CN115617827A - A method and system for jointly updating business models based on parameter compression - Google Patents

Abstract

In the method for updating the service model, when the server judges that the iteration turn of the current iteration turn is a preset target turn, the server judges whether the model updating process enters a key stage or not based on the learning rate of two adjacent turns or the variation range of the parameters. When the key stage is not entered, the server sends k indication information corresponding to k dimensions to each participant, wherein each indication information is determined by the server according to the convergence condition of each dimension of the n local parameter vectors received from the n participants and is used for indicating whether the corresponding dimension is compressed in the current round and the subsequent rounds of iterations. Therefore, each participant compresses the dimension which is converged or close to the converged dimension in the determined local parameter vector in the current iteration and the subsequent iterations, and provides the obtained target parameter vector with the reduced data quantity to the server.

Description

A method and system for jointly updating business models based on parameter compression

technical field

One or more embodiments of this specification relate to the field of machine learning, and in particular to a method and system for jointly updating service models based on parameter compression.

Background technique

Commodity supply chain integration service enterprise groups have accumulated a large amount of business data and financial data due to their industrial diversification and large business scale, covering the entire process of integrated services such as procurement, manufacturing, distribution, warehousing, logistics, distribution, and financing. In order to fully tap the value of data to empower business operations and improve enterprise efficiency, we will use distributed learning methods to update business models based on massive business and financial data. In this case, the multi-participant joint update model is the most widely adopted method, which needs to be combined with each gradient or model parameters (collectively referred to as parameters) update in each round of iterations. However, passing parameters frequently in such a large parameter range will affect the update efficiency of the model. Therefore, how to efficiently update the model in a large-scale cluster has become a key concern. In order to alleviate the communication bottleneck, some schemes propose to increase the size of the sample batch, that is, each participant calculates parameters based on a large batch of samples, thereby reducing the communication frequency of each iteration. In this solution, the key stages in the model update process are identified, so as to ensure the accuracy of the model while greatly reducing the communication traffic.

Contents of the invention

One or more embodiments of this specification describe a method and system for jointly updating business models based on parameter compression, which can greatly reduce communication traffic while ensuring model accuracy.

In the first aspect, a method for jointly updating business models based on parameter compression is provided, including:

Each participant i determines a local parameter vector with k dimensions according to the local sample set and the local model parameters of the business model, and provides it to the server;

The server, when the t is equal to the preset target round, performs the judgment of the first condition and the dimension convergence detection in parallel;

The first condition includes: the ratio of the learning rate of the t+1th round to the learning rate of the tth round is less than the learning rate threshold, or the target distance between the aggregation parameter vector of the tth round and the aggregation parameter vector of the t-1th round is not less than the distance Threshold; the t-th round of aggregation parameter vectors is obtained by aggregating n parts of local parameter vectors sent by the n participants;

The dimensional convergence detection includes: averaging and calculating the variance of the n element values corresponding to each dimension j of the n parts of the local parameter vector, and determining according to the average value and variance value calculated for each dimension j Corresponding to the signal-to-noise ratio of each dimension j, the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension; and comparing the signal-to-noise ratio corresponding to each dimension j with the proportion threshold to obtain corresponding indication information, the The indication information is used to indicate whether to compress the dimension j in the current round and subsequent rounds of iterations;

The server, when the first condition is not satisfied, sends k pieces of indication information corresponding to the k dimensions to each participant i;

Each participant i, according to the k indication information, compresses or does not compress each dimension j of the corresponding local parameter vector to obtain a target parameter vector, and provides it to the server;

The server, based on the n target parameter vectors sent by the n participants, obtains the first updated parameter of the business model, and sends it to each participant i for each participant i to use based on The first update parameter updates its local model parameters for the next iteration.

In the second aspect, a system for jointly updating business models based on parameter compression is provided, including:

Each participant i is used to determine a local parameter vector with k dimensions according to the local sample set and the local model parameters of the business model, and provide it to the server;

The server is configured to perform the judgment of the first condition and the dimension convergence detection in parallel when the t is equal to the preset target round;

The first condition includes: the ratio of the learning rate of the t+1th round to the learning rate of the tth round is less than the learning rate threshold, or the target distance between the aggregation parameter vector of the tth round and the aggregation parameter vector of the t-1th round is not less than the distance Threshold; the t-th round of aggregation parameter vectors is obtained by aggregating n parts of local parameter vectors sent by the n participants;

The dimensional convergence detection includes: averaging and calculating the variance of the n element values corresponding to each dimension j of the n parts of the local parameter vector, and determining according to the average value and variance value calculated for each dimension j Corresponding to the signal-to-noise ratio of each dimension j, the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension; and comparing the signal-to-noise ratio corresponding to each dimension j with the proportion threshold to obtain corresponding indication information, the The indication information is used to indicate whether to compress the dimension j in the current round and subsequent rounds of iterations;

The server is further configured to send k pieces of indication information corresponding to the k dimensions to each participant i when the first condition is not met;

Each participant i is further configured to compress or not compress each dimension j of the corresponding local parameter vector according to the k indication information, obtain a target parameter vector, and provide it to the server;

The server is further configured to obtain the first update parameter of the business model based on the n target parameter vectors sent by the n participants, and send it to each participant i for each participant Party i updates its local model parameters based on the first update parameters for the next iteration.

In the method and system for jointly updating business models based on parameter compression provided by one or more embodiments of this specification, when the server judges that the current round of iteration is the preset target round, it Or the change range of the parameters to judge whether the model update process has entered a critical stage. Before entering the critical stage, the server sends k indication information corresponding to k dimensions to each participant, each of which is based on each of n local parameter vectors received from n participants. Determined by the convergence of the dimension, it is used to indicate whether to compress the corresponding dimension in the current round and subsequent rounds of iterations. This enables each participant to compress the converged or near-converged dimensions of the determined local parameter vectors in the current round and several subsequent iterations, and provide the obtained target parameter vectors with a smaller amount of data to the server. In this way, the communication traffic can be greatly reduced while ensuring the accuracy of the model.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following will briefly introduce the drawings that need to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of this specification. Those of ordinary skill in the art can also obtain other drawings based on these drawings without making creative efforts.

Fig. 1 shows one of the schematic diagrams of a system for jointly updating business models based on parameter compression according to an embodiment;

Fig. 2 shows an interaction diagram of a method for jointly updating business models based on parameter compression according to an embodiment;

Fig. 3 shows an interaction diagram of a method for jointly updating product recommendation models based on parameter compression according to an embodiment;

Fig. 4 shows the second schematic diagram of the system for jointly updating service models based on parameter compression according to an embodiment.

detailed description

The solutions provided in this specification will be described below in conjunction with the accompanying drawings.

Fig. 1 shows one of schematic diagrams of a system for jointly updating service models based on parameter compression according to an embodiment. In Fig. 1, the system includes a server and n participants, where n is a positive integer. Each participant can be implemented as any device, platform, server or device cluster with computing and processing capabilities. The server includes critical stage identification means, convergence detection means, compression means and update means.

In Figure 1, each participant i among participant 1-participant n can determine the corresponding local parameter vector with k dimensions according to its local sample set and local model parameters of the business model, and provide it to server. Among them, k is a positive integer. Here the business model is used to predict the classification or regression value of the business object. The business objects can be pictures, texts, users or commodities, for example. Wherein, i is a positive integer, and 1≤i≤n.

After receiving n copies of local parameter vectors sent by n participants, the server can first judge whether the round of the current iteration is the preset target round, and if so, perform the first condition through the key stage identification device judgment, and the dimension convergence detection is performed through the convergence detection device therein. It should be understood that the judgment of the first condition and the detection of dimensionality convergence here can be executed in parallel.

It should be noted that the judgment of the above-mentioned first condition is used to identify the key stage of the model update process (the model parameter update at this stage has a greater impact on the convergence of the model parameters, and is usually called the core link of the model update). The first condition includes: the ratio of the learning rate of the t+1th round to the learning rate of the tth round is less than the learning rate threshold, or the target distance between the aggregated parameter vector of the tth round and the aggregated parameter vector of the t-1th round is not less than the distance threshold . The t-th round of aggregation parameter vector is obtained by aggregating n parts of local parameter vectors sent by n participants. That is to say, this solution is based on the learning rate of two adjacent rounds or the variation range of parameters to identify whether the model update process has entered a critical stage.

The above-mentioned dimensional convergence detection may specifically include: averaging and calculating the variance of the n element values corresponding to each dimension j of n local parameter vectors, and determining Corresponding to the signal-to-noise ratio of each dimension j, the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension. Wherein, j is a positive integer, and 1≤j≤k.

The above-mentioned dimensional convergence detection further includes: acquiring corresponding indication information based on the determined signal-to-noise ratio for each dimension j, so as to obtain k pieces of indication information corresponding to k dimensions.

Taking any first dimension among the k dimensions as an example, if the signal-to-noise ratio corresponding to the first dimension is smaller than the ratio threshold, it is determined that the indication information corresponding to the first dimension is compression. And if the signal-to-noise ratio corresponding to the first dimension is not less than the ratio threshold, it is determined that the indication information corresponding to the first dimension is not to be compressed.

When the first condition is not met, or in other words, when the model update process does not enter a critical stage, the k pieces of indication information are sent to each participant i through the compression device.

Each participant i, according to the received k pieces of indication information, compresses or does not compress each dimension j of the corresponding local parameter vector to obtain a target parameter vector, and provides it to the server.

The server obtains the first updated parameter of the business model based on the n target parameter vectors sent by n participants through the update device therein, and sends it to each participant i for each participant i to use based on the first A parameter update updates its local model parameters for the next iteration.

It should be understood that after entering the next iteration, each participant i, after calculating the local parameter vector, compresses or does not compress each dimension j of the local parameter vector according to the received k indication information until reaching Compression end condition. The compression end condition here includes receiving updated k indication information or stopping the compression indication, or satisfying the iteration end condition, and the like.

It should be noted that when the server has not entered the critical stage through the identification of the critical stage identification device, the server sends k pieces of indication information corresponding to k dimensions to each participant to indicate that each participant is in the current round and subsequent rounds The corresponding dimension is compressed during the iteration. Since the amount of parameters after compression is smaller than the amount of parameters before compression, this scheme can reduce the amount of data transmitted between the participant and the server. In addition, since the above-mentioned compression process is performed in a non-critical stage, it will not affect the accuracy of the model. That is to say, this scheme can greatly reduce the communication traffic while ensuring the accuracy of the model.

Fig. 2 shows an interaction diagram of a method for jointly updating service models based on parameter compression according to an embodiment. It should be noted that this method involves multiple rounds of iterations. Figure 2 shows the interaction steps included in the tth (t is a positive integer) round of iterations, and because the interaction process between the participants participating in the tth round of iterations and the server is similar, Therefore, Figure 2 mainly shows the interaction steps between any participant participating in the t-th iteration (for convenience of description, referred to as the first participant) and the server, and the interaction steps between other participants participating in the iteration and the server. Refer to the interaction steps between the first participant and the server. As shown in Figure 2, the method may include the following steps:

Step 202, each participant i determines a local parameter vector with k dimensions according to the local sample set and the local model parameters of the business model, and provides it to the server.

Taking any participant as an example, the business objects corresponding to the samples in the local sample set maintained by it may include any of the following: pictures, texts, users, and commodities.

In addition, the above business model may be a classification model or a regression model, and is used to predict the classification or regression value of the above business object. In one embodiment, the business model can be realized based on decision tree algorithm, Bayesian algorithm, etc. In another embodiment, the business model can be realized based on neural network.

It should be noted that when the t-th iteration is the first iteration, the above-mentioned local model parameters may be initialized by the server before the multi-round iteration starts, and then the initialized model parameters are issued or provided to each participant, so that Each participant can use the above initialized model parameters as local model parameters. Of course, in practical applications, it is also possible for each participant to agree on the structure of the model (such as which model to use, the number of layers of the model, the number of neurons in each layer, etc.), and then perform the same initialization. to get the respective local model parameters.

When the t-th iteration is not the first iteration, the above local model parameters may be updated in the t-1th iteration.

Finally, the above-mentioned local parameter vector may include a gradient vector or a model parameter vector, and reference may be made to the prior art for determination of the local parameter vector. Taking the gradient vector as an example, its determination method can be as follows: first determine the prediction result according to the local sample set and local model parameters, and then determine the prediction loss according to the prediction result and sample labels. Finally, according to the prediction loss, and using the backpropagation method, the gradient vector corresponding to the local model parameters is determined.

It should be understood that, in practical applications, the above method for determining the local parameter vector also includes multiple rounds of iterations, and the specific iteration rounds may be preset.

In step 204, when the server judges that t is equal to the preset target round, the judgment of the first condition and the detection of dimension convergence are performed in parallel.

It should be understood that the above t is the round of the current iteration.

In practical applications, there may be multiple preset target rounds. In the case of multiple preset target rounds, t may be compared with each preset target round in turn to determine whether t is equal to the preset target round.

The judgment of the above-mentioned first condition can also be understood as identifying a key stage of the model updating process. The first condition may include: the ratio of the learning rate of the t+1th round to the learning rate of the tth round is less than the learning rate threshold, or the target distance between the aggregated parameter vector of the tth round and the aggregated parameter vector of the t-1th round is not less than the distance threshold.

In machine learning, gradient descent is a parameter optimization algorithm widely used to minimize model error. The gradient descent method estimates model parameters through multiple iterations and minimizes the loss function in each iteration. In the gradient descent method, a uniform learning rate is usually given, which is used to control the learning progress of the model during the iterative process.

Generally speaking, in the early iterations, the learning rate is larger, so the forward step size will be longer, and then the gradient can be descended at a faster speed. In the later iterations, the learning rate is gradually reduced to reduce the learning step size, which helps the convergence of the model and makes it easier to approach the optimal solution.

Therefore, this scheme identifies the key stage based on the change range of the learning rate between two adjacent rounds. Specifically, when the ratio of the learning rate of the t+1th round to the learning rate of the tth round is less than the learning rate threshold, it is determined that the first condition is satisfied, that is, entering a critical stage of the model update process. Referring to the aforementioned learning rate setting method, it can be concluded that this solution uses the later iterations as a key stage in the model update process.

In addition, this scheme also identifies critical stages based on the variation range of parameters between two adjacent rounds. Specifically, when the target distance between the aggregation parameter vector of the t-th round and the aggregation parameter vector of the t-1th round is not less than the distance threshold, it is determined that the first condition is satisfied, that is, entering a critical stage of the model update process. That is to say, this scheme takes the iteration with large parameter changes as the key stage of the model update process.

The aforementioned t-th round of aggregation parameter vectors is obtained by aggregating n parts of local parameter vectors sent by n participants. In one example, the t-th round aggregation parameter vector is obtained by summing n local parameter vectors.

The calculation method of the aggregation parameter vector of the t-1th round above is similar to the calculation method of the aggregation parameter vector of the t-th round, which is obtained by aggregating n parts of local parameter vectors sent by n participants in the t-1 round.

In an example, the above target distance is calculated by calculating the second-order norm distance between the t-th round of aggregation parameter vector and the t-1-th round of aggregation parameter vector, and the second-order norm of the t-1-th round of aggregation parameter vector, and then It is obtained by calculating the ratio of the second-order norm distance to the second-order norm. The specific calculation formula can be as follows:

（公式1）

(Formula 1)

Among them, d is the target distance, grad _t is the aggregation parameter vector of the t-th round, and grad _t-1 is the aggregation parameter vector of the t-1 round.

In short, when the ratio of the learning rate between two adjacent rounds is less than the learning rate threshold, or when the target distance of the aggregated parameter vectors between two adjacent rounds is not less than the distance threshold, it is judged that the first condition is satisfied, that is, entering the model update process critical stage, otherwise it is judged that the first condition is not satisfied.

The dimensional convergence detection in step 204 may include: averaging and calculating the variance of the n element values corresponding to each dimension j of n local parameter vectors, and according to the average and variance values calculated for each dimension j, A signal-to-noise ratio corresponding to each dimension j is determined, and the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension.

That is, for each dimension in the k dimensions, the n element values corresponding to the dimension of the n local parameter vectors are averaged and the variance is calculated to obtain k average values and k variance values corresponding to the k dimensions . After that, the ratio of 1 mean value and 1 variance value corresponding to each dimension j can be calculated to obtain the signal-to-noise ratio corresponding to each dimension j. Here, k signal-to-noise ratios corresponding to k dimensions can form a signal-to-noise ratio vector.

Next, for the signal-to-noise ratio corresponding to each dimension j, the signal-to-noise ratio can be compared with the corresponding proportion threshold. If it is less than the proportion threshold, it means that the dimension has entered a stable stage and is close to convergence, so the corresponding The indication information in this dimension is determined as a compression indication. If it is not less than the proportion threshold, it means that the dimension is still changing dynamically and has not converged, so the indication information corresponding to the dimension is determined as an indication of no compression.

After determining the corresponding indication information for each dimension j, k indication information corresponding to k dimensions can be obtained, and each indication information is used to indicate whether to perform Compression processing. In other words, each participant i determines whether to compress or not compress each dimension of the determined n local parameter vectors according to the k indication information in the current round and subsequent rounds of iterations.

Step 206, when the first condition is not met, the server sends k pieces of indication information corresponding to k dimensions to each participant i.

In other words, when the identification model update process has not entered a critical stage, the server sends k indication information corresponding to k dimensions to each participant i, so that each participant i can, according to the k indication information, Each dimension j of the corresponding local parameter vector is compressed or not compressed. However, when the recognition model update process enters a critical stage, k pieces of indication information corresponding to the k dimensions are not sent. That is, the results of dimensionality convergence detection (that is, the k pieces of indication information) are used only when the first condition is not satisfied.

Step 208 , each participant i compresses or does not compress each dimension j of the corresponding local parameter vector according to the k indication information, obtains a target parameter vector, and provides it to the server.

In one example, the above target parameter vector is obtained by the following steps:

For each dimension j in the k dimensions, it is judged whether the indication information corresponding to dimension j is a compression indication or an uncompression indication. If the instruction information is a compression instruction, perform quantization processing on the element value of the local parameter vector corresponding to the dimension j to obtain a quantization value as a processing result. The quantized value contains a smaller number of bits than the element value. In the case that the indication information is an indication of no compression, the element value corresponding to the dimension j of the local parameter vector is reserved as a processing result. A target parameter vector is formed based on processing results corresponding to respective dimensions in the k dimensions.

In an example, the above quantization process may include: converting the element value corresponding to dimension j (generally a floating point number containing 32 bits) into an integer (for example, multiplying by 10 to the 8th power), and then intercepting the predetermined value from the end A number of bits (for example, 4 bits or 8 bits) are used as corresponding quantization values.

Of course, in practical applications, quantization values corresponding to any dimension may also be obtained through other quantization processing methods, which is not limited in this specification.

In another example, the above target parameter vector is obtained by the following steps:

For each dimension j in the k dimensions, it is judged whether the indication information corresponding to dimension j is a compression indication or an uncompression indication. If the indication information is a compression indication, it is judged whether the element value corresponding to the dimension j of the local parameter vector is smaller than the element value threshold, and if yes, 0 is taken as the processing result, otherwise, the element value itself is taken as the processing result. In the case that the indication information is an indication of no compression, the element value corresponding to the dimension j of the local parameter vector is reserved as a processing result. A target parameter vector is formed based on processing results corresponding to respective dimensions in the k dimensions.

Step 210, the server obtains the first update parameter of the business model based on n target parameter vectors sent by n participants, and sends it to each participant i for each participant i to update based on the first update parameter , to update its local model parameters for the next iteration.

In an example, the acquisition of the first update parameters of the business model includes: aggregating n target parameter vectors to obtain an updated t-th round of aggregation parameter vectors. The first updated parameter of the business model is obtained by subtracting the product of the updated aggregation parameter vector of the t-th round and the learning rate of the t-th round from the global model parameters of the business model.

Wherein, the above-mentioned aggregating n target parameter vectors may include: averaging or weighting the n target parameter vectors to obtain an updated t-th round of aggregation parameter vectors.

In an example, the first update parameter of the business model can be obtained by referring to the following formula.

（公式2）

(Formula 2)

Wherein, w _t is the first update parameter, and w _t-1 is the global model parameter currently maintained by the server. η _t is the learning rate of the t-th round, also known as the learning step size, n is the number of target parameter vectors, and m _i is the i-th target parameter vector.

It should be understood that after obtaining the above-mentioned first update parameters, the server can use the first update parameters to update (or replace) the global model parameters it currently maintains, so as to obtain updated global model parameters for the next iteration .

In addition, after receiving the above-mentioned first update parameters, each participant i can use the first update parameters to update (or replace) its locally maintained local model parameters, so as to obtain updated local model parameters for the following One iteration.

It should be understood that since k indications are received in the t-th iteration, after entering the next iteration, each participant i determines the local parameters of the t+1th round according to the local sample set and the updated local model parameters After vectoring, based on the k indication information received in the tth round, each dimension j of the local parameter vector of the t+1th round is compressed or not compressed, and the target parameter vector of the t+1th round is obtained and provided to to the server. Afterwards, the server can aggregate the n copies of the t+1 round target parameter vectors sent by n participants, obtain the t+1 round update parameters, and send them to each participant i, so that the server and each participant i updates the model parameters of the t+1 round (including the global model parameters maintained by the server and the local model parameters maintained by each participant i); and so on until the compression end condition is reached. The compression end condition here includes receiving updated k indication information or stopping the compression indication, or satisfying the iteration end condition (for example, the number of iterations reaches a predetermined number of rounds or the global model parameters converge) and the like.

It should be noted that since each participant i sends a target parameter vector with a smaller amount of data than the initial local parameter vector to the server in each round of iterations starting from the tth round, this scheme can reduce the time spent on model update During the process, the amount of data transmission between the participants and the server can improve the efficiency of data transmission, which helps to improve the efficiency of model updating.

The above describes the updating method of the model parameters (including the global model parameters maintained by the server and the local model parameters maintained by each participant i) in the t-th round when the first condition is satisfied, that is, in the non-critical stage. When the first condition is not satisfied, that is, in the critical stage, the local parameter vector sent by each participant i is directly used to update the model parameters of the t-th round. That is to say, in the critical stage, each dimension of the local parameter vector of each participant is not compressed, so that the local parameter vector contains more useful information, which will not affect the process of model updating.

The above directly uses the local parameter vector sent by each participant i to update the t-th round of model parameters, including: the server obtains the second update parameter of the business model based on n local parameter vectors, and sends it to each Participant i, for each participant i to update its local model parameters based on the second update parameters for the next round of iteration.

Of course, in practical applications, if it is identified as a critical stage, the server may instruct each participant to perform compression with a low compression ratio for their respective local parameter vectors, which is not limited in this specification.

Here, the method for obtaining the second update parameter may refer to the above formula 2, which will not be repeated here in this specification.

It should also be noted that in step 204, if the server judges that t is not equal to the preset target round, or in other words, judges that the current iteration round is not equal to the preset target round, then steps 206-210 can be replaced by : The server obtains other update parameters of the business model based on n local parameter vectors sent by n participants, and sends them to each participant i for each participant i to update its local Model parameters to use for the next iteration.

Finally, after multiple rounds of iterations, each participant i uses the local model parameters it obtains as its business model for collaborative update with other participants.

Taking any participant as an example, if the business object corresponding to the sample in its local sample set is a picture, then the business model that it cooperates with other participants to update can be a picture recognition model. In the case that the business object corresponding to the samples in its local sample set is text, the business model that it coordinates with other parties to update may be a text recognition model. In the case that the business objects corresponding to the samples in its local sample set are commodities and users, then the business model that it coordinates with other participants to update can be a commodity recommendation model and so on.

To sum up, the business model joint update method based on parameter compression provided by the embodiment of this specification can identify the key stages in the model update process. Compression of parameters to be transmitted. When the recognition enters a non-critical stage, the compression of the parameters to be transmitted by the participants should be improved as much as possible to reduce the communication traffic. And the compression from recognition to parameters is performed automatically, thereby improving the automatic degree of decision-making. In addition, based on the result of dimensionality convergence detection, the server sends k pieces of instruction information to each participant, which can realize automatic control of participant parameter compression.

In addition, this solution automatically identifies the key stage of model update, so as to take appropriate compression measures, avoiding the loss of model accuracy caused by blind compression, and even the model parameters do not converge due to blind compression. Through the automatic identification method of this scheme, the participants can compress the communication when necessary, reduce the communication traffic, improve the model update speed, and improve the scalability of the model update, which can support larger clusters and larger model size.

In short, this scheme is an adaptive and automatic switching method of compressing and non-compressing parameters, and by selecting low compression in critical stages and high compression in other places, the traffic can be greatly reduced, and the accuracy and accuracy of model parameter convergence can be guaranteed. Effect.

The following takes the business model as a product recommendation model as an example to illustrate this solution.

Fig. 3 shows an interaction diagram of a method for jointly updating product recommendation models based on parameter compression according to an embodiment. It should be noted that this method involves multiple rounds of iterations. Figure 3 shows the interaction steps included in the tth (t is a positive integer) round of iterations, and because the interaction process between each participant participating in the tth iteration and the server is similar, Therefore, Figure 3 mainly shows the interaction steps between any participant participating in the t-th iteration (for convenience of description, referred to as the first participant) and the server, and the interaction steps between other participants participating in the iteration and the server. Refer to the interaction steps between the first participant and the server. As shown in Figure 3, the method may include the following steps:

Step 302, each participant i determines a local parameter vector with k dimensions according to the local sample set and the local model parameters of the commodity recommendation model, and provides it to the server.

The business objects corresponding to the samples in the above local sample set include users and commodities, and the characteristics of the samples include user attributes (such as occupation, hobbies, and education, etc.), operation behavior (such as browsing, clicking, and closing, etc.) and Product attributes (for example, it can be product category, product price, product details, etc.).

In step 304, when judging that t is equal to the preset target round, the server performs the judging of the first condition and the dimension convergence detection in parallel.

The above-mentioned first condition includes: the ratio of the learning rate of the t+1 round to the learning rate of the t-th round is less than the learning rate threshold, or the target distance between the aggregation parameter vector of the t-th round and the aggregation parameter vector of the t-1 round is not less than the distance threshold . The t-th round of aggregation parameter vector is obtained by aggregating n parts of local parameter vectors sent by n participants.

The above-mentioned dimensional convergence detection includes: averaging and calculating the variance of the n element values corresponding to each dimension j of n local parameter vectors, and determining the corresponding The signal-to-noise ratio of a dimension j, the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension; and the signal-to-noise ratio corresponding to each dimension j is compared with the proportion threshold to obtain the corresponding indication information, the indication information is used Indicates whether to perform compression processing on dimension j in the current round and subsequent rounds of iterations.

Step 306, when the first condition is not met, the server sends k pieces of indication information corresponding to k dimensions to each participant i.

Step 308 , each participant i compresses or does not compress each dimension j of the corresponding local parameter vector according to the k indication information, obtains a target parameter vector, and provides it to the server.

In step 310, the server obtains the first updated parameter of the product recommendation model based on n target parameter vectors sent by n participants, and sends it to each participant i for each participant i to use based on the first update parameter. Update parameters, which update their local model parameters for the next iteration.

After multiple rounds of iterations, each participant i uses the local model parameters it obtains as its product recommendation model to be updated collaboratively with other participants.

To sum up, the method for jointly updating commodity recommendation models based on parameter compression provided by the embodiment of this specification can update commodity recommendation models while saving communication resources.

Corresponding to the above method for jointly updating business models based on parameter compression, an embodiment of this specification also provides a system for jointly updating business models based on parameter compression. As shown in FIG. 4 , the system may include: a server 402 and n Participant 404 .

Each participant 404 is configured to determine a local parameter vector with k dimensions according to the local sample set and the local model parameters of the service model, and provide it to the server 402 .

Wherein, the local parameter vector includes a gradient vector or a model parameter vector.

The server 402 is configured to perform the judgment of the first condition and the dimension convergence detection in parallel when t is equal to the preset target round.

Among them, the first condition includes: the ratio of the learning rate of the t+1 round to the learning rate of the t-th round is less than the learning rate threshold, or the target distance between the aggregation parameter vector of the t-th round and the aggregation parameter vector of the t-1 round is not less than the distance threshold. The t-th round of aggregation parameter vectors is obtained by aggregating n parts of local parameter vectors sent by n participants 404 .

Among them, the above-mentioned target distance is calculated by calculating the second-order norm distance between the t-th round aggregation parameter vector and the t-1-th round aggregation parameter vector, and the second-order norm of the t-1-th round aggregation parameter vector, and then calculating the second-order norm. It is obtained by the ratio of the first-order norm distance to the second-order norm.

The above-mentioned dimensional convergence detection includes: averaging and calculating the variance of the n element values corresponding to each dimension j of n local parameter vectors, and determining the corresponding The signal-to-noise ratio of a dimension j, the signal-to-noise ratio is used to indicate the convergence of the corresponding dimension; and the signal-to-noise ratio corresponding to each dimension j is compared with the proportion threshold to obtain the corresponding indication information, and the indication information is used Indicates whether to perform compression processing on dimension j in the current round and subsequent rounds of iterations.

Wherein, the above determination corresponds to the signal-to-noise ratio of each dimension j, including:

Calculate the ratio of the mean value to the variance value corresponding to each dimension j, and obtain the signal-to-noise ratio corresponding to each dimension j.

In addition, each of the above indication information is a compression indication when the signal-to-noise ratio of the corresponding dimension is less than the proportion threshold, and is a non-compression indication when the signal-to-noise ratio of the corresponding dimension is not less than the proportion threshold.

The server 402 is further configured to send k pieces of indication information corresponding to the k dimensions to each participant 404 when the first condition is not met.

Each participant 404 is further configured to compress or not compress each dimension j of the corresponding local parameter vector according to the k indication information to obtain a target parameter vector and provide it to the server 402 .

In an example, each participant 404 is specifically configured to: for each dimension j in the k dimensions, determine whether the indication information corresponding to dimension j is a compression indication or an uncompression indication. In the case that the instruction information is a compression instruction, quantization processing is performed on the element value of the local parameter vector corresponding to the dimension j, and a quantization value is obtained as a processing result. The number of bits contained in the quantized value is less than the number of bits contained in the value of the element. In the case that the indication information is an indication of no compression, the element value corresponding to the dimension j of the local parameter vector is reserved as a processing result. A target parameter vector is formed based on processing results corresponding to respective dimensions in the k dimensions.

In another example, each participant 404 is specifically configured to: for each dimension j in the k dimensions, determine whether the indication information corresponding to dimension j is a compression indication or an uncompression indication. When the instruction information is a compression instruction, it is judged whether the element value corresponding to the dimension j of the local parameter vector is smaller than the element value threshold, if yes, 0 is taken as the processing result, otherwise, the element value is taken as the processing result. In the case that the indication information is an indication of no compression, the element value corresponding to the dimension j of the local parameter vector is reserved as a processing result. A target parameter vector is formed based on processing results corresponding to respective dimensions in the k dimensions.

The server 402 is further configured to obtain the first update parameter of the business model based on n target parameter vectors sent by n participants 404, and send it to each participant 404 for each participant 404 to use based on the first update parameter A parameter update updates its local model parameters for the next iteration.

Server 402 is specifically used for:

Aggregate n target parameter vectors to obtain the updated round t aggregation parameter vector. The first updated parameter of the business model is obtained by subtracting the product of the updated aggregation parameter vector of the t-th round and the learning rate of the t-th round from the global model parameters of the business model.

Optionally, the server 402 is further configured to obtain the second updated parameter of the business model based on n local parameter vectors when the first condition is met, and send it to each participant 404 for each A participant 404 updates its local model parameters based on the second update parameters for the next iteration.

The functions of each functional module of the device in the above embodiment of this specification can be realized through the steps of the above method embodiment. Therefore, the specific working process of the device provided by one embodiment of this specification will not be repeated here.

The service model joint update system based on parameter compression provided by an embodiment of this specification can update the service model while saving communication resources.

Each embodiment in this specification is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for relevant parts, refer to part of the description of the method embodiment.

The steps of the methods or algorithms described in conjunction with the disclosure of this specification can be implemented in the form of hardware, or can be implemented in the form of a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, mobile hard disk, CD-ROM or any other form of storage known in the art medium. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be a component of the processor. The processor and storage medium can be located in the ASIC. Alternatively, the ASIC may be located in the server. Of course, the processor and the storage medium can also exist in the server as discrete components.

Those skilled in the art should be aware that, in the above one or more examples, the functions described in the present invention may be implemented by hardware, software, firmware or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a general purpose or special purpose computer.

The foregoing describes specific embodiments of this specification. Other implementations are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. Multitasking and parallel processing are also possible or may be advantageous in certain embodiments.

The above-mentioned specific implementation modes further describe the purpose, technical solutions and beneficial effects of this specification in detail. Scope of protection: Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions in this specification shall be included in the scope of protection of this specification.

Claims (10)

1. A service model joint updating method based on parameter compression relates to a server and n participants; the method comprises a plurality of iterations, wherein any tth iteration comprises:

each participant i determines a local parameter vector with k dimensions according to the local sample set and the local model parameters of the business model, and provides the local parameter vector with k dimensions to the server;

the server parallelly judges a first condition and detects the dimension convergence under the condition that the t is equal to a preset target turn;

the first condition includes: the ratio of the t +1 th round learning rate to the t-1 th round learning rate is smaller than a learning rate threshold, or the target distance between the t-th round aggregation parameter vector and the t-1 th round aggregation parameter vector is not smaller than a distance threshold; the t-th round aggregation parameter vector is obtained by aggregating n local parameter vectors sent by the n participants;

the dimension convergence detection comprises: averaging and calculating the variance of the n element values of the n local parameter vectors corresponding to each dimension j, and determining the signal-to-noise ratio corresponding to each dimension j according to the average value and the variance value calculated for each dimension j, wherein the signal-to-noise ratio is used for indicating the convergence condition of the corresponding dimension; comparing the signal-to-noise ratio corresponding to each dimension j with an occupation ratio threshold value to obtain corresponding indication information, wherein the indication information is used for indicating whether the dimension j is compressed in the current iteration and a plurality of subsequent iterations;

the server, in case the first condition is not satisfied, sending k pieces of indication information corresponding to the k dimensions to each participant i;

each participant i compresses or decompresses each dimension j of the corresponding local parameter vector according to the k pieces of indication information to obtain a target parameter vector, and provides the target parameter vector to the server;

and the server acquires a first updating parameter of the service model based on the n target parameter vectors sent by the n participants and sends the first updating parameter to each participant i, so that each participant i updates the local model parameter of the participant i based on the first updating parameter for the next iteration.

2. The method of claim 1, further comprising:

and under the condition that the first condition is met, the server acquires second updating parameters of the service model based on the n local parameter vectors and sends the second updating parameters to each participant i, so that each participant i updates the local model parameters of the participant i based on the second updating parameters for the next iteration.

3. The method of claim 1, wherein the target distance is obtained by calculating a second-order norm distance between the t-th aggregation parameter vector and the t-1 st aggregation parameter vector, and a second-order norm of the t-1 st aggregation parameter vector, and then calculating a ratio of the second-order norm distance to the second-order norm.

4. The method of claim 1, wherein the determining a signal-to-noise ratio corresponding to each dimension j comprises:

the ratio of the mean to the variance values corresponding to each dimension j is calculated to obtain the signal-to-noise ratio corresponding to each dimension j.

5. The method of claim 1, wherein each of the k indicators is a compression indicator if the signal-to-noise ratio of the corresponding dimension is less than a duty ratio threshold, and is an uncompression indicator if the signal-to-noise ratio of the corresponding dimension is not less than the duty ratio threshold.

6. The method of claim 1, wherein the target parameter vector is obtained by:

for each dimension j in the k dimensions, judging whether the indication information corresponding to the dimension j is a compression indication or an uncompression indication;

under the condition that the indication information is a compression indication, carrying out quantization processing on the element value of the local parameter vector corresponding to the dimension j to obtain a quantization value as a processing result; the bit number contained in the quantization value is smaller than the bit number contained in the element value;

under the condition that the indication information is the non-compression indication, retaining the element value of the local parameter vector corresponding to the dimension j as a processing result;

and forming the target parameter vector based on the processing result corresponding to each dimension in the k dimensions.

7. The method of claim 1, wherein the target parameter vector is obtained by:

for each dimension j in the k dimensions, judging whether the indication information corresponding to the dimension j is a compression indication or an uncompression indication;

under the condition that the indication information is a compression indication, judging whether an element value of the local parameter vector corresponding to the dimension j is smaller than an element value threshold value, if so, taking 0 as a processing result, otherwise, taking the element value as the processing result;

under the condition that the indication information is the non-compression indication, retaining the element value of the local parameter vector corresponding to the dimension j as a processing result;

and forming the target parameter vector based on the processing result corresponding to each dimension in the k dimensions.

8. The method of claim 1, wherein the obtaining of the first updated parameter of the business model comprises:

aggregating the n target parameter vectors to obtain an updated t-th aggregation parameter vector;

and subtracting the product of the updated t-th round aggregation parameter vector and the t-th round learning rate from the global model parameter of the business model to obtain a first updating parameter of the business model.

9. The method of claim 1, wherein the local parameter vector comprises a gradient vector or a model parameter vector.

10. A business model joint updating system based on parameter compression comprises a server and n participants;

each participant i is used for determining a local parameter vector with k dimensions according to the local model parameters of the local sample set and the business model and providing the local parameter vector with k dimensions to the server;

the server is used for judging a first condition and performing dimensionality convergence detection in parallel under the condition that the t is equal to a preset target turn;

the first condition includes: the ratio of the t +1 th round learning rate to the t-1 th round learning rate is smaller than a learning rate threshold, or the target distance between the t-th round aggregation parameter vector and the t-1 th round aggregation parameter vector is not smaller than a distance threshold; the t-th round aggregation parameter vector is obtained by aggregating n local parameter vectors sent by the n participants;

the dimension convergence detection comprises: averaging and calculating the variance of n element values of the n local parameter vectors corresponding to each dimension j, and determining the signal-to-noise ratio corresponding to each dimension j according to the average value and the variance value calculated for each dimension j, wherein the signal-to-noise ratio is used for indicating the convergence condition of the corresponding dimension; comparing the signal-to-noise ratio corresponding to each dimension j with an occupation ratio threshold value to obtain corresponding indication information, wherein the indication information is used for indicating whether the dimension j is compressed in the current iteration and a plurality of subsequent iterations;

the server is further configured to send k pieces of indication information corresponding to the k dimensions to each participant i if the first condition is not satisfied;

each participant i is further configured to perform compression or non-compression processing on each dimension j of the corresponding local parameter vector according to the k pieces of indication information to obtain a target parameter vector, and provide the target parameter vector to the server;

the server is further configured to obtain a first update parameter of the service model based on the n target parameter vectors sent by the n participants, and send the first update parameter to each participant i, so that each participant i updates the local model parameter of the participant i based on the first update parameter for the next iteration.

Images