Abstract:An efficient customer service management system hinges on precise forecasting of service volume. In this scenario, where data non-stationarity is pronounced, successful forecasting heavily relies on identifying and leveraging similar historical data rather than merely summarizing periodic patterns. Existing models based on RNN or Transformer architectures may struggle with this flexible and effective utilization. To tackle this challenge, we initially developed the Time Series Knowledge Base (TSKB) with an advanced indexing system for efficient historical data retrieval. We also developed the Retrieval Augmented Cross-Attention (RACA) module, a variant of the cross-attention mechanism within Transformer's decoder layers, designed to be seamlessly integrated into the vanilla Transformer architecture to assimilate key historical data segments. The synergy between TSKB and RACA forms the backbone of our Retrieval-Augmented Time Series Forecasting (RATSF) framework. Based on the above two components, RATSF not only significantly enhances performance in the context of Fliggy hotel service volume forecasting but also adapts flexibly to various scenarios and integrates with a multitude of Transformer variants for time-series forecasting. Extensive experimentation has validated the effectiveness and generalizability of this system design across multiple diverse contexts.

What problem does this paper attempt to address?

The paper primarily aims to address key challenges in customer service quality prediction, particularly in improving prediction accuracy in scenarios where data non-stationarity is significant. Specifically, the paper proposes a new framework called RATSF (Retrieval-Augmented Time-Series Forecasting) to tackle the following issues: 1. **Limitations of Existing Models**: Traditional time-series forecasting models based on RNN or Transformer architectures have limitations in utilizing historical data, making it difficult to effectively identify and leverage similar historical data patterns. 2. **Handling Non-Stationary Data**: In scenarios such as hotel service volume forecasting, data is influenced by various factors like order status, weather, and international events, leading to non-stationary characteristics. This makes it challenging for traditional univariate time-series forecasting methods to be effectively applied. 3. **Need for Flexibility and Adaptability**: There is a need for a general and flexible time-series forecasting system design that can efficiently utilize historical sequence information to address complex forecasting challenges. To solve the above problems, the paper contributes in the following aspects: - **Time-Series Knowledge Base (TSKB)**: Developed a time-series knowledge base for efficient storage and indexing of historical data. It employs an advanced indexing system that can quickly retrieve relevant historical data segments. - **Retrieval-Augmented Cross-Attention (RACA) Module**: This is a variant of the cross-attention mechanism in the Transformer decoder layer, designed to be seamlessly integrated into the standard Transformer architecture to facilitate the integration of key historical data segments. - **RATSF Framework**: The combination of TSKB and RACA forms the core of the RATSF framework, which not only significantly improves the performance of Fliggy hotel service volume forecasting but also flexibly adapts to various scenarios and integrates with multiple Transformer variants for time-series forecasting. Experimental results show that in real-world application scenarios (such as Fliggy hotel service volume forecasting) and multiple public datasets, RATSF demonstrates significant performance improvements compared to baselines and other methods. For example, in the Fliggy application, the prediction error was reduced from approximately 15% to about 8%, significantly lowering personnel management costs. Additionally, the paper validates the effectiveness and importance of each component of RATSF through a series of ablation studies.

RATSF: Empowering Customer Service Volume Management through Retrieval-Augmented Time-Series Forecasting

TFEformer: Temporal Feature Enhanced Transformer for Multivariate Time Series Forecasting

Foreformer: an Enhanced Transformer-Based Framework for Multivariate Time Series Forecasting

Generalizable Memory-driven Transformer for Multivariate Long Sequence Time-series Forecasting

Flight Demand Forecasting with Transformers

Enhancing Time Series Forecasting: A Hierarchical Transformer with Probabilistic Decomposition Representation

Transformers with Attentive Federated Aggregation for Time Series Stock Forecasting

Leveraging 2D Information for Long-term Time Series Forecasting with Vanilla Transformers

TCLN: A Transformer-based Conv-LSTM Network for Multivariate Time Series Forecasting

Multivariate Resource Usage Prediction With Frequency-Enhanced and Attention-Assisted Transformer in Cloud Computing Systems

Inter-Series Transformer: Attending to Products in Time Series Forecasting

TS2ARCformer: A Multi-Dimensional Time Series Forecasting Framework for Short-Term Load Prediction

DRFormer: Multi-Scale Transformer Utilizing Diverse Receptive Fields for Long Time-Series Forecasting

STTM: A New Approach Based Spatial-Temporal Transformer And Memory Network For Real-time Pressure Signal In On-demand Food Delivery

RSMformer: an efficient multiscale transformer-based framework for long sequence time-series forecasting

Recurrent Tensor Factorization for Time-Aware Service Recommendation

From Known to Unknown: Knowledge-guided Transformer for Time-Series Sales Forecasting in Alibaba

A Differential Attention Fusion Model Based on Transformer for Time Series Forecasting

ETSformer: Exponential Smoothing Transformers for Time-series Forecasting

Spatiotemporal Fusion Transformer for large-scale traffic forecasting

Survey and Taxonomy: The Role of Data-Centric AI in Transformer-Based Time Series Forecasting