ems123/MyAnchor-Daily-dataset

Part 1 -

Project Proposal: "MyAnchor" - AI Memory Assistant Student Name: Emilie Levenbach - 208935841 Project Type: Personal Idea (AI for Social Good / Health)

1. The Problem Living with a father who is navigating the early stages of dementia, I see his daily struggle firsthand. I watch him and realize that his main difficulty is not just forgetting facts, but losing his sense of time. He often forgets if he ate lunch an hour ago or what is planned for tomorrow. This "temporal disorientation" causes him anxiety and leads to repetitive questions that can be exhausting for the family.

My Goal: To create a personalized AI voice assistant that acts as an "External Memory Anchor," answering questions about the user's daily schedule and history based on a logged diary.

2. The Solution (App Concept)

A simple AI application where the user (my father) can ask natural language questions. • Did I go swimming today?" • "What do I have tomorrow morning?" • "How was my mood last week?"

The system will act as a central hub, aggregating data from two simulated sources:

Digital Calendar: For future plans and scheduled appointments (e.g., "Doctor at 17:00").
Family Reporter: For real-time important updates logged by family members (e.g., "Dad took his pills")

3. Data Strategy (Synthetic Data Generation - 10%) Since I cannot use my father's real private medical data for a public project, I will generate a Synthetic Dataset representing one full year of a senior citizen's life. • Generation Tool: I will use an LLM (Gamini) to generate a realistic CSV log. • Data Structure: o Date & Time o Activity (e.g., Swimming, Pills, Lunch, Nap, Doctor Appointment) o Mood_Score (1-10) - Designed to show high correlation with physical activity. o Notes (Free text description, e.g., "The water was cold but refreshing"). o Arrangements
4. Exploratory Data Analysis (EDA - 10%)

I will generate a Synthetic Dataset representing one full year of a senior citizen's life. Data Structure: • Date & Time • Source (Google Calendar vs. Family Update) • Activity (e.g., Swimming, Pills, Lunch) • Mood_Score (1-10) • Status (Planned, Completed, Cancelled)

I will analyze the synthetic data to find patterns and visualize them: • Graph 1: Mood Score vs. Activity Type (Proving that swimming improves mood). • Graph 2: Activity distribution by hour of the day. • Graph 3: Weekly routine clusters.

5. AI & Tech • Embeddings: I will convert the text diary entries into vectors using a Sentence-Transformer model (e.g., all-MiniLM-L6-v2). • Vector Search: Using FAISS or simple Cosine Similarity to find relevant diary entries based on the user's question. • LLM: Using a text-generation model (like Flan-T5 or GPT) to take the relevant entries and phrase them into a human-like answer. • App UI: Hugging Face Spaces (Gradio) with a focus on accessibility (Audio input/output).

System Architecture: How it Works

My Anchor is not just a static analysis tool; it is a live system that connects historical data with real-time inputs.

The system is built on three pillars:

1. The "Memory" (Historical Data)

• Source: CSV Dataset (7 years of activities & mood scores).
• Function: Uses the Embedding Model (MiniLM) to understand what usually makes Dad happy and what his long-term habits are.

2. The "Live Context" (Google Calendar)

• Source: Real-time Google Calendar API.
• Function: The app pulls current events to know where Dad is right now. It prevents the system from suggesting "Go to the pool" if the calendar says "Doctor Appointment at 10:00".

3. The "Guardian" (Family Input)

• Source: A dedicated input interface (Bot/Form).
• Function: Family members update the daily schedule and inject new "Anchors" (e.g., "Grandkids visiting today"). The system combines this new info with his routine to keep him oriented.

---

Flow: Family Input + Calendar Data ➡️ AI Processing (RAG) ➡️ Personalized Recommendation for Dad.

---

Part 2 -

My Anchor - Daily Routine Dataset

This dataset was created for the "My Anchor" project. The goal is to help my father, who is dealing with early signs of memory loss, to maintain a stable daily routine and stay oriented in time and place.

The data tracks his daily activities from 2018 to 2025, including meals, medication, exercise, work, and his mood.

Data Analysis (EDA)

Here are the insights from the data that help us understand his routine and well-being.

1. Daily Anchors: Structure of the Day

This graph shows the "heartbeat" of the day. You can clearly see the peaks at specific hours (08:00, 13:00, 19:00). These are the "Anchors" – fixed times for food and medication that keep him oriented in time.

2. Emotional Memory: What makes him happy?

We analyzed which activities give the highest mood score. As seen below, Swimming has the most positive impact on his well-being. This helps the app recommend the right activity when he feels low.

3. Spatial Orientation: Safety Zones

This chart shows where he spends his time. While most of the time is spent safely at Home, he still maintains independence by going to the Pool and Haifa (for appointments).

---

Part 3: Embeddings & Semantic Search

In this step, I transformed the raw textual data into a machine-readable format to enable Semantic Search. Unlike a simple keyword search, this allows the app to understand the intent and context behind a user's query (e.g., understanding that "feeling down" is related to "Mood Score: 3").

1. Data Preparation

Before feeding the data into the models, I combined the relevant columns (Activity, Location, Mood) into a single descriptive sentence for each row.

• Example: "Activity: Swimming at Pool with Mood Score: 9"

2. Model Selection Experiment

To choose the best model for the app, I conducted a benchmarking experiment comparing three Hugging Face models:

• Model A: all-mpnet-base-v2 (The "Mercedes") - High accuracy, but slow.
• Model B: all-MiniLM-L6-v2 (The "Subaru") - Fast & efficient.
• Model C: paraphrase-albert-small-v2 (The "Scooter") - Lightweight.

The Results: I built a script to measure the processing time for 16,000 records.

1. Environment: The test was executed on a Google Colab environment.
2. Process:
   • Loaded the dataset (16,000 text entries).
   • Iterated through the list of candidate models.
   • Used the time library to record the start and end timestamps for the encoding process (model.encode()).
3. Code Logic:

   start_time = time.time()
   embeddings = model.encode(text_list)
   duration = time.time() - start_time
   

As part of this project, I chose to use the all-MiniLM-L6-v2 model as the foundation for the semantic search engine, based on the Transformer model optimization principles learned in the course. While large-scale models (such as BERT-base) generate vector representations in a 768-dimensional space, they often suffer from "noise"—excess information that does not contribute to understanding the semantic meaning of the sentence. MiniLM performs dimensionality reduction to just 384 dimensions, enabling more dense and effective Representation Learning. As we learned regarding the "Curse of Dimensionality," a compact but high-quality representation ensures that each dimension in the vector "works harder" and carries more semantic weight. This directly improves the accuracy of the Cosine Similarity calculation and makes the retrieval sharper and more focused.

Additionally, the model applies a technique called Attention Head Pruning. In this process, the model learns to refine the most critical Attention Weights between words and discard weak connections or meaningless filler words. The model was optimized for maximum Short-text Precision at the expense of the ability to handle long documents, a feature that perfectly matches the application's Use Case, which is based on short memories and journal entries. This choice allows for high performance while maintaining low Latency, which is critical for the user experience in a real-time system.

• MPNet was the most accurate but took significantly longer (high latency).
• MiniLM was 3x faster while maintaining excellent accuracy.

Conclusion: Since "My Anchor" is designed for an elderly user requiring real-time feedback, low latency is critical. Therefore, I selected all-MiniLM-L6-v2 as the engine.

Deliverable: The embeddings and the original dataframe were serialized and saved into my_anchor_embeddings.pkl.

---

Created by: Emilie Levenbach