Side information fusion sequence recommendation model for frequency domain collaborative enhancement

Abstract

The invention relates to the technical field of sequence recommendation, discloses a frequency domain collaborative enhancement side information fusion sequence recommendation model, and aims to solve the problem that high frequency or periodic noise on a frequency domain is difficult to relieve due to the fact that side information is fused on a time domain by an existing method in a current sequence recommendation task. The method comprises the following steps: performing adaptive filtering on a commodity sequence in a frequency domain to capture deep interests and preferences of a user, and adopting an adaptive threshold to reduce noise influence; the commodity and side information sequence is converted to a frequency domain for feature fusion through Fourier transform, and important features can be highlighted and irrelevant noise can be suppressed through the sparse feature of a frequency spectrum; besides, the model introduces frequency domain prediction loss to further optimize the prediction precision of the model, in contrast experiments on four public data sets, compared with an optimal base line, the average Recall (at) 10 and Recall (at) 20 of the method are improved by 4.35% and 4.57% respectively, the average NDCG (at) 10 and NDCG (at) 20 are improved by 2.45% and 3.02% respectively, the optimal performance expression is achieved, and the effectiveness of the method provided by the invention is proved.

Description

Technical Field

[0001] The present invention relates to the technical field of sequence recommendation, and in particular to a frequency domain collaboratively enhanced side information fusion sequence recommendation model. Background Art

[0002] Sequential Recommendation (SR) aims to model the historical behavior of users to capture their interests and preferences, so as to recommend potential products of interest to users. Early methods usually extract local or global features of product sequences by mapping product IDs into a high-dimensional latent space and using convolutional neural networks, position-based attention mechanisms, etc. to capture users' interests and preferences. Since the side information of products (such as categories, brands, etc.) contains rich product characteristics, these methods ignore the side information of products, resulting in limited depth and accuracy of user preference modeling. To overcome this limitation, recent studies have enhanced user preference representation by fusing side information to recommend more suitable products to users. Yuan et al. constructed a heterogeneous graph between products and their side information to obtain fused embeddings. Liu et al. first defined the information intrusion problem caused by early fusion and proposed to alleviate this problem by only adding attributes in the calculation of attention scores. Zhou et al. designed four auxiliary self-supervised tasks to further extract important information from side information. Li et al. captured user behavior by sparse modeling from coarse-grained to fine-grained. Xie et al. generated more refined and accurate user representations by optimizing the attention distribution in product sequences using side information. Hao et al. used independent attention layers to capture fine-grained interests in product sequences and coarse-grained interests in side information, and introduced contrastive learning to enhance the representation ability of the model. Lin et al. introduced a multi-sequence integrated attention layer to learn multiple representations of users by adaptively modeling interactions within and outside the sequence.

[0003] Although this type of method combines product sequences and side information in the time domain to enhance user preference representation and has achieved good results, there are various noises in the user interaction sequence, such as irregular noise and high-frequency noise. Existing methods ignore the potential noise in the product sequence, lack effective noise suppression, and are difficult to fully utilize the frequency characteristics of user behavior. Even if side information is introduced to enhance user preference representation, the impact of noise may still be amplified.

[0004] Therefore, it is necessary to provide a new frequency domain collaboratively enhanced side information fusion sequence recommendation model to solve the above technical problems. Summary of the invention

[0005] To solve the above technical problems, the present invention proposes a frequency-domain cooperative enhancement model for side information fusion in sequential recommendation (FCE-SR). First, an adaptive filtering layer is used to apply adaptive filtering to the product sequence to capture user preferences and reduce the influence of noise through an adaptive threshold. Secondly, a frequency domain fusion network is used to convert the product sequence and the side information sequence to the frequency domain for preliminary fusion. The spectrum in the frequency domain can not only enhance the features related to accuracy and suppress irrelevant features, but also provide a global perspective so that each spectrum component can focus on all features in the spatial domain. Next, the output of the frequency domain fusion network is non-invasively fused with the output of the adaptive filtering layer. Prediction loss is also introduced in the frequency domain to improve the quality of the prediction results. Overall experiments on four public general datasets, Amazon Beauty, Sports, Toys and Yelp, show that the FCE-SR method outperforms the existing basic sequential recommendation method and the sequential recommendation method that integrates side information.

[0006] The frequency domain collaboratively enhanced side information fusion sequence recommendation model provided by the present invention includes:

[0007] The embedding layer converts the user's interaction sequence V u ={v1,v2,...,v n} Input to the product embedding layer and attribute embedding layer, and then get the product embedding E ID and the embedding of product side information A ;

[0008] A frequency domain collaborative enhancement layer, wherein the frequency domain collaborative enhancement layer is stacked in multiple layers to alleviate the impact of noise in the user interaction sequence;

[0009] The prediction layer is used to predict the type of side information that the user is interested in.

[0010] Preferably, the frequency domain collaborative enhancement layer includes:

[0011] The frequency domain adaptive filtering layer is used to apply adaptive filtering to the product sequence to capture user preferences and reduce the impact of noise through adaptive thresholds;

[0012] Frequency domain fusion network, used to convert product sequences and side information sequences into frequency domain for preliminary fusion, where the sparse characteristics of the spectrum can highlight important features and suppress irrelevant noise;

[0013] Multi-head attention layer to further utilize the fused information;

[0014] The feedforward neural network is used to feed the output of the multi-head attention layer into the feedforward neural network layer to obtain a new product sequence representation.

[0015] Preferably, in the frequency domain adaptive filtering layer, the spatial features are first converted into spectral features by discrete Fourier transform, and this process is implemented by fast Fourier transform. Next, an adaptive filter is constructed, and then two sets of learnable filters are used to further process it: a global filter that learns global interaction features from the original frequency domain data and a local filter that learns local interaction features from the adaptive filter.

[0016] Preferably, in the prediction layer, a frequency domain loss is proposed to supervise network training to improve the quality of the prediction result, that is, the given representation and the representation of the target product are respectively converted to the frequency domain by Fourier transform to obtain their representations in the frequency domain, and the mean square error of the head and the imaginary part are respectively calculated, and then the sum is obtained to obtain the frequency domain prediction loss of the product;

[0017] For the prediction of products in the spatial domain, the cross entropy is used to calculate the spatial domain loss to measure the difference between the prediction and the true value of the product;

[0018] Based on the spatial domain prediction loss of the product, the frequency domain prediction loss of the product, and the side information prediction loss, the overall loss is obtained.

[0019] Compared with related technologies, the frequency domain collaboratively enhanced side information fusion sequence recommendation model provided by the present invention has the following beneficial effects:

[0020] The model can capture users' deep interest preferences by adaptively filtering the product sequence in the frequency domain, and uses adaptive thresholds to reduce the impact of noise, thereby improving the accuracy of recommendations.

[0021] The model uses Fourier transform to convert product and side information sequences into the frequency domain for feature fusion. The sparse characteristics of the spectrum can highlight important features and suppress irrelevant noise, thereby achieving more effective side information fusion.

[0022] This model outperforms other models when integrating different types of side information such as categories, cities, and brands, showing higher accuracy and stability;

[0023] The model further optimizes the prediction accuracy of the model by introducing frequency domain prediction loss. The model can effectively suppress the influence of noise and capture effective user interest features through frequency domain processing and filtering.

[0024] The model can handle sparse data as well as noisy data well. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1The overall framework diagram of a frequency domain collaboratively enhanced side information fusion sequence recommendation model provided by the present invention;

[0026] Figure 2 This is a result diagram of the ablation experiment in the present invention;

[0027] Figure 3 A bar graph showing the influence of different k values ​​on the recommended performance in the present invention;

[0028] Figure 4 A bar graph showing the impact of different frequency domain prediction losses on recommendation performance in the present invention;

[0029] Figure 5 This is a bar chart showing the impact of fusing different side information on recommendation performance in the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0031] In view of the problem that existing methods fuse side information in the time domain but are difficult to alleviate high-frequency or periodic noise, the present invention proposes a frequency-domain cooperative enhancement model for side information fusion in sequential recommendation (FCE-SR). Figure 1 As shown, the method of the present invention is mainly divided into three modules: 1) embedding layer; 2) frequency domain collaborative enhancement layer; 3) prediction layer.

[0032] 1.1 Embedding Layer

[0033] In the embedding layer, the user’s interaction sequence V u ={v1,v2,...,v n} Input to the product embedding layer and attribute embedding layer, and then get the product embedding E ID and the embedding of product side information A :

[0034] E ID =Φ ID ([I1,I2,...,I n ]). (1)

[0035] E A =Φ A ([a1,a2,...,a n ]). (2)

[0036] Where Φ(·) represents the corresponding embedding layer, which encodes the product ID and its side information into a vector. The embedding lookup table of the product and its side information can be expressed as M ID ∈R |I|×d , M A ∈R |A|×d .

[0037] 1.2 Frequency Domain Collaborative Enhancement Layer

[0038] In order to alleviate the impact of noise in user interaction sequences, a frequency domain collaborative enhancement layer is proposed, which mainly includes a frequency domain adaptive filtering layer (Frequency Domain Adaptive Filtering Layer, FDAF), a frequency domain fusion network (Frequency Domain Fusion Network, FDFN), a multi-head attention layer and a feedforward neural network. Multiple layers of frequency domain collaborative enhancement layers are stacked in the FCE-SR model. One of the layers is taken as an example below.

[0039] 1.2.1 Frequency Domain Adaptive Filtering Layer

[0040] Considering the existence of noise in the product sequence, inspired by FMLP-Rec, a frequency domain adaptive filtering layer is proposed, which uses Fourier analysis to enhance the feature representation of the product sequence and capture user preferences, while mitigating noise through adaptive thresholding.

[0041] First, the spatial features are converted into spectral features through Discrete Fourier Transform (DFT), and this process is implemented using Fast Fourier Transform (FFT). Given the representation of the product sequence E ID ∈R n×d , converting it to the frequency domain:

[0042] X ID =F(E ID )∈C n×d . (3)

[0043] Where F(·) represents the one-dimensional discrete Fourier transform, X ID Indicates E ID In the frequency domain, it is a complex tensor, C represents the complex space, n represents the sequence length, and d represents the hidden layer size.

[0044] Next, we construct the adaptive filter E f , first calculate X ID The power spectrum p of X is calculated by IDThe power spectrum p of helps to identify the main frequency components. The power spectrum p is calculated as the square of the frequency component size: p = |X ID | 2 , which gives a measure of the intensity of different frequencies in the time series data. Then the high-frequency components are adaptively filtered from the power spectrum p through a trainable threshold θ, which is the key to achieving effective noise reduction. This threshold θ is set as a learnable parameter, which is optimized during training through back propagation based on the spectral characteristics of the data. The basic signal frequency and noise are identified through θ. Adaptive filter E f The expression is as follows:

[0045] E f =X ID ⊙(p>θ). (4)

[0046] Among them, ⊙ represents element-wise multiplication, p>θ is a binary mask, and the frequencies with power higher than the threshold θ are retained while filtering out other frequencies.

[0047] After adaptive filtering of the frequency domain data, two sets of learnable filters are used to further process it. One is from the original frequency domain data X ID A global filter that learns global interaction features and an adaptive filter E f Learn local filters for local interaction features. G and W L are learnable global and local filters respectively. The use of these filters is as follows:

[0048] E G =W G ⊙X ID . (5)

[0049] E L =W L ⊙E f . (6)

[0050] E GL =E G +E L . (7)

[0051] Then, the inverse discrete Fourier transform (IDFT) is used to transform the representation E of the enhanced product GL Convert back to the spatial domain and use the Inverse Fast Fourier Transform (IFFT) to achieve this process:

[0052]

[0053] Among them, F -1 (·) represents the inverse one-dimensional discrete Fourier transform, which converts a complex tensor into a real tensor. Indicates E ID The output of the frequency domain adaptive filter layer. The frequency domain adaptive filter layer can reduce the noise in the product sequence and obtain a pure product sequence representation with stronger representation ability. In addition, residual connection, layer normalization (LayerNormalization, LayerNorm) and random dropout (Dropout) operations are added to alleviate the gradient disappearance and unstable training problems:

[0054]

[0055] 1.2.2 Frequency Domain Fusion Network

[0056] Previous fusion methods mainly perform feature modeling in the time domain, lack effective noise suppression, and have difficulty in fully utilizing the frequency characteristics of user behavior. The present invention shifts the perspective to the frequency domain and designs a more effective method to represent and fuse products and their side information. The Fourier transform usually produces a sparse spectrum. The spectrum provides a more effective means for the representation and fusion of products and their side information, in which the characteristic of significant frequency components close to zero helps to achieve the distinguishing representation of features, enhance features related to accuracy, and suppress irrelevant features. In addition, the spectral method can provide a global perspective, allowing each spectral component to focus on all features in the spatial domain, thereby effectively capturing multi-scale information. First, the enhanced product sequence representation is and side information representation E A Converted to the frequency domain, the formula is as follows:

[0057]

[0058] X A They are and E A In the frequency domain, they are all complex tensors. Through Fourier transform, the spatial features are effectively integrated in each frequency element, which makes it possible to extract important features from the product and its side information through dot product. Then, Introducing filter banks For X A Introducing filter banks To compress the spectral features, k represents the number of filters. The compressed spectral features help reveal important information related to the target product. By using the filter bank for element-level multiplication, these features are compressed, important components are highlighted, and irrelevant noise is reduced:

[0059]

[0060]

[0061] in, and |X A | 2 They are and X A The power spectrum of Length, and |X A | 2 The operation smoothes the spectrum and highlights the main components of the spectrum. It also facilitates subsequent learning of feature compression. It can better compress important features and aggregate information related to the target product. Its combination with the filter bank makes frequency domain feature compression more efficient.

[0062] By focusing on the key information components in the spectrum, the information is effectively enhanced while irrelevant components are suppressed, thus achieving the purpose of deep fusion. ID and H A Perform average pooling operations separately, and then apply convolution operations separately, so that a filter based on the product spectrum and a filter based on the side information spectrum can be obtained. These filters work together to filter and enhance useful features. Next, the two filters and the corresponding spectral features are cross-spectral selected by interleaving and multiplying the product spectrum-based filter (side information spectrum-based filter) with the compressed side information spectrum (compressed product spectrum features) features to achieve more refined feature fusion. In this process, the side information and product spectrum are organically combined to obtain a more expressive sequence embedding:

[0063]

[0064] Among them, Θ ID and θ A are one dimension and H respectively. ID and H A The trainable parameters are of the same size, Conv(·) is a 1×1 convolution layer, and Avg(·) is an average pooling function. Finally, the inverse discrete Fourier transform (IDFT) is used to transform the representation of the product and its side information back to the spatial domain:

[0065]

[0066] Among them, Z ID yes In the time domain, Z A yes Representation in the time domain. In addition, residual connections, layer normalization and dropout operations are added:

[0067]

[0068] Z ID =LayerNorm(E A +Dropout(Z A )). (17)

[0069] Then Z ID Send it to the multi-head attention layer for further learning, and Z A It is sent to the next frequency domain collaborative enhancement layer as new side information embedded in E A .

[0070] 1.2.3 Multi-Head Attention Layer

[0071] In order to further utilize the fused information, the product sequence fused with the side information is represented as Z ID It is sent to the multi-head attention layer. Different from the traditional multi-head self-attention layer, this layer uses Z ID Generated query (Query) Q ID and key K ID , using the pure commodity sequence after self-enhancement in the frequency domain Generate Value V ID , taking a head as an example, the attention calculation method is:

[0072]

[0073] Among them, Q ID =Z ID W Q , K ID =Z ID W K , d is the hidden layer dimension, W Q , W K , W V is the projection matrix. The output of each head is concatenated to get the output of the multi-head self-attention layer:

[0074]

[0075] Among them, s is the number of heads, W H is a learnable parameter, and Concat(·) is a concatenation operation.

[0076] 3.2.4 Feedforward Neural Network The output M of the multi-head attention layer is sent to the feedforward neural network layer to obtain a new product sequence representation:

[0077]

[0078] Among them, W1, W2, b1, b2 are learnable parameters, and ReLU (·) is an activation function. The fusion feature obtained later Input to the next frequency domain collaborative enhancement layer as the new input E ID In the final stage of the model, a final layer of sequence representation is used The last item in E p , used to perform subsequent related prediction tasks.

[0079] 1.3 Prediction Layer

[0080] 1.3.1 Prediction Layer

[0081] Based on the above fused information, the present invention predicts the type of information that the user is interested in. The prediction formula for the opposite side information A is as follows:

[0082]

[0083] in, represents the predicted value of the side information A, σ is the sigmoid activation function, and W A and b A is a learnable parameter, and then the cross entropy loss is used to calculate the prediction loss L of the side information AP :

[0084]

[0085] In addition, in order to improve the quality of the prediction results, a frequency domain loss is proposed to supervise the network training. p and the target product representation E y , convert them into the frequency domain through Fourier transform to get their representation in the frequency domain and F = F real +iF imag , calculate the mean square error of the head and imaginary parts respectively, and then sum them up to get the frequency domain prediction loss L of the product FP :

[0086]

[0087] L FP =MSE real +MSE imag . (25)

[0088] Where N is the number of samples, and They are E p The real and imaginary parts expressed in the frequency domain, F real and Fimag They are E y The real and imaginary parts are expressed in the frequency domain, i is the imaginary unit, MSE real is the real part of the computational loss, MSE imag is the computational loss of the imaginary part.

[0089] For the prediction of goods in the spatial domain, the cross entropy is used to calculate the spatial domain loss L SP , to measure the forecast of the product The difference between the true value y is:

[0090]

[0091] Prediction loss L based on the spatial domain of commodities SP , the frequency domain prediction loss of the product L FP , side information prediction loss L AP , the overall loss of the design of the present invention is as follows:

[0092] L Totall =L SP +L FP +L AP . (27)

[0093] 2. Experimental Analysis

[0094] 2.1 Dataset

[0095] In order to compare the performance of the models, four public datasets are used: Amazon Beauty, Sports, Toys, and Yelp, referring to the work of Lin et al. The statistical results of these four datasets after preprocessing are summarized in Table 1.

[0096] Table 1. Dataset statistics

[0097]

[0098] 2.2 Evaluation indicators

[0099] In the experiment of the present invention, referring to Lin et al., a leave-one-out cross-validation method was used for evaluation. For each user-item interaction sequence, the last two items were retained as validation data and test data, respectively, and the remaining items were used to train the sequence recommendation model. The performance of the sequence recommendation model is evaluated by the top-K recall rate (RECALL@K) and the top-K normalized discounted cumulative gain (NDGC@K) in an overall ranking manner, so as to make a fair comparison. This study takes the sum.

[0100] 2.3 Baseline Model and Implementation Details

[0101] Two groups of baseline models are selected to compare the recommendation performance, including the sequential recommendation baseline model and the sequential recommendation baseline model integrating side information.

[0102] Sequence recommendation baseline models: 1) Caser uses horizontal and vertical convolutional filters to learn multi-level patterns and user preferences. 2) BERT4Rec uses a bidirectional self-attention network to model user behavior sequences. 3) SASRec uses a self-attention mechanism to learn user interest preferences from product sequences. 4) DuoRec studies the problem of representation degradation in recommendation and proposes a solution using contrastive learning technology.

[0103] Sequential recommendation baseline models that integrate side information: 1) SASRecF extends SASRec and concatenates the embeddings of items and their side information and feeds them into the model for learning. 2) FDSA-CL is an improved version of FDSA that uses contrastive learning to exploit the beneficial interactions of representations from the side information level and the item level. 3) CAFE predicts the next item by adding and fusing user representations at the item and side information levels. 4) S3-Rec designs four pre-training tasks based on contrastive learning that maximizes mutual information. 5) ICAI feeds the fused embeddings obtained by constructing heterogeneous graphs into the recommendation model. 6) NOVA-SR designs a non-invasive attention mechanism to learn user representations. 7) DIF-SR fuses items and their side information by decoupling attention calculations and fusing attention matrices. 8) MSSR designs a multi-sequence integrated attention layer to learn multiple representations of users.

[0104] The FCE-SR models are implemented based on the recommended framework RecBole, and they are trained for 200 epochs using the Adam optimizer, with a learning rate of 1e-4 and hidden sizes set to 256. For other hyperparameters, the best settings mentioned in the previous paper are followed. NVIDIA GeForce RTX 3090 with 24GB memory is used for training.

[0105] 2.4 Results comparison and analysis

[0106] This section conducts overall performance experiments; ablation experiments; experiments on the impact of hyperparameter K; experiments on the impact of different sequence lengths; experiments on the impact of using different frequency domain losses; experiments on the impact of fusing different side information; the relevant experimental results and their experimental analysis are as follows.

[0107] 2.4.1 Overall performance experiment

[0108] Experiments were conducted on four public datasets, and the recorded experimental results are the average values ​​of the runs under 5 different random seeds. Table 2 shows the overall performance of all baseline models and the FCE-SR proposed in this invention. The performance data of the baseline models are cited from MSSR and DIF, and the datasets used are exactly the same as theirs.

[0109]

[0110]

[0111] Table 2. Overall performance experiment. The best performing model is marked in bold, and the second best performing model is marked with an underline.

[0112] Among the sequence recommendation methods, the following observations are made: 1) On all datasets, SASRec and BERT4Rec outperform Caser in all indicators, which proves the superiority of the attention mechanism. 2) Compared with other SR methods, DuoRec achieves the best results in most cases, except for SASRec which performs better in some cases on Yelp, which illustrates the effectiveness of contrastive learning denoising and also shows that suppressing the presence of noise in the data helps improve the performance of the model.

[0113] Among the sequence recommendation methods with side information fusion, the following observations are made: 1) Simple side information fusion, such as SASrec, FDSA-CL, etc., is not always effective, and the results are worse than SASRec in most indicators of all datasets. This phenomenon is attributed to the intrusive fusion they adopt, which has the disadvantage of too large an embedding space. 2) Although S3Rec and ICAI improve the complexity of fusion and have better performance than SASRec, they still use intrusive fusion and do not consider the noise problem in the data. Compared with DuoRec, the performance difference is not much. 3) NOVA-SR, DIF-SR, and MSSR are all based on non-intrusive fusion. Although they improve the performance of the side information fusion model, they do not consider the noise problem in the data. 4) FCE-SR achieves the best performance in all indicators on four public datasets. Compared with the best baseline MSSR, FCE-SR's average Recall@10 and Recall@20 are improved by 4.35% and 4.57% respectively, and the average NDCG@10 and NDCG@20 are improved by 2.45% and 3.02% respectively, which proves the effectiveness of the proposed method.

[0114] 2.4.2 Ablation Experiment

[0115] In order to analyze the contribution of different modules in the FCE-SR model, this section conducts a detailed ablation study and presents the results in Figure 2 1) W / O FDAF: Remove the frequency domain adaptive filter layer in the model, do not perform noise reduction on the product sequence, and use the original product sequence directly. After removing the frequency domain adaptive filter layer (W / O FAMF) in the model, the performance degradation shows that using the original product sequence will introduce noise interference, reflecting the necessity of noise reduction on the product sequence. 2) W / O FDFN: Remove the frequency domain fusion network, use the fully connected layer, extract the features of the product and its side information, and fuse them. After removing the frequency domain fusion network (W / OFFN), the performance degradation shows the effectiveness of side information fusion in the frequency domain, and highlights the importance of frequency features in user behavior. 3) W / O FP: Remove the frequency domain prediction loss. After removing the frequency domain prediction loss (W / O FP), the performance degradation shows that the frequency domain prediction loss plays a role in improving the prediction accuracy of the model, which shows that introducing prediction constraints in the frequency domain helps capture the frequency features in user behavior. 4) W / OAP: Remove the side information prediction loss. After removing the side information prediction loss (W / OAP), the performance decreases, indicating that the side information prediction loss plays a role in enhancing the model's utilization of side information. By introducing the side information prediction loss, the characteristics of side information can be more effectively captured and fused.

[0116] 2.4.3 Impact of Hyperparameters

[0117] This experiment aims to study the impact of the number of filters in the filter bank on the model performance. This experiment uses the same number of filters for the product and its side information, and the number of filters k∈{2,4,6,8}, Figure 3 The experimental results of the model using different numbers of filters in four datasets are shown, and Recall@20 and NDCG@20 are used as evaluation indicators. The experimental results show that the number of filters has different effects on different datasets. In the Beauty and Sport datasets, a smaller number of filters (k=2) can achieve better results, while increasing the number of filters does not significantly improve the performance. On the contrary, in the Toys and Yelp datasets, a larger number of filters can achieve better results. The Toys dataset performs best when k=6, and the Yelp dataset performs best when k=8, indicating that more filters help capture the features in the dataset. In general, different datasets have different requirements for the number of filters, reflecting the diversity of feature distribution in each dataset.

[0118]

[0119] Table 3. Model performance (Recall@k) at different sequence lengths

[0120]

[0121] Table 4. Model performance under different sequence lengths (NDCG@k)

[0122] 2.4.4 Effect of different sequence lengths

[0123] This experiment explores the performance of the model under different data sparsity conditions by setting different maximum sequence lengths (values ​​are {10, 20, 30, 40, 50}). The experiment uses four indicators, Recall@10, Recall@20, NDCG@10 and NDCG@20, to evaluate on four datasets: Beauty, Sport, Toys and Yelp. The results are shown in Tables 3 and 4. As the sequence length increases (from 10 to 50), the Recall and NDCG indicators on most datasets show a steady improvement. Longer sequences not only provide the model with richer historical interaction information, which helps to capture user preferences more comprehensively, but also inevitably introduce more noise. When the maximum sequence length reaches 50, the model still achieves the best performance, and it can be found that the model can handle sparse data and noisy data well.

[0124] 2.4.5 Effects of using different frequency domain losses

[0125] This experiment uses different losses as frequency domain losses and compares the performance of L1 and L2 loss functions on four datasets. The evaluation indicators are Recall@20 and NDCG@20. From the experimental results, it can be seen that the performance of L2 loss is better than L1 loss on all datasets, especially on Toys and Yelp datasets, which suggests that the model has a certain noise reduction ability. In theory, L2 loss is usually sensitive to large errors and is more susceptible to noise interference. However, the experimental results show that when using L2 loss, the model still maintains a relatively robust performance, indicating that the model can effectively suppress the influence of noise and capture effective user interest features through frequency domain processing and filtering mechanisms.

[0126] 2.4.6 The impact of integrating different side information

[0127] In order to evaluate the ability of the proposed model in fusing different types of side information, this experiment explores the performance of multiple sequence recommendation models that fuse different side information. This experiment compares the effects of NOVA-SR, DIF-SR, MSSR and FCE-SR models when fusing categories, cities and brands. The evaluation indicators are Recall@20 and NDCG@20. The results are shown in Figure 2. Figure 5As shown in the figure, the FCET-SR model performs better than other models in fusing categories, cities, and brands, especially in fusing categories and city information. This shows that the FCE-SR model exhibits higher accuracy and stability when dealing with diversified side information fusion. In contrast, although the DIF-SR and MSSR models perform worse in different dimensions, they are still better than the NOVA-SR model. The NOVA-SR model performed relatively poorly in all test results, especially in fusing categories and city information. At the same time, it can be observed from the figure that all models perform better than categories when fusing brand information, and perform worst when fusing city information. This phenomenon can be attributed to differences in information granularity, relevance, and data characteristics:

[0128] Difference in information granularity: Brand information usually has more fine-grained features and can better capture users' personalized preferences and needs. In contrast, category information is more generalized and difficult to reflect the subtle differences of users. City information is broader and less instructive to users' specific interests and purchasing behaviors. Especially in recommendation tasks, city information has a low relevance and is difficult to directly reflect users' preferences for products.

[0129] Differences in the relevance of information: Brand and category information are more directly related to the characteristics of the product and can more accurately describe the attributes of the product and user behavior patterns. However, city information varies greatly among different users and may be data sparse, making it difficult for the model to extract useful features from it. These factors together lead to the situation where brand information fusion has the best effect, followed by category, and city information fusion has the lowest effect.

[0130] Compared with related technologies, the frequency domain collaboratively enhanced side information fusion sequence recommendation model provided by the present invention has the following beneficial effects:

[0131] The present invention proposes a frequency domain collaboratively enhanced side information fusion sequence recommendation model, aiming to improve the fusion effect of side information in the sequence recommendation task. This method uses a frequency domain adaptive filter layer to capture user preferences and reduces noise through an adaptive threshold. A frequency domain fusion network is used to perform mutual enhancement fusion of information in the frequency domain, and prediction loss is introduced in the frequency domain to improve the quality of the prediction results. Related experiments show the effectiveness and rationality of the FCE-SR method proposed in the present invention. In future work, we will explore the integration of more diverse side information into the model and further promote the progress of sequence recommendation using side information.

[0132] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A frequency domain collaboratively enhanced side information fusion sequence recommendation model, characterized in that: include: Embedding layer, the embedding layer includes a product embedding layer and an attribute embedding layer Input the user's interaction sequence into the product embedding layer and attribute embedding layer, and then obtain the embedding of the product and the embedding of the product side information; A frequency domain collaborative enhancement layer, wherein the frequency domain collaborative enhancement layer is stacked in multiple layers to alleviate the impact of noise in the user interaction sequence; The prediction layer is used to predict the type of side information that the user is interested in.

2. The frequency domain collaboratively enhanced side information fusion sequence recommendation model according to claim 1, characterized in that: The frequency domain collaborative enhancement layer comprises: The frequency domain adaptive filtering layer is used to apply adaptive filtering to the product sequence to capture user preferences and reduce the impact of noise through adaptive thresholds; Frequency domain fusion network, used to convert product sequences and side information sequences into frequency domain for preliminary fusion, where the sparse characteristics of the spectrum can highlight important features and suppress irrelevant noise; Multi-head attention layer to further utilize the fused information; The feedforward neural network is used to feed the output of the multi-head attention layer into the feedforward neural network layer to obtain a new product sequence representation.

3. The frequency domain collaboratively enhanced side information fusion sequence recommendation model according to claim 2, characterized in that: In the frequency domain adaptive filtering layer, the spatial features are first converted into spectral features by discrete Fourier transform, and this process is implemented by fast Fourier transform. Next, an adaptive filter is constructed, and then two sets of learnable filters are used to further process it: a global filter that learns global interaction features from the original frequency domain data and a local filter that learns local interaction features from the adaptive filter.

4. The frequency domain collaboratively enhanced side information fusion sequence recommendation model according to claim 1, characterized in that: In the prediction layer, a frequency domain loss is proposed to supervise network training to improve the quality of prediction results, that is, the given representation and the representation of the target product are converted to the frequency domain through Fourier transform to obtain their representations in the frequency domain, and the mean square errors of the head and the imaginary part are calculated respectively, and then the frequency domain prediction loss of the product is obtained by summing them up; for the prediction of the product in the spatial domain, the spatial domain loss is calculated using cross entropy to measure the difference between the prediction and the true value of the product; based on the spatial domain prediction loss of the product, the frequency domain prediction loss of the product, and the side information prediction loss, the overall loss is obtained.

Images