein, the user change information code r is mapped to the feature space by using a linear transformation mode, and the formula is as follows:
fterm＝Wt Tr
wherein, WtIs a mapping matrix;
(2.6) according to the user conversation and flow use behavior data after the missing value completion and the abnormal value elimination, constructing a conversation behavior time sequence ξ with the time interval of 1 day and the span of T dayscallAnd traffic usage behavior timing ξdataCoarsening and normalizing the granularity g;
(2.7) constructing two multi-layer LSTM networks with the same structure, wherein each layer comprises T/g LSTM units, and carrying out coarsening and normalization on conversation behavior time sequence ξcallAnd traffic usage behavior timing ξdataRespectively input into two LSTM networks, and extracting the time sequence characteristic f of the conversation behaviorcallAnd flow usage behavior timing characteristics fdata。
6. The method according to claim 5, wherein the method comprises the following steps: in the step (2.1), the method for obtaining the user natural attribute code v includes:
(2.1.1) directly performing one-hot encoding on the discrete attributes;
(2.1.2) dividing continuous attributes except age into 5 intervals through equal frequency sub-boxes and then carrying out one-hot coding;
(2.1.3) regarding the age attribute, considering that users in different age groups have different preference of changing machines, dividing the age into 8 intervals, and then carrying out unique hot coding; wherein 8 intervals are under 16 years old, 16 to 21 years old, 22 to 27 years old, 28 to 33 years old, 34 to 39 years old, 40 to 45 years old, 46 to 51 years old, and over 51 years old, respectively;
and (2.1.4) splicing the one-hot codes of all the user natural attributes to obtain a user natural attribute code v.
7. The method according to claim 5, wherein the method comprises the following steps: in the step (2.3), the method for obtaining the terminal attribute code l includes:
(2.3.1) directly one-hot coding discrete attributes other than brand;
(2.3.2) counting the use distribution of the terminal brands, dividing the terminal brands into 10 categories for coding, wherein the brands of 9 bits before the use proportion are divided into 9 categories, and the rest terminal brands are classified into other categories;
(2.3.3) dividing the continuous attribute into 5 intervals through equal frequency sub-boxes and then carrying out one-hot coding;
and (2.3.4) splicing the one-hot codes of all the terminal attributes to obtain a terminal attribute code l.
8. The method according to claim 5, wherein the method comprises the following steps: in the step (2.4), the method for constructing the user history change machine information code r is as follows:
(2.4.1) assume that user u's historical switch machine records are: < u, s1，t1＞，＜u，s2，t2＞，…，＜u，sk，tk>. The current date is tnow(ii) a Corresponding the user history to the change recordIs divided into user history use terminals s1，…，sk-1And currently used terminal sk；
(2.4.2) encoding a plurality of terminal attributes corresponding to the terminals used by the user history
Maximum pooling processing is carried out to obtain historical use terminal codes lhisIndicating a user's historical preference for the terminal;
(2.4.3) encoding the terminal attribute corresponding to the terminal currently used by the user
As the currently used terminal code lnowIndicating the current preference of the user for the terminal;
(2.4.4) calculating the historical average change time interval p of the userhisAnd the current terminal usage duration pnowThe formula is as follows:
pnow＝interval(tk，tnow)
wherein interval (·,) represents the number of days between two dates;
(2.4.5) terminal code l used by user historyhisCurrently used terminal code lnowNormalized historical average change machine time interval phisAnd the normalized current terminal use time pnowAnd splicing to obtain the historical user change machine information code r.
9. The method according to claim 5, wherein the method comprises the following steps: in the step (2.6), the specific steps of the user behavior sequence coarsening and normalization processing include:
(2.6.1) performing maximum and minimum normalization processing on each sequence to normalize the processed data to[0，1]X is the original value, xmaxIs the maximum value of the sequence in which the value lies, xminFor the minimum value of the sequence in which the value lies, the formula is as follows:
(2.6.2) calculating the average value of the behavior time sequence with the time span of T days every g days, and coarsening the sequence with the original step size of T to the step size of T/g.
10. The method according to claim 5, wherein the method comprises the following steps: in the step (2.7), three-layer LSTM network is adopted to extract the time sequence characteristics of the user conversation and flow using behaviors; the LSTM network is a cyclic neural network, and each LSTM unit comprises a memory unit ctAnd three gates: input door itOutput gate otAnd forget door ftRespectively controlling the input, the output and the updating of data; with xtFor input at time t, ht-1And ct-1The calculation formula for the hidden state and the memory unit state at the previous moment is as follows:
it＝sigm(Wxixt+Whiht-1+bi)
ft＝sigm(Wxfxt+Whfht-1+bf)
ot＝sigm(Wxoxt+Whoht-1+Wcoct-1+bo)
wherein the operator
The point multiplication operation is represented, W and b represent a weight matrix and a bias vector respectively, and sigm and tanh represent a sigmoid function and a hyperbolic tangent function respectively.
11. The method according to claim 1, wherein the method comprises the following steps: the characteristic fusion and prediction stage is specifically as follows:
(3.1) characterizing the natural attributes of the user by fattrAnd f, representing the information characteristics of the user history change machinetermTime sequence characteristic f of conversation behaviorcallAnd flow usage behavior timing characteristics fdataSplicing to obtain a depth characteristic f;
and (3.2) sending the depth feature f into the full-connection network, and predicting whether the user will change the machine within the future T' days.
12. The method according to claim 11, wherein the method comprises the following steps: in the step (3.2), the depth feature x is processed by using the full-connection network, and the probability p of changing the machine in the future T' days of the user is output; the activation function of a hidden layer of the full-connection network is ReLU, the neuron number of an output layer is 1, the activation function is Sigmoid, the range of an output result is [0, 1], and the probability of changing the machine in the future T' day of a user is represented; the nonlinear capability of the model can be enhanced by stacking a plurality of fully-connected layers, and in order to balance the fitting capability and the complexity of the model, a three-layer fully-connected network is adopted; wherein the full connectivity layer maps the input h using a non-linear function ReLU activation according to:
z＝ReLU(WTh)
where W is the full link layer weight matrix and z is the output of the full link layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911135219.1A CN110991711A (en) | 2019-11-19 | 2019-11-19 | Multi-factor perception terminal switching prediction method based on deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911135219.1A CN110991711A (en) | 2019-11-19 | 2019-11-19 | Multi-factor perception terminal switching prediction method based on deep neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110991711A true CN110991711A (en) | 2020-04-10 |
Family
ID=70085086
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911135219.1A Pending CN110991711A (en) | 2019-11-19 | 2019-11-19 | Multi-factor perception terminal switching prediction method based on deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110991711A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112785069A (en) * | 2021-01-29 | 2021-05-11 | 北京明略昭辉科技有限公司 | Prediction method and device for terminal equipment changing machine, storage medium and electronic equipment |
| CN116611860A (en) * | 2022-11-30 | 2023-08-18 | 天翼数字生活科技有限公司 | Method and system for predicting terminal machine changing based on width learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH11290532A (en) * | 1998-04-06 | 1999-10-26 | Tatsunoko:Kk | Sales prediction control device and machine change and business change control device for game machine |
| CN106845731A (en) * | 2017-02-20 | 2017-06-13 | 重庆邮电大学 | A kind of potential renewal user based on multi-model fusion has found method |
| CN109583949A (en) * | 2018-11-22 | 2019-04-05 | 中国联合网络通信集团有限公司 | A kind of user changes planes prediction technique and system |
| CN109583659A (en) * | 2018-12-07 | 2019-04-05 | 南京富士通南大软件技术有限公司 | User's operation behavior prediction method and system based on deep learning |
-
2019
- 2019-11-19 CN CN201911135219.1A patent/CN110991711A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH11290532A (en) * | 1998-04-06 | 1999-10-26 | Tatsunoko:Kk | Sales prediction control device and machine change and business change control device for game machine |
| CN106845731A (en) * | 2017-02-20 | 2017-06-13 | 重庆邮电大学 | A kind of potential renewal user based on multi-model fusion has found method |
| CN109583949A (en) * | 2018-11-22 | 2019-04-05 | 中国联合网络通信集团有限公司 | A kind of user changes planes prediction technique and system |
| CN109583659A (en) * | 2018-12-07 | 2019-04-05 | 南京富士通南大软件技术有限公司 | User's operation behavior prediction method and system based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 刘刚, 人民交通出版社 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112785069A (en) * | 2021-01-29 | 2021-05-11 | 北京明略昭辉科技有限公司 | Prediction method and device for terminal equipment changing machine, storage medium and electronic equipment |
| CN116611860A (en) * | 2022-11-30 | 2023-08-18 | 天翼数字生活科技有限公司 | Method and system for predicting terminal machine changing based on width learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11531867B2 (en) | User behavior prediction method and apparatus, and behavior prediction model training method and apparatus | |
| CN109034398A (en) | Feature selection approach, device and storage medium based on federation's training | |
| CN108573399B (en) | Merchant recommendation method and system based on transition probability network | |
| US20170091810A1 (en) | Brand engagement touchpoint attribution using brand engagement event weighting | |
| CN109741114A (en) | A kind of user under big data financial scenario buys prediction technique | |
| CN106997549A (en) | The method for pushing and system of a kind of advertising message | |
| CN110008408A (en) | A kind of session recommended method, system, equipment and medium | |
| CN110955828B (en) | Multi-factor embedded personalized package recommendation method based on deep neural network | |
| CN110825968B (en) | Information pushing method, device, storage medium and computer equipment | |
| CN112508118B (en) | Target object behavior prediction method aiming at data offset and related equipment thereof | |
| CN116010684A (en) | Article recommendation method, device and storage medium | |
| CN107507036A (en) | The method and terminal of a kind of data prediction | |
| CN109242555B (en) | Voice-based advertisement playing method and related product | |
| US20210081755A1 (en) | Method, apparatus, electronic device and storage medium for predicting user attribute | |
| CN112633962A (en) | Service recommendation method and device, computer equipment and storage medium | |
| CN108133390A (en) | For predicting the method and apparatus of user behavior and computing device | |
| CN110991711A (en) | Multi-factor perception terminal switching prediction method based on deep neural network | |
| CN110766438A (en) | Method for analyzing user behaviors of power grid users through artificial intelligence | |
| CN116468460B (en) | Consumer finance customer image recognition system and method based on artificial intelligence | |
| CN106909560A (en) | Point of interest sort method | |
| CN110704510A (en) | User portrait combined question recommendation method and system | |
| CN107784511A (en) | A kind of customer loss Forecasting Methodology and device | |
| CN112149352A (en) | Prediction method for marketing activity clicking by combining GBDT automatic characteristic engineering | |
| CN118134630A (en) | Credit risk level assessment method and device and electronic equipment | |
| CN112465565A (en) | User portrait prediction method and device based on machine learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200410 |