Source-linked AI summary
Temporal Meta-path Guided Explainable Recommendation
Hongxu Chen, Yicong Li, Xiangguo Sun, Guandong Xu, Hongzhi Yin
TL;DR
Path-based explainable recommendation often overlooks the temporal order of user–item interactions, even though recent purchases can provide a more appealing explanation. TMER models consecutive purchases with item–item paths and attention, and the paper reports strong results across three real-world datasets.
Problem
Many path-based explainable recommendation methods treat knowledge graphs as static, while purchase order can affect both recommendation precision and the explanation of a user’s intent.
Source: §1 · pp. 1–2Method
TMER models sequential user–item dynamics using attention over item–item paths between consecutive purchases, alongside user–item path context.
Source: §1 · p. 2, §3.4.1 · p. 5Results
TMER achieves the best scores among the compared state-of-the-art recommendations, with reported advantages on three real-world datasets.
Source: §1 · p. 3, §4.3 · p. 6Takeaways & Limitations
Abstract
from arXiv · showhide
This paper utilizes well-designed item-item path modelling between consecutive items with attention mechanisms to sequentially model dynamic user-item evolutions on dynamic knowledge graph for explainable recommendations. Compared with existing works that use heavy recurrent neural networks to model temporal information, we propose simple but effective neural networks to capture user historical item features and path-based context to characterise next purchased item. Extensive evaluations of TMER on three real-world benchmark datasets show state-of-the-art performance compared against recent strong baselines.
1 INTRODUCTION
Path-based explainable recommenders often treat knowledge graphs as static, overlooking how purchase order can affect both recommendations and their explanations. TMER models temporal user–item dynamics with item–item paths and attention, and evaluates the approach on three real-world datasets.
1 INTRODUCTION
- Static path-based methods can miss that a recent purchase may better explain a subsequent item than a co-purchase relationship.The paper’s example contrasts a phone case explained by another customer’s purchase with one explained by the user’s recent phone purchase.Source: §1 · pp. 1–2
- TMER integrates users’ dynamic interactions over time into path-based knowledge-aware recommendation to improve recommendation performance and explainability.The paper frames this as addressing prior approaches that consider sequence information only within paths or treat a user’s interactions as a separate whole.Source: §1 · p. 2
- TMER models consecutive purchased items using item–item paths as contextual features for sequential recommendation.These paths represent possible factors connecting consecutive purchases, and their features are combined with features of the previous item to predict the next item.Source: §1 · p. 2
- Attention over diverse user–item and item–item meta-paths learns path weights that can serve as explanations of recommendations.The two attention units characterize users’ general shopping tastes and temporal purchasing motivations, respectively.Source: §1 · p. 2
- The paper reports extensive evaluations on three real-world datasets, with results described as showing TMER’s superiority, effectiveness, and temporal explainability.The authors contrast its approach with sequential methods relying on recurrent neural networks, describing TMER as light and simple.Source: §1 · pp. 2–3
2 PRELIMINARIES
The preliminaries define the graph and path concepts used to represent recommendation knowledge, then state the sequential recommendation task as predicting items and possible reasons for a user.
2 PRELIMINARIES
- An information network is a graph whose edges encode relations between typed entities; a relation and its reverse are distinct unless the relation is symmetric.The paper represents an edge as a relation connecting a head entity to a tail entity.Source: §2 · p. 3
- A heterogeneous information network contains multiple entity types and multiple relation types.The paper gives a product recommendation network as an example, with users, items, categories, colors, and brands among its entity types.Source: §2 · p. 3
- A meta-path composes relations into a typed path, such as user→item→brand→item in the product recommendation network.The paper notes that it will discuss its chosen meta-paths in later sections.Source: §2 · p. 3
- Sequential knowledge-aware explainable recommendation predicts a user’s top-k items from the item set, transaction sequence, and associated information network, along with possible reasoning for recommendations.The task is defined for a user and that user’s transaction sequence.Source: §2 · p. 3
3 MODEL
TMER initializes user and item representations, models path-based context, and uses attention to carry sequential information between consecutive items before making recommendations. Its architecture combines user-item and item-item path instances with item and user embeddings in a recommendation network.
3.1 Overview of TMER Architecture
- TMER embeds users and items, mines sequential and nonsequential meta-path instances, uses attention to update item representations, and feeds the resulting information to a recommendation network.The architecture passes the previous item's feature to the next one.Source: §3.1 · p. 3
3.2 Initialize User and Item Representations
- DeepWalk initializes user and item representations, with the authors preferring its path-focused embeddings over local-neighborhood aggregation for capturing global higher-order relationships.The paper notes that alternative graph-based initializers can also be applied.Source: §3.2 · pp. 3–4
3.3 Incorporating Meta-path based Context
- TMER models user-item and item-item path instances; item-item paths can express complementary, substitutable, and co-purchase relations as well as temporal dependencies between consecutive purchases.Because the number of meta-paths can be too large to handle, the method samples instances using similarity between adjacent nodes.Source: §3.3 · pp. 4–5
3.4 Parameterizing Combinational Features of Meta-paths as Recommendation Context
- Multi-head self-attention combines multiple sampled paths to represent the potentially mixed reasons connecting a user-item or item-item pair.Paths are treated as sentences and nodes as tokens to learn path embeddings before attention.Source: §3.4.1 · p. 5
- TMER models each transition using information passed from the previous item and a sampled path connecting the consecutive items, rather than relying on recurrent networks.The carried information is an attentive aggregation of previous item-item connectivity, and the model can capture long- and short-range sequence memory.Source: §3.4.1 · p. 5
- A two-layer attention mechanism updates each item representation with information from the preceding item and the paths arriving at the current item.For the first item, the update instead incorporates the user-item instance from the user to that item.Source: §3.4.3 · p. 5
3.5 The Complete Recommendation Model
- The recommendation model concatenates user, item, and path-instance information before an MLP generates user-item prediction scores.The MLP has two hidden layers with ReLU and a sigmoid output, and training uses implicit-feedback loss with negative sampling from a uniform noise distribution.Source: §3.5 · p. 6
4 EXPERIMENT
The experiments use three Amazon datasets, evaluate ranked recommendations and explanation case studies, and compare TMER with established recommendation baselines.
4.1.1 Datasets.
- The experiments use Amazon Musical Instruments, Automotive, and Toys and Games, each with brand and category metadata.User, item, brand, and category entities form the heterogeneous information network; meta-path instances are generated and sampled.Source: §4.1.1 · p. 6
4.1.2 Evaluations.
- For each positive test item, evaluation ranks it with 500 random negatives using HR@K and NDCG@K at K=1, 5, 10, and 20.Recommendation explanations are evaluated through case studies.Source: §4.1.2 · p. 6
4.1.3 Baselines.
- The comparison includes session-based GRU4Rec, attention-based NARRE, meta-path-context MCRec, and the sparse-prediction methods NFM and FMG.The passage describes each baseline’s recommendation approach.Source: §4.1.3 · p. 6
4.2 Parameter Settings
TMER uses a fixed per-user split and selected user-item and item-item meta-paths, with its learning rate tuned by dataset.
4.2 Parameter Settings
- Each user’s first two items are bridge items, the remaining four non-test items are training items, and the last six items are test positives.The hyperparameters are carefully tuned for optimal results in all experiments.Source: §4.2 · p. 6
- TMER’s learning rate is 5e−6 for Musical Instruments, 5e−5 for Automotive, and 1e−4 for Toys and Games.Other baseline parameters are searched for their best results.Source: §4.2 · p. 6
- TMER uses UIBI, UICI, UIBICI, and UICIBI user-item paths, plus seven listed item-item paths: ICIBI, IBICI, ICICI, IBIBI, IUIUI, ICIUI, and IBIUI.The selected paths are described as nearly all existing 4- to 6-hop meta-paths for these user, item, brand, and category datasets.Source: §4.2 · p. 6
4.3 Effectiveness Analysis on Recommendation Results
TMER achieves the best scores against the compared recommendation baselines, with reported advantages on all three datasets.
4.3 Effectiveness Analysis on Recommendation Results
- At least 1%, 8%, 6%, and 6% HR@K gains over second place occur on Toys and Games at K={1, 5, 10, 20}.The comparison reports these improvements in hit ratio.Source: §4.3 · p. 6
- On Toys and Games, NDCG@K gains over second place are 0.9%, 3%, 6%, and 3% at K={1, 5, 10, 20}.The reported advantages also hold on the other datasets.Source: §4.3 · p. 6
- TMER achieves the best scores against existing state-of-the-art recommendations across the three datasets.Table 3 presents the baseline performance comparison.Source: §4.3 · p. 6, Table 3 · p. 7
- MCRec’s comparison differs because it uses user-item interactions for recommendation while ignoring intrinsic item-item relations.Source: §4.3 · p. 7
4.4 Effectiveness Analysis on Attention
Ablations indicate that both user-item and item-item instance self-attention modules contribute to recommendation performance.
4.4 Effectiveness Analysis on Attention
- Removing either the user-item or item-item instance self-attention module reduces the performance boost seen with TMER.The variants are TMER-RUI, which removes user-item instances and their self-attention module, and TMER-RII, which removes item-item self-attention.Source: §4.4 · p. 7
- The self-attention modules adaptively adjust the importance of different user-item and item-item instance paths.The self-learned paths contribute to recommendation performance.Source: §4.4 · p. 7
- The item-item meta-path instances self-attention module improves Musical Instruments results by 6.7% on HR@20 and 12.3% on NDCG@20.Source: §4.4 · p. 7
4.5 Effectiveness Analysis on meta-paths
The analysis compares meta-path schemes using HR@10 and NDCG@10 on Musical Instruments and Automotive, with 100 sampled negative items per positive pair. All-attribute meta-paths perform better on NDCG@10, while HR@10 is nearly unchanged; category-related paths improve more than brand-related paths.
4.5 Effectiveness Analysis on meta-paths
- All-attribute meta-paths outperform the other schemes on NDCG@10 for Musical Instruments and Automotive.The analysis samples 100 negative items per user-item positive pair and evaluates HR@10 and NDCG@10.Source: §4.5 · p. 7
- HR@10 is almost the same across the meta-path schemes on Musical Instruments and Automotive.Source: §4.5 · p. 7
- Category-related meta-paths improve more than brand-related meta-paths in the reported trends.Source: §4.5 · p. 7
4.6 Case Study: generating explanation for recommendation
The case study illustrates how TMER uses purchase-history-linked item-item paths and their attention weights to explain recommendations. The examples include attribute-based reasons between consecutive purchases and note that sparse item metadata can leave only one path between some items.
4.6 Case Study: generating explanation for recommendation
- TMER generates multiple item-item instance paths from purchase history, then learns their weights with attention and aggregates them for each item-item pair.Source: §4.6 · p. 7
- Figure 4 places user u273’s historical purchases above the reasoning item-item paths, with category attributes in green and brand attributes in blue.Black blocks lack meta information in the dataset.Source: Figure 4 · p. 8
- For item i2280, the highest-weight stated reason links prior musical-instrument-category purchases i2954 and i2237, with attention weight 0.18.Source: §4.6 · p. 8
- For item i4514, the example cites prior item 2280’s matching brand and category; sparse item attributes mean some pairs have only one instance path.The learned paths feed the item attention module, supporting recommendations with purchase-sequence information.Source: §4.6 · p. 8
5 RELATED WORK
Related work spans collaborative filtering, knowledge-graph embeddings, meta-path similarity, and instance-path approaches. The paper positions TMER as combining sequential user-item behavior with item-item path information to address limitations in prior approaches.
5 RELATED WORK
- Collaborative filtering relies on similar user histories but faces sparsity and cold-start problems; knowledge-graph methods add side information and can support recommendation performance and explainability.Examples of side information include user and item attributes, item contents, and external information.Source: §5 · p. 8
- Embedding-based methods model user-item interactions through knowledge-graph embeddings, but their performance is sensitive to graph quality.Examples include TranE used in and a shared user-item embedding space.Source: §5 · p. 8
- Meta-path similarity methods learn structured relations between node types, but do not capture information from particular path instances, limiting explanation specificity.Examples include MF and FM over meta-path similarities and aspect-level similarity matrices [10].Source: §5 · p. 8
- Meta-path-instance context supports fine-grained explanations, but the described approach ignores sequential interaction dynamics.Source: §5 · p. 8
- The LSTM approach models sequences within user-item paths but omits the importance of users’ clicked-history sequences.Source: §5 · p. 8
- The approach in models user behavior sequences and user-item path connectivity, but omits intrinsic item-item relations and their complex semantics.The paper proposes differentiating user-item and sequential item-item instances to improve explainable recommendation.Source: §5 · p. 9
6 CONCLUSION
The conclusion presents TMER as a path-based explainable recommender that models dynamic user-item interactions over time. It explores attention-weighted item-item paths between consecutive purchases and identifies human-readable explanations and causal path discovery as future directions.
6 CONCLUSION
- TMER models dynamic user-item interactions over time with path-based knowledge-aware explainable capabilities.Source: §6 · p. 9
- It uses attention mechanisms over item-item paths between consecutive items to model users’ sequential behavior.Source: §6 · p. 9
- Future work may generate human-readable explanations with NLP and explore causal learning to discover more appealing explanation paths.Source: §6 · p. 9