A method for adaptively setting the representation modulus length of a recommendation system

Abstract

A method for adaptively setting the representation modulus length of a recommendation system, comprising the following steps: calculating a reference temperature coefficient τ0 according to the current state of the model; calculating, for each user, a personalized temperature coefficient τ based on τ0 u ; according to the above τ u , obtaining a representation vector with adjusted modulus length of the user item. The advantages of the present invention are: it can adaptively find the optimal representation modulus length at a fine-grained level, and improve the recommendation performance of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of collaborative filtering of recommendation systems, especially the technical field of collaborative filtering based on representation learning. Background Art

[0002] To achieve a more accurate and efficient design of recommendation systems, existing recommendation system models are constantly updated and iterated, aiming to design models with stronger adaptability to real scenarios. With the rapid development of hardware in the past decade and the in-depth cultivation of big data technology, recommendation systems based on representation learning are gradually replacing the models based on rules or matrix factorization techniques in the past. Among them, the collaborative filtering task realizes relatively accurate user recommendations only through the data of users' historical click records without relying on other additional auxiliary information, which is one of the most extensive and basic tasks in recommendation systems.

[0003] Although existing recommendation models based on representation learning have achieved relatively good performance improvement by building increasingly large and complex network architectures, existing recommendation models lack the exploration of the length of representations, and the length of representations plays a crucial role in the performance of recommendations. Existing methods either use regularization terms to implicitly control the length of representations, however, this will lead to the problem of popularity bias in recommendation results and also affect the convergence speed of model training; or directly adopt the method of length normalization in a brute-force manner and use a hyperparameter, i.e., the temperature coefficient, to control the normalized length of representations. However, the model performance is very sensitive to the temperature coefficient, and it is often difficult to find a suitable temperature coefficient manually. At the same time, the unified temperature coefficient lacks flexibility and is not conducive to the training of recommendation models. To overcome this problem, a method for setting an adaptive temperature coefficient must be designed. Summary of the Invention

[0004] The present invention aims to overcome the above problems of the prior art and proposes a method for adaptively setting the length of representations of a recommendation system.

[0005] A method for adaptively setting the length of representations of a recommendation system according to the present invention includes the following steps:

[0006] 1) Calculate the baseline temperature coefficient according to Equation (1) according to the state of the current model:

[0007]

[0008] where μ + , the μ distribution represents the mean of the predicted scores of the recommendation model for positive samples and all samples, that is Here, n, m represent the number of users and items in the recommendation system; D represents the set of positive samples, and each element of it is a user-item pair; |D| represents the number of positive samples; f uiis the predicted value of the sample formed by user \(u\) and item \(i\). Any recommendation model based on representation learning can be used, i.e., \(f\) ui = F(h u , h i ), where F(.,.) represents the model used, and h u , h i represent the original unadjusted magnitude representation vectors of the user and the item respectively. For example, for a typical matrix factorization model,

[0009] 2) According to the above \(\tau_0\), for each user, calculate its personalized temperature coefficient:

[0010]

[0011] where \(L(u)\) is the cumulative loss mean of each user, that is where \(\delta(.,.)\) represents the loss function selected for use, which can be cross-entropy loss, squared loss, etc.; \(r\) ui is the observed label, representing whether the user likes this item; \(q\) is the loss threshold of the entire dataset. For the convenience of calculation, \(q\) is set to the mean of \(L(u)\) here, that is \(W(.)\) is the Lambert-W function; \(max(.,.)\) represents taking the maximum of the two; \(\beta\) is a coefficient, which can be simply set to 1 here.

[0012] 3) According to the above \(\tau\) u , obtain the adjusted magnitude representation vectors of the user and the item:

[0013]

[0014] where \(|.| \) represents the L2 norm of the vector.

[0015] Based on the commonly used model matrix factorization MF and collaborative graph neural network LGN as the basic recommendation models, two publicly available real datasets Yelp and MovieLens are selected for experiments. Starting from the evaluation principle of the implicit feedback recommendation system, the dataset is divided into a training set, a validation set, and a test set. Recall@20 and NDCG@20 are used as metrics.

[0016] Table 1 Performance comparison of different methods on two datasets

[0017]

[0018] As can be seen from Table 1: On two different types of datasets, the proposed norm length control methods of the present invention (i.e., MF+Adapt-τ, LGN+Adapt-τ) can both outperform the existing methods based on L2 or brute-force search by hand (i.e., MF+L2, LGN+L2, MF+Norm, LGN+Norm).

[0019] The present invention proposes a method for adaptively setting the norm length of the recommendation system's representation, and realizes the adaptive fine-grained setting of the norm length by examining statistical indicators such as the predicted value and cumulative loss of the recommendation model.

[0020] Compared with the existing methods based on L2 regularization and brute-force search by hand, the advantages of the present invention are: it can adaptively find the fine-grained optimal representation norm length and improve the recommendation performance of the model. Brief Description of the Drawings

[0021] Figure 1 is the overall flowchart of the present invention. Detailed Embodiments

[0022] Now, various exemplary embodiments of the present invention will be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and implementation schemes of the present invention. It should be understood that the terms described in the present invention are only for describing specific embodiments and are not used to limit the present invention. Additionally, for the numerical ranges in the present invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Each intermediate value within any stated value or stated range, as well as each smaller range between any other stated value or intermediate value within the stated range, is also included in the present invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0023] Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention pertains. Although the present invention only describes preferred methods and materials, any methods and materials similar or equivalent to those described herein can also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods and / or materials related to the documents. In case of conflict with any incorporated document, the content of this specification shall prevail.

[0024] Without departing from the scope or spirit of the present invention, various improvements and changes can be made to the specific embodiments of the present invention's specification, which are obvious to those skilled in the art. Other embodiments obtained from the specification of the present invention are obvious to those skilled in the art. The specification and embodiments of this application are only exemplary.

[0025] As used herein, terms such as "comprising", "including", "having", "containing", etc. are all open-ended terms, meaning including but not limited to.

[0026] In the present invention, unless otherwise specified, "parts" are all calculated by mass parts.

[0027] The present invention provides a network commodity recommendation method based on adaptively setting the representation modulus length, as Figure 1 shown, and its steps include:

[0028] Step S1: Obtain and process data.

[0029] Specifically: In an application scenario of a recommendation system, taking the commodity review platform Yelp as an example, collect the interaction data between users and commodities on the platform, denoted as the positive sample set D. Each of its elements is a user-item pair (u, i) ∈ D, representing that user u likes item i. We also use the label r ui to represent whether the user likes this item.

[0030] Step S2: Train a recommendation model by using a method based on adaptively setting the representation modulus length.

[0031] This step is the core of the present invention, and the following elaborates on it and takes the classical recommendation model based on matrix factorization as an example for illustration. Specifically, it includes the following sub-steps:

[0032] 1) According to the state of the current model, calculate the reference temperature coefficient according to the following formula:

[0033]

[0034] where μ + , the μ distribution represents the mean of the predicted scores of the recommendation model for positive samples and all samples, that is Here, n and m represent the number of users and items in the recommendation system; D represents the positive sample set, and each of its elements is a user-item pair; |D| represents the number of positive samples; f ui represents the predicted value of the sample composed of user u and item i, and any recommendation model based on representation learning can be used. For example, for a typical matrix factorization model, here h u , h i respectively represent the unadjusted representation vectors of the original moduli lengths of the user and the item.

[0035] 2) According to the above τ0, calculate the personalized temperature coefficient for each user:

[0036]

[0037] Here, L(u) is the cumulative loss mean for each user, that is Here, δ(.,.) represents the loss function selected for use, which can be cross-entropy loss, squared loss, etc.; r ui is the observed label, representing whether the user likes this item; q is the loss threshold for the entire dataset. For the convenience of calculation, q is set to the mean of L(u) here, that is W(.) is the Lambert-W function; max(.,.) represents taking the maximum of the two; β is a coefficient, which can be simply set to 1 here.

[0038] 3) According to the above τ u , obtain the representation vector of the user item after adjusting the norm:

[0039]

[0040] Here, |.| represents the L2 norm of the vector.

[0041] 4) Use the representation after adjusting the norm to train the model, that is, use the following formula to calculate the loss function:

[0042]

[0043] And use the gradient descent algorithm to update the model.

[0044] Step S3: Implement commodity recommendation according to the learned representation after adjusting the norm.

[0045] Specifically: For each user u, calculate the interest score of the user for a certain item i according to and recommend the top K items with the highest scores to the user according to this interest score.

[0046] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that those skilled in the art can think of according to the inventive concept of the present invention.

Claims

1. A method for adaptively setting the representation modulus length of a recommendation system, characterized in that: It includes the following steps: 1) Calculate the reference temperature coefficient τ0 according to the current state of the recommendation model. The calculation process is as follows: where μ + , the μ distribution represents the mean of the predicted scores of the recommendation model for positive samples and all samples, that is Here, n and m represent the number of users and items in the recommendation system; D represents the set of positive samples, and each of its elements is a user-item pair; |D| represents the number of positive samples; f ui represents the predicted value of the model for the sample composed of user u and item i; 2) Based on τ0, for each user, calculate their personalized temperature coefficient τ u ; The calculation process is as follows: For each user u, calculate its personalized temperature coefficient according to the following formula: Here, L(u) is the cumulative loss mean for each user, that is Here, δ(.,.) represents the loss function selected for use; r ui is the observed label, representing whether the user likes this item; q is the loss threshold for the entire dataset, and here q is set to the mean of L(u), that is W(.) is the Lambert-W function; max(.,.) represents taking the maximum value of the two; β is a coefficient; 3) According to the above τ u , obtain the representation vector after adjusting the modulus length of the user item, and realize commodity recommendation.

2. The method for adaptively setting the representation modulus length of a recommendation system according to claim 1, characterized in that: h u ,h i respectively represent the original unadjusted magnitude representation vectors of the user and the item.

3. The method for adaptively setting the representation modulus length of a recommendation system according to claim 1, characterized in that: The selected loss function δ(.,.) is the cross-entropy loss or the squared loss.

4. The method for adaptively setting the representation modulus length of a recommendation system according to claim 1, characterized in that: β is set to 1.

5. The method for adaptively setting the representation modulus length of a recommendation system according to claim 1, characterized in that: The calculation process in step 3) is as follows: Here, |.| represents the L2 norm of the vector.

Images