CN111159242B

CN111159242B - Client reordering method and system based on edge calculation

Info

Publication number: CN111159242B
Application number: CN201911390108.5A
Authority: CN
Inventors: 范俊; 顾湘余; 李文杰
Original assignee: Hangzhou Xiaoying Innovation Technology Co ltd
Current assignee: Hangzhou Xiaoying Innovation Technology Co ltd
Priority date: 2019-12-27
Filing date: 2019-12-27
Publication date: 2023-04-25
Anticipated expiration: 2039-12-27
Also published as: CN111159242A

Abstract

The invention discloses a client reordering method and a system based on edge calculation, wherein the reordering method comprises the following steps: s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information; s2, randomly arranging the K pieces of information to obtain K-! Ordering the second information; s3, extracting characteristics of recommended information and context information of the recommended information in each second information sequence; and S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate predicted values, and selecting the second information sequence corresponding to the highest click through rate predicted value for display. The method and the system fully consider the display characteristics of each client, reorder the sequences at the clients, and improve the user experience and the recommendation effect.

Description

Client reordering method and system based on edge calculation

Technical Field

The invention relates to the field of recommendation ordering, in particular to a client reordering method and system based on edge calculation.

Background

Rank Learning (LTR) is mainly used in the information retrieval field, and comprehensively considers a plurality of ranking features To Rank search results. LTR is a widely and deeply studied problem, whether in the search, advertising, or recommendation fields. Since the goals of the LTR in most cases are click-through rates, the LTR problem is also referred to as click-through rate (Click Through Rate, CTR) prediction, which refers to the click-through arrival rate of a web advertisement (picture advertisement/text advertisement/keyword advertisement/rank advertisement/information advertisement, etc.), i.e., the actual number of clicks of the advertisement (strictly speaking, the number of reached target pages) divided by the advertisement's presentation amount (Show content).

The common practice for recommending scenes is: the application requests a recommendation list from the server; the server side recalls and sorts the information according to the user characteristics, the information characteristics and the context characteristics, returns the information to an information list which is sorted from high to low according to the click rate, and displays the information to the user according to the order of the list. Whether the recommendation information is clicked is closely related to the order and position when presented to the user. Thus, LTRs are generally classified into the following three types: the advantage of this algorithm is simplicity, assuming that each recommended information is clicked on only itself, irrespective of other information around the information; the parilwise optimizes the partial order relation of the click probabilities of different information, the relationship between each message and the other messages in the list is considered; lisdwise optimizes the global order of the entire list, which has the disadvantage of high algorithm complexity. In the existing recommended environment, considering the requirement of online algorithm performance, most LTRs adopt a pointwise algorithm, and common algorithms include: logistic regression, gbdt, factorization machine, dnn which are currently popular, etc. However, for the pointwise algorithm, documents of the same class cannot be ranked; the relative order between the documents is not considered, calculated entirely from the classification point of view of the single documents.

Furthermore, as previously stated, whether the recommendation information is clicked is related to the location in addition to whether it meets the user interest preferences. The same recommendation list shows a wide variety of styles on different devices. As shown in fig. 1, the same list of recommendations, when presented on different devices, gives the user a completely different visual experience and experience. The same recommended video list is displayed on the left device, and videos of different categories are arranged at intervals, so that people feel more diversified. And the video is displayed on the right side equipment, and the similar videos are gathered on one side, so that the user experience is influenced and then clicking is influenced. However, the display form server side of the recommendation result cannot be determined, especially some H5 pages, allow the user to zoom the display window at any time on the side, so that personalized reordering of the recommendation list on the client side has very important practical significance. However, almost all CTR estimation in the prior art is completed at the server, so that the advantage of strong computing power of the server is fully utilized, and the same recommendation ordering is adopted for different clients, so that the influence of different display forms of different clients is not considered.

Therefore, how to implement information ordering suitable for each client presentation form is a problem to be solved in the art.

Disclosure of Invention

The invention aims at overcoming the defects of the prior art and provides a client reordering method and a system based on edge calculation. The method and the system fully consider the display characteristics of each client, reorder the sequences at the clients, and improve the user experience and the recommendation effect.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

a client reordering method based on edge calculation comprises the following steps:

s1, acquiring a first information sequence which is recommended by a server and comprises k pieces of information;

s2, randomly arranging the K pieces of information to obtain K-! Ordering the second information;

s3, extracting characteristics of recommended information and context information of the recommended information in each second information sequence;

and S4, sequentially inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model to obtain corresponding click through rate predicted values, and selecting the second information sequence corresponding to the highest click through rate predicted value for display.

Further, the features of the recommendation information include: the click passing rate estimation result is output by the information semantic vector, the information category and the server sequencing model; the characteristics of the context information include information surrounding the current recommendation information in the current user interface presentation.

Further, the Attention-LSTM-MLP model is generated at the server and sent to the client.

Further, the generation of the Attention-LSTM-MLP model is specifically as follows:

s41, constructing an Attention-LSTM-MLP network;

s42, extracting characteristics of recommended information and context information thereof, calculating click passing rate according to the click number of the user of the recommended information, and training the Attention-LSTM-MLP network based on the characteristics and the click passing rate.

Further, the Attention-LSTM-MLP network is specifically:

connecting an attribute network behind an hidden layer in the LSTM network; the hidden layer of the LSTM outputs the extracted characteristics and inputs the characteristics into the Attention network; the attribute network is converted into a weight coefficient of each node through a softmax function, the value of each node in the attribute network is multiplied by the weight coefficient to be used as the output of the node, the final code output of the attribute-LSTM is obtained, and the code output is input into the MLP network to obtain the click through rate estimated value.

Further, the loss function of the Attention-LSTM-MLP network is as follows:

wherein y is _i Is the calculated true click through rate, p _i Is a predicted value calculated according to the attribute-LSTM-MLP network, n is the number of samples, y _i The specific calculation is as follows:

and num is the number of information clicked by the user in the recommended information sequence which is displayed and exposed at one time.

Further, the feature of the context information is set to 8 positions, and the missing is filled with default value 0.

The invention also provides a client reordering system based on edge calculation, which comprises a client, and specifically comprises the following steps:

the acquisition module is used for acquiring a first information sequence which is recommended by the server and comprises k pieces of information;

a random ordering module for randomly arranging the K pieces of information to obtain K-! Ordering the second information;

the extraction module is used for extracting the characteristics of the recommended information and the context information of the recommended information in each second information sequence;

the estimating module is used for inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model in sequence to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display.

Further, the reordering system further includes a server, which specifically includes:

the construction module is used for constructing the Attention-LSTM-MLP network;

and the training module is used for extracting characteristics of the recommended information and the context information thereof, calculating the click passing rate according to the click number of the user of the recommended information, and training the Attention-LSTM-MLP network based on the characteristics and the click passing rate.

Compared with the prior art, the invention has the following effects:

(1) According to the method and the system, the recommendation results are reordered on the terminal by combining the display form of the specific client, the display characteristics of the client are fully considered, the same server results are different in ordering of different clients, and the user experience and the recommendation effect are improved;

(2) According to the method, a lisdwise algorithm is adopted to perform global target optimization on the whole list, and because the candidate information to be ordered, which is reordered on the end, is very small, the extra cost introduced by the lisdwise algorithm on the basis of optimizing the ordering is small;

(3) In view of the fact that large-scale users and information features cannot be transmitted to the client, the method and the device have fewer utilized features, mainly the features affecting the vision of the users and the context features, and further optimize the sorting performance;

(4) According to the invention, the preference of the human brain for dispersing and gathering different contents is learned by adding the attention mechanism, so that the data processing amount in the reordering process is reduced, and the data processing efficiency is improved;

(5) The invention fully utilizes the characteristics of the server and the client, trains and generates the Attention-LSTM-MLP model at the server to send to the client, and reduces the burden of the client.

Drawings

FIG. 1 is an example of a display of recommendation information on different devices;

FIG. 2 is a flowchart of a method for reordering clients based on edge computation according to an embodiment;

FIG. 3 is a diagram showing a comparison of recommended information on a web page with a mobile terminal;

FIG. 4 is a block diagram of an Attention-LSTM-MLP network;

fig. 5 is a diagram of a client reordering system based on edge computation according to the second embodiment.

Detailed Description

Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which describes the embodiments of the present invention with reference to specific examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention. It should be noted that the following embodiments and features in the embodiments may be combined with each other without conflict.

It should be noted that the illustrations provided in the following embodiments merely illustrate the basic concept of the present invention by way of illustration, and only the components related to the present invention are shown in the drawings and are not drawn according to the number, shape and size of the components in actual implementation, and the form, number and proportion of the components in actual implementation may be arbitrarily changed, and the layout of the components may be more complicated.

The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.

Example 1

As shown in fig. 2, this embodiment proposes a client reordering method based on edge computation, including:

in the field of information recommendation, a server performs information recall and sorting according to user characteristics, recommended information characteristics and context characteristics, and returns an information column which is sorted from high to low according to the click rate to a client. In order to adapt to the display characteristics of different clients and realize optimal information ordering for each client, the invention reorders the received recommended information ordering at the client.

Aiming at the information ordering of different clients, the display characteristics of the clients need to be considered, and corresponding ordering is generated for each client. The server has a large number of clients connected, which is certainly a huge burden if the respective ordering is customized for each client. With the popularity of mobile devices and the development of the future internet of things, terminal devices are becoming more and more diversified, and the performance on clients is also more and more powerful, which makes it possible to migrate part of the computing tasks from the server to the mobile terminal. Therefore, the invention takes the client as the edge node and uses the edge calculation to realize the reordering of the recommendation information. Edge computation is a distributed computing architecture that handles the computation of applications, data and services by hub nodes moving to edge nodes on the network logic.

Specifically, after receiving the information ordering recommended by the server, the client side of the invention does not immediately display the information ordering, but intercepts the information for further reordering.

in order to comprehensively consider the mutual influence of ordering among all information and overcome the defects of the conventional pointwise, pairwise, the method adopts a listwise algorithm to order the information so as to perform global target optimization on the whole list. Since the candidate information to be sorted reordered on the client is very small (typically 6-12), it is possible to use listwise algorithm, and there is no obvious sort performance degradation.

Specifically, for ordering the information returned from a server arriving at the client, a sequence of length K, such as a sequence ordered by click rate, is used to randomly arrange all the information in the sequence, and all possible obtained rearrangement numbers are K-! .

the invention considers the position relation among the information besides the characteristics of the recommended information. The second information ordering for each candidate is different in the style it presents on different clients. Thus, the present invention is now performing the process of obtaining K-! After the second information is ordered, the context information of each piece of recommended information is acquired according to the display style of the client.

The final feature dimension of the ordering model of the server may be hundreds of millions in order to improve the learning ability of the model. The feature dimension of the client is not too high, limited by the computing power and the amount of network transmission data. In view of the large-scale user, recommendation information features that cannot be transmitted to clients, the invention utilizes fewer features, mainly features that affect the user's vision, and context features. The recommended information features used in the reordering of the invention include: information semantic vector, information category, CTR of information. The characteristics of the context information include information surrounding the current information item. The method comprises the following steps:

(1) Information semantic vector: the floating point unit vector with the dimension of 128 dimensions is calculated by the server and transmitted to the client;

(2) Information category: discrete features are transmitted to the client by the server, and one-hot represents;

(3) CTR of information: a CTR estimated result output by the server sequencing model;

(4) Contextual information: in the current user interface presentation form, information surrounding the current information item.

As shown in fig. 3, the left graph is typical context information under a web page, the middle dark color is a target information item, and the surrounding 8 light squares represent its context information item; the right diagram is typical context information on a cell phone device. To unify feature dimensions, the contextual features of the present invention are all set to 8 positions, and the missing are all filled with default value 0. For example, for dark target information in the right graph, the left information item is missing due to the display restriction of the client, and the present invention adopts 0 to represent the corresponding left information.

According to the invention, the recommendation information returned by the server side is reordered at the client side, and the click through rate estimated value is estimated through the Attention-LSTM-MLP model generated through training. Specifically, the invention encodes the returned information sequence by using a neural network, and the RNN is a recurrent neural network, which is a generic term for a series of neural networks capable of processing sequence data. The long-short-term memory network (Long short term memory, LSTM) is a specific form of RNN, has strong adaptability in time sequence data analysis, hidden layer neurons are connected, and the LSTM can memorize previous information and apply the previous information to current calculation output, so that the problem of long-term dependence in time sequence information is solved. Thus, the present invention utilizes LSTM neural networks.

The LSTM unit is used for learning the input sequence and encoding the second information ordered features into a vector representation with a fixed length. Thus, for shorter length input sequences, LSTM neural networks can learn a corresponding reasonable vector representation. However, LSTM neural networks have difficulty learning reasonable vector representations when the input sequences are very long. Based on this, the invention introduces an Attention (Attention) mechanism that enables a model to view different parts of the data in a targeted way. An attention mechanism is added to learn the preferences of the human brain for different content distributions and aggregations.

The Attention mechanism is implemented by preserving the intermediate output results of LSTM neural networks on input sequences, then training a model to selectively learn these inputs and correlate the output sequences with the model output.

After the characteristic of the second information sequence is encoded through the Attention-LSTM network, the invention adopts a Multi-Layer Perceptron (MLP) to estimate the click through rate of the input code, and the code is used as the input to obtain a score through an MLP network. Eventually we will choose the highest scoring sequence as the final result after reordering.

It should be noted that the execution subject of the steps S1-S4 is a client. The training of the Attention-LSTM-MLP model needs to process a large amount of data, the resource limitation of the client is considered, and the Attention-LSTM-MLP model is suitable for a large amount of clients, so that the training of the whole model is performed at the server. After training, the model is transmitted to the client, and click through rate prediction is performed on the client. The specific steps for generating the attribute-LSTM-MLP model are as follows:

s41, constructing an Attention-LSTM-MLP network;

as shown in fig. 4, the present invention joins an Attention network before the hidden layer in the conventional LSTM network. The LSTM network encodes the feature of each second information sequence and converts the feature into a high-level feature, and the hidden layer output vector of the LSTM network is the extracted feature h _i . In the forward calculation process of the attribute network, the layer is converted into a weight coefficient w of each node through a softmax function _i The value of each node in the attention network is multiplied by the weight coefficient to be the output of the node. Determining the influence degree of each dimension in the hidden layer on the result by adding an attribute network, wherein the larger the weight coefficient is, the larger the influence of the result is, so that the network is focused on a certain orA change in a few dimensions. And obtaining the final coding output y of the Attention-LSTM through the output of each node.

The coded output y of the Attention-LSTM is input to an MLP network, which is an artificial neural network of forward structure that maps a set of input vectors to a set of output vectors. The MLP can be seen as a directed graph, consisting of multiple layers of nodes, each layer being fully connected to the next. Except for the input nodes, each node is a neuron (or processing unit) with a nonlinear activation function. The present invention adopts a conventional MLP network structure, and will not be described in detail herein.

After the Attention-LSTM-MLP network is constructed, the invention adopts a large amount of sample data to train the network. As with the sequences to be ordered, for each sequence of recommended information, the characteristics of the recommended information and its context information are extracted. The sample data is a displayed and exposed recommended information sequence, and the actual user click number of the user is obtained, so that the click through rate is calculated.

LTR is one of supervised learning, assuming that training samples are represented as { (x) _i ，y _i ),i∈[1,n]}. Here, unlike the pointwise class algorithm, x _i Not characteristic of an information item, but x _i And its contextual characteristics. Click through rate y _i Calculated with the following formula:

where num is the number of information clicked by the user in the recommended information sequence that has been displayed, exposed at one time. The range of values for this function is 0, 1.

For the loss function, the invention uses the root mean square error, the formula is as follows:

wherein y is _i Is the calculated true click through rate, p _i Is a predicted value calculated according to the Attention-LSTM-MLP network, and n is the number of samples.

According to the invention, sample data is input into the Attention-LSTM-MLP network, the Attention-LSTM-MLP model is optimized by calculating the loss function of the whole reordering model, and the Attention-LSTM-MLP model is generated by training.

Example two

As shown in fig. 5, this embodiment proposes a client reordering system based on edge computation, where the client specifically includes:

It is noted that the training of the Attention-LSTM-MLP model requires processing a large amount of data, considering the resource limitation of the client, and the Attention-LSTM-MLP model is applicable to a large amount of clients, so the training of the whole model of the present invention is performed at the server. After training, the model is transmitted to the client, and click through rate prediction is performed on the client. Therefore, the client reordering system based on edge calculation of the present invention further includes a server, which specifically includes:

the present invention joins an Attention network before the hidden layer in the conventional LSTM network. The LSTM network encodes the feature of each second information sequence and converts the feature into a high-level feature, and the hidden layer output vector of the LSTM network is the extracted feature h _i . In the forward calculation process of the attribute network, the layer is converted into a weight coefficient w of each node through a softmax function _i The value of each node in the attention network is multiplied by the weight coefficient to be the output of the node. The influence degree of each dimension in the hidden layer on the result is judged by adding an attribute network, and the larger the weight coefficient is, the larger the influence of the result is, so that the network is focused on the change of a certain dimension or a plurality of dimensions. And obtaining the final coding output y of the Attention-LSTM through the output of each node.

LTR is one of supervised learning, assuming that training samples are represented as { (x) _i ,y _i ),i∈[1,n]}. Here, unlike the pointwise class algorithm, x _i Not characteristic of an information item, but x _i And its contextual characteristics. Click through rate y _i Calculated with the following formula:

Therefore, the client reordering method and system based on edge calculation provided by the invention reorder the recommended results on the end by combining the display form of the specific client, fully considers the display characteristics of the client, ensures that the same server results are different in the ordering of different clients, and improves the user experience and the recommended effect; by adopting the lisdwise algorithm, global target optimization is carried out on the whole list, and because the candidate information to be ordered, which is reordered on the end, is very small, the extra cost introduced by adopting the lisdwise algorithm on the basis of optimizing the ordering is small; in view of the fact that large-scale users and information features cannot be transmitted to the client, the method and the device have fewer utilized features, mainly the features affecting the vision of the users and the context features, and further optimize the sorting performance; according to the invention, the preference of the human brain for dispersing and gathering different contents is learned by adding the attention mechanism, so that the data processing amount in the reordering process is reduced, and the data processing efficiency is improved; the invention fully utilizes the characteristics of the server and the client, trains and generates the Attention-LSTM-MLP model at the server to send to the client, and reduces the burden of the client.

Note that the above is only a preferred embodiment of the present invention and the technical principle applied. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, while the invention has been described in connection with the above embodiments, the invention is not limited to the embodiments, but may be embodied in many other equivalent forms without departing from the spirit or scope of the invention, which is set forth in the following claims.

Claims

1. The client reordering method based on edge calculation is characterized by comprising the following steps:

s4, sequentially inputting the characteristics of each second information sequence into an attribute-LSTM-MLP model to obtain corresponding click through rate predicted values, and selecting the second information sequence corresponding to the highest click through rate predicted value for display;

the generation of the Attention-LSTM-MLP model is specifically as follows:

s41, constructing an Attention-LSTM-MLP network;

s42, extracting characteristics of recommended information and context information thereof, calculating click passing rate according to the click number of users of the recommended information, and training the Attention-LSTM-MLP network based on the characteristics and the click passing rate;

the attribute-LSTM-MLP network specifically comprises:

connecting an attribute network behind an hidden layer in the LSTM network; the hidden layer of the LSTM outputs the extracted characteristics and inputs the characteristics into the Attention network; the attribute network is converted into a weight coefficient of each node through a softmax function, the value of each node in the attribute network is multiplied by the weight coefficient to be used as the output of the node, the final encoded output of the attribute-LSTM is obtained, the encoded output is input into the MLP network to obtain a click through rate pre-estimated value, and the loss function of the attribute-LSTM-MLP network is as follows:

wherein y is _i Is the calculated true click through rate, p _i Is according to AtteThe predictive value calculated by the ntion-LSTM-MLP network, n is the number of samples, y _i The specific calculation is as follows:

2. The client reordering method of claim 1, wherein the characteristics of the recommendation information comprise: the click passing rate estimation result is output by the information semantic vector, the information category and the server sequencing model; the characteristics of the context information include information surrounding the current recommendation information in the current user interface presentation.

3. The client reordering method of claim 1, wherein the Attention-LSTM-MLP model is generated at a server and transmitted to a client.

4. A client reordering method as claimed in claim 3, characterized in that the characteristics of the context information are set to 8 positions, the absence being filled with a default value of 0.

5. The client reordering system based on edge calculation comprises a client, and is characterized by comprising the following specific components:

the estimating module is used for inputting the characteristics of each second information sequence into the Attention-LSTM-MLP model in sequence to obtain corresponding click through rate estimated values, and selecting the second information sequence corresponding to the highest click through rate estimated value for display;

the method further comprises a server, and specifically comprises the following steps:

the training module is used for extracting characteristics of the recommended information and the context information thereof, calculating click passing rate according to the click quantity of the user of the recommended information, and training the Attention-LSTM-MLP network based on the characteristics and the click passing rate;

the attribute-LSTM-MLP network specifically comprises:

connecting an attribute network behind an hidden layer in the LSTM network; the hidden layer of the LSTM outputs the extracted characteristics and inputs the characteristics into the Attention network; the attribute network is converted into a weight coefficient of each node through a softmax function, the value of each node in the attribute network is multiplied by the weight coefficient to be used as the output of the node, the final code output of the attribute-LSTM is obtained, and the code output is input into the MLP network to obtain the click through rate predicted value

The loss function of the Attention-LSTM-MLP network is as follows:

6. The client reordering system of claim 5, wherein the features of the recommendation information comprise: the click passing rate estimation result is output by the information semantic vector, the information category and the server sequencing model; the characteristics of the context information include information surrounding the current recommendation information in the current user interface presentation.