ECML-PKDD2021 "Single Model for Influenza Forecasting of Multiple Countries by Multi-task Learning"

Transcript

1. Single Model for Influenza Forecasting of Multiple Countries by Multi-task Learning Taichi Murayama, Shoko Wakamiya, Eiji Aramaki Nara Institute of Science and Technology (NAIST) ECML-PKDD 2021
2. 2 Importance of forecasting infectious diseases l Social Impact: Influenza and COVID-19 l Influenza has resulted in between 12,000 - 61,000 deaths annually l The rapid increase of infected people causes the medical crisis l The necessity of preventive measures of infectious diseases by public health Background https://www.bcbswny.com/content/wny/provider/news/blue-bulletin/an-early-start-to-flu-season.html
3. 3 Research of forecasting infectious disease (mainly Influenza) l For more accurate forecasting, they utilize ML and NN-based model, not compartment models such as SIR l Recently, the forecasting using user-generated content (UGC) has become the mainstream of research l Google Flu [Ginsberg et al. 2009] l Twitter [Paul et al., 2014] l Most research target one country’s epidemic situation Background
4. 4 Research of forecasting infectious disease (mainly Influenza) l For more accurate forecasting, they utilize ML and NN-based model, not compartment models such as SIR l Recently, the forecasting using user-generated content (UGC) has become the mainstream of research l Google Flu [Ginsberg et al. 2009] l Twitter [Paul et al., 2014] l Most research target one country’s epidemic situation Background We propose a single model targeting at forecasting flu in multiple countries
5. 5 The ILI (Influenza like illness) rates of multiple countries l Same behavior: epidemic in winter and calm down in summer Background: Can we make the forecasting model for multiple countries? min max 2017/01/05 2018/01/04 2019/01/03 The tendency of ILI rates in each country is similar ⇒ Are the useful features for forecasting ILI rates in future also similar?
6. 6 Building a single model for forecasting influenza in multiple countries Overview l Method l We treat the forecasting task in multiple countries as multi-task problem l We utilize Encoder-Decoder model to be useful for forecasting ILI rates l Input: Time series of search queries (Google) + ILI rates in past l Output: ILI rates in future (1-week to 5-week ahead) l Advantage l Build the forecasting model with high accuracy l Treating as multi-task problem covers the problem of little data (In many countries, data on infectious diseases are small) Purpose of our research       
    
    
    
    
    
        
      ILI rates Search queries
7. 7 Building a single model for forecasting influenza in multiple countries Overview l Problem l How do we find suitable search queries in multiple languages? l Input: Time series of search queries (Google) + ILI rates in past l How do we effectively utilize two types of data (ILI rates and search queries) for the forecasting? l Search queries have possibility not to be useful for the forecasting [Emily L. Aiken, 2019] Purpose of our research
8. 8 Building a single model for forecasting influenza in multiple countries Overview l Problem l How do we find suitable search queries in multiple languages? l Input: Time series of search queries (Google) + ILI rates in past l How do we effectively utilize two types of data (ILI rates and search queries) for the forecasting? l Search queries have possibility not to be useful for the forecasting [Emily L. Aiken, 2019] Purpose of our research Find the useful search queries leveraging by Word-Alignment method Propose the novel forecasting model leveraging by Attention mechanism
9. 9 Building a single model for forecasting influenza in multiple countries Overview l Problem l How do we find suitable search queries in multiple languages? l Input: Time series of search queries (Google) + ILI rates in past l How do we effectively utilize two types of data (ILI rates and search queries) for the forecasting? l Search queries have possibility not to be useful for the forecasting [Emily L. Aiken, 2019] Purpose of our research Find the useful search queries leveraging by Word-Alignment method Propose the novel forecasting model leveraging by Attention mechanism
10. 10 Dataset l ILI (Influenza like Illness) rates Data l Target: 5 countries United States(U.S.), Japan(JP), England(UK), France(FR), Australia (AU) l Term: 26th week in 2013 – 29th week in 2020 l Report ILI rates per a week (52 points in a year) l Search queries l We utilize Google Trend (the term is the same as ILI rates data) l Method for selecting search queries l English: List of previous research [Zou B, 2018] l Other languages (fr and jp): WT-based Preparation: Dataset
11. 11 Find search queries other than in English Word-alignment + Time-series correlation-based (WT-based) Preparation: Dataset l Word-Alignment l Method of building multi-lingual embedding l Extract the candidate of search queries by calculating of cosine similarities between English queries and words of other languages l Obtaining words that are difficult to obtain by translation ex. Orthographical variant ʮΠϯϑϧʯorʮΠϯϑϧΤϯβʯ l Time-series correlation l Obtaining words with high correlation between time series of Google Trend and ILI rates Select search queries with higher values, which is the combination of the similarity score by word alignment and the correlation score of time series
12. 12 Preparation: Dataset Examples of the selected search queries
13. 13 Proposed Model (One country) l Input: Two types of data (Time series of search queries, Historical ILI rates) Output: ILI rates in future (1-week to 5-week ahead) l Encoder-Decoder based Model l Problem: How do we effectively input two types of data (ILI rates and search queries) ? Proposed Model (One country) ILI rates in past Time series of search queries Proposed Model ILI rates in future       
     
     
     
     
     
         
       ILI rates Search queries
14. 14 Proposed Model (One country) Proposed Model (One country)
15. 15 Proposed Model (One country) Proposed Model (One country) Input
16. 16 Proposed Model (One country) Proposed Model (One country) Encoder-Decoder Model Input
17. 17 Output Encoder-Decoder Model Input Our proposed Encoder-Decoder architecture to forecast ILI rates Proposed Model (One country) Proposed Model (One country)
18. 18 Proposed Model (One country) Proposed Model (One country) Point 1: Split seasonalized and deseasonalized component into the historical ILI rates Point 2: Utilize the Attention Mechanism
19. 19 Point 1: Split seasonalized and deseasonalized component l Search queries are useful features for the forecasting of non-seasonal component. l Our model only forecasts non-seasonal of time series, assuming that seasonal of time series is constant throughout a year. l Split the historical ILI data by Seasonal-trend decomposition using LOESS (STL) min max 2017/01/05 2018/01/04 2019/01/03 Proposed Model (One country)
20. 20 l Combining search queries and the non-seasonal part of the ILI rate by Attention mechanism l Model con consider the useful search queries by weighting heavily Query 1 Query 2 Query L ・・・ Non-seasonal ILI rates Image of Attention in our model Point 2: Utilize the Attention Mechanism Proposed Model (One country)
21. 21 Proposed Model (One country) Proposed Model (One country) Point 1: Split seasonalized and deseasonalized component into the historical ILI rates Point 2: Utilize the Attention Mechanism
22. 22 One country Multiple countries Proposed Model (Multiple countries) Proposed Model (Targeting at Multiple Countries)
23. 23 One country Multiple countries Proposed Model (Multiple countries) v v Proposed Model (Targeting at Multiple Countries)
24. 24 l Treat ILI forecasting in multiple countries as separate task l Basic components have shared parameters for capturing general features (circled by Blue) l Particular components to each country, such as Attention, have different parameters (circled by Red) Proposed Model (Targeting at Multiple Countries) Proposed Model (Multiple countries)
25. 25 l Set the initial representations of Encoder (GRU) as learnable embedding for each country: Country Embedding Proposed Model (Targeting at Multiple Countries) Proposed Model (Multiple countries)
26. 26 &YQFSJNFOUBM
    4FUUJOHT l No. of search queries: 5 countries (JP, US, FR, UK, AU) º10 queries l Test term: 30th-week in 2017 – 29th-week in 2018 (Today’s Presentation) l Evaluate the forecasting score in 1-5 week ahead l Evaluation Metrics l RMSE: Smaller value indicating better performance l !" : Higher value indicating better performance l Comparable methods: GRU / Transformer / Multi-task Elastic Net Experiments
27. 27 &YQFSJNFOUBM
    3FTVMUT
    64 Learning from U.S. Variations of proposed model • Proposed w/o sq: 1 country without search queries • Proposed_single: 1 country • Proposed_multi2: Learning from two countries (JP and US) • Proposed_multi5: Learning from five countries Experiments Learning from multiple countries
28. 28 &YQFSJNFOUBM
    3FTVMUT
    64 Learning from U.S. Experiments Learning from multiple countries l Proposed model achieves the highest accuracy among models targeting at one country
29. 29 l Proposed model achieves the highest accuracy among models targeting at multiple countries Experiments Learning from U.S. Learning from multiple countries &YQFSJNFOUBM
    3FTVMUT
    64
30. 30 Experiments &YQFSJNFOUBM
    3FTVMUT
    'PVS
    DPVOUSJFT
31. 31 )PX
    QSPQPTFE
    NPEFM
    VUJMJ[F
    TFBSDI
    RVFSJFT
    Discussion l Degree of improvement between with and without search queries (GRU vs Proposed) l GRU w/o sq ⇒ GRU: Average improvement: 0.007 in RMSE and 0.001 in "# l Proposed w/o sq ⇒ Proposed_single Average improvement: 0.091 in RMSE and 0.017 in "# l Inputting simply search queries does not contribute much to the improvement of accuracy, while the introduction of Attention mechanism makes a significant improvement of accuracy.
32. 32 7JTVBMJ[BUJPO
    UIF
    XFJHIUJOH
    PG
    "UUFOUJPO 2017/30th 2018/28th 2018/50th 2018/8th 2018/18th 2018/40th fever flu flu and fever the flu flu cough cough fever symptoms of flu flu headache the flu virus flu and cold contagious flu l The weighting of attention of each search query: Red indicates bigger value and blue indicates smaller value. l Attention changes the weights of search queries at the beginning of year, when the flu pandemic. Discussion
33. 33 l Build a single model to forecast influenza epidemics in multiple countries l Treating the forecasting task in multiple countries as multi-task problem l Utilize search queries by Attention mechanism l Proposed model achieves higher accuracy than other models in influenza forecast task l Our experimental results show that Attention mechanism has possibility to be useful for finding suitable search queries Conclusion