TME-PSR: Time-aware, Multi-interest, and Explanation Personalization for Sequential Recommendation
arXiv:2604.09439v1 [cs.IR] 10 Apr 2026
Qingzhuo Wang1 , Leilei Wen1 , Juntao Chen1 , Kunyu Peng1 , Ruiyang Qin1 , Zhihua Wei1 , Wen Shen1 1 Tongji University [email protected], [email protected], {2432047, 2534000}@tongji.edu.cn, [email protected], {zhihua wei, wenshen}@tongji.edu.cn, Abstract In this paper, we propose a sequential recommendation model that integrates Time-aware personalization, Multi-interest personalization, and Explanation personalization for Personalized Sequential Recommendation (TME-PSR). That is, we consider the differences across different users in temporal rhythm preference, multiple fine-grained latent interests, and the personalized semantic alignment between recommendations and explanations. Specifically, the proposed TME-PSR model employs a dual-view gated time encoder to capture personalized temporal rhythms, a lightweight multihead Linear Recurrent Unit architecture that enables fine-grained sub-interest modeling with improved efficiency, and a dynamic dual-branch mutual information weighting mechanism to achieve personalized alignment between recommendations and explanations. Extensive experiments on real-world datasets demonstrate that our method consistently improves recommendation accuracy and explanation quality, at a lower computational cost.
1
Introduction
In recent years, personalized recommendation has attracted increasing attention as user expectations rise in recommendation systems. Personalized recommendation aims to select items (information or products) that best suit individual users’ preferences based on their historical behaviors. Although recent advances in sequential recommendation (SR) models have significantly improved recommendation performance, most methods still fall short in delivering comprehensive personalization. In this paper, we aim to model three types of personalization simultaneously, including the following three perspectives. • Temporal rhythm preference. Given a set of historical interactions of a specific user, the interaction time can reflect the personalized temporal rhythm preference of the user. For example, as shown in fig. 1, User 1 watches two romantic dramas on one day, and then, after a long gap of nearly a month, suddenly watches two action movies, which suggests a clear shift in temporal rhythm. In contrast, User 2
Long Time ago
Recently
(a)
User 1
(b)
User 2
Figure 1: Examples of movie interaction sequences of two different users, which show different preferences for time and interests. User 1 liked romance movies a long time ago, but recently the interest has shifted to superhero movies. User 2 has always liked superhero movies, but also follows both Marvel series and DC series two subinterests.
watches superhero movies regularly. To encode such longshort-term temporal rhythms across different users, some studies design specific network architectures [Yu et al., 2019; Zhang et al., 2021; Zheng et al., 2022; Dang et al., 2023; Li et al., 2020; Chen et al., 2022a; Liu et al., 2024b; Shi et al., 2023; Tran et al., 2023] to improve the recommendation performance, such as using two attention branches to encode long-term and short-term temporal rhythms, respectively. In comparison, we propose a flexible module to adaptively encode the differences in long-short-term temporal rhythms among different users, which can be plugged into existing models without changing the model’s original architecture. More crucially, the proposed module is not limited to recommendation tasks, it also encodes temporal rhythm preferences for the explanation task, separately for the same user. • Multiple fine-grained latent interests. A single sequence may contain multiple latent interests of a user, such as the interest in superhero movies and the interest in romance movies. Moreover, a user’s latent interest can be further divided into multiple fine-grained sub-interests, such as the sub-interest in DC series movies and the sub-interest in Marvel series movies, as fig. 1(b) shows. Existing studies introduce an additional attention layer or an additional capsule network [Li et al., 2019; Chen et al., 2021; Cen et al., 2020; Zhang et al., 2022; Chai et al., 2022; Xie et al., 2023; Li et al., 2024] to learn multiple interests for different users. However, the time and space complexity of these methods
increases with the number of multiple interests. In comparison, we propose a method to make the model automatically learn multiple fine-grained sub-interests with parallel optimized network branches, thereby explicitly disentangling them while improving efficiency and parameter economy. • Personalized semantic alignment between recommendations and explanations. As explainability gains importance in recommendation systems, more users desire a textual rationale explaining why the item is recommended to them. Many studies, such as collaborative explanation [Li et al., 2021; Li et al., 2023a; Yu et al., 2024], fuse recommendation features and explanation features to achieve semantic alignment between recommendations and explanations, thereby enabling explainable recommendation. However, these methods assume that the degree of semantic alignment between recommendations and explanations is uniform across users. But different users have different requirements for the alignment degree. For example, if a user often watches Marvel’s superhero movies in the past, giving an explanation such as “Iron Man 3, Robert Downey Jr. makes a powerful comeback” can effectively entice the user to choose this movie, this user has high requirements for the semantic alignment degree between recommendations and explanations. In contrast, for those users having a wide range of interests, this explanation might be less appealing, as they may not be familiar with “Iron Man” or “Robert Downey Jr.” In this case, generating an explanation like “An excellent action movie” might be more likely to entice them to watch. These users have low requirements for the semantic alignment degree. Therefore, it is necessary to consider differences in the degree of semantic alignment between recommendations and explanations among different users, which is termed explanation personalization. To this end, in this paper, we propose a sequential recommendation model to achieve the above three types of personalization. Specifically, as shown in fig. 2, the overall structure of our model is given as follows: • To realize the time-aware personalization, we propose a dual-view gated time encoder module. First, the module adaptively balances short-term and long-term behavior patterns based on each user’s temporal rhythm preference. Second, for the same user, the module independently considers the temporal rhythm preference of the recommendation task and the explanation task. The proposed module is plug-andplay, requiring no changes to existing backbones. • To realize the multi-interest personalization, we propose to divide the representations of intermediate layers into multiple subspaces and track independent sub-interests in each subspace. Specifically, we extend the Linear Recurrent Unit (LRU) [Orvieto et al., 2023] to a multihead LRU architecture, so as to disentangle multiple fine-grained latent sub-interests of a certain user while keeping lightweight. • To realize the explanation personalization, we design a dynamic dual-branch mutual information (MI) weighting mechanism, inspired by [Yu et al., 2024], so as to adaptively adjust the degree of semantic alignment between recommendations and explanations.
2
RELATED WORK
Time-Aware Sequential Recommendation. Time-aware methods studies incorporate temporal information to improve recommendation performance. Early works utilized time matrix factorization [Chen et al., 2020; Ding and Li, 2005; Xiong et al., 2010] or attention mechanisms [Li et al., 2020; Zhang et al., 2021; Chen et al., 2022a] to encode time intervals. Later, some methods [Dang et al., 2023; Tran et al., 2023; Liu et al., 2024b] have improved model performance by enhancing the uniformity of data through time intervals to obtain more regular historical sequences. Recent methods also study temporal phenomena such as repeat consumption [Dai et al., 2024], temporal graph contrastive learning [Zhang et al., 2024] or review-driven [Shi et al., 2023]. Above methods all propose a specific model structure, but many require architecture-coupled designs, which limits portability. Multi-Interest Sequential Recommendation. To model users’ multifaceted preferences, many methods learn multiple user vectors. Capsule/attention aggregators [Li et al., 2019; Cen et al., 2020] and their variants [Chen et al., 2021; Chai et al., 2022] extract several interests for matching/ranking. Recent methods [Xia et al., 2022; Zhang et al., 2022; Xie et al., 2023; Du et al., 2024; Li et al., 2024] have improved the model’s ability to decompose multiple interests and robustness. However, complexity often scales with interest count, leading to overlapping/redundancy of interests and higher computational costs. Explainable Sequential Recommendation. Explainable SR provides textual rationales. Approaches utilize attention weights [Chen et al., 2023; Yang et al., 2021; Gholami et al., 2022; Li et al., 2023d], knowledge graphs [Cui et al., 2022; Geng et al., 2022a; Huang et al., 2019; Xian et al., 2019], or collaborative generation [Li et al., 2021; Li et al., 2023a; Yu et al., 2024]. But existing collaborative explanation methods typically assume a fixed alignment degree between recommendations and explanations across users. Broader advances in general sequential recommendation are discussed in App. A.
3
METHODOLOGY
Problem Formulation. Let U denote the set of users, V denote the set of items, and E denote the set of explanations.1 For each user u ∈ U, let Su = [(v1 , e1 , t1 ), (v2 , e2 , t2 ), . . . , (vn , en , tn )] denote the interaction sequence of the user u, where at time step i ∈ {1, 2, ..., n}, the user action is a tuple of (vi , ei , ti ), vi ∈ V is the item interacted at time step i, ei ∈ E is the associated explanation, and ti ∈ R+ is the interaction timestamp. We propose a dual-branch model TME-PSR, as fig. 2 shows, one branch predicts the next item vi+1 that user u is likely to interact with, while the other generates the corresponding explanation ei+1 for this item. The proposed TME-PSR model achieves time-aware personalization, multi-interest personalization, and explanation personalization, as follows. 1
The explanations are extracted from user reviews and clustered into representative phrases through semantic hashing techniques [Li et al., 2021], where each item is assigned a unique explanation.
Inputs
Dual-view gated time encoder
Multihead LRU
Base Embedding
items …
Dynamic dual branch MI weighting
Romance
...
Superhero
…
explanations
REC Predictor
LRU Head
MLP
…
… MLP
…
Time-aware Encoder Short-term GRU
timestamps
Romance
...
Superhero
LRU Head
…
MLP …
Long-term GRU
Personalized Time-aware Embedding
EXP Predictor
Figure 2: Overall architecture of our TME-PSR. At the input layer, we propose a dual-view gated time encoder. Specifically, we use a timeaware encoder to extract short-term behavior patterns and long-term behavior patterns from input timestamps. Then, we adaptively weight the two patterns based on different user’s temporal rhythm preference, so as to extract personalized time-aware embeddings. At the encoding layer, we propose a multihead LRU architecture, each head captures a fine-grained sub-interest. At the prediction layer, we design a dynamic dual-branch MI weighting mechanism, so as to adaptively adjust the degree of semantic alignment between the recommendation branch and the explanation branch.
3.1
Time-Aware Personalization
To model user-specific temporal rhythm preference, we propose a dual-view gated time encoder to extract the personalized time-aware embedding from the timestamps of the user sequence Su . Specifically, we first encode two types of time intervals: the interval between adjacent items and the interval between each item and the first item in the sequence, thereby obtaining features containing short-term and long-term temporal information, respectively. Then, we employ a gating mechanism to adaptively fuse the short-term features and the long-term features, producing the personalized time-aware embedding of the user sequence Su . Base embedding construction. Let MV ∈ R|V|×d and ME ∈ R|E|×d denote the embedding matrices for all items V and all explanations E, respectively, where d is the embedding size. Given a user sequence Su with n interactions, we first apply a look-up operation from MV and ME to form the user’s item embedding matrix EV ∈ Rn×d and explanation embedding matrix EE ∈ Rn×d . We construct the recommendation base embeddings2 Erec ∈ Rn×d and the explanation base embeddings2 Eexp ∈ Rn×d as follows, which will serve as a basis for constructing personalized time-aware embedding: rec
E
= α · EV + (1 − α) · EE ,
E
exp
= α · EE + (1 − α) · EV , (1)
where α ∈ [0, 1] is a hyperparameter that controls the relative importance of each item embedding and its corresponding explanation embedding. Dual-view gated time encoder. Let T = [t1 , t2 , ..., tn ] ∈ Rn denote the timestamp sequence corresponding to the interaction sequence of the user u2 , we compute two types of time 2
rec
The embeddings (e.g. E ) and hidden states extracted in the METHODOLOGY section are all for a single user sequence Su . The subscript u is omitted in these symbols for simplicity.
interval to extract short-term and long-term temporal features as follows: (1) We compute the difference between adjacent timestamps ti and ti−1 to obtain the adjacent time interval ∆adj i = log(1 + ti − ti−1 ) ∈ R+ , for i = 2, 3, ..., n, where the log operation is to ensure numerical smoothness; the constant 1 is added to avoid taking the log of zero. (2) We compute the difference between each timestamp ti and the first timestamp t1 to obtain the absolute time interval + ∆abs i = log(1 + ti − t1 ) ∈ R , for i = 2, 3, ..., n. Therefore, we construct the adjacent interval vector ∆adj = adj n and the absolute interval vector [0, ∆adj 2 , ..., ∆n ] ∈ R abs abs abs ∆ = [0, ∆2 , ..., ∆n ] ∈ Rn , where the constant 0 is prepended to each vector to align with the length n of the user sequence Su to facilitate subsequent matrix operations. To extract short-term temporal features from ∆adj and extract long-term temporal features from ∆abs , we employ two shared GRU [Chung et al., 2014] networks, GRUadj (·) : Rn → Rn×d and GRUabs (·) : Rn → Rn×d , as time encoding networks: Hadj = GRUadj (∆adj ), 2
adj
Habs = GRUabs (∆abs ),
(2)
n×d
where features H ∈ R represent the short-term temporal rhythm and features2 Habs ∈ Rn×d represent the longterm temporal rhythm. To learn personalized time-aware embeddings, we propose to learn two gating parameters γ rec and γ exp for recommendation embeddings and explanation embeddings, respectively. These gates adaptively fuse the two features: Ẽrec = γ rec · Hadj + (1 − γ rec ) · Habs , Ẽ
exp
=γ
exp
·H
adj
+ (1 − γ
exp
abs
)·H ,
(3) (4)
where the gating parameters γ rec and γ exp are learned as follows: γ rec = σ(MLPrec (Ērec )),
γ exp = σ(MLPexp (Ēexp )),
(5)
where MLPrec (·) : Rd → R1 and MLPexp (·) : Rd → R1 are two-layer perceptrons; Ērec = Mean1(Erec ) ∈ R1×d and Ēexp = Mean1(Eexp ) ∈ R1×d ; σ is sigmoid activation. Finally, the personalized time-aware embeddings are obtained by combining the base embeddings2 Erec and Eexp and the temporal features2 Ẽrec and Ẽexp : Erec-time = Erec + β · Ẽrec ,
Eexp-time = Eexp + β · Ẽexp ,
(6)
where β is a scaling hyperparameter that controls the strength of temporal information.
3.2
Multi-interest Personalization
In this section, we further extract fine-grained sub-interests of users from the obtained personalized time-aware embeddings Erec-time and Eexp-time . To this end, we divide Erec-time and Eexp-time into several sub-embeddings, and we extend LRU [Orvieto et al., 2023] to a multihead variant to extract sub-interests from these sub-embeddings. Specifically, given interaction (vi , ei , ti ) of the i-th time step and its corresponding time-aware personalized recommendation embedding Erec-time ∈ R1×d and explanation embedding Eexp-time ∈ i i 1×d R , we divide the recommendation embedding Erec-time i d (rec-time,1) (rec-time,H) into H sub-embeddings, ei , ..., ei ∈ R1× H , exp-time and divide the explanation embedding Ei into H subd (exp-time,1) (exp-time,H) embeddings, ei , ..., ei ∈ R1× H . Then, we propose a multihead LRU architecture with H independent recurrent streams and we apply the architecture to both the recommendation branch and the explanation branch. d d d d exp exp H × H and W H × H denote Let Whrec , Urec h ∈ R h , Uh ∈ R the learnable matrices for the h-th head in the recommendation branch and explanation branch, respectively. Each head learns to focus on a specific sub-interest. The hidden state for each head at time step i is computed as follows: Xi d (rec-time,h) 1× H = e (Whrec )i−k Urec , h ∈ R k=1 k X i d (exp,h) (exp-time,h) 1× H = Hi ek (Whexp )i−k Uexp . h ∈ R (rec,h)
Hi
k=1
(rec,h)
H
(exp,h)
(rec,h)
d
To measure the personalized degree of semantic alignment between recommendations and explanations, we extend the work of [Yu et al., 2024] to design a dynamic dual-branch MI weighting mechanism. Dynamic dual-branch MI weighting mechanism. Given exp d ∈ Rd at time step the hidden features2 Zrec i ∈ R and Zi i, which contain time-aware personalized information and fine-grained multi-interest information, we use two learnable exp weights µrec i and µi to adaptively learn MI contributions for the recommendation branch and the explanation branch, respectively, as follows: exp µexp i = σ(MLPexp (Zi )),
rec µrec i = σ(MLPrec (Zi )),
d
(13)
where MLPrec (·) : R → R and MLPexp (·) : R → R1 , with sigmoid activation σ(·). We design the following objective function to make the model learn the semantic alignment information from the recommendation branch and the explanation branch as follows: JMI =
1
1 Xn exp exp rec [µrec i · JMI + µi · JMI ], i=1 n
d
(14)
P|V| exp(Sirec ) exp 1 rec where JMI = − |V| i=1 log P|V| exp(S rec ) and JMI = j=1 j P|E| exp(Siexp ) 1 [ − |E| exp , which are proposed in Yu et i=1 log P|E| j=1 exp(Sj
)
al., 2024]. Here, Sirec = (MV )i Λrec Zexp measures the simi ilarity between Zexp and the i-th item embedding (MV )i , i rec Siexp = (ME )i Λexp Zrec i measures the similarity between Zi and the i-th explanation embedding (ME )i . These similarities are calculated by the two learnable matrices Λrec and Λexp . rec In this way, the term JMI encourages each Zexp to be closer i to its paired recommendation embedding i.e. (MV )i than to exp unpaired ones. Same for JMI and each Zrec i .
3.4
(8)
d Prediction layer. Given the hidden features Zrec i ∈ R from exp d eq. (11) and Zi ∈ R from eq. (12) at time step i, the predicd |V| tion layer contains two functions Pred(Zrec i ) : R → R exp d |E| rec and Pred(Zi ) : R → R . The function Pred(Zi ) produces the probabilities of the user u selecting the next item; the function Pred(Zexp i ) produces the degree of alignment between the item and all explanations:
(rec,h) , ..., Hn ) ∈ Rn× H ,
(9)
d (exp,h) (exp,h) (exp,h) , ..., Hn ) ∈ Rn× H . = Stack(H1 , H2
(10)
, H2
Explanation Personalization
(7)
We stack the hidden states of all time steps to get the complete hidden state matrices H(rec,h) and H(exp,h) of the h-th head for recommendation branch and explanation branch, respectively, as follows: H(rec,h) = Stack(H1
3.3
Finally, the personalized multi-interest representations2 Zrec for the recommendation branch and Zexp for the explanation branch are obtained by concatenating all heads: Zrec = Concat(H(rec,1) , H(rec,2) , ..., H(rec,H) ) ∈ Rn×d ,
(11)
Zexp = Concat(H(exp,1) , H(exp,2) , ..., H(exp,H) ) ∈ Rn×d .
(12)
The proposed multihead LRU architecture not only captures multiple fine-grained sub-interests but also reduces the time and space complexity of the model. The time complexity of ideal parallel computing of all heads achieves 2 d 2 O(log(n) · ( H ) ). The space complexity achieves O( dH ). d2 The per-step incremental inference complexity is O( H 2 ), independent of n. Please see App. F for the detailed complexity analysis. We demonstrate the advantages of the proposed multihead LRU on efficiency in fig. 5.
Optimization
exp exp rec rec exp Pred(Zrec , i ) = MV Zi + b , Pred(Zi ) = ME Zi + b (15)
where brec ∈ R|V| and bexp ∈ R|E| are bias terms. Thus, the complete optimization objective is given as follows: L = Lrec + Lexp + JMI .
(16)
The recommendation objective Lrec and the explanation objective Lexp are computed via cross-entropy loss based on the prediction scores Pred(Zrec ) and Pred(Zexp ).3 By optimizing the final objective L, the proposed TMEPSR model achieves time-aware personalization, multiinterest personalization and explanation personalization for sequential recommendation, thereby improving both recommendation and explanation performance. 3
The formulations of Lrec and Lexp are provided in App. B.1.
AM-Electronics Recommendation Explanation R@10 N@10 R@10 N@10 PITF 0.0286 0.0139 0.0880 0.0547 0.0307 0.0154 0.0901 0.0553 BPER SCEMIM-GRU4Rec 0.0290 0.0152 0.0948 0.0629 SCEMIM-SASRec 0.0369 0.0195 0.1011 0.0681 SCEMIM-FMLPRec 0.0388 0.0204 0.1032 0.0702 SCEMIM-LRURec 0.0327 0.0164 0.1087 0.0727 SCEMIM-CL4SRec 0.0367 0.0183 0.1090 0.0713 SCEMIM-DiffuRec 0.0402 0.0212 0.1186 0.0794 SCEMIM-ComiRec 0.0391 0.0185 0.1229 0.0725 0.0390 0.0194 0.1170 0.0784 SCEMIM-Re4 SCEMIM-TEDDY 0.0385 0.0206 0.1173 0.0735 SCEMIM-TLSAN 0.0383 0.0154 0.0957 0.0574 SCEMIM-MOJITO 0.0399 0.0201 0.1156 0.0690 SCEMIM-UniRec 0.0401 0.0211 0.1201 0.0790 TME-PSR 0.0550 0.0291 0.1510 0.0979 Improvement 36.82% 37.26% 22.86% 23.30% Model
AM-Movies Recommendation Explanation R@10 N@10 R@10 N@10 0.0235 0.0101 0.0984 0.0679 0.0281 0.0131 0.1028 0.0687 0.0432 0.0216 0.1151 0.0771 0.0541 0.0257 0.1229 0.0809 0.0558 0.0261 0.1192 0.0822 0.0549 0.0269 0.1171 0.0830 0.0515 0.026 0.1213 0.0831 0.0544 0.0282 0.1352 0.0897 0.0509 0.0262 0.1247 0.0817 0.0487 0.0271 0.1207 0.0815 0.0514 0.0275 0.1201 0.0854 0.0457 0.0244 0.1012 0.0707 0.0508 0.0280 0.1245 0.0840 0.0560 0.0301 0.1339 0.0924 0.0620 0.0339 0.1359 0.0931 10.71% 12.62% 0.52% 0.76%
Yelp Recommendation Explanation R@10 N@10 R@10 N@10 0.0204 0.0094 0.0333 0.0185 0.0212 0.0106 0.0393 0.0214 0.0452 0.0228 0.0417 0.0233 0.0592 0.0364 0.0419 0.0231 0.0639 0.0403 0.0408 0.0249 0.0650 0.0397 0.0422 0.0230 0.0536 0.0320 0.0424 0.0259 0.0624 0.0377 0.0498 0.0276 0.0583 0.0358 0.0449 0.0273 0.0608 0.0354 0.0455 0.0258 0.0599 0.0367 0.0432 0.0260 0.0492 0.0271 0.0375 0.0215 0.0572 0.0354 0.0424 0.0262 0.0597 0.0386 0.0473 0.0270 0.0740 0.0455 0.0533 0.0309 13.87% 12.90% 7.03% 11.96%
Table 1: Recommendation and explanation performance, best results are marked in bold, second best results underlined. The improvement row is calculated as the percentage increase of the best result relative to the second best result. AM-Electronics Rec Exp Time-aware Multi-interest Explanation R@10 N@10 R@10 N@10 0.0327 0.0164 0.1087 0.0727 ! 0.0402 0.0197 0.1371 0.0912 0.0361 0.0184 0.1312 0.0849 ! ! 0.0324 0.0160 0.1252 0.0844 ! ! 0.0498 0.0247 0.1455 0.0938 0.0450 0.0240 0.1423 0.0920 ! ! ! ! 0.0401 0.0234 0.1378 0.0915 ! ! ! 0.0550 0.0291 0.1510 0.0979 Strategy
Table 2: Ablation of key strategies the AM-Electronics dataset, best results are marked in bold, second best results underlined.
4
EXPERIMENTS
Models and datasets. We compare the proposed TME-PSR model with 4 types of SR approaches: (1) General models: GRU4Rec [Hidasi et al., 2015], SASRec [Kang and McAuley, 2018], FMLPRec [Zhou et al., 2022], CL4SRec [Xie et al., 2022], LRURec [Yue et al., 2024], and DiffuRec [Li et al., 2023e]. (2) Models with temporal modeling: TLSAN [Zhang et al., 2021], MOJITO [Tran et al., 2023] and UniRec [Liu et al., 2024b]. (3) Models with multiinterest modeling: ComiRec [Cen et al., 2020], Re4 [Zhang et al., 2022] and TEDDY [Li et al., 2024]. (4) Explainable models: PITF [Rendle and Schmidt-Thieme, 2010] and BPER [Li et al., 2023a]. Since the first three types of models lack explainability, we insert these 12 models as backbones into the explainable SR framework SCEMIM [Yu et al., 2024] (The experiments described in SCEMIM have demonstrated that it can enable the model to acquire explainability while maintaining or even improving recommendation performance), thus obtaining 12 explainable SR models, denoted by SCEMIM-X (where X represents each model). We conduct experiments on three widely used datasets for explainable SR, including the Amazon Movies dataset, the Amazon Electronics dataset, and the Yelp [Li et al., 2021]
dataset. In these datasets, each user-item interaction is associated with a timestamp, an item ID, and an explanation ID. The explanations are extracted from user reviews and clustered into representative phrases through semantic hashing techniques [Li et al., 2021]. We follow the previous sequential recommendation work [Kang and McAuley, 2018; Yu et al., 2024] to divide each user’s interaction history chronologically into three parts: the most recent interaction for testing, the second most recent for validation, and the remaining for training.4 We evaluate both recommendation and explanation performance using two standard ranking metrics: Recall@K and NDCG@K, where K = 10. We refer to Recall@K as R@K and NDCG@K as N@K. We set hyperparameters α = 0.9, β = 0.1, d = 50 and H = 2.4 Recommendation and explanation performance comparison. Tab. 1 shows the overall performance of all methods in recommendation and explanation across three datasets. As shown, our method consistently outperforms all baseline models across all evaluation metrics for both the recommendation task from 10.71% to 37.26% and the explanation tasks from 0.52% to 25.73%. These results verify the wide applicability and stability of the proposed model. A comparison with the baseline models without SCEMIM are provided in App. C. Results show that our method still provides a significant improvement. Ablation Study. To further investigate the effectiveness of the three proposed personalized strategies, we evaluate every non-empty combination of the three personalization strategies on the AM-Electronics dataset.5 The main findings are as follows: (1) All three strategies are individually beneficial. Activating any single component (rows 2–4) already lifts both R@10 and N@10 above the backbone with no personalization (row 1). (2) Pairwise combinations show clear 4 The statistics of the preprocessed datasets, the hyperparameter analysis and more implementation details are provided in App. B. 5 The ablation study on all three datasets are provided in App. D.
Model only focuses on long term temporal features
Model equally focuses on long term and short term features
Model only focuses on short term temporal features
Model adaptively focuses on long term and short term features
Recommendation
Explanation
d=60
d=120
Figure 3: Comparison of four gating strategies. Results show that the model that adaptively focuses on both long-term and short-term features consistently achieves the best performance, which proves the effectiveness of the proposed method.
synergy. For example, combining time-aware personalization and multi-interest personalization (row 5) outperforms any single component. (3) The full model is consistently the best. Enabling all three strategies (bottom row) attains the top performance on every metric. This confirms that the three types of personalization have significant mutual synergy. In summary, the complete TME-PSR configuration reliably delivers the strongest recommendation and explanation quality across all settings. Detailed analysis of the effectiveness of the time-aware personalization. In eq. (3) we use a gating parameter γ rec to control the contribution of short-term and long-term temporal features to the recommendation time-aware embedding Ẽrec . A larger γ rec means a greater contribution of short-term temporal features to Ẽrec , while a smaller γ rec emphasizes longterm temporal features in Ẽrec . Same for the gating parameter γ exp and the explanation time-aware embedding Ẽexp in eq. (4). Such a dual-view gated time encoder helps us realize the time-aware personalization. To further validate the effectiveness of our dual-view gated time encoder, we compare the following four baselines: (1) We set Ẽrec = Ẽexp = Habs to obtain a model that only focuses on long-term temporal features. (2) We set Ẽrec = Ẽexp = Hadj to obtain a model that only focuses on short-term temporal features. (3) We set Ẽrec = Ẽexp = 0.5 × (Hadj + Habs ) to obtain a model that focuses on short-term and longterm temporal features equally. (4) A model uses the proposed dual-view gated time encoder, which focuses on shortterm and long-term temporal features adaptively as eq. (3) and eq. (4) show. As shown in fig. 3, models that focus on both short-term and long-term temporal patterns outperform those that only focus on a single pattern, and dynamically balancing the two patterns further improves performance, validating the effectiveness of the proposed dual-view gated time encoder. We also found a linear relationship between the values of the two gating parameters and the average time interval of each user sequence, please see App. E for details. Detailed analysis of the effectiveness of the multiinterest personalization. To further explore the effect of the head number H of the proposed multihead LRU architecture, we evaluate the impact of different head numbers H under varying embedding sizes d ∈ {60, 120, 240}, ensuring that the head number H is a divisor of d. As fig. 4 shows, the
d=240
Figure 4: Left two columns show recommendation and explanation performance with different head numbers H under different embedding sizes d. Results show that the optimal head number increases with the increase of d. The third column shows the relationship between H and the encoder parameters. As H increases, the number of parameters decreases inversely. GRU4Rec SASRec LRURec Multihead LRU H=2 Multihead LRU H=4 Multihead LRU H=6 Multihead LRU H=8
Figure 5: Inference and training efficiency of GRU4Rec, SASRec, LRURec and the proposed multihead LRU. Results show that the proposed method outperforms the other three methods, and its efficiency increases with the head number H.
optimal head number H ∗ consistently increases with the performance on both the recommendation branch and the explanation branch: H ∗ ≈ 2, 6, 10 for d = 60, 120, 240, and as the embedding size d increases, the overall performance improves. These findings indicate that higher dimensions of the representation space provide richer capabilities for interest disentanglement, and higher dimensions require more finegrained disentanglement. When H exceeds H ∗ , performance begins to degrade, likely due to each head being allocated too narrow a subspace, limiting its expressiveness. We also explored the relationship between the number of encoder parameters and the head number H. As the third column of fig. 4 shows, when d remains fixed, the model parameters gradually decrease as H increases, which follows an inverse 2 proportional relationship dH and achieving stronger disentanglement with fewer parameters as subspaces get finer. We further explored incremental inference and training efficiency across GRU4Rec [Hidasi et al., 2015], SASRec [Kang and McAuley, 2018], LRURec [Yue et al., 2024], and the proposed multihead LRU (H = 2, 4, 6, 8) on embedding size d = 240, batch size= 256. We perform only one batch of inference at each step and continuously expand the input length after each prediction, and test the training time required for a batch. As fig. 5 shows, the proposed method outperforms the compared three methods in terms of incremental inference and training efficiency, and the performance improvement is more significant with a larger number of heads.
(a)
(b)
Balanced
Recommendation-dominant
Explanation-dominant
Figure 6: Visualization of multihead LRU output representation at d = 120, H = 6 with t-SNE. Model with a fixed weight
Model with one dynamic weight
Model with two dynamic weights
Figure 7: Comparison of three MI weighting strategies. Results show that the model with two dynamic weights consistently achieves the best performance, which proves the effectiveness of the proposed dynamic dual-branch MI weighting mechanism.
In summary, as d increases, accordingly increases the number of heads H simultaneously deliver: (1) higher accuracy (2) reduced parameters (3) faster encoding. These results provide converging evidence that our approach achieves fine-grained sub-interest disentanglement while remaining lightweight and efficient. To further explore whether the proposed multihead LRU architecture truly capture distinct and fine-grained subinterests, we performed a t-SNE [Maaten and Hinton, 2008] analysis on the output representations of each head under the conditions of d = 120 and H = 6. As fig. 6 shows, the six heads form well-separated clusters; each head covers a complementary region in the embedding space. Detailed analysis of the effectiveness of the explanation personalization. In eq. (14), we use two learnable weights µrec and µexp to dynamically control the contriburec tion of the recommendation MI loss JMI and the explanaexp tion MI loss JMI to the total MI loss JMI . To further validate the effectiveness of such a dynamic dual-branch MI weighting mechanism, we compare the following three baseexp rec lines: (1) We set JMI = 0.001 × (JMI + JMI ) to obtain a exp rec model that uses a fixed weight for JMI and JMI . (2) We set exp exp rec JMI = µ · (JMI + JMI ) with µ = σ(MLP([Zrec i , Zi ])) to rec obtain a model that uses one learnable weight for JMI and exp JMI . (3) A model using the proposed dynamic MI weighting mechanism, which applies two dynamic weights µrec and exp rec µexp to JMI and JMI , respectively, as eq. (14) shows. As fig. 7 shows, the model with learnable weights outperform those with a fixed weight, and dynamically learning two taskspecific MI weights further improves performance, validating the effectiveness of the proposed dynamic dual-branch MI weighting mechanism. Let µrec (u,i) represent the MI weight of the recommendation task of the i-th interaction of user sequence Su , and µexp (u,i) represent the MI weight of the explanation task. We com-
Figure 8: (a) Visualization of distributions of all interaction sequences average dynamic MI weights µ̄rec and µ̄exp . (b) K-means clustering based on (µ̃rec , µ̃exp ).
Pn exp rec pute the average weight µ̄rec = u = i=1 µ(u,i) /n and µ̄u Pn exp /n. fig. 8(a) reports the distributions of the averµ i=1 (u,i) exp aged weights µ̄rec u and µ̄u of all interaction sequences. Results show that all distributions are highly left-skewed and concentrated near zero, indicating that most sequences are assigned low semantic alignment strengths between recommendations and explanations in both the recommendation task and the explanation task.6 To further investigate user-level differences in the degree of semantic alignment strength, we apply an enhanced K-means clustering [De Amorim, 2016] on the following 2D points set, exp exp rec rec (µ̃rec u , µ̃u ), where µ̃u and µ̃u are normalized values of µ̄u exp exp and µ̃ exhibit and µ̄u . Results show that user-level µ̃rec u u three well-separated clusters shown in fig. 8(b): (1) A balanced group near the diagonal, this type of users has similar semantic alignment requirements in the recommendation task and the explanation task. (2) A recommendation-dominant group, this type of users has higher semantic alignment requirements in the recommendation task. (3) An explanationdominant group, this type of users has higher semantic alignment requirements in the explanation task.6 This proves that even for the same user, the recommendation task and the explanation task have different requirements for semantic alignment, which confirms the necessity of personalization.
5
CONCLUSION
This paper proposes TME-PSR, a model for sequential recommendation that simultaneously addresses three types of personalization: time-aware, multi-interest, and explanation personalization. Our approach advances the field by demonstrating that personalization should be considered jointly across multiple dimensions. Extensive experiments demonstrate that our model consistently improves recommendation accuracy, explanation quality and efficiency across multiple datasets. For future work, we plan to extend personalization to cross-domain or multi-modal recommendation settings, while maintaining explainability, represents an exciting opportunity to generalize the proposed model further.
6
Specific examples of the two types of user alignment requirements discussed in fig. 8(a) and the three types of users in fig. 8(b) are provided in App. G.
References [Cen et al., 2020] Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang. Controllable multi-interest framework for recommendation. In Proceedings of the 26th ACM SIGKDD international conference, pages 2942–2951, 2020. [Chai et al., 2022] Zheng Chai, Zhihong Chen, Chenliang Li, Rong Xiao, Houyi Li, Jiawei Wu, Jingxu Chen, and Haihong Tang. User-aware multi-interest learning for candidate matching in recommenders. In Proceedings of the 45th SIGIR conference on research and development in information retrieval, pages 1326–1335, 2022. [Chen et al., 2020] Tong Chen, Hongzhi Yin, Quoc Viet Hung Nguyen, Wen-Chih Peng, Xue Li, and Xiaofang Zhou. Sequence-aware factorization machines for temporal predictive analytics. In 2020 IEEE 36th ICDE, pages 1405–1416, 2020. [Chen et al., 2021] Gaode Chen, Xinghua Zhang, Yanyan Zhao, Cong Xue, and Ji Xiang. Exploring periodicity and interactivity in multi-interest framework for sequential recommendation. arXiv preprint arXiv:2106.04415, 2021. [Chen et al., 2022a] Lihua Chen, Ning Yang, and Philip S Yu. Time lag aware sequential recommendation. In Proceedings of the 31st ACM international conference on information & knowledge management, pages 212–221, 2022. [Chen et al., 2022b] Qi Chen, Guohui Li, Quan Zhou, Si Shi, and Deqing Zou. Double attention convolutional neural network for sequential recommendation. ACM Transactions on the Web, 16(4):1–23, 2022. [Chen et al., 2023] Huiyuan Chen, Kaixiong Zhou, Zhimeng Jiang, Chin-Chia Michael Yeh, Xiaoting Li, Menghai Pan, Yan Zheng, Xia Hu, and Hao Yang. Probabilistic masked attention networks for explainable sequential recommendation. In IJCAI, pages 2068–2076, 2023. [Chung et al., 2014] Jun-Young Chung, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv: Neural and Evolutionary Computing, Dec 2014. [Cui et al., 2022] Zhihong Cui, Hongxu Chen, Lizhen Cui, Shijun Liu, Xueyan Liu, Guandong Xu, and Hongzhi Yin. Reinforced kgs reasoning for explainable sequential recommendation. World Wide Web, 25(2):631–654, 2022. [Dai et al., 2024] Sunhao Dai, Changle Qu, Sirui Chen, Xiao Zhang, and Jun Xu. Recode: Modeling repeat consumption with neural ode. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2599–2603, 2024. [Dang et al., 2023] Yizhou Dang, Enneng Yang, Guibing Guo, Linying Jiang, Xingwei Wang, Xiaoxiao Xu, Qinghui Sun, and Hong Liu. Uniform sequence better: Time interval aware data augmentation for sequential recommendation. In Proceedings of the AAAI conference on artificial intelligence, pages 4225–4232, 2023.
[De Amorim, 2016] Renato Cordeiro De Amorim. A survey on feature weighting based k-means algorithms. Journal of Classification, 33(2):210–242, 2016. [de Souza Pereira Moreira et al., 2021] Gabriel de Souza Pereira Moreira, Sara Rabhi, Jeong Min Lee, Ronay Ak, and Even Oldridge. Transformers4rec: Bridging the gap between nlp and sequential/session-based recommendation. In Proceedings of the 15th ACM conference on recommender systems, pages 143–153, 2021. [Ding and Li, 2005] Yi Ding and Xue Li. Time weight collaborative filtering. In Proceedings of the 14th ACM international conference on Information and knowledge management, pages 485–492, 2005. [Ding et al., 2021] Yujuan Ding, Yunshan Ma, Wai Keung Wong, and Tat-Seng Chua. Leveraging two types of global graph for sequential fashion recommendation. In Proceedings of the 2021 international conference on multimedia retrieval, pages 73–81, 2021. [Donkers et al., 2017] Tim Donkers, Benedikt Loepp, and Jürgen Ziegler. Sequential user-based recurrent neural network recommendations. In Proceedings of the eleventh ACM conference on recommender systems, pages 152– 160, 2017. [Du et al., 2024] Yingpeng Du, Ziyan Wang, Zhu Sun, Yining Ma, Hongzhi Liu, and Jie Zhang. Disentangled multiinterest representation learning for sequential recommendation. In Proceedings of the 30th ACM SIGKDD Conference, pages 677–688, 2024. [Geng et al., 2022a] Shijie Geng, Zuohui Fu, Juntao Tan, Yingqiang Ge, Gerard De Melo, and Yongfeng Zhang. Path language modeling over knowledge graphsfor explainable recommendation. In Proceedings of the ACM web conference 2022, pages 946–955, 2022. [Geng et al., 2022b] Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM conference on recommender systems, pages 299–315, 2022. [Gholami et al., 2022] Ehsan Gholami, Mohammad Motamedi, and Ashwin Aravindakshan. Parsrec: Explainable personalized attention-fused recurrent sequential recommendation using session partial actions. In Proceedings of the 28th ACM SIGKDD Conference, pages 454–464, 2022. [Hidasi and Karatzoglou, 2018] Balázs Hidasi and Alexandros 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, Oct 2018. [Hidasi et al., 2015] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Sessionbased recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939, 2015. [Hou et al., 2022] Yupeng Hou, Shanlei Mu, WayneXin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. To-
wards universal sequence representation learning for recommender systems. Jun 2022. [Huang et al., 2019] Xiaowen Huang, Quan Fang, Shengsheng Qian, Jitao Sang, Yan Li, and Changsheng Xu. Explainable interaction-driven user modeling over knowledge graph for sequential recommendation. In proceedings of the 27th ACM international conference on multimedia, pages 548–556, 2019. [Järvelin and Kekäläinen, 2002] Kalervo Järvelin and Jaana Kekäläinen. Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS), 20(4):422–446, 2002. [Ji et al., 2024] Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. Genrec: Large language model for generative recommendation. In European Conference on Information Retrieval, pages 494–502. Springer, 2024. [Jiang et al., 2024] Yiheng Jiang, Yuanbo Xu, Yongjian Yang, Funing Yang, Pengyang Wang, Chaozhuo Li, Fuzhen Zhuang, and Hui Xiong. Trimlp: A foundational mlp-like architecture for sequential recommendation. ACM Transactions on Information Systems, 42(6):1– 34, 2024. [Kang and McAuley, 2018] Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), pages 197–206. IEEE, 2018. [Li et al., 2019] Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. Multi-interest network with dynamic routing for recommendation at tmall. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 2615– 2623, 2019. [Li et al., 2020] Jiacheng Li, Yujie Wang, and Julian McAuley. Time interval aware self-attention for sequential recommendation. In Proceedings of the 13th international conference on web search and data mining, pages 322–330, 2020. [Li et al., 2021] Lei Li, Yongfeng Zhang, and Li Chen. Extra: Explanation ranking datasets for explainable recommendation. In Proceedings of the 44th International ACM SIGIR conference on Research and Development in Information Retrieval, pages 2463–2469, 2021. [Li et al., 2023a] Lei Li, Yongfeng Zhang, and Li Chen. On the relationship between explanation and recommendation: Learning to rank explanations for improved performance. ACM Transactions on Intelligent Systems and Technology, 14(2):1–24, 2023. [Li et al., 2023b] Lei Li, Yongfeng Zhang, and Li Chen. Prompt distillation for efficient llm-based recommendation. In Proceedings of the 32nd ACM international conference on information and knowledge management, pages 1348–1357, 2023.
[Li et al., 2023c] Xinhang Li, Chong Chen, Xiangyu Zhao, Yong Zhang, and Chunxiao Xing. E4srec: An elegant effective efficient extensible solution of large language models for sequential recommendation. arXiv preprint arXiv:2312.02443, 2023. [Li et al., 2023d] Yicong Li, Hongxu Chen, Yile Li, Lin Li, Philip S Yu, and Guandong Xu. Reinforcement learning based path exploration for sequential explainable recommendation. IEEE Transactions on Knowledge and Data Engineering, 35(11):11801–11814, 2023. [Li et al., 2023e] Zihao Li, Aixin Sun, and Chenliang Li. Diffurec: A diffusion model for sequential recommendation. ACM Transactions on Information Systems, 42(3):1– 28, 2023. [Li et al., 2024] Zihao Li, Yunfan Xie, Wei Emma Zhang, Pengfei Wang, Lixin Zou, Fei Li, Xiangyang Luo, and Chenliang Li. Disentangle interest trend and diversity for sequential recommendation. Information Processing & Management, 61(3):103619, 2024. [Lin et al., 2018] Xiang Lin, Shuzi Niu, Yiqiao Wang, and Yucheng Li. K-plet recurrent neural networks for sequential recommendation. In The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, page 1057–1060, Jun 2018. [Lin et al., 2022] Qianying Lin, Wen-Ji Zhou, Yanshi Wang, Qing Da, Qing-Guo Chen, and Bing Wang. Sparse attentive memory network for click-through rate prediction with long sequences. Aug 2022. [Liu et al., 2023a] Qidong Liu, Fan Yan, Xiangyu Zhao, Zhaocheng Du, Huifeng Guo, Ruiming Tang, and Feng Tian. Diffusion augmentation for sequential recommendation. In Proceedings of the 32nd ACM International conference on information and knowledge management, pages 1576–1586, 2023. [Liu et al., 2023b] Sijia Liu, Jiahao Liu, Hansu Gu, Dongsheng Li, Tun Lu, Peng Zhang, and Ning Gu. Autoseqrec: Autoencoder for efficient sequential recommendation. In Proceedings of the 32nd ACM international conference on information and knowledge management, pages 1493–1502, 2023. [Liu et al., 2024a] Qidong Liu, Xian Wu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Feng Tian, and Yefeng Zheng. Large language models enhanced sequential recommendation for long-tail user and item. arXiv e-prints, pages arXiv–2405, 2024. [Liu et al., 2024b] Yang Liu, Yitong Wang, and Chenyue Feng. Unirec: A dual enhancement of uniformity and frequency in sequential recommendations. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1483–1492, 2024. [Liu et al., 2024c] Zihan Liu, Yupeng Hou, and Julian McAuley. Multi-behavior generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1575– 1585, 2024.
[Liu et al., 2025] Qidong Liu, Xian Wu, Wanyu Wang, Yejing Wang, Yuanshao Zhu, Xiangyu Zhao, Feng Tian, and Yefeng Zheng. Llmemb: Large language model can be a good embedding generator for sequential recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12183–12191, 2025. [Long et al., 2024] Chao Long, Huanhuan Yuan, Junhua Fang, Xuefeng Xian, Guanfeng Liu, Victor S Sheng, and Pengpeng Zhao. Learning global and multi-granularity local representation with mlp for sequential recommendation. ACM Transactions on Knowledge Discovery from Data, 18(4):1–15, 2024. [Maaten and Hinton, 2008] Laurensvander Maaten and GeoffreyE. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, Jan 2008. [Orvieto et al., 2023] Antonio Orvieto, Samuel L Smith, Albert Gu, Anushan Fernando, Caglar Gulcehre, Razvan Pascanu, and Soham De. Resurrecting recurrent neural networks for long sequences. In International Conference on Machine Learning, pages 26670–26698. PMLR, 2023. [Peng et al., 2024] Bo Peng, Ziqi Chen, Srinivasan Parthasarathy, and Xia Ning. Modeling sequences as star graphs to address over-smoothing in self-attentive sequential recommendation. ACM Transactions on Knowledge Discovery from Data, 18(8):1–24, 2024. [Pi et al., 2019] Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. Practice on long sequential user behavior modeling for click-through rate prediction. In Proceedings of the 25th ACM SIGKDD International Conference, Jul 2019. [Qin et al., 2020] Jiarui Qin, Weinan Zhang, Xin Wu, Jiarui Jin, Yuchen Fang, and Yong Yu. User behavior retrieval for click-through rate prediction. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, Jul 2020. [Qin et al., 2024] Xiuyuan Qin, Huanhuan Yuan, Pengpeng Zhao, Guanfeng Liu, Fuzhen Zhuang, and Victor S Sheng. Intent contrastive learning with cross subsequences for sequential recommendation. In Proceedings of the 17th ACM international conference on web search and data mining, pages 548–556, 2024. [Qiu et al., 2022] Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. Contrastive learning for representation degeneration problem in sequential recommendation. In Proceedings of the fifteenth ACM international conference on web search and data mining, pages 813–823, 2022. [Rajput et al., 2023] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems, 36:10299–10315, 2023. [Rendle and Schmidt-Thieme, 2010] Steffen Rendle and Lars Schmidt-Thieme. Pairwise interaction tensor factorization for personalized tag recommendation. In
Proceedings of the third ACM international conference on Web search and data mining, pages 81–90, 2010. [Rendle et al., 2010] Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. Factorizing personalized markov chains for next-basket recommendation. In Proceedings of the 19th international conference on World wide web, pages 811–820, 2010. [Shi et al., 2023] Xiaoyu Shi, Quanliang Liu, Yanan Bai, and Mingsheng Shang. Rtisr: a review-driven time intervalaware sequential recommendation method. Journal of Big Data, 10(1):32, 2023. [Sun et al., 2019] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441–1450, 2019. [Tang and Wang, 2018] Jiaxi Tang and Ke Wang. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining, pages 565–573, 2018. [Tran et al., 2023] Viet Anh Tran, Guillaume Salha-Galvan, Bruno Sguerra, and Romain Hennequin. Attention mixtures for time-aware sequential recommendation. In Proceedings of the 46th SIGIR conference, pages 1821–1826, 2023. [Wang et al., 2023] Jinpeng Wang, Ziyun Zeng, Yunxiao Wang, Yuting Wang, Xingyu Lu, Tianxiang Li, Jun Yuan, Rui Zhang, Hai-Tao Zheng, and Shu-Tao Xia. Missrec: Pre-training and transferring multi-modal interest-aware sequence representation for recommendation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6548–6557, 2023. [Wang et al., 2024] Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al. Eager: Two-stream generative recommender with behavior-semantic collaboration. In Proceedings of the 30th ACM SIGKDD Conference, pages 3245–3254, 2024. [Wu et al., 2023] Zihao Wu, Xin Wang, Hong Chen, Kaidong Li, Yi Han, Lifeng Sun, and Wenwu Zhu. Diff4rec: Sequential recommendation with curriculumscheduled diffusion augmentation. In Proceedings of the 31st ACM international conference on multimedia, pages 9329–9335, 2023. [Xia et al., 2022] Lianghao Xia, Chao Huang, Yong Xu, and Jian Pei. Multi-behavior sequential recommendation with temporal graph transformer. IEEE Transactions on Knowledge and Data Engineering, 35(6):6099–6112, 2022. [Xian et al., 2019] Yikun Xian, Zuohui Fu, Shan Muthukrishnan, Gerard De Melo, and Yongfeng Zhang. Reinforcement knowledge graph reasoning for explainable recommendation. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, pages 285–294, 2019.
[Xie et al., 2022] Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui. Contrastive learning for sequential recommendation. In 2022 IEEE 38th international conference on data engineering (ICDE), pages 1259–1273. IEEE, 2022. [Xie et al., 2023] Yueqi Xie, Jingqi Gao, Peilin Zhou, Qichen Ye, Yining Hua, Jae Boum Kim, Fangzhao Wu, and Sunghun Kim. Rethinking multi-interest learning for candidate matching in recommender systems. In Proceedings of the 17th ACM conference on recommender systems, pages 283–293, 2023. [Xiong et al., 2010] Liang Xiong, Xi Chen, Tzu-Kuo Huang, Jeff Schneider, and Jaime G Carbonell. Temporal collaborative filtering with bayesian probabilistic tensor factorization. In Proceedings of the 2010 SIAM international conference on data mining, pages 211–222. SIAM, 2010. [Yang et al., 2021] Zuoxi Yang, Shoubin Dong, and Jinlong Hu. Gfe: General knowledge enhanced framework for explainable sequential recommendation. Knowledge-Based Systems, 230:107375, 2021. [Yang et al., 2023] Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. Generate what you prefer: Reshaping sequential recommendation via guided diffusion. Advances in Neural Information Processing Systems, 36:24247–24261, 2023. [Yu et al., 2019] Zeping Yu, Jianxun Lian, Ahmad Mahmoody, Gongshen Liu, and Xing Xie. Adaptive user modeling with long and short-term preferences for personalized recommendation. In IJCAI, volume 7, pages 4213– 4219, 2019. [Yu et al., 2024] Yi Yu, Kazunari Sugiyama, and Adam Jatowt. Sequential recommendation with collaborative explanation via mutual information maximization. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1062–1072, 2024. [Yuan et al., 2019] Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M Jose, and Xiangnan He. A simple convolutional generative network for next item recommendation. In Proceedings of the twelfth ACM international conference on web search and data mining, pages 582–590, 2019. [Yue et al., 2024] Zhenrui Yue, Yueqi Wang, Zhankui He, Huimin Zeng, Julian McAuley, and Dong Wang. Linear recurrent units for sequential recommendation. In Proceedings of the 17th ACM international conference on web search and data mining, pages 930–938, 2024. [Zhang et al., 2021] Jianqing Zhang, Dongjing Wang, and Dongjin Yu. Tlsan: Time-aware long-and short-term attention network for next-item recommendation. Neurocomputing, 441:179–191, 2021. [Zhang et al., 2022] Shengyu Zhang, Lingxiao Yang, Dong Yao, Yujie Lu, Fuli Feng, Zhou Zhao, Tat-Seng Chua, and Fei Wu. Re4: Learning to re-contrast, re-attend, reconstruct for multi-interest recommendation. In Proceed-
ings of the ACM web conference 2022, pages 2216–2226, 2022. [Zhang et al., 2024] Shengzhe Zhang, Liyi Chen, Chao Wang, Shuangli Li, and Hui Xiong. Temporal graph contrastive learning for sequential recommendation. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 9359–9367, 2024. [Zheng et al., 2022] Yu Zheng, Chen Gao, Jianxin Chang, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. Disentangling long and short-term interests for recommendation. In Proceedings of the ACM web conference 2022, pages 2256–2267, 2022. [Zhou et al., 2022] Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. Filter-enhanced mlp is all you need for sequential recommendation. In Proceedings of the ACM web conference 2022, pages 2388–2399, 2022.
A
Sequential recommendation predicts a user’s next interaction by modeling their historical behaviors. Early models used Markov chains [Rendle et al., 2010] to model sequential data. Early studies adopted CNN [Chen et al., 2022b; Tang and Wang, 2018; Yuan et al., 2019] and RNN methods [Donkers et al., 2017; Hidasi et al., 2015; Hidasi and Karatzoglou, 2018; Lin et al., 2018] to capture complex patterns. More recently, Transformer-based models [Kang and McAuley, 2018; Sun et al., 2019; de Souza Pereira Moreira et al., 2021] use self-attention for long-range dependencies with better parallelism. Some studies used contrastive learning [Xie et al., 2022; Qiu et al., 2022; Qin et al., 2024] to improve recommendation performance. Diffusion brings a generative view to SR by treating the target item as a distribution and denoising it from noise conditioned on a user’s history, which helps capture uncertainty and multi-aspect signals, representative methods [Li et al., 2023e; Yang et al., 2023; Wu et al., 2023; Liu et al., 2023a] show competitive accuracy with a practical denoising pipeline. GNN-based SR [Liu et al., 2023b; Peng et al., 2024; Ding et al., 2021] constructs session/transition graphs to encode higher-order item transitions and local dependencies. All-MLP designs [Zhou et al., 2022; Jiang et al., 2024; Long et al., 2024] replace self-attention with learnable filter blocks for high throughput and low latency. Multi-modal SR pretrains on text/images (etc.) and transfers to downstream SR: [Wang et al., 2023] adopts a pretrain-and-transfer encoder–decoder with contrastive objectives, while [Hou et al., 2022] learns ID-agnostic universal sequence/item representations from descriptions for cross-domain transfer. Generative retrieval replaces ANN with sequence models that decode semantic IDs of items; [Rajput et al., 2023; Liu et al., 2024c; Wang et al., 2024] demonstrates strong performance and unifies retrieval/ranking under a single autoregressive model. LLMs unify diverse rec tasks via text-to-text formulations [Geng et al., 2022b; Ji et al., 2024; Li et al., 2023b], or use LLMs for learning semantic embeddings [Liu et al., 2025; Liu et al., 2024a; Li et al., 2023c]. Handling ultra-long histories (thousands of events) motivates scalable encoders and training/serving optimizations; recent industry-scale methods [Yue et al., 2024; Qin et al., 2020; Lin et al., 2022; Pi et al., 2019] design long-sequence-optimized Transformers and serving pipelines, and surveys position ultra-long SR as a key frontier.
B
Implementation Details
All experiments are conducted on a server equipped with 4×NVIDIA RTX 4090 GPUs (24GB each), an Intel Xeon Silver 4314 CPU @ 2.40GHz, and 256GB RAM. The software environment includes Ubuntu 18.04.4 LTS, Python 3.9.21, PyTorch 2.6.0+cu124, and CUDA 12.2.
B.1
Dataset Movies Electronics Yelp # Users 22,248 18,129 90,742 # Items 15,458 10,435 46,372 # Explanations 31,451 23,541 119,874 # Actions 305,801 144,772 1,204,515 Avg. actions/users 13.75 7.95 13.27 Avg. actions/items 19.78 13.87 25.98 Avg. actions/expls 9.72 6.15 10.05 User-item density 0.089% 0.076% 0.029% User-expl density 0.063% 0.034% 0.022% User-item-expl density 2.83e-6% 3.23e-6% 2.39e-7%
Additional Related Work
Complete formulations of the recommendation objective function and the explanation objective function. Lrec = − log P
rec exp([Pred(Zrec i )]yi ) , rec exp([Pred(Z i )]v ) v∈V
(17)
Table 3: The statistics of the preprocessed datasets. exp exp([Pred(Zexp i )]yi ) , Lexp = − log P exp e∈E exp([Pred(Zi )]e )
(18)
where yirec is the target label at time step i in the recommendation branch; yiexp is the target label at time step i in the explanation branch.
B.2
Baselines
We compare our method with several representative sequential recommendation baselines: • PITF [Rendle and Schmidt-Thieme, 2010]: A pairwiseinteraction tensor-factorization model using triadic (user, item, tag) embeddings to support explainable ranking. • BPER [Li et al., 2023a]: A Bayesian-personalized explanation-ranking model using joint BPR optimization to align item recommendations with textual rationales. • GRU4Rec [Hidasi et al., 2015]: A gated-recurrent-unit [Chung et al., 2014] model using pairwise ranking losses to learn within-session user–item transitions. • SASRec [Kang and McAuley, 2018]: A transformer-based sequential model leveraging self-attention to capture useritem transition patterns. • FMLPRec [Zhou et al., 2022]: A filter-based all-MLP model using learnable Fourier kernels to model sequential signals with low latency. • LRURec [Yue et al., 2024]: A linear-recurrence model using closed-form state transitions to approximate RNN behavior while enabling parallel computation. • CL4SRec [Xie et al., 2022]: A contrastive-learning SR model that augments sequences and jointly optimizes nextitem prediction with self-supervised contrastive signals to learn stronger user representations. • DiffuRec [Li et al., 2023e]: A diffusion-based generative SR model that represents target items as distributions and denoises them from Gaussian noise conditioned on the user’s history, improving recommendation under preference uncertainty. • ComiRec [Cen et al., 2020]: A controllable multi-interest framework that extracts multiple intent vectors (via capsule/attention variants) and matches candidates to the most relevant interest for each user. • Re4 [Zhang et al., 2022]: A multi-interest framework using re-contrast, re-attend and re-construct regularisers to maintain distinct intent vectors. • TEDDY [Li et al., 2024]: A disentangled SR method that separates interest trend and interest diversity (via adaptive masking and dedicated objectives) to improve robustness and
AM-Electronics AM-Movies Yelp Recommendation Recommendation Recommendation R@10 N@10 R@10 N@10 R@10 N@10 GRU4Rec 0.0259 0.0134 0.0431 0.0208 0.0451 0.0228 0.0342 0.0182 0.0536 0.0256 0.0604 0.0364 SASRec FMLPRec 0.0357 0.0186 0.0558 0.0253 0.0620 0.0403 LRURec 0.0345 0.0159 0.0560 0.0259 0.0625 0.0397 CL4SRec 0.0371 0.0194 0.0523 0.0262 0.0510 0.0318 DiffuRec 0.0389 0.0201 0.0548 0.0273 0.0599 0.0325 ComiRec 0.0395 0.0198 0.0515 0.0271 0.0562 0.0333 0.0355 0.0185 0.0540 0.0265 0.0643 0.0400 Re4 TEDDY 0.0392 0.0208 0.0537 0.0279 0.0628 0.0411 TLSAN 0.0372 0.0191 0.0464 0.0245 0.0571 0.0323 MOJITO 0.0386 0.0198 0.0519 0.0302 0.0575 0.0364 UniRec 0.0403 0.0200 0.0557 0.0317 0.0578 0.0362 TME-PSR 0.0550 0.0291 0.0620 0.0339 0.0740 0.0455 Improvement 36.47% 39.90% 10.71% 6.94% 15.09% 10.71%
Recommendation
Explanation
Model
Table 4: overall with non-SCEMIM models
personalization. • TLSAN [Zhang et al., 2021]: A dual-stream attention model using time-position embeddings for temporal drift. • MOJITO [Tran et al., 2023]: A time-aware Transformer that mixes temporal contexts with Gaussian-mixture attention, enabling better next-item prediction under varying temporal conditions. • UniRec [Liu et al., 2024b]: A uniformity–frequency enhanced SR model that exploits regularized time spacing and item-frequency cues via dual objectives to mitigate drift and sparsity. In addition, SCEMIM [Yu et al., 2024] is an explainable sequential recommendation framework that enhances sequential recommendation with mutual information maximization between recommendation and explanation representations. table 3 shows the statistics of the datasets.
Figure 9: Impact of hyperparameters across three datasets. Results show that when set α = 0.9 and set β = 0.1, the proposed model achieve the best performance.
B.3
The ideal DCG, obtained by perfect ranking, is
Evaluation Metrics
Recall measures whether the ground-truth item (or explanation) appears in the top-k predictions. NDCG (Normalized Discounted Cumulative Gain) [Järvelin and Kekäläinen, 2002] further considers the position of correct predictions, assigning higher rewards to items ranked higher in the prediction list. Higher Recall and NDCG [Järvelin and Kekäläinen, 2002] values indicate better recommendation and explanation performance. Recall@K. For a user u, let Gu be the set of relevant items bu(K) the top-K recommendations. The in the test set and R per-user recall is (K)
Recall@K(u) =
bu |R
∩ Gu | . |Gu |
(19)
Averaging over all test users U yields Recall@K =
bu(K) ∩ Gu | 1 X |R . |U| |Gu |
(20)
u∈U
When |Gu | = 1 (next-item prediction), Recall@K reduces to Hit@K.
NDCG@K. Given the ranked list, let ru,i denote the relevance (1/0) of the item at position i for user u. The discounted cumulative gain is DCGu @K =
K X
ru,i . log2 (i + 1) i=1
min(K,|Gu |)
IDCGu @K =
(21)
1 . log2 (i + 1)
(22)
DCGu @K ∈ [0, 1], IDCGu @K
(23)
X i=1
We normalise DCG to NDCGu @K =
and report the average over all users: 1 X NDCG@K = NDCGu @K. |U|
(24)
u∈U
B.4
Hyperparameter Sensitivity Analysis
Figure 9 plots NDCG@10 on the three datasets while sweeping the two scalar hyperparameters that control feature fusion in TME-PSR. The hyperparameter α determines the degree of fusion between the recommendation base embedding EV and the explanation base embedding EE features both in the recommendation branch and the explanation branch. The hyperparameter β controls the contribution of the temporal features Ẽrec and Ẽexp in the final personalized time-aware embeddings Erec-time and Eexp-time . When varying the value of
AM-Electronics AM-Movies Yelp Rec Exp Rec Exp Rec Exp Time-aware Multi-interest Explanation R@10 N@10 R@10 N@10 R@10 N@10 R@10 N@10 R@10 N@10 R@10 N@10 0.0327 0.0164 0.1087 0.0727 0.0549 0.0269 0.1171 0.0830 0.0650 0.0397 0.0422 0.0230 ! 0.0402 0.0197 0.1371 0.0912 0.0563 0.0266 0.1287 0.0841 0.0672 0.0410 0.0454 0.0250 ! 0.0361 0.0184 0.1312 0.0849 0.0550 0.0274 0.1300 0.0897 0.0652 0.0399 0.0428 0.0234 0.0324 0.0160 0.1252 0.0844 0.0567 0.0272 0.1242 0.0868 0.0654 0.0399 0.0437 0.0238 ! ! ! 0.0498 0.0247 0.1455 0.0938 0.0583 0.0295 0.1294 0.0872 0.0699 0.0425 0.0476 0.0262 ! ! 0.0450 0.0240 0.1423 0.0920 0.0592 0.0286 0.1274 0.0898 0.0689 0.0409 0.0488 0.0275 0.0401 0.0234 0.1378 0.0915 0.0587 0.0300 0.1288 0.0913 0.0685 0.0412 0.0469 0.0259 ! ! ! ! ! 0.0550 0.0291 0.1510 0.0979 0.0620 0.0339 0.1359 0.0931 0.0740 0.0455 0.0533 0.0309 Strategy
Table 5: Ablation of key strategies on three datasets, best results are marked in bold, second best results underlined. AM Electronics
E
AM Movies
We further explore the relationships between the two gating parameters γ rec and γ exp and the average time interval t̄ of each user sequence. Specifically, for each user sequence Su , Pn 1 we compute the average time interval t̄ = n−1 (t − i i=2 ti−1 ). Figure 10 visualizes 2D points (t̄, γ rec ) and 2D points (t̄, γ exp ) for different user sequences. Results show that as the average time interval t̄ increases, the two gating parameters γ rec and γ exp decrease linearly. That is, sequences with short intervals tend to have larger γ rec and γ exp . This aligns with our design intuition: sequences with short intervals reflect recent, dense behaviors and thus benefit more from short-term temporal modeling, because large γ rec and γ exp mean a greater contribution of short-term temporal features to Ẽrec and Ẽexp . In contrast, sequences with long intervals rely more on longterm temporal modeling.
Yelp
Figure 10: Relationship between the gating parameters γ rec and the average time interval t̄ (left) and relationship between the gating parameters γ exp and t̄ (right). Each point represents a single user sequence. Results show that as t̄ increases, the two gating parameters decrease linearly.
one hyperparameter, we keep the other fixed. With β fixed to 0.1, we turn over α in {0.7, 0.8, 0.9, 1.0}. Increasing α from 0.7 to 0.9 consistently improves both recommendation performance and explanation performance. Then performances drops obviously when α = 1.0, indicating the recommendation embedding and the explanation embedding are not fused. With α fixed to 0.9, we turn over β in {0.05, 0.1, 0.2, 0.3}. We observed that β = 0.1 yields the best results. Performance drop significantly when it goes beyond 0.1. On the other hand, when the β is too small, like 0.05, will cause the temporal feature to have too little impact to work.
C
Complete comparisons with baseline models
We compare the proposed TME-PSR with the origin models without SCEMIM. As table 4 shows, the proposed method still achieved significant improvements in recommendation metrics across all three datasets.
D
Complete ablation study
We evaluate every non-empty combination of the three personalization strategies on the three datasets, as table 5 shows. The main findings are entirely consistent with the ablation study described in sec 4. In summary, while the relative contribution of each component varies with the dataset characteristics, the complete TME-PSR configuration reliably delivers the strongest recommendation and explanation quality across all settings.
F
Time-aware Personalization Details
Complexity analysis of multihead LRU
We analyze a single-layer LRU for full-length forward with recursive parallelization. For one head with hidden size d, the time complexity is O(log(n) · d2 ) [Yue et al., 2024] for a sequence of length n. We ignore constant factors and lineartime terms due to normalization/residual wiring and do not account for additional FFN/projection costs here. Then, the time complexity of serially computing all heads is O(H · 2 log(n) · (d/H)2 ) = O(log(n) · dH ). The time complexity d 2 of ideal parallel computing of all heads is O(log(n) · ( H ) ). d d ×H H With per-head Wh , Uh ∈ R , the space complexity of d 2 per head is O(( H ) ) and the total space complexity of the en2 tire architecture is O( dH ), which is inversely proportional to the number of heads H. In comparison, typical transformerbased recommenders has a time complexity of O(n2 d + nd2 ) and a space complexity of O(d2 ) and typical RNN-based recommenders has a time complexity of O(nd2 ) and a space complexity of O(d2 ).
G
Examples of three representative user
Fig. 8(a) shows that all distributions are highly left-skewed and concentrated near zero, indicating that most sequences are assigned low semantic alignment strengths between recommendations and explanations in both the recommendation task and the explanation task. For example, User 1 in
(a) User 1 =4.3e-5 =1.8e-5
Wonderful romance
Fantastic animation
Good science fiction
Spectacul action
Great superhero movie
Best marvel movie yet
Redemption Philosophy
Era-Life Metaphor
Loneliness and Freedom
(b) User 2 =3e-3 =3.1e-3
(c) User 3 =3.3e-5 =5.4e-5
Figure 11: Examples of three representative user types, correspond to the three clusters in Figure 7(b).
fig. 11(a) and User 3 in fig. 11(c) have low values of µ̄rec and µ̄exp . Meanwhile, the long tail of the distribution ensures a portion of sequences still receive strong semantic alignment signals when needed. For example, User 2 in fig. 11(b) has high values of µ̄rec and µ̄exp . exp Fig. 8(b) shows that user-level µ̃rec u and µ̃u exhibit three well-separated clusters: (1) The user shown in fig. 11(b) belongs to the balanced group, this type of users has similar semantic alignment requirements in the recommendation task and the explanation task. (2) The user shown in fig. 11(a) belongs to the recommendation-dominant group, this type of users has higher semantic alignment requirements in the recommendation task. (3) The user shown in fig. 11(c) belongs to the explanation-dominant group, this type of users has higher semantic alignment requirements in the explanation task.