• TDA Literature Review

Meeting: Jan 17, 2022

• Presented results of experiments on Volunteer data and Movielens
• Discussed the metrics for measuring performance on the two datasets
• Showed graphs of two methods and their performance on the two datasets

• Key: the BOA method still outperforms LDA in both datasets.

• Comments and Future work
• To compute metrics as a whole and not aggregate them separately for each user.
• To apply BOA method on a recommendation paper and compare the results
• To compile a representative code and send to the professor
• To read literature on latest work in RS.

Meeting: Jan 24, 2022

• Discussed the possible alternatives of embedding the data, social net embedding
• Refined the life document to include user history pertaining to time, location and participation history
• Discussed a few papers that have used GNN for prediction

• Comments and Future work
• To read the papers related to Social Recommendation eg, Song et al (2019)
• Understand the nature of embeddings they used for the input data
• Summarise findings of papers (using GNN) whose data is similar to ours

Meeting: Feb 18, 2022

• Presented a Survey on GNN Social recommendation
• Presented Item-to-Item KNN method used by Amazon
• Discussed two papers on Social Recommendation and presented their findings
• Paper 1: GNN for Social Recommendation - Fan, Ma, Li, He,Zhao,Tang, and Yin. 2019. (ACM, WWW)
• Paper 2: Session-based Social Recommendation via Dynamic Graph Attention Networks, Song, Wang and Xiao (2019) – (ACM, WSDM)

• Comments and Future work
• To begin working on experiments related to dynamic session-based recommendation
• To make a list of the features and embeddings that can be used from our data,
• Determine session and neighbours based on our data, eg location based or participation based.

Meeting: March 1, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (19:00-20:00)
• Discussion
• Tim presented a proposal on GNN

• Comments and Future work
• To clarify the research problem by formulating the underlying mathematical nitty gritties
• To consider geographic location and tasks types as input features and determine how to incorporate them

Meeting: March 23, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (11:00-12:00)
• Discussion
• Shutong presented and discussed the plan on how the experiments will be conducted.
• Tim presented the findings on the first experiment for Session based GNN

• Comments and Future work
• Tim to write a brief methodology discussing his implementation of the algorithm
• Need for providing details for experimental plan and hyper-parameter tuning

Meeting: March 31, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (11:00-12:00)
• Discussion
• Tim presented a write up of methodology section that also included experiment plan
• Comments and Future work
• Tim to give more clarity on the network architecture diagram
• To consider node embeddings as one hot key vectors
• To also include task type information on the volunteer preference
• Tim to remind Shutong and the professor every week before each meeting.
• Shutong suggested Tim to use PPT to draw flowchart and diagrams
• Shutong suggested to provide reference code for matching task types

Meeting: April 7, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (11:00-12:00)
• Discussion
• Shutong presented her preliminary experimental results
• Tim discussed the DGRec framework.
• Comments and Future work
• Shutong to continue with her experimental plan
• Shutong to consider narrowing the research towards a more specific problem
• Tim to conduct experiment and reproduce the results in the paper
• Attachments:
• shutong_experiments.pptx
• graph_attention_network.docx

Meeting: April 14, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (11:00-12:00)
• Discussion
• Shutong presented statistical data for Pioneers to be used in an article
• Tim presented the reproduced results from the paper
• Comments and Future work
• Shutong to write a summary of her latest findings
• Tim to consider a naive baseline approach that will be compared against the current method
• Attachments:
• tim_progress_4.14.pptx

Meeting: April 21, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (11:00-12:00)
• Discussion
• Shutong presented a summary of her proposal and latest findings
• Tim compared the performance of voting process/UserKNN against the model of interest
• Comments and Future work
• Tim to explore at least one or two graph models for further comparison
• Shutong to make recommended adjustments to the presentation
• Attachments:
• tim_progress_4.21pptx.pptx

Meeting: May 16, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Shutong presented an updated version of her research findings, challenges and proposal. Greedy algorithm solves some of the common challenges in community detection.
• Shutong discussed latest findings on NCE and community detection for Futian and Nanshan districts
• Tim showed comparative results for DGRec and GR4URec. DGRec still outperforms baselines such as UserKNN
• Comments and Future work
• Shutong to determine the relationship between SHS and NCE results
• Shutong to summarize charts and send them to Pioneers in order to get feedback on changes that correspond to the detected time slots
• Shutong to make minor recommended adjustments to the presentation such as illustrating the important next steps as a pipeline instead of explaining them in text.
• Tim to analyze data to determine the optimal time window that gives dynamic sequences
• Tim to conduct experiment at district level, for instance using Futian and/or Nanshan data

Meeting: May 23, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Tim presented statistical information to determine appropriate time window
• Based on standard deviation and entropy, Futian and Nanshan had highest metrics at k=7 and k= 28, respectively.
• Shutong discussed the relationship between NCE, SHS and S-NCE
• Shutong reported that C-NCE has positive correlation with the number of tasks and volunteers, meaning that when the number of tasks/volunteers increase, the system becomes unstable. C-NCE also is positively correlated to the number of SHS [for more details, see attachment]

• Comments and Future work
• Shutong encountered an efficiency problem where its taking longer to compute results for Futian
• Shutong to show her computational method to pave way for alternative solutions
• Tim to consider double checking the data since the number of unique organisers keeps decreasing/constant as time window increases.
• Tim to conduct experiment at district level with induced structural breaks in the data

• Attachments
• Shutong: shutong_cnce.pptx

Meeting: May 30, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Tim presented updated statistical information to determine appropriate time window. The average of Std and Entropy yielded a prediction window of 14 days. Structural breaks were used to determine breakpoints. DGRec with 14 day window was then trained on the breakpoints and results were presented for Futian district.
• Model performance during low volunteer turnout period is low, implying that when participation is low, volunteers behavior is not deterministic.
• Shutong discussed the meeting minutes from her previous presentation with another collaborating professor.

• Comments and Future work
• Tim to come up with hypothesis on factors that influence the participation of volunteers during low turnout period, ie, breakpoint 4.In the long run, to think about volunteer retention.
• Tim to extract weights from the model and analyze them
• Tim to keep doing experiments and compare the performance of DGRec with GR4URec
• Also report statistics of the total number of participants per each breakpoint.
• Instead of improving the model, think of how we can use it to analyze volunteer behaviour

Meeting: June 13, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Tim presented an analysis of hypotheses on volunteer participation. H1: The distribution of task types impacts daily volunteer participation, H2: Distribution of task types impacts model prediction.
• It was also noted that task types that appeared as 'noise' were not significant in OLS BP4 regression
• Shutong shared an outline of the experiments that she intends to do in the coming days/weeks.

• Comments and Future work
• Tim to construct hypotheses around improving model prediction accuracy and not sway far away from the main problem.
• Tim to consider directly testing if and how task types or volunteer participation impacts model prediction instead of applying indirect convolutions
• Tim to write a draft paper
• Shutong to follow up and execute the next tasks according to the experiment plan
• Shutong and Tim to work on PPTs for weekly group meeting scheduled for June 23rd

Meeting: June 27, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Tim shared his plan for the next steps. It includes Paper writing and Ablation study.
• Ablation study will focus on determining the impact of long v short term interests, impact of increasing neighbours and difference between individual and neighbours' interests
• Shutong shared her plan for the next experiments. Some of the outstanding work include in-depth analysis of SHS, stability and finding benchmarks

• Comments and Future work
• Tim to separate ablation study from parameter based experiments such as the impact of increasing neighbourhood.
• Tim to prioritise experiments and research conferences for submission.

Meeting: July 11, 2022

• Attendees:
• Professor Yang Li, Shutong and Tim (14:00-15:00)
• Discussion
• Shutong shared latest experiments including an overview of updated dataset (2022 data) with labels.
• Shutong showed an extensive study on the relationship between SHS and evolution of communities over time. Using interesting visuals, she also discussed a case study of Nanshan district. Unfortunately Nanshan district seem to not have volunteer data for a prolonged period of time, potentially affecting the results.
• Tim briefly showed that the impact of increasing neighbors on recommendation accuracy is not very high. Tim also pointed a challenge regarding determining the impact of tasks types on recommendation.

• Comments and Future work
• Shutong
• Shutong to inquire more about the missing participation data for Nanshan district (2020.10.1-10.20)
• Shutong to determine how does community size affect the relationship between SHS number and evolution type
• Shutong to fine-tune FaceNet community detection algorithm parameter, for instance, by increasing alpha
• On the Hypothesis: SHS increases merge probability:
• Shutong to conduct a case study on two merging communities in order to:
• Determine whether task type or other behaviors are becoming similar after the merging event
• Analyze the position of SHS, for example, does it connect to higher degree nodes than other users?
• Finally, Shutong to review 2-3 papers on recent SHS and dynamic community detection
• Tim
• Tim to check the effect of task type on the recommendation result
• Tim to continue the ablation study experiments to determine:
• The effect of neighbors in GNN
• The effect of long term interest and short term interest of neighbours
• The effect of individual interests without neighbourhood information

• Thanks to Professor Li for helping jotting down the notes

Meeting: July 18, 2022

• Attendees:
• Professor Yang Li, Shutong, Zhang Anping, Kong Dexu and Tim (14:40-16:00)
• Discussion
• Shutong presented two papers on community detection. Detecting the evolving community structure in dynamic social networksExternal Link
• Structural holes and managerial performance: Identifying the underlying mechanisms External Link
• Tim presented presented results from the paper that were produced by a different method: RNN Session
• Using the same approach, Tim showed the effect of teasing apart individual interests from neighbours interests.
• Adding neighbours improves the accuracy. However, continuously adding neighbours after 1 does not improve the results significantly.

• Comments and Future work
• Tim to do experiments and test the model at district level, particularly where the performance is not high. For example, Breakpoint Analysis showed that the model does not perform well under certain circumstances.
• Thus, Tim should conduct similar experiments on district level to determine the effect of task types on model accuracy.
• Tim to furnish the WOLT poster and send it to professor at least 2 days before the deadline (07.25)