Multi-aspect features of items for time-ordered sequential recommendation

Abstract

The key to sequential recommendation modeling is to capture dynamic users’ interests. Existing sequential recommendation methods (e.g., self-attention mechanism) have achieved extraordinary success in modeling users’ interests. However, these models ignore that users have different levels of preferences for different aspects of items, failing to capture users’ most concerning aspects. In addition, they are highly dependent on the quality of training data, which may lead to overfitting of the model when the training data is insufficient. To address the above issues, we propose a novel sequence-aware model (Multi-Aspect Features of Items for Time-Ordered Sequential Recommendation, MFITSRec), which combines the features of items with user behavior sequences to learn more complex item-item and item-attribute relationships. Moreover, the model uses a self-attention network based on an absolute time relationship, which can better represent the changes in users’ interests and capture users’ preferences for particular aspects of items. Extensive experiments on five datasets demonstrate that our model outperforms various baseline models. In particular, the model’s prediction accuracy has been significantly improved on sparse datasets.

Keywords

Sequential recommendation multi-aspect preferences of users data sparsity absolute time relationship

1 Introduction

With the increase of information on the Internet, personalized recommendations are vital. It is a constant challenge to effectively capture the dynamically changing interests of users and to make accurate recommendations for their next interaction. Since, in each platform, sequences composed of users’ historical interaction behaviors are inherently sequential, a large number of research capture users’ dynamic interest changes from sequences and predict the next possible interactive items for users [1, 2].

In order to model the dynamic interests of users, researchers have proposed many types of sequential recommendation methods [3–5]. Markov Chains (MCs) [3, 6] modeled the user behavior sequence as a classical approach and predicted the next interaction by transferring probabilities. FPMC [7] combined matrix decomposition and MCs for simulating personalized transfer probabilities. Usually, both approaches only consider low-order interactions (i.e., first-order and second-order) and ignore higher-order interaction information. Deep learning-based approaches have become the dominant methods for modeling the user behavior sequence [8–10]. In addition, methods based on the self-attention mechanism [5, 11–13] adaptively adjust the weights, resulting in a substantial improvement in the performance of the sequential recommendation model. It should be noted that these deep learning methods generally require a large amount of high-quality data to achieve performance.

Traditional sequential recommendation models rely on sequential relationships and emphasize the sequentiality of interaction actions. However, some users will not frequently engage in interactive behaviors in real life, which tends to result in historical data sparsity. Additionally, for different items, users have different preferences for various aspects of them. For example, two users interact with the same type of items; One of them interacts with each item from a different brand, while another one interacts with all items from the same brand. We can find that the user has a strong preference for the brand, i.e., when interacting with the items, the user has a specific preference for the brand. Ignoring the multi-aspect features of items and relying only on the user’s historical interaction sequence for the prediction cannot accurately capture the dynamic changes in users’ interests and preferences for a particular aspect of items.

In this paper, we proposed MFITSRec, a time-ordered sequential recommendation model with multi-aspect features of items, which combined the unique features of items and user interaction sequences. It enabled a finite-length sequence to carry helpful information about various aspects of items. When the interaction data is sparse, MFITSRec can better capture users’ preferences for a particular aspect of an item and recommend items to users that match the particular preference. In addition, we considered the time interval as in TiSASRec [11] and introduced the absolute time relationship and the relative position of the interaction sequence into the self-attention mechanism. In summary, our contributions are as follows:

We propose embedding the multi-aspect features of items in the user interaction sequence to model users’ preferences for multiple features so as to explore the degree of preference of different users for different items.

We combine unique features of items and absolute temporal relationships to enhance the user interaction sequence and demonstrate that the model can improve recommendation performance on sparse datasets.

We conducted extensive experiments on five datasets from three platforms to demonstrate the effectiveness of our model. The results show that our model outperforms competitive baselines significantly in terms of recommendation performance on both sparse and dense datasets and effectively captures users’ preferences for various aspects of items.

2 Related Work

General recommendation methods [14–17] focus on modeling users’ interests using vast amounts of implicit [18] interaction data to obtain the most relevant items for the user. However, these methods ignore the sequentiality between user interactions and have difficulty capturing the dynamic changes in users’ preferences. As a result, sequential recommendations that model users’ dynamic preferences have received much attention. Early sequential recommendations relied heavily on MCs [3, 7], which modeled the user-item interaction data in a sequence and predicted the next interaction by transferring the probability. Although MC-based models take into account the dependencies between historical interactions, they still cannot capture more associations between behaviors in a relatively long sequence [5, 19]. Most of the existing models use a deep learning approach to enable the models to capture the long-term preferences of users. Convolutional neural networks (CNNs) [4, 9, 20] treated historical interactions as "images" and captured local features of sequences. Recurrent neural networks (RNNs) [8, 21–23] utilized gated recurrent units (GRU) to model user behavior sequences. Graph Neural Networks (GNN) [10, 24, 25] modeled sequences as graph-structured data and used GNN to extract the embedding vectors of items. Recently, the sequential recommendation model FMLP-Rec [26], based on the MLP architecture, encoded user sequences by learnable filters and achieved the best performance on sparse datasets. However, these sequence recommendation methods can still have limitations when dealing with long sequences, may suffer from gradient disappearance or explosion, and are less interpretable.

Fig. 1

The overall framework of MFITSRec. The model mainly consists of embedding layers, an absolute time relationship layer, the self-attention network layers, and a prediction layer. Firstly, obtaining the historical interaction sequence of a target user and enhancing the interaction representation by obtaining the features of each aspect of the corresponding items in the sequence and the timestamps of the interactions, at the same time, deriving the absolute time relationship between the interactions. Secondly, the absolute time relationship and the interaction sequence are input to a multi-layer self-attention sequential recommendation model based on the absolute time relationship for recommendation learning to obtain a multi-aspect preference representation of the user. Finally, the user’s score for each item is predicted using the representation and the combination embeddings of items.

The attention mechanism is first applied to machine translation tasks, which can make the output dependent on the relevant parts of the input, enhancing the model’s ability to handle long sequences while improving interpretability. Therefore, it is also naturally applied to the sequential recommendation [5, 27–29]. Li et al. [27] combined RNNs with the attention mechanism to capture users’ primary goals in the sequence. Wang et al. [30] integrated attention mechanisms into a shallow network to construct contextual representations without strict sequential relationships. RepeatNet [31] incorporated the repeated exploration mechanism with the attention mechanism into RNN. It is worth noting that these models essentially introduce attentional ideas into some original models (e.g., CNN, RNN, etc.) and still rely on these original models. As Transformer [32], a purely attention-based mechanism, achieves state-of-the-art performance in machine translation tasks, the self-attention mechanism is widely used in sequence recommendation [5, 11–13]. One of its advantages is that it has captured long-term dependencies by computing the attention weights between each pair of items in a sequence and is easily parallelized. SASRec [5] first used self-attention for the sequential recommendation, adapting weights to items. Since the position of items in a sequence is unknown by the self-attentive model, some approaches embed relative position relationships as inputs to the model [5, 12, 13]. However, relying only on relative position information is insufficient for determining the impact of items with different time intervals on the next item. To overcome the problem, some researchers embed timestamp information into the model to get the temporal relationship between each item [11, 33, 34]. TiSASRec [11] modeled interaction sequences as sequences with time intervals and enhanced SASRec by adding timestamp information. As the information proliferated, Stai et al. [35] extended the recommendation process to the set of accompanying information except for the item itself, while they explored the examined research topic by introducing computationally efficient solutions. Recently, the watchlist recommendation (WLR) modeled Trans2D [36] uses items’ attributes to recommend against eBay users’ watchlists. Most existing models fail to capture users’ preferences for specific aspects of items and suffer from overfitting due to data sparsity. Our work embedded multi-aspect features of items into the user’s historical sequence to capture multi-aspect preferences of users and enhance interaction representations while combining relative location and absolute time relationships to model the timestamp sequence.

3 Methodology

We propose a sequential recommendation model MFITSRec that combines multi-aspect features of items with timestamps, mainly including the embedding layer, the absolute temporal relationship layer, the self-attention network layers, and the prediction layer. This section describes the specific construction of the model, and Figure 1 shows the architecture of MFITSRec.

In the sequential recommendation task of this section, the set of users and items are denoted as $U$ and $I$ respectively. We give a user interaction sequence $L^{u} = (l_{1}^{u}, l_{2}^{u}, \dots, l_{| L^{u} |}^{u})$ , where $u \in U$ , $l_{i}^{u} \in I$ , and a sequence of timestamps $T^{u} = (t_{1}^{u}, t_{2}^{u}, \dots, t_{| T^{u} |}^{u})$ , corresponding to each interaction. In addition, we extract the feature set $F = (f_{1}, f_{2}, \dots, f_{k})$ of items, where f_i is the i-th feature. The goal of the task is to predict the next item that the user may interact with. We use the user interaction sequence $(l_{1}^{u}, l_{2}^{u}, \dots, l_{| L^{u} | - 1}^{u})$ , the corresponding timestamp sequence $(t_{1}^{u}, t_{2}^{u}, \dots, t_{| T^{u} | - 1}^{u})$ , and the multi-aspect features lookup table of all items as inputs to the model. The following sequence of items $(l_{2}^{u}, l_{3}^{u}, \dots, l_{| L^{u} |}^{u})$ corresponding to the interaction is taken as the desired output at each time step. Our notation is summarized in Table 1.

Table 1
Notation

Notation Description

$U, I$ user and item set

$F$ feature set for item

L ^u historical interaction sequence for user u

T ^u timestamp sequence of user u corresponding to the interaction sequence

$R \in ℕ^{n \times n}$ absolute time relationship matrix

$k \in ℕ$ number of features

$n \in ℕ$ maximum sequence length

$b \in ℕ$ number of stacked self-attention network layers

$d_{i} \in ℕ$ latent dimension of item

$d_{g}, d_{b} \in ℕ$ latent dimensions of category and brand

$d \in ℕ$ latent vector dimension

$I \in ℝ^{| I | \times d_{i}}$ item embedding matrix

$A \in ℝ^{| I | \times k}$ multi-aspect features lookup table for the items

$A^{g} \in ℝ^{| I | \times d_{g}}, A^{b} \in ℝ^{| I | \times d_{b}}$ embedding matrix of category and brand

$P^{K}, P^{V} \in ℝ^{n \times d}$ embedding matrix of relative positions for key and value

$R^{K}, R^{V} \in ℝ^{n \times n \times d}$ embedding matrix of absolute time relationship for key and value

$E^{I} \in ℝ^{n \times d}$ input embedding matrix

$S_{i} \in ℝ^{d}$ embedding matrix after self-attention layer as time step i

$F_{i} \in ℝ^{d}$ embedding matrix after point-wise feed-forward network as time step i

Notation	Description
$U, I$	user and item set
$F$	feature set for item
L ^u	historical interaction sequence for user u
T ^u	timestamp sequence of user u corresponding to the interaction sequence
$R \in ℕ^{n \times n}$	absolute time relationship matrix
$k \in ℕ$	number of features
$n \in ℕ$	maximum sequence length
$b \in ℕ$	number of stacked self-attention network layers
$d_{i} \in ℕ$	latent dimension of item
$d_{g}, d_{b} \in ℕ$	latent dimensions of category and brand
$d \in ℕ$	latent vector dimension
$I \in ℝ^{\| I \| \times d_{i}}$	item embedding matrix
$A \in ℝ^{\| I \| \times k}$	multi-aspect features lookup table for the items
$A^{g} \in ℝ^{\| I \| \times d_{g}}, A^{b} \in ℝ^{\| I \| \times d_{b}}$	embedding matrix of category and brand
$P^{K}, P^{V} \in ℝ^{n \times d}$	embedding matrix of relative positions for key and value
$R^{K}, R^{V} \in ℝ^{n \times n \times d}$	embedding matrix of absolute time relationship for key and value
$E^{I} \in ℝ^{n \times d}$	input embedding matrix
$S_{i} \in ℝ^{d}$	embedding matrix after self-attention layer as time step i
$F_{i} \in ℝ^{d}$	embedding matrix after point-wise feed-forward network as time step i

3.1 Embedding Layer

Since the number of interactions varies from user to user, we fix the length of each user interaction sequence to a specific value $n \in ℕ$ . When the number of user interactions exceeds n, the effect of the first few interaction items on predicting the next item for the user can be ignored. For user $u \in U$ , a new input sequence l = (l₁, l₂, ⋯ , l_n) is formed after fixing the length of the user interaction sequence $(l_{1}^{u}, l_{2}^{u}, \dots, l_{| L^{u} | - 1}^{u})$ . If lvertL^urvert - 1 is longer than n, only the n most recent interactions are selected; if lvertL^urvert - 1 is shorter than n, a padding item 0 is added to the left side of the sequence until the length is n. At the same time, the timestamp sequence $(t_{1}^{u}, t_{2}^{u}, \dots, t_{| T^{u} | - 1}^{u})$ corresponding to the user interaction sequence is transformed accordingly to obtain t = (t₁, t₂, ⋯ , t_n). Different from the input sequence, the filler of the timestamp sequence is the timestamp corresponding to the earliest interactive item l₁ in l.

We build an item embedding matrix $I \in ℝ^{| I | \times d_{i}}$ , where d_i is the latent dimension of item, and retrieve the item embedding of the user input sequence to obtain the input embedding matrix $E \in ℝ^{n \times d_{i}}$ :

$E = [\begin{matrix} I_{l_{1}} \\ I_{l_{2}} \\ \dots \\ I_{l_{n}} \end{matrix}]$ (1)

For the items in $I$ , we use their attributes to represent the multi-aspect features of items while enhancing the item representation. After extracting the feature set $F$ of items, we build a multi-aspect features lookup table $A \in ℝ^{| I | \times k}$ of items, expressed as:

$A = [\begin{matrix} A_{I_{1}}^{f_{1}} & A_{I_{1}}^{f_{2}} & \dots & A_{I_{1}}^{f_{k}} \\ A_{I_{2}}^{f_{1}} & A_{I_{2}}^{f_{2}} & \dots & A_{I_{2}}^{f_{k}} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ A_{I_{| I |}}^{f_{1}} & A_{I_{| I |}}^{f_{2}} & \dots & A_{I_{| I |}}^{f_{k}} \end{matrix}]$ (2)

Where $A_{I_{i}}^{f_{j}}$ is the j-th feature value of item I_i. For each feature in $F$ , we define embedding matrices of k features separately:

$A^{f_{1}} = [\begin{matrix} a_{I_{1}}^{f_{1}} \\ a_{I_{2}}^{f_{1}} \\ \dots \\ a_{I_{| I |}}^{f_{1}} \end{matrix}], A^{f_{2}} = [\begin{matrix} a_{I_{1}}^{f_{2}} \\ a_{I_{2}}^{f_{2}} \\ \dots \\ a_{I_{| I |}}^{f_{2}} \end{matrix}], \dots, A^{f_{k}} = [\begin{matrix} a_{I_{1}}^{f_{k}} \\ a_{I_{2}}^{f_{k}} \\ \dots \\ a_{I_{| I |}}^{f_{k}} \end{matrix}]$ (3)

Where $A^{f_{i}} \in ℝ^{| I | \times d_{f_{i}}}$ is the embedding matrix of feature f_i, $a_{I_{j}}^{f_{i}}$ is the embedding of feature f_i for item I_j, and d_{f
_i} is the embedding dimension of feature f_i. We experiment with multiple features of items and show the user’s attention to them in Section 4.3. Constrained by time complexity, we remove some items’ attributes that get little attention from users and consider only categories and brands whose embedding matrices are denoted as $A^{g} \in ℝ^{| I | \times d_{g}}$ and $A^{b} \in ℝ^{| I | \times d_{b}}$ respectively:

$A^{g} = [\begin{matrix} a_{I_{1}}^{g} \\ a_{I_{2}}^{g} \\ \dots \\ a_{I_{| I |}}^{g} \end{matrix}], A^{b} = [\begin{matrix} a_{I_{1}}^{b} \\ a_{I_{2}}^{b} \\ \dots \\ a_{I_{| I |}}^{b} \end{matrix}]$ (4)

Where $a_{I_{i}}^{g} \in ℝ^{d_{g}}$ and $a_{I_{i}}^{b} \in ℝ^{d_{b}}$ are the category and brand embedding of item I_i, d_g and d_b are the latent dimensions of category and brand respectively. We retrieve the feature embeddings of the corresponding items of the sequence in the embeddings of items’ categories and brands respectively, and fuse them with the item embeddings contained in the input embedding matrix, denoted as $E^{I} \in ℝ^{n \times d}$ :

$E^{I} = [\begin{matrix} [I_{l_{1}}; a_{I_{l_{1}}}^{g}; a_{I_{l_{1}}}^{b}] \\ [I_{l_{2}}; a_{I_{l_{2}}}^{g}; a_{I_{l_{2}}}^{b}] \\ \dots \\ [I_{l_{n}}; a_{I_{l_{n}}}^{g}; a_{I_{l_{n}}}^{b}] \end{matrix}]$ (5)

Where $[I_{l_{i}}; a_{I_{l_{i}}}^{g}; a_{I_{l_{i}}}^{b}]$ is the connection of item embedding $I_{l_{i}} \in ℝ^{d_{i}}$ with the category feature embedding $a_{I_{l_{i}}}^{g} \in ℝ^{d_{g}}$ and brand feature embedding $a_{I_{l_{i}}}^{b} \in ℝ^{d_{b}}$ corresponding to item l_i, and d = d_i + d_g + d_b. Compared to [11], the introduction of features of items allows a finite-length sequence to carry more helpful information and capture users’ preferences for specific aspects of the items.

We still use the learnable relative position embedding as in [5] because multiple user interactions may occur in the same absolute timestamp, and it may not be possible to accurately model the sequentiality in the input sequence with only the absolute timestamp. For the key and value in the self-attention network layer, we use $P^{K} \in ℝ^{n \times d}$ and $P^{V} \in ℝ^{n \times d}$ to denote their relative positions respectively:

$P^{K} = [\begin{matrix} p_{K 1} \\ p_{K 2} \\ \dots \\ p_{Kn} \end{matrix}], P^{V} = [\begin{matrix} p_{V 1} \\ p_{V 2} \\ \dots \\ p_{Vn} \end{matrix}]$ (6)

3.2 Absolute time relationship layer

We introduced an absolute time relationship, as in [11], to model the effect of user interactions at different intervals on predicting the next item. Different users generate interactions with different frequencies, which makes modeling their interactions difficult because the intervals between them do not have the same magnitude, so we express the absolute time relationship between two interactions as their relative time interval. For the time sequence t = (t₁, t₂, ⋯ , t_n) of fixed length, the absolute time relationship between item i and item j is:

$r_{ij} = ⌊ \frac{| t_{i} - t_{j} |}{t_{\min}} ⌋$ (7)

Where $t_{\min} \in ℕ$ is the minimum value of the absolute time interval for users, and $r_{ij} \in ℕ$ . We assume that when the time between two interactions of a user is too long, the effect on predicting the next item of the user is minimal, so we assign a maximum time interval t_max. When the relative time interval is too large, it is limited to t_max. The absolute time relationship between all interactions after the restriction forms the absolute time relationship matrix of a user $R \in ℕ^{n \times n}$ :

$R = [\begin{matrix} r_{11} & r_{12} & \dots & r_{1 n} \\ r_{21} & r_{22} & \dots & r_{2 n} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ r_{n 1} & r_{n 2} & \dots & r_{nn} \end{matrix}]$ (8)

Where the main diagonal elements are all 0, similar to the relative position embedding, we use $R^{K} \in ℝ^{n \times n \times d}$ and $R^{V} \in ℝ^{n \times n \times d}$ to denote the absolute time relationship of key and value in the self-attention network layer respectively:

$R^{K} = [\begin{matrix} r_{11}^{K} & r_{12}^{K} & \dots & r_{1 n}^{K} \\ r_{21}^{K} & r_{22}^{K} & \dots & r_{2 n}^{K} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ r_{n 1}^{K} & r_{n 2}^{K} & \dots & r_{nn}^{K} \end{matrix}], R^{V} = [\begin{matrix} r_{11}^{V} & r_{12}^{V} & \dots & r_{1 n}^{V} \\ r_{21}^{V} & r_{22}^{V} & \dots & r_{2 n}^{V} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ r_{n 1}^{V} & r_{n 2}^{V} & \dots & r_{nn}^{V} \end{matrix}]$ (9)

Where $r_{ij}^{K} \in ℝ^{d}$ , $r_{ij}^{V} \in ℝ^{d}$ .

3.3 Self-Attention Network Layer

Since the self-attention mechanism can capture sequential relationships in a sequence, we use it to model the historical interaction behavior of users. In calculating the scaled dot product attention [32] process, inspired by [11], we consider both the relative position in the input sequence and the absolute temporal relationship between each item in the sequence.

Self-Attention Layer: For the given input sequence E^I, we use self-attention weights to compute a new sequence S = (S₁, S₂, ⋯ , S_n), where $S_{i} \in ℝ^{d}$ is the weighted sum of the linearly transformed input elements with the relative position embedding and the absolute time-relative embedding:

$S_{i} = \sum_{j = 1}^{n} m_{ij} ([I_{l_{j}}; a_{I_{l_{j}}}^{g}; a_{I_{l_{j}}}^{b}] W^{V} + r_{ij}^{V} + p_{Vj})$ (10)

Where $W^{V} \in ℝ^{d \times d}$ is the projection matrix of values, m_ij is the self-attention weight, defined as: $m_{ij} = \frac{\exp (e_{ij})}{\sum_{k = 1}^{n} \exp (e_{ik})}$ (11)

$e_{ij} = \frac{[I_{l_{i}}; a_{I_{l_{i}}}^{g}; a_{I_{l_{i}}}^{b}] W^{Q} ([I_{l_{j}}; a_{I_{l_{j}}}^{g}; a_{I_{l_{j}}}^{b}] W^{K} + r_{ij}^{K} + p_{Kj})}{\sqrt{d}}$ (12)

Where $W^{Q} \in ℝ^{d \times d}$ and $W^{K} \in ℝ^{d \times d}$ are the projection matrices of the query and key respectively, the scale factor d is used to scale the dot product attention to avoid too large of the inner product values.

To avoid information leakage, i.e., the output at time step t contains the effect of subsequent items on predicting the next item, we use masks, as in [5, 11], to disable the linking of subsequent keys to the current query at time step t.

Point-Wise Feed-Forward Network: As self-attention is a linear model, to make the model nonlinear, we apply the same two-layer point-wise feed-forward network after the self-attention layer, with a ReLU activation function in between:

$F_{i} = FFN (S_{i}) = ReLU (S_{i} W_{1} + b_{1}) W_{2} + b_{2}$ (13)

Where $W_{1} \in ℝ^{d \times d}$ and $W_{2} \in ℝ^{d \times d}$ are weight matrices, $b_{1} \in ℝ^{d}$ and $b_{2} \in ℝ^{d}$ are bias vectors.

Stacking Layers: Inspired by [5], we stack b self-attention layers and feed-forward network layers to learn more complex item transformations. In order to avoid overfitting and gradient disappearance caused by stacking, we perform layer normalization, residual concatenation, and Dropout operations for each layer as in [5]:

$g (x) = x + Dropout (g (LayerNorm (x)))$ (14)

Where g (x) indicates self-attention layer or feed-forward network layer. That is to say, layer g first normalizes the input x using layer normalization before processing it, then applies dropout to its output, and finally adds the initial input x to the processed output.According to [37], layer normalization is defined as:

$LayerNorm (x) = α ⊙ \frac{x - μ}{σ^{2} + ϵ} + β$ (15)

Where ⊙ is point-wise product, μ and σ are the mean and variance of x, α and β are the learned scale factor and deviation term respectively.

3.4 Prediction Layer

After b self-attention layers and feed-forward network layers, we obtain a multi-aspect preferences representation of users incorporating item’s attributes and absolute temporal relationships and calculate the user’s prediction score for the item i at time step t:

$s_{i, t} = F_{t} {[I_{l_{i}}; a_{I_{l_{i}}}^{g}; a_{I_{l_{i}}}^{b}]}^{T}$ (16)

Where s_i,t is the probability that the following item is i after giving the first t items (l₁, l₂, ⋯ , l_t), $[I_{l_{i}}; a_{I_{l_{i}}}^{g}; a_{I_{l_{i}}}^{b}]$ is a combined embedded representation of the item. We generate a list of the highest scoring K (e.g., 10) items to recommend to the user.

3.5 Model Training

For the user interaction sequence l = (l₁, l₂, ⋯ , l_n) after a fixed length, our desired output sequence is O = (o₁, o₂, ⋯ , o_n), where the element o_t at time step t is defined as:

$o_{t} = {\begin{matrix} < pad > & if l_{t} is padding item \\ l_{t + 1} & 1 \leq t < n \\ L_{| L^{u} |}^{u} & t = n \end{matrix}$ (17)

Where <pad> indicates the filler item. In the training process, we negatively sample the corresponding desired output sequence O = (o₁, o₂, ⋯ , o_n) to obtain the negative sample sequence N = (n₁, n₂, ⋯ , n_n), where n_i ∉ L^u, and using binary cross-entropy as the objective function:

$- \sum_{L^{u} \in L} \sum_{t \in [1, 2, \dots, n]} [log (σ (s_{o_{t}, t})) + log (1 - σ (s_{n_{t}, t}))]$ (18)

Where σ (·) is the sigmoid function. We optimize the model using the ADAM optimizer [38], a stochastic gradient descent (SGD) variant with adaptive moment estimation.

4 Experiments

This section evaluates our model on a large dataset of different domains. Our experiments aim to answer the following research questions:

RQ1: Can our model outperform other baselines on dense and sparse datasets?

RQ2: Can our model capture user’s preferences for specific item aspects? How would the different features of items affect the model’s performance?

RQ3: How would parameters and modules affect the model’s performance, such as dimension of item, the maximum length of sequences, maximum time interval, the dimension of each attribute, and absolute time relationships?

4.1 Experimental Settings

Datasets. We conducted experiments on five datasets from three real application scenarios. These datasets all include items’ features and timestamps of interactions and differ significantly in sparsity as follows:

Amazon: This is a dataset of user purchases and ratings collected by McAuley et al. [39]. from the e-commerce platform Amazon. They divide the dataset into different sub-datasets based on the higher-ranked categories on Amazon. We used "Beauty" and "Video Games". These two datasets are sparse.

MovieLens: They come from the GroupLens Research group at the University of Minnesota and are divided into four different versions depending on the number of users. We used two versions: "ML-100k" and "ML-1M". These two datasets are dense.

Steam: This dataset is crawled from Steam, a prominent online video game distribution platform. The dataset is very sparse.

We preprocessed these data in advance. As in [5, 11, 13], for all datasets, we treated reviews or ratings as implicit feedback (i.e., users interact with the items) and sorted the interactions using timestamps. For the Amazon and Steam datasets with high sparsity, we discard the data of users with less than five interactions and the corresponding items’ features. Table 2 provides detailed statistics for the datasets, where MovieLens are the densest, and Amazon and Steam have a lower average number of user interactions and a higher sparsity. On all datasets, we used the last item in each user interaction sequence as the test set, the previous item of the test item as the validation set, and used all remaining items as the training set.

Table 2
Dataset statistics

Dataset #users #items #avg sequence len #sparsity

Beauty 22363 12101 6.88 99.94%

Games 24303 10672 7.54 99.92%

MovieLens-1M 6040 3706 163.6 95.58%

MovieLens-100k 943 1682 104.04 93.81%

Steam 144051 11153 3.49 99.97%

Dataset	#users	#items	#avg sequence len	#sparsity
Beauty	22363	12101	6.88	99.94%
Games	24303	10672	7.54	99.92%
MovieLens-1M	6040	3706	163.6	95.58%
MovieLens-100k	943	1682	104.04	93.81%
Steam	144051	11153	3.49	99.97%

Evaluation Metrics. To evaluate the performance of all models, we used Hit Ratio (HR@N) and Normalized Discounted Cumulative Gain (NDCG@N) [16] as evaluation metrics, defined as:

$HR @ N = \frac{1}{M} \sum_{i = 1}^{M} hits (i)$ (19)

$NDCG @ N = \frac{1}{M} \sum_{i = 1}^{M} \frac{1}{{log}_{2} (p_{i} + 1)}$ (20)

Where M is the number of users, hits (i) is whether the item that the i-th user interacts with is in the recommendation list of length N, if in hits (i) =1, otherwise, hits (i) =0. In addition, p_i indicates the position of the i-th user-interactive item in the recommendation list. If the item does not exist in the list, it is denoted as p_i→ ∞. We set N to 10 in the experiment. For each user, we added 100 randomly selected negative samples after the positive samples and calculated HR@10 and NDCG@10 based on the ranking of these 101 items [40]; the higher these two metrics are (ideally as close to 1 as possible), the better the performance.

Baselines. To validate the effectiveness of our approach, we compared MFITSRec with the following classical methods and state-of-the-art models:

BPR [41]: It optimizes the implicit feedback matrix decomposition using a Bayesian personalized ranking loss function.

FMC/FPMC [7]: FPMC combines matrix decomposition with first-order Markov Chain to capture the general preferences of users and the Sequentiality of interaction behaviors; FMC is a simplified version of FPMC and does not include user personalization actions.

TransRec [42]: It is a first-order sequential recommendation method where items are embedded in the transition space. The user is modeled as a translation vector to obtain the transition from the current item to the next item.

Caser [4]: It is based on CNN and models higher-order MCs in horizontal and vertical directions.

SASRec [5]: It uses the self-attention mechanism to capture the user interaction sequence, considering sequential relative position relations.

SSE-PT [13]: It incorporates the user ID embedding vector and utilizes a new regularization method.

TiSASRec [11]: It models the specific timestamps of interaction behaviors and captures the dynamic preferences of users using a self-attention mechanism based on time intervals.

TARN [43]: It jointly models users’ static and dynamic preferences through feature interaction network and convolutional neural network.

BAR [44]: It adds behavior information to the representation module and applies the new modules to various backbone models.

FMLP-Rec [26]: It is a model based on the MLP architecture. It encodes user sequences by learnable filters, achieving state-of-the-art performance on more sparse datasets.

TLSRec [45]: It integrates a hierarchical modeling of user preference and a time lag sensitive fine-grained fusion of the long-term and short-term preferences.

Implementation Details. In our experiments, we use BAR with SASRec as the backbone model. We considered the latent dimension of item d_i from among {10, 20, 30, 40, 50}. We considered the regularization parameters between {0.0001, 0.001, 0.01, 0.1, 1}, and the learning rates are chosen from {0.1, 0.01, 0.001, 0.0001}. For the other hyperparameters and initialization strategies, we followed the authors’ recommendations of the corresponding methods and adjusted them during the experiments based on the validation. The training is terminated if the validation performance does not improve within 20 epochs.

Fig. 2

Visualization of the attention value assigned by the model in the Games dataset to the various features of items in a user’s historical sequence. The last row is the recommended item, and the darker the color, the higher the attention weight.

We implemented MFITSRec 1 using PyTorch and optimized the model using the Adam optimizer, with the learning rate set to 0.001 and the batch size set to 128. We used two attention network layers and point-by-point feedforward network layers. For the MovieLens, we set the maximum sequence length to 200; for the Amazon, the maximum sequence length was 50; and for the Steam, the maximum sequence length was 10. For all datasets, we set the latent dimensions of category and brand to 50, and the impact of each attribute was discussed in Section 4.3. The maximum time interval was set to 512 for the Movielens and 256 for the other datasets. The dropout rate was 0.5 for all datasets.

4.2 Performance Comparison (RQ1)

Table 3 and Table 4 show the recommended performance of all baseline methods and our model on the five benchmark datasets. We used bold to indicate the best results for each column, and underline to indicate the second best results. Our proposed MFITSRec outperforms all previous methods on all datasets. In the baseline approach, Caser outperforms TransRec on denser datasets, and the opposite is true for sparse datasets. This is because the neural network-based approach captures long sequences using higher-order transitions, while the MC-based approach focuses on the dynamic transfer of items. However, the orders in high-order models tend to be set very small and cannot be dynamically adjusted, so the performance of Caser is worse than that of the self-attention-based SASRec. In dense datasets, TiSASRec performs better than FMLP-Rec, and FMLP-Rec performs better in sparse datasets because FMLP-Rec replaces the Transformer with an MLP layer in the frequency domain, which solves overfitting problems caused by too few parameters of the model on sparse data. TARN combines users’ dynamic and static preferences and outperforms SASRec, which considers only one of them on all datasets. BAR outperforms its backbone model SASRec on all datasets, indicating that splitting a user’s historical interaction sequence into an item sequence and an action sequence is an effective modeling approach. TLSRec performs best on the four datasets but slightly worse than TiSASRec on Steam. Despite this, TLSRec still achieves the best level on Steam. Presumably, this is because although TLSRec can learn the user’s long-term preferences by capturing the dependencies between sessions, it cannot divide enough effective sessions for an overly sparse dataset like Steam.Compared with TiSASRec, we improved NDCG@10 by more than 2% and more than 8% on the sparser datasets. Compared with TLSRec, we improved NDCG@10 by more than 2%.

Table 3
Performance Comparison for sparse datasets

Methods Beauty Games Steam

HR@10 NDCG@10 HR@10 NDCG@10 HR@10 NDCG@10

BPR 0.3775 0.2183 0.4853 0.2875 0.7061 0.4436

FMC 0.3771 0.2477 0.6358 0.4456 0.7731 0.5193

FPMC 0.4310 0.2891 0.6082 0.4680 0.7710 0.5011

TransRec 0.4607 0.3020 0.6838 0.4557 0.7624 0.4852

Caser 0.4264 0.2547 0.5282 0.3214 0.7874 0.5381

SASRec 0.4663 0.3080 0.6843 0.4602 0.7867 0.5108

SSE-PT 0.4963 0.3159 0.6955 0.4677 0.7885 0.5369

TiSASRec 0.4981 0.3329 0.7080 0.4670 0.8053 0.5523

TARN 0.4979 0.3324 0.6996 0.4698 0.7985 0.5476

BAR 0.4995 0.3336 0.7073 0.4704 0.8039 0.5492

FMLP-Rec 0.5029 0.3351 0.7091 0.4773 0.8031 0.5470

TLSRec 0.5138 0.3465 0.7168 0.4816 0.8048 0.5521

MFITSRec 0.5293 0.3630 0.7332 0.5030 0.8396 0.5685

Methods	Beauty	Games	Steam
BPR	0.3775	0.2183	0.4853	0.2875	0.7061	0.4436
FMC	0.3771	0.2477	0.6358	0.4456	0.7731	0.5193
FPMC	0.4310	0.2891	0.6082	0.4680	0.7710	0.5011
TransRec	0.4607	0.3020	0.6838	0.4557	0.7624	0.4852
Caser	0.4264	0.2547	0.5282	0.3214	0.7874	0.5381
SASRec	0.4663	0.3080	0.6843	0.4602	0.7867	0.5108
SSE-PT	0.4963	0.3159	0.6955	0.4677	0.7885	0.5369
TiSASRec	0.4981	0.3329	0.7080	0.4670	0.8053	0.5523
TARN	0.4979	0.3324	0.6996	0.4698	0.7985	0.5476
BAR	0.4995	0.3336	0.7073	0.4704	0.8039	0.5492
FMLP-Rec	0.5029	0.3351	0.7091	0.4773	0.8031	0.5470
TLSRec	0.5138	0.3465	0.7168	0.4816	0.8048	0.5521
MFITSRec	0.5293	0.3630	0.7332	0.5030	0.8396	0.5685

Table 4

Performance Comparison for dense datasets

Methods	ML-1M		ML-100k
	HR@10	NDCG@10	HR@10	NDCG@10
BPR	0.5781	0.3287	0.4675	0.2467
FMC	0.6983	0.4676	0.4763	0.2498
FPMC	0.7599	0.5176	0.4942	0.2531
TransRec	0.6413	0.3969	0.4932	0.2390
Caser	0.7886	0.5538	0.6726	0.3912
SASRec	0.8285	0.5982	0.7338	0.4464
SSE-PT	0.8346	0.6163	0.7381	0.4611
TiSASRec	0.8359	0.6156	0.7593	0.4828
TARN	0.8325	0.6139	0.7496	0.4728
BAR	0.8351	0.6192	0.7524	0.4811
FMLP-Rec	0.8291	0.5333	0.7306	0.4109
TLSRec	0.8365	0.6202	0.7632	0.4866
MFITSRec	0.8369	0.6233	0.7735	0.4941

In real life, the length of the recommendation list of different platforms is different. Therefore, we have modified K (i.e., the number of items recommended to users) for a more thorough comparison. Table 5 shows the recommendation performance of some state-of-the-art methods and our model when K = {5, 20}. Similar to NDCG@10 and Hit@10, MFITSRec performs best on all datasets. On average, MFITSRec improves TiSASRec by 10%, 5%, 7%, and 4% in terms of NDCG@5, Hit@5, NDCG@20, and Hit@20 respectively. At the same time, MFITSRec improves FLSRec, the current best baseline, by 4%, 2%, 2%, and 1% in terms of NDCG@5, Hit@5, NDCG@20, and Hit@20 respectively. Presumably, this is because, with the addition of multi-aspect features of items to MFITSRec, on the one hand, we can improve the accuracy of recommendations by capturing users’ preferences for specific aspects of items and making personalized recommendations for each user. On the other hand, we can mitigate overfitting by enhancing the representation of users’ historical sequence, which allows MFITSRec to learn more useful information on sparse data. We will discuss the impact of each feature in Section 4.3. In addition, we retained the absolute temporal relationship between interactions and adjusted the weights using time intervals. From the experimental results, we argue that distinguishing the degree of users’ preferences for different aspects and considering the absolute time interval between interactions are essential for personalized recommendations.

Table 5

Performance comparison with different lengths of recommendation list

Datasets	Metric	SASRec	SSE-PT	TiSASRec	TARN	BAR	FMLP-Rec	TLSRec	MFITSRec
Beauty	HR@5	0.3697	0.3821	0.3931	0.3965	0.3991	0.4037	0.4061	0.4123
	NDCG@5	0.2265	0.2624	0.2716	0.2754	0.2865	0.3031	0.3158	0.3177
	HR@20	0.6154	0.6191	0.5865	0.5916	0.5987	0.6197	0.6339	0.6529
	NDCG@20	0.3386	0.3583	0.3296	0.3361	0.3466	0.3646	0.3735	0.3985
Games	HR@5	0.5212	0.5452	0.5461	0.5602	0.5601	0.5839	0.6086	0.6336
	NDCG@5	0.3462	0.3735	0.3782	0.3899	0.3962	0.4248	0.4602	0.4816
	HR@20	0.7379	0.7652	0.7789	0.7879	0.7906	0.8052	0.8318	0.8624
	NDCG@20	0.4742	0.4869	0.4922	0.4956	0.5001	0.5182	0.5256	0.5432
ML-1M	HR@5	0.6403	0.6965	0.7098	0.7106	0.7091	0.6765	0.7289	0.7320
	NDCG@5	0.4207	0.4966	0.5502	0.5518	0.5526	0.4724	0.5691	0.5759
	HR@20	0.8516	0.8836	0.8897	0.8921	0.8965	0.8702	0.9152	0.9213
	NDCG@20	0.5814	0.6042	0.6116	0.6118	0.6112	0.5381	0.6243	0.6264
ML-100k	HR@5	0.5529	0.5864	0.5936	0.5963	0.5997	0.5133	0.6026	0.6068
	NDCG@5	0.3821	0.4067	0.4129	0.4196	0.4183	0.3321	0.4299	0.4316
	HR@20	0.8399	0.8521	0.8670	0.8679	0.8705	0.8420	0.8729	0.8782
	NDCG@20	0.4532	0.4924	0.5057	0.5076	0.5059	0.4295	0.5102	0.5126
Steam	HR@5	0.6421	0.6543	0.6697	0.6609	0.6621	0.6642	0.6765	0.6834
	NDCG@5	0.4659	0.4825	0.5001	0.4926	0.4957	0.4996	0.5012	0.5076
	HR@20	0.8531	0.8764	0.8849	0.8921	0.8926	0.9031	0.9216	0.9347
	NDCG@20	0.5326	0.5547	0.5796	0.5603	0.5616	0.5702	0.5753	0.5821

4.3 Multi-Aspect Preferences (RQ2)

It is well known that users focused on different aspects of items differently when interacting with them (e.g., a user likes a certain brand of electronics while preferring a suspenseful type of movie). Our proposed MFITSRec can capture users’ preferences for specific aspects of items, allowing for better performance of the model. We visualize the attention values assigned by the model to various features of items in user’s historical sequence. Figure 2 shows the interaction sequence of a user in the Games dataset and the corresponding item multi-attribute list, where the color shades represent the magnitude of attention weights. We used the last item and its attributes that the user interacted with as the item to be predicted by the model (i.e., s^true = 1). During the test, we observed that the item obtained the highest prediction score (s^predict = 0.8627) among all positive and negative samples. It is clear that the user interacted most frequently with items of brand = 14 and categories = 183, which means that the user had a specific preference for brand = 14 or categories = 183. For MFITSRec, it gives a larger weight to these two attributes and considered that the user was more inclined to focus on items’ brands in Games.

Fig. 3

Impact of item’s latent dimension d_i on model’s performance (NDCG@10).

The reality is that not all items’ features strongly correlate with users’ interests (e.g., almost no users pay attention to where the item is produced), and as a result, they cannot improve recommendation performance significantly. Meanwhile, we will inevitably introduce many parameters into the model if we choose many features, making the time complexity too high and the experiments difficult to perform. We discuss the training efficiency of MFITSRec in Section 4.5. To avoid these issues, we conducted separate experiments on the features of each aspect of the items. We found that, in all datasets, users paid more attention to the category and brand of items compared to other features. Therefore, as introduced in Section 3.1, we discarded these features which are less attractive to users and considered only the category and brand of items. To verify the effect of various aspects of features on the model’s performance, we removed the brand attribute and the category attribute separately to obtain two new models. We compared these two models, TiSASRec and MFITSRec, on all datasets. Table 6 shows the results of the four models on the Games dataset. We found that both models with only category features and brand features outperformed TiSASRec, indicating that the inclusion of features of items enhances the model’s representation. The main reason limiting the performance of TiSASRec is that there are many identical timestamps (i.e., the same time interval) in the user sequence, making it impossible to model the absolute time relationships well [11]. Including features of items alleviate this problem from the information enhancement perspective. Among category and brand features, the model using only brand features achieved better results, meanings that users pay more attention to the items’ brand during the interaction. Adding more features was undoubtedly better for sparse datasets, and considering two features at the same time can lead to better performance. Our results reflected that capturing multi-aspect preferences of users using multi-aspect features of items is crucial on sparse datasets.

Table 6

The comparison of each attribute

Methods	Games
	HR@10	NDCG@10
MFITSRec_no_features	0.6870	0.4661
MFITSRec_no_category	0.6965	0.4675
MFITSRec_no_brand	0.7067	0.4782
MFITSRec	0.7332	0.5030

4.4 Ablation Study (RQ3)

Fig. 4

Impact of maximum sequence length n on model’s performance (NDCG@10).

Influence of item’s latent dimension d_i: Figure 3 shows the variation of NDCG@10 for some neural network sequential methods when item’s latent dimension d_i is changed from 10 to 50, with other optimal hyperparameters constant. In most cases, the model performed better as the dimension increased; however, sparse datasets like Steam, SASRec, and SSE-PT performed worse when d_i got larger because the lack of interaction information led to overfitting. For MFITSRec, we obtained satisfactory results at d_i = 50, so we chose 50 as the value of d_i in the subsequent experiments.

Influence of maximum sequence length n: For all datasets with different average sequence lengths, the choice of the maximum sequence length significantly affected the recommendation performance. A larger sequence length generally means the model learned more historical interaction data and more accurate recommendation results. However, the model’s performance decreased when the maximum sequence length chosen was much longer than the average length. Figure 4 shows the variation of NDCG@10 when we choose different maximum sequence lengths for the methods described above. For Amazon datasets, each model presented superior performance when n = 30, and for MovieLens datasets, the models achieved satisfactory performance at n = 200. The best results for the Steam dataset with the smallest average sequence length are obtained on n = 10, except for SASRec and SSE-PT. When n is large, more noise will be introduced in the model.

Influence of the number of stacked self-attention network layers b: In general, more stacked self-attention layers mean better model-fitting ability. However, when the number of layers exceeds a certain threshold, the model will be overfitted, leading to its performance decrease and an increase in time complexity. Figure 5 shows the effect of the model on NDCG@10 using different numbers of attention modules on Beauty and ML-1M. When b = 2, the performance of models reached its optimum, and when b > 2, the performance of models started to decline and converge as the number increased because too much stacking led to overfitting.

Fig. 5

Impact of the number of stacked self-attention network layers b on model’s performance.

Influence of maximum time interval t_max: When the time interval between two interactions exceeds a certain threshold, there is almost no association between them. As the threshold increases, the ability of the model to capture the temporal relationship between the two items increases. However, when the value of t_max is too large, some unrelated items may start to impact the model’s accuracy. Figure 6 shows the NDCG@10 of MFITSRec and TiSASRec on Beauty and ML-1M using different maximum time intervals. For Beauty, the performance of MFITSRec was optimized when t_max = 256, and TiSASRec was optimized when t_max = 512. For ML-1M, both models achieved satisfactory performance at t_max = 512. When t_max exceeded the optimal value, the performance of both models gradually decreased as the interval increased.

Fig. 6

Impact of maximum time interval t_max on model’s performance (NDCG@10).

Influence of latent dimensions of features d_g, d_b: The latent dimension of each feature has an impact on the model’s performance. Table 7 shows the impact of latent dimensions of category and brand on our model’s performance at Beauty and Games. We chose the dimension of each feature from {25, 50, 100} and combined them with one another. We found that MFITSRec performed better when both d_g and d_b were 50. However, too large latent dimensions of features on two datasets are not friendly to the model and are prone to overfitting. For Beauty, the results are suboptimal when d_g = d_b = 25; for Games, the smaller features’ latent dimensions similarly limited the model’s performance.

Table 7

Influence of latent dimensions of features d_g, d_b on model’s performance

Datasets	d of attributes		HR@10	NDCG@10
	d _g	d _b
Beauty	25	25	0.5250	0.3587
	25	50	0.5198	0.3498
	50	25	0.5140	0.3479
	50	50	0.5293	0.3630
	100	50	0.5115	0.3456
	50	100	0.5143	0.3510
	100	100	0.5216	0.3520
Games	25	25	0.7009	0.4705
	25	50	0.7123	0.4871
	50	25	0.7133	0.4862
	50	50	0.7332	0.5030
	100	50	0.7201	0.4937
	50	100	0.7058	0.4806
	100	100	0.7102	0.4855

Influence of absolute temporal relationships and relative position: The method that combines both absolute time relationship and relative position performed better than the method that uses only one of them. We removed the absolute time relationship and the relative position from MFITSRec separately to obtain two new models. Table 8 shows the comparison results of TiSASRec, M_no_ti, M_no_pos, and MFITSRec on the five datasets. M_no_ti and M_no_pos significantly outperformed TiSASRec on the Amazon and Steam datasets. However, they underperformed on the MovieLens dataset, implying that it is more helpful to focus on other useful information about the item than on the interaction when there is insufficient interaction data. In addition, M_no_pos outperformed M_no_ti on all datasets, indicating that absolute temporal relationships can more accurately capture the associations between items.

Table 8

Influence of absolute time relationship and relative position on model’s performance

Datasets	Metric	TiSASRec	M_no_ti	M_no_pos	MFITSRec
Beauty	HR@10	0.4981	0.5040	0.5190	0.5293
	NDCG@10	0.3329	0.3323	0.3472	0.3630
Games	HR@10	0.7080	0.7147	0.7180	0.7332
	NDCG@10	0.4670	0.4909	0.4969	0.5030
ML-1M	HR@10	0.8359	0.8311	0.8279	0.8369
	NDCG@10	0.6156	0.6042	0.6057	0.6233
ML-100k	HR@10	0.7593	0.7508	0.7664	0.7735
	NDCG@10	0.4898	0.4656	0.4881	0.4941
Steam	HR@10	0.8053	0.8089	0.8338	0.8396
	NDCG@10	0.5523	0.5535	0.5548	0.5552

4.5 Training Efficiency

Figure 7 compares the training efficiency of MFITSRec with different numbers of items’ features on the Beauty dataset. We consider the number of features k used by the model from {1, 2, 3, 4, 5}. Obviously, the increase in the number of features will lead to an exponential increase in the number of parameters of the model, so the training speed of the model will gradually decrease. When k ≤ 2, the performance of MFITSRec improves with the increase in the number of features. When k = 2 and k = 3, the performance of MFITSRec is comparable, but there is a significant decrease in training speed when k = 3. When k > 3, the performance of MFITSRec decreases slightly while the training time increases, which may be due to overfitting caused by too many parameters. In general, training efficiency greatly affects the time cost of recommender model implementation. Therefore, we comprehensively consider the performance and time cost of the model and input two features into MFITSRec.

Figure 8 compares the training efficiency of SASRec, TiSASRec, and MFITSRec on the Beauty dataset, where k = 2. Since the parameters of MFITSRec and TiSASRec are larger than SASRec, the training speed is slightly longer than that of SASRec, but it is still on a comparable level. Obviously, our model has the best performance within the same amount of time.

Fig. 7

Training efficiency of MFITSRec with different numbers of item’s features on the Beauty dataset.

Fig. 8

Training efficiency of SASRec, TiSASRec, and MFITSRec on the Beauty dataset.

5 Conclusion

In this work, we proposed a new sequence-aware model (MFITSRec), which combined multi-aspect features of items with user behavior sequences to capture users’ preferences for a particular aspect of items and augment the input interaction sequence. This special design is due to the observation that users may place different levels of importance on different aspects of the same item. Our experiments on sparse and dense datasets showed that MFITSRec outperformed other competitive baselines. In addition, we investigated the relationship between each feature and the absolute time of items. We demonstrated that embedding multi-aspect features of items could effectively capture users’ preferences for various aspects of items and improve the data sparsity problem.

Footnotes

Acknowledgments

The work is supported by the Science and Technology Research Program of Chongqing Municipal Education Commission (No.KJZD-K202101105, KJQN202001136), Humanities and Social Sciences Research Program of Chongqing Municipal Education Commission (No.22SKGH302), the National Natural Science Foundation of China (No.61702063), the Action Plan for High-Quality Development of Graduate Education of Chongqing University of Technology (NO.gzlcx20232104).

References

Wang

, Cao

, Wang

, Sheng

Q.Z.

, Orgun

M.A.

and Lian

, A survey on session-based recommender systems, ACM Computing Surveys (CSUR) 54(7) (2021), 1–38.

Chen

, Xu

, Zhang

, Tang

, Cao

, Qin

and Zha

, Sequential recommendation with user memory networks, in Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining 2018, pp. 108–116.

Shani

, Heckerman

, Brafman

R.I.

and Boutilier

, An mdp-based recommender system, Journal of Machine Learning Research 6(9) (2005).

Tang

and Wang

, Personalized top-n sequential recommendation via convolutional sequence embedding, in Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining 2018, pp. 565–573.

Kang

W.-C.

and McAuley

, Self-attentive sequential recommendation, in 2018 IEEE International Conference on Data Mining (ICDM) IEEE 2018, pp. 197–206.

Hosseinzadeh

, Aghdam, N. Hariri, B. Mobasher and R. Burke, Adapting recommendations to contextual changes using hierarchical hidden markov models, in Proceedings of the 9th ACM Conference on Recommender Systems 2015, pp. 241–244.

Rendle

, Freudenthaler

and Schmidt-Thieme

, Factorizing personalized markov chains for next-basket recommendation, in Proceedings of the 19th International Conference on World Wide Web 2010, pp. 811–820.

Hidasi

and Karatzoglou

, Recurrent neural networks with top-k gains for session-based recommendations, in Proceedings of the 27th ACM International Conference on Information and Knowledge Management 2018, pp. 843–852.

Yuan

, He

, Jiang

, Guo

, Xiong

, Xu

and Xiong

, Future data helps training: Modeling future contexts for session-based recommendation, in Proceedings of The Web Conference 2020 2020, pp. 303–313.

10.

, Tang

, Zhu

, Wang

, Xie

and Tan

, Session-based recommendation with graph neural networks, vol. 33, no. 01, in Proceedings of the AAAI Conference on Artificial Intelligence 2019, pp. 346–353.

11.

, Wang

and McAuley

, Time interval aware self-attention for sequential recommendation, in Proceedings of the 13th International Conference on Web Search and Data Mining 2020, pp. 322–330.

12.

Zhang

, Tay

, Yao

, Sun

and An

, Next item recommendation with self-attentive metric learning, in Thirty-Third AAAI Conference on Artificial Intelligence, vol. 9, 2019.

13.

, Li

, Hsieh

C.-J.

and Sharpnack

, Sse-pt: Sequential recommendation via personalized transformer, in Fourteenth ACM Conference on Recommender Systems 2020, pp. 328–337.

14.

Linden

, Smith

, York

and com

Amazon.

, recommendations: Item-to-item collaborative filtering, IEEE Internet Computing 7(1) (2003), 76–80.

15.

Koren

, Rendle

and Bell

, Advances in collaborative filtering, Recommender Systems Handbook (2022), 91–142.

16.

, Liao

, Zhang

, Nie

, Hu

and Chua

T.-S.

, Neural collaborative filtering, in Proceedings of the 26th International Conference on World Wide Web 2017, pp. 173–182.

17.

Chen

, Yu

, Lu

, Qian

and Li

, Ir-rec: An interpretive rules-guided recommendation over knowledge graph, Information Sciences 563 (2021), 326–341.

18.

, Koren

and Volinsky

, Collaborative filtering for implicit feedback datasets, in 2008 Eighth IEEE International Conference on Data Mining IEEE 2008, pp. 263–272.

19.

Wang

, Hu

, Wang

, Cao

, Sheng

Q.Z.

and Orgun

, Sequential recommender systems: Challenges, progress and prospects, arXiv preprint arXiv:2001.04830, 2019.

20.

Tuan

T.X.

and Phuong

T.M.

, 3d convolutional networks for session-based recommendation with content features, in Proceedings of the Eleventh ACM Conference on Recommender Systems 2017, pp. 138–146.

21.

Hidasi

, Karatzoglou

, Baltrunas

and Tikk

, Session-based recommendations with recurrent neural networks, arXiv preprint arXiv:1511.06939, 2015.

22.

Bogina

and Kuflik

, Incorporating dwell time in session-based recommendations with recurrent neural networks, in RecTemp@ RecSys 2017, pp. 57–59.

23.

Quadrana

, Karatzoglou

, Hidasi

and Cremonesi

, Personalizing session-based recommendations with hierarchical recurrent neural networks, in Proceedings of the Eleventh ACM Conference on Recommender Systems 2017, pp. 130–137.

24.

Wang

, Wei

, Cong

, Li

X.-L.

, Mao

X.-L.

and Qiu

, Global context enhanced graph neural networks for session-based recommendation, in Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval 2020, pp. 169–178.

25.

Song

, Xiao

, Wang

, Charlin

, Zhang

and Tang

, Session-based social recommendation via dynamic graph attention networks, in Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining 2019, pp. 555–563.

26.

Zhou

, Yu

, Zhao

W.X.

and Wen

J.-R.

, Filter-enhanced mlp is all you need for sequential recommendation, in Proceedings of the ACM Web Conference 2022 2022, pp. 2388–2399.

27.

, Ren

, Chen

, Ren

, Lian

and Ma

, Neural attentive session-based recommendation, in Proceedings of the 2017 ACM on Conference on Information and Knowledge Management 2017, pp. 1419–1428.

28.

Liu

, Zeng

, Mokhosi

and Zhang

, Stamp: Shortterm attention/memory priority model for session-based recommendation, in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining 2018, pp. 1831–1839.

29.

Santana

M.R.

and Soares

, Hybrid model with time modeling for sequential recommender systems, arXiv preprint arXiv:2103.06138, 2021.

30.

Wang

, Hu

, Cao

, Huang

, Lian

and Liu

, Attention-based transactional context embedding for next-item recommendation, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018.

31.

Ren

, Chen

, Li

, Ren

, Ma

and De

, Rijke, Repeatnet: A repeat aware neural recommendation machine for session-based recommendation, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 4806–4813.

32.

Vaswani

, Shazeer

, Parmar

, Uszkoreit

, Jones

and N.

, Gomez, L. Kaiser and I. Polosukhin, Attention is all you need, Advances in Neural Information Processing Systems 30 (2017).

33.

Cho

S.M.

, Park

and Yoo

, Meantime: Mixture of attention mechanisms with multi-temporal embeddings for sequential recommendation, in Fourteenth ACM Conference on Recommender Systems 2020, pp. 515–520.

34.

Fan

, Liu

, Zhang

, Xiong

, Zheng

and Yu

P.S.

, Continuous-time sequential recommendation with temporal graph collaborative transformer, in Proceedings of the 30th ACM International Conference on Information & Knowledge Management 2021, pp. 433–442.

35.

Stai

, Kafetzoglou

, Tsiropoulou

E.E.

and Papavassiliou

, A holistic approach for personalization, relevance feedback & recommendation in enriched multimedia content, Multimedia Tools and Applications 77 (2018), 283–326.

36.

Singer

, Roitman

, Eshel

, Nus

, Guy

, Levi

, Hasson

and Kiperwasser

, Sequential modeling with multiple attributes for watchlist recommendation in e-commerce, in Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining 2022, pp. 937–946.

37.

J.L.

, Kiros

J.R.

and Hinton

G.E.

, Layer normalization, arXiv preprint arXiv:1607.06450, 2016.

38.

Kingma

D.P.

and Ba

, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980, 2014.

39.

McAuley

, Targett

, Shi

and Van Den Hengel

, Image-based recommendations on styles and substitutes, in Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval 2015, pp. 43–52.

40.

Koren

, Factorization meets the neighborhood: A multifaceted collaborative filtering model, in Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 2008, pp. 426–434.

41.

Rendle

, Freudenthaler

, Gantner

and Schmidt-Thieme

, Bpr: Bayesian personalized ranking from implicit feedback, arXiv preprint arXiv1205.2618, 2012.

42.

, Kang

W.-C.

and McAuley

, Translation-based recommendation, in Proceedings of the Eleventh ACM Conference on Recommender Systems 2017, pp. 161–169.

43.

Zhang

, Cao

, Shi

and Niu

, Neural time-aware sequential recommendation by jointly modeling preference dynamics and explicit feature couplings, IEEE Transactions on Neural Networks and Learning Systems 33(10) (2021), 5125–5137.

44.

, Pan

and Ming

, Bar: Behavior-aware recommendation for sequential heterogeneous one-class collaborative filtering, Information Sciences 608 (2022), 881–899.

45.

Chen

, Yang

and Yu

P.S.

, Time lag aware sequential recommendation, in Proceedings of the 31st ACM International Conference on Information & Knowledge Management 2022, pp. 212–221.

Multi-aspect features of items for time-ordered sequential recommendation

Abstract

Keywords

1 Introduction

2 Related Work

4.1 Experimental Settings

Table 2 Dataset statistics Dataset #users #items #avg sequence len #sparsity Beauty 22363 12101 6.88 99.94% Games 24303 10672 7.54 99.92% MovieLens-1M 6040 3706 163.6 95.58% MovieLens-100k 943 1682 104.04 93.81% Steam 144051 11153 3.49 99.97%

Footnotes

Acknowledgments

References

Table 2
Dataset statistics

Dataset #users #items #avg sequence len #sparsity

Beauty 22363 12101 6.88 99.94%

Games 24303 10672 7.54 99.92%

MovieLens-1M 6040 3706 163.6 95.58%

MovieLens-100k 943 1682 104.04 93.81%

Steam 144051 11153 3.49 99.97%