Submit final solution for Code Challenge 2025

[bahman-farhadian]

• Implemented robust parent-child kinship matching using likelihood ratios
• Added allele frequencies for all 21 loci
• Per-locus classification: identical, partial, mutation, mismatch, missing
• CLR based on allele frequencies with penalties for mutations/mismatches
• Filter: require ≥3 partial matches to confirm parent-child relationship
• Prune unrelated candidates (>2 mismatches)
• Score boosted by number of partial matches
• Added detailed explanation in Hossein_Hamzehei_Bahman_Farhadian_Explanation.md
• Included test results and data batch

[bahman-farhadian]

Hi Ali,

Thanks for reviewing my PR! The CI result was:

Correct matches: 27/35
Accuracy       : 77.1%
Execution time : 3.86 seconds
Final score    : 97.1/120

However, my local testing shows better results:

Metric	Highest	Average	Lowest
Accuracy	100% (35/35)	88.5% (31/35)	71.4% (25/35)
Execution time	8.55 seconds	~9.22 seconds	10.77 seconds
Final score	120/120	108.5/120	91.4/120

Is this variance due to random dataset generation, or could my approach be improved? If there's room for improvement, I'd appreciate any suggestions on how to make the function more robust and reliable with less variance.

Also, would a CI re-run be possible?

Thank you.

[tavallaie]

Did you run it over 500k dataset, like the problem requiring?
CI only runs over 5k

[bahman-farhadian]

Yes, I ran it using the make all command, which generates a data directory containing the str_database.csv file. Each time, this file includes 500,000 rows. Did I make a mistake here during execution?

[tavallaie]

You made several unrelated changes that introduced errors:

First, you should not have modified the function that calls the match function.
Second, according to the challenge requirements, you should not perform simple pairwise matching. Instead, you must build an index and perform matching using that index.
You should not have changed the README or uploaded any ZIP files. Your pull request should include only the updated match function.

[bahman-farhadian]

Hi Ali,
Sorry about that. I thought including my AI interaction logs and a short report would be helpful for transparency, but I understand now that was out of scope.
I've cleaned up my commit and addressed all points:

✓ Removed ZIP files
✓ Only participant_solution.py is changed
✓ No modifications to find_matches()
✓ Uses index-based matching with allele_index for fast candidate lookup

Thanks for the feedback!

[bahman-farhadian]

I also noticed my posterior probability calculation wasn't following the Bayesian formula with a 50% prior — I've fixed that now.

I tested the function 100 times against the 50k dataset:

Max Accuracy	Average Accuracy	Min Accuracy	Variance
100.0%	89.5%	77.1%	23.93

The average accuracy looks promising, but I'm still seeing quite a bit of variance across runs (min 77.1%, max 100%). I'm trying to make the matching behavior more consistent and reliable.

I know you're busy, but if you have any suggestions or ideas on how to stabilize the performance and reduce the variance, I would really appreciate your guidance.

Thanks a lot for your time and help!

[bahman-farhadian]

Just a quick update, Ali: I've updated the matching algorithm. I added a filter to reject self-matches and identical twins using a ratio-based approach: if a candidate has >90% identical loci AND fewer than 3 partial matches, it's rejected as likely same-person/twin rather than parent-child.

The filter logic:

if identical_ratio > 0.9 and partial_count < 3:
    return None  # Reject self-match/twin

This is based on the fact that true parent-child pairs have many partial matches (different strings but shared allele), while self-matches have almost all identical matches.