ash-shar / cmap

Repository containing Parameter Inference and Posterior Computation codes of CMAP

CRP-based Multifacet Activity Profiling Model (CMAP)

CMAP is a compact, generative framework for learning participant behavior representations on social media platforms. It jointly models the temporal evolution of content, behavior and inter-participant links in a unified latent framework, integrated with the Chinese Restaurant Process (Pitman-Yor) to effectively deal with behavior skew and data sparsity.

If this code is helpful in your research, please cite the following publication

Krishnan, Adit, Ashish Sharma, and Hari Sundaram. "Insights from the Long-Tail: Learning Latent Representations of Online User Behavior in the Presence of Skew and Sparsity." Proceedings of the 27th ACM International Conference on Information and Knowledge Management. ACM, 2018.

Getting Started

These instructions will help you get the model up and running on your system.

Platforms Supported

• Unix, MacOS, Windows (with appropriate compilers/environment)

Prerequisites

Please ensure that the following dependencies are installed:

• g++ (>=4.8.0)

Development Setup

Run the following command for compiling the project

$ make

Variations of Model Supported

This repository supports the following variations of CMAP:

• Unified - Modelling words and actions jointly using the same set of topics.
• Factored - Modelling words and actions using different sets of topics.

Both the Unified and Factored models have a hierarchical counterparts where we introduce a hierarchy over set of topics. In total, the repository supports 4 different variations of CMAP. All the 4 variations can either be run with parallelization (using threads, if multiple cores are available) or without parallelization (linear execution of code).

Please refer the paper for more details.

Input File Format

For a paricular dataset, the model requires 2 input files:

• _pre_processed.txt: Each row of this file corresponds to one data point in your dataset and has 4 columns - Text, UserId, Behaviour and Timestamp (all tab separated). The columns are described below:
  • Text: The text in your data point. Pre-process the text for efficient use.
  • UserId: A user index between 0 to num_users-1 corresponding to the user of the data point.
  • Behaviour: The action observed in the data point. Eg, questioning, answering, commenting, etc. for Stack-Exchanges
  • Timestamp: The normalized value of time of data point. The value must be between 0.01 to 0.99 and should be truncated to 2 decimal places.
• _links.txt: This file is optional and can be provided if you have social interaction information as part of your dataset. Each row corresponds to a link from a data point i to data point j and has 2 colums - i and j (tab-separated). The data points are zero-indexed and indexing is defined from the _pre_processed file.

Both the files should be placed inside the Data folder.

Running the Model

The model can be executed using the following command.

$ ./CMAP <options>

where possible options include:

--dataset <name>:         Name of the dataset to use (required)
--model <type>:           Model-Type. 0 for unified, 1 for factored (default 0)
--hr:                     Use the hierarachical version of the model
--thread <num_thread>:    Use the threaded version of the model. Specify number of threads to use
--G <num_groups>:         Specify the value of number of groups (default: 20)
--K <num_topics>:         Specify number of topics (Use only for unified model) (default: 20)
--K_w <num_text_topics>:  Specify number of text topics (Use only for factored model) (default: 20)
--K_b <num_behav_topics>: Specify number of behavior topics (Use only for factored model) (default: 5)
--scale <s>:              Specify the value of scale parameter (default: 1.5)
--discount <d>:           Specify the value of dicount parameter (default: 0.5)
--iter <num_iter>:        Specify the number of iterations to run (default: 500)
--help:                   Print help

Providing the dataset name is mandatory. Please refer to the paper for optimal values of these parameters.

Sample Run

A sample dataset named biology is present in the Data folder. For running the hierarchical variation of the unified model with G = 20 and K = 10 for 100 iterations, execute the following command:

$ ./CMAP --dataset biology --model 0 --hr --G 20 --K 10 --iter 100

Output

For each full run of CMAP, a folder named <CMAP_Model_Type>___ will be created in the Output Folder. The files of the output folder are described below:

• vocab-mapping.txt: Words to indices Mapping.
• behavior-mapping.txt: Actions to indices Mapping.
• group-priors.txt: Prior probability of groups.
• group-topic-distribution.txt: Distribution of groups over topics.
• group-user-distribution.txt: Distribution of groups over users.
• link-prob.txt: Probability of having a link from group i to group j.
• group-time-alpha.txt: Group-Time Alpha values.
• group-time-beta.txt: Group-Time Beta values.
• topic-word-distribution.txt: Topic to Word Distribution.
• topic-behavior-distribution.txt: Topic to Behavior Distribution.
• table-assignment-status.txt: Status of Data points seating.
• top-topic-words.txt: Top 20 words in a topic.
• top-topic-behav.txt: Top behaviors in a topic.
• top-group-users.txt: Top users in a group.

Generating Latent Representations from the trained model

The latent representations of the users can be generated by using the codes provided in Posterior-Computation folder. The folder contains 2 scripts corresponding to the 2 main variants (unified and factored) of our model. Both the codes require the following inputs:

--output_path <path>:     Path to CMAP output folder
--corpus_path <path>:     Path to <dataset>_pre_processed file
--links_path <path>:      Path to <dataset>_links file
--G <num_groups>:         Specify the value of number of groups (default: 20)
--K <num_topics>:         Specify number of topics (Use only for unified model) (default: 20)
--K_w <num_text_topics>:  Specify number of text topics (Use only for factored model) (default: 20)
--K_b <num_behav_topics>: Specify number of behavior topics (Use only for factored model) (default: 5)
--discount <d>:           Specify the value of dicount parameter (default: 0.5)
--help:                   Print help

Both the scripts output posteriors in the following file <CMAP_Output>/posteriors-user-interactions.txt. Each line of the file contains UserId and list of posteriors (tab-separated). UserId is the id of user as provided in the _pre_processed file. List of posteriors contain posteriors of all interactions (data points) of the user. For generating the latent representation of users, their posterior list is aggregated using either L1-norm or L2-norm along the interaction axis. We use L2-norm over interaction posteriors for generating user representations in all the tasks (Certificate Earner, Reputed Users, Question Recommendation, and Future Activity Prediction). Moreover, interaction level posterior (without any aggregation) are utilized for the tasks of Question Recommendation and Future Activity Prediction. Refer to the paper for details regarding the tasks.