[0054]The invention is further illustrated by detail a preferred embodiment in detail below with reference to the accompanying drawings.
[0055]Such asfigure 1 As shown, for the present invention, a movie click rate prediction method based on the field interactive information intensity factor dispeckler, which can process the MOVIELENS disclosure data set data, and according to the modeling user feature information, film category characteristics, etc. according to the present invention. Based on the interaction between the characteristics and features based on the field interaction information, it is estimated.
[0056]Specifically, the establishment method of the model includes: S1, selected data set as a data sample, pre-processing data samples, dividing pre-processed data samples into training sets and test sets.
[0057]Specifically, in the present embodiment, the MOVIELENS-1M data set is selected as a data sample. Before the data sample is preprocessing, the data in the selected data set is integrated, and the isnull () function is used to see if there is a missing value, and the data samples of the absence of too much data sample are removed to avoid predictive deviations. Then, the above data is processed into data that conforms to the model input format, and the pre-processing of data samples during this process.
[0058]Example, integration of data in the selected data set can be specifically: The original data file directory is shown in Table 1, including:
[0059]Table 1
[0060] file name Data category User.data User_id, gender, agn, occuptation Ratings.dat User_id, name_id, rating, timestamp Movies.dat Movie_id, Title, Genres
[0061]Integrate multiple data files together, and the format is as follows:
[0062]Table 2
[0063]
[0064]Further, the pretreatment operation specifically includes: converting discrete data in the data sample to a single-hot vector, converting the continuous data in the data sample into a sequence of sequence length, and it is obtained to obtain a total pre-pre-processed data set, This data set is in line with the input format of the movie click rate. Wherein, the unique thermal encoding in the unique vector uses 0 and 1 to indicate the parameters, and N-bit registers are encoded using the n-bit state register.
[0065]In the present embodiment, data samples are prepared using the SklearN library in Python and the Feature_Column method in the DeepCtr library. Of course, the method used by the pretreatment is not limited to the above, but also methods of other pretreatment purposes.
[0066]In the present embodiment, 80% of the data samples in the pre-treated data sample are used as the training set for the training set, 20% of data as the test set, and evaluate the predicted results of the model.
[0067]S2, data integration of the data of the training set to obtain a low-dimensional thickening.
[0068]Further, the step S2 specifically: compresses the data of the training set through the model embedding layer to the low-dimensional thickening direction. Among them, iffigure 2 As shown, the model is embedded in a fully connected neural network, the model embedded layer for converting the encoded sparse data and a dense source of the specified low dimension. The low-dimensional thickening moiety can be represented as: a(0)= [E1, E2..., Em-1, Em]. In the present embodiment, the dimension of the low-dimensionally thickening vector is 4 dimension, ie M = 4.
[0069]S3, using the low-dimensional thickening vector training based on the field interactive information intensity factor breakdown model.
[0070]The film click rate predictive model (DEEPFWFM) of the present invention is a deep width prediction model, which includes two partial parallel processing, including the FWFM module and DNN module, and two parts share the same input data.
[0071]Such asimage 3 As shown, specifically, the step S3 specifically includes:
[0072]In the FWFM module, the low-dimensional thickening vectors obtained by step S2 are input to the factor-based fractal machine model FWFM based on the domain interactive intensity information (ie, the model feature intertteral layer) performs low-order feature interaction, and outputs the output data to Attention mechanism (attention mechanism) layer is weighted to obtain the prediction result of the FWFM module. (Please seeFigure 4 )
[0073]Compared with the ordinary feature combination model, the movie click rate estimates of the present invention adds self-focus mechanisms and interaction interaction between domain information, and can significantly improve the accuracy of the hit rate estimate.
[0074]Among them, the mechanism of the Attention mechanism is: the model of joining the Attention mechanism layer can learn the degree of influence of user history tendency to preference for user current behavior. For example, the user browses furniture and clothes in the previous period, and the desk advertisement presentation to the user will be greatly affected by the user's furniture, which is less affected by the user.
[0075]Implementation of the attention, using the multiplier care mechanism, learn the weight matrix H by learning user history preferencesT, W indicates the initial weight of the attention mechanism model, AIJ Indicates the attention value, which can be interpreted as the degree of influence on the target predictive value after the later, and use the RELU function to activate, and then use the SoftMax function to use the SoftMax function to normalize the resultant value.
[0076]For the importance of predictive target, the Attention mechanism calculation method of the Attention mechanism layer is:
[0077]
[0078]
[0079]Among them, AIJ Indicates the final attention value, which can be interpreted as the interaction Weight Weight We in the feature component I and the feature component j.IJ A 'IJ Indicates the activation value after the feature component passes through the attention network activation function (RELU function); HTFor the weight matrix; W is an Attention mechanism layer initial weight; Xi, XjThe Group I and the JU-J column are respectively, that is, the characteristic component; Vi, VjRepresents X, respectivelyi, XjCorresponding hidden vector, i, VjIndicates the volume of the hidden vector; F (i), f (j) represents the field of feature components I, feature component J, RF (i), f (j) For weight, it is used to build the interaction interaction between the model F (i) and the domain f (j); b∈Rt, B is the model parameters, b is the T-Tenni set, and R is a real set, t is the model implicit layer.
[0080]Further, the factor of the field-based interactive intensity information is specifically that the interaction interaction between different domains is considered on the basis of the conventional FM algorithm, and the interaction interaction between different domains is given a unified weight. To reduce parameters. The factor of field-based interactive intensity information is used to add the interaction interaction of domain information on the basis of a conventional FM algorithm, and the formula of Factor Decomposition Machine Model FWFM based on domain interactive interaction information is:
[0081]
[0082]Where w0For the bias item weight, W0∈R, For linear combination, it is used to extract a single-order feature and a separate weight of the domain information; For interaction part, used to consider interaction between field information and feature information, w∈RnFor the one-time coefficient, m is the total characteristic dimension.
[0083]Meanwhile, in the DNN module, the low-dimensional thickening vectors obtained by step S2 are input to the depth neural network DNN, and high-order feature intense interaction is performed by the feedforward neural network, the depth neural network DNN, to obtain the prediction result of the DNN module.
[0084]In the present embodiment, the depth of the depth neural network DNN is two layers, and the number of neurons per layer is 128 and 128, respectively, and we use the RELU activation function at each fully attachment layer. The number of output vectors of the depth neural network DNN is the number of its last layer of neurons, in the present embodiment, the last number of neurons is 128.
[0085]The prediction result of the FWFM module and the predicted result of the DNN module activates the predicted result of the movie click rate predictive model, specifically, the predicted result of the FWFM module and the predicted result of the DNN module form a new vector, and then The processing is performed via the activation function SIGMOID as the final prediction result. The predicted result of the movie click rate estimated model is:
[0086]
[0087]among them, Indicates the predicted result of the movie click rate predictive model, YFWFM represents the prediction result of the FWFM module, and YDNN represents the predicted result of the DNN module, and the Sigmoid function is an existing function.
[0088]S4, using the test set and evaluation index to perform the movie click rate estimated model test for step S3 training.
[0089]The evaluation indicators include: Accuracy, 000 root error (MSE), and Log-Cosh loss functions, using the above indicators to measure the advantages and disadvantages of the model.
[0090]Wherein, the accuracy is calculated:
[0091]
[0092]Among them, Accuracy is the calculated accuracy, P is expressed as the total amount of actual spoiler, and the data sample is specified in the user, n is expressed as the actual negative sample, and the negative sample means that the user does not click on the data sample, TP representation prediction The correct and predicted values are positive samples, TN represents the predicted correct and the predicted value is negative sample.
[0093]The average square root error is a mean of the prediction data and the square error of the original data corresponding point error, and the calculation formula of the normal root error is:
[0094]
[0095]Among them, MSE is calculated, the root error, YiIndicates the true value of the data set: Yi= {Y1Y2, ... yn} Indicates the prediction result obtained by the model:
[0096]The Log-Cosh loss function is a loss function applied to the regression task, which is smoother than the existing L2 regression loss function, which is in line with the formula:
[0097]
[0098]Among them, log-cosh () is the logarithm of a hypoth of the predictive error.
[0099]In summary, a film click rate estimation method based on the domain interactive information intensity factor decomposition machine is based on the field-based interactive information intensity factor decomposition model, considering different fields. The application of interaction in the movie click, first in the data pretreatment portion, divides the data into discrete and continuous type, thereby converting discrete data into a numerical type, and then converts the continuous data into a sequence of sequence lengths. Input, the pre-processed data is compressed by the model embedding layer to the low dimensional thickening vector, which obtains a new vector by the movie click rate. Finally, the test set and evaluation indicators are used to verify the movie click rate of the training result to obtain the final training or predictive results. This method considers the communication-related feature field as a weighted feature, and consistently consistently consistently consolidation of interaction interaction relationship between different domains to facilitate the correlation of user interest and movie characteristics.
[0100]Further, this method considers the interaction strength between the fields and the interaction interaction training between the field, indicating the strength of interaction between the field to enhance the accuracy of the model of the movie hits.
[0101]Although the content of the present invention has been described in detail by the preferred embodiments, it should be appreciated that the above description should not be construed as limiting the invention. The above will be apparent to those skilled in the art, and various modifications and alternatives to the present invention will be apparent. Accordingly, the scope of the invention should be limited by the appended claims.